For the Targeting Model, report-generated outputs can be easily applied to a campaign, but users are not limited to the exact variables and combinations identified by the model. Users can also create their own custom JSON file using the same general structure as the model outputs.

This allows users to define variables and combinations based on their own campaign learnings, business rules, historical performance, or optimization goals. For example, a user may already know that a certain geography, publisher, device type, content genre, or day of week is strategically important and can build those values into a custom model. The complete model file must be formatted as a JSON array.

Before creating a custom model, review Model Scoring Overview to understand how model entries are matched, weighted, combined into a request-level score, and evaluated against the threshold selected on the line.

Score Requirements

• Each entry must include a numeric score between -1 and 1.
  • Scores entered in the JSON file use the −1-to-1 model scale. After matching entries are combined, the result is converted to the 0-to-1 bidder scale used by the line threshold. See Model Scoring Overview for details.

• Supported Fields and Expected Values
• The required fields depend on whether the entry is a single variable, a 2-way combination, or a 3-way combination. The following fields can be used in custom Targeting Models:

Field	Group	How to write the value
geo_region	Geography	Lowercase, exactly as it appears in the logs.
geo_zip	Geography	Postal code as it appears in the logs. US ZIPs must keep their leading zero (‘04501’). Non-US formats are fine (‘1000-001’, ‘SW1A 1AA’).
publisher_name	Inventory	Lowercase, exactly as it appears in the logs.
site_domain	Inventory	Lowercase, exactly as it appears in the logs.
app_bundle	Inventory	Lowercase, exactly as it appears in the logs.
exchange	Inventory	Lowercase, exactly as it appears in the logs.
deal_id	Inventory	Lowercase, exactly as it appears in the logs.
content_channel	Content	Lowercase, exactly as it appears in the logs.
content_network	Content	Lowercase, exactly as it appears in the logs.
content_series	Content	Lowercase, exactly as it appears in the logs.
content_title	Content	Lowercase, exactly as it appears in the logs.
content_genre	Content	Lowercase, exactly as it appears in the logs.
content_livestream	Content	‘0’ or ‘1’.
device_type	Device	IAB code as a whole number — ‘3’, never ‘3.0’ and never ‘Connected TV’.
day	Time	Lowercase three-letter code (mon, tue, wed, thu, fri, sat, sun), in UTC.
hour	Time	Whole number 0-23, in the user’s LOCAL time.

Values should match the format used in the underlying bid request data.

For example

{
  "feature": "device_type",
  "value": "3",
  "score": 0.80
}

Use “3”, not “Connected TV” or “CTV”.

Single Variable

A single-variable entry must include:

• feature: The name of the feature.
• value: The specific value associated with the feature.
• score: A numeric value between -1 and 1.

Example:

{
  "feature": "geo_zip",
  "value": "10025",
  "score": 0.95
}

This example assigns a score of 0.95 to ZIP Code 10025.

2-Way Combination

A 2-way combination must include:

• feature1: The name of the first feature.
• feature2: The name of the second feature.
• combination: The exact values associated with feature1 and feature2, provided in the same order and separated by ×.
• score: A numeric value between -1 and 1.

Example 1:

{
  "feature1": "geo_zip",
  "feature2": "device_type",
  "combination": "10025 × 4",
  "score": 0.90
}

This represents ZIP Code 10025 + Device Type 4 with a score of 0.90.

Example 2:

{
  "feature1": "content_genre",
  "feature2": "day",
  "combination": "news/talk radio × thu",
  "score": 0.75
}

This represents Content Genre = news/talk radio + Day = Thursday with a score of 0.75.

The values in combination must match the order of the corresponding feature fields.

3-Way Combination

A 3-way combination must include:

• feature1: The name of the first feature.
• feature2: The name of the second feature.
• feature3: The name of the third feature.
• combination: The exact values associated with feature1, feature2, and feature3, provided in the same order and separated by ×.
• score: A numeric value between -1 and 1.

Example:

{
  "feature1": "geo_zip",
  "feature2": "publisher_name",
  "feature3": "device_type",
  "combination": "10025 × iheartmedia × 4",
  "score": 0.85
}

The order of the values in combination must match the order of the corresponding feature fields.

For example:

feature1 = geo_zip
feature2 = publisher_name
feature3 = device_type

must correspond to:

10025 × iheartmedia × 4

Users can include one or multiple single variables, 2-way combinations, and 3-way combinations in a custom JSON file to define the optimization logic they want to apply.

Model-generated metrics such as conversion rate, lift, score, and statistical significance apply only to variables and combinations produced and evaluated by the model. User-created variables or combinations should not be assumed to have the same performance unless they are separately analyzed and validated.

Reach out to your Account Manager or contact us through the Pontiac Help Center for assistance creating and implementing a custom JSON model.

Field Weights (Optional)

By default, every field in a User Defined Model contributes equally. For example, a match on site_domain carries the same influence as a match on hour.

Optional field weights allow users to change that balance when there is a specific reason to give one attribute more influence than another.

A model file without field weights behaves as an equally weighted model.

How Field Weights Affect Scoring

When a bid request matches multiple model entries, their scores are combined into a single result. With field weights, the classifier calculates a weighted average:

Combined model score =
Sum of (each matching score × its weight)
÷
Sum of all matching weights

A field with a higher weight has more influence over the combined result than a field with a lower weight.

If all weights are equal, the calculation is equivalent to the plain average used by an unweighted User Defined Model.

JSON

To provide field weights in JSON, wrap the model entries in an object and add a field_weights map:

{
  "entries": [
    {
      "feature": "site_domain",
      "value": "espn.com",
      "score": 0.80
    },
    {
      "feature": "hour",
      "value": "20",
      "score": -0.40
    }
  ],
  "field_weights": {
    "site_domain": 3.0,
    "hour": 1.0
  }
}

The existing plain-array JSON format remains valid when no custom field weights are needed:

[
  {
    "feature": "site_domain",
    "value": "espn.com",
    "score": 0.80
  },
  {
    "feature": "hour",
    "value": "20",
    "score": -0.40
  }
]

CSV

For CSV models, add a weight column. Add one separate row for each field weight, leaving value and score empty:

feature,value,score,weight
site_domain,espn.com,0.80,
hour,20,-0.40,
site_domain,,,3.0
hour,,,1.0

The JSON and CSV examples above produce the same model.

Field Weight Example

Assume a request matches:

• site_domain = espn.com: score 0.80, weight 3.0
• hour = 20: score -0.40, weight 1.0

The weighted combined model score is:

((0.80 × 3.0) + (-0.40 × 1.0)) ÷ (3.0 + 1.0)
= 0.50

The combined model score of 0.50 converts to a bidder score of:

(0.50 + 1) ÷ 2 = 0.750

Without custom weights, the two entries contribute equally:

(0.80 + -0.40) ÷ 2 = 0.20

which produces a bidder score of:

(0.20 + 1) ÷ 2 = 0.600

Giving site_domain more weight therefore shifts the result toward its stronger positive score.

Only the Ratios Matter

Field weights are normalized so that the average field weight is 1 before the model is built.

For example:

Weights supplied	Normalized weights
site_domain: 3.0, hour: 1.0	site_domain: 1.5, hour: 0.5
site_domain: 30, hour: 10	site_domain: 1.5, hour: 0.5

These configurations are equivalent because the relative weighting between the fields is the same.

Increasing all weights by the same factor does not make the model more aggressive. Field weights control the relative influence of fields when the request-level score is calculated. How the resulting bidder scores affect bidding is controlled separately by the line’s threshold settings.

Combination Weights

A two- or three-variable combination uses the mean of its component fields’ weights.

For example, if:

site_domain weight = 1.5
hour weight = 0.5

then a matching site_domain × hour combination receives a weight of:

(1.5 + 0.5) ÷ 2 = 1.0

A combination does not receive additional weight simply because it contains more variables.

Field Weight Rules

Rule	Behavior
Field not listed	Uses the default field weight of 1
Unknown field name	File is rejected
Weight is not numeric	File is rejected
Negative weight	File is rejected
Weight of 0	Accepted with a warning; the field stops contributing to the combined score
One field far above the rest	File is rejected if, after normalization, a single field would overwhelm the score

If the intent is for a field to have no influence, removing its entries is generally clearer than assigning the field a weight of 0.

When to Use Field Weights

Use field weights when there is a specific reason to believe one supported attribute should have more influence than another.

For example, a campaign may have strong evidence that placement domain is more important to performance than time of day. Increasing the relative weight of site_domain allows domain matches to contribute more heavily to the combined score.

If there is no clear reason to weight fields differently, leave field weights out. Equal weighting is the default.

Pontiac Analytics-generated models may also contain field weights, but those weights are measured automatically from the campaign’s training data rather than supplied by the user. When there is not enough data to measure field importance reliably, the Analytics model falls back to equal weighting.

For more information about how field weights are applied during scoring, see Model Scoring Overview.

Threshold Settings

The model file defines the model’s scoring signals and optional field weights, but it does not define how the resulting bidder score affects bidding.

Threshold settings are selected when the model is associated with a Campaign or Line.

• Single Value applies a hard cutoff.
• Spread uses separate minimum and maximum thresholds to create a probabilistic bidding band.

For complete threshold behavior, including how unmatched requests scoring 0.5 are handled, see Model Scoring Overview.

Next Steps

For details about how matching entries are weighted and averaged, why unmatched requests receive a neutral bidder score of 0.5, and how Single Value and Spread threshold settings affect bidding, continue to Model Scoring Overview.

To upload your UDM file, follow the following steps:

1. Navigate to the UDM tab and click New Model.
2. Enter the model’s name and then click Create.

3. Upload a file that meet’s the following conditions:
   • File Type: JSON or CSV
   • Limit: 15,000 entries.
   • File Detection: A first character of [ or { is read as JSON. Anything else is read as CSV. The file extension is ignored.
   • Scores: Must be between -1 and 1. Scores are relative, not absolute. Reserve values near ±1 for the strongest signals.
   • Score of 0: Valid and not ignored. It counts as a match and pulls the combined result toward the neutral midpoint. If no influence is intended, omit the entry instead.
   • Formatting Entries: Values are lowercased and trimmed on both sides, so casing and surrounding spaces never matter. Nothing else is normalised — the comparison is otherwise exact.
   • Missing Entries: Entries whose value is “_missing_”, “missing”, “nan”, “none”, “null” or empty are dropped — the bidder never sends those, so they could never match.
   • Field Weights: Optional. JSON models may include a field_weights map. CSV models may include field-weight rows using the weight column.
   • See the sections above for accepted fields and model-entry formats.

JSON Example

Without custom field weights, the existing JSON array format remains unchanged:

[
  {
    "feature": "geo_zip",
    "value": "10025",
    "score": 0.95
  },
  {
    "feature": "device_type",
    "value": "4",
    "score": 0.3
  },
  {
    "feature": "publisher_name",
    "value": "iheartmedia",
    "score": 0.4
  },
  {
    "feature1": "geo_zip",
    "feature2": "device_type",
    "combination": "10025 × 4",
    "score": 0.9
  },
  {
    "feature1": "geo_zip",
    "feature2": "publisher_name",
    "feature3": "device_type",
    "combination": "10025 × iheartmedia × 4",
    "score": 0.85
  }
]

Notes

• Combination values are separated by x and must appear in the same order as the feature keys.

CSV Example

feature,value,feature1,feature2,feature3,combination,score
geo_zip,10025,,,,,0.95
device_type,4,,,,,0.30
publisher_name,iheartmedia,,,,,0.40
,,geo_zip,device_type,,10025 × 4,0.90
,,geo_zip,publisher_name,device_type,10025 × iheartmedia × 4,0.85

Notes

• Extra columns are ignored, so a value_rankings or 2way_combinations export from any report can be edited and uploaded as-is — add a ‘feature’ column to a value_rankings file to say which field its values belong to.
• Combination values are separated by x and must appear in the same order as the feature keys.

Field Groups and Values

Field	Group	How to write the value
geo_region	Geography	Lowercase, exactly as it appears in the logs.
geo_zip	Geography	Postal code as it appears in the logs. US ZIPs must keep their leading zero (‘04501’). Non-US formats are fine (‘1000-001’, ‘SW1A 1AA’).
publisher_name	Inventory	Lowercase, exactly as it appears in the logs.
site_domain	Inventory	Lowercase, exactly as it appears in the logs.
app_bundle	Inventory	Lowercase, exactly as it appears in the logs.
exchange	Inventory	Lowercase, exactly as it appears in the logs.
deal_id	Inventory	Lowercase, exactly as it appears in the logs.
content_channel	Content	Lowercase, exactly as it appears in the logs.
content_network	Content	Lowercase, exactly as it appears in the logs.
content_series	Content	Lowercase, exactly as it appears in the logs.
content_title	Content	Lowercase, exactly as it appears in the logs.
content_genre	Content	Lowercase, exactly as it appears in the logs.
content_livestream	Content	‘0’ or ‘1’.
device_type	Device	IAB code as a whole number — ‘3’, never ‘3.0’ and never ‘Connected TV’.
day	Time	Lowercase three-letter code (mon, tue, wed, thu, fri, sat, sun), in UTC.
hour	Time	Whole number 0-23, in the user’s LOCAL time.

4. Click the Validate & upload button.
5. Confirm upload was valid and accepted.
   • Success Message: Uploaded. Tells the user how many records and how many 1-field, 2-field, and 3-field records the file contains. The model is now live wherever it is already in use.
   • Replace file: Users can upload a new model that will replace the model everywhere it is already in use. Nothing needs to be re-attached, and a file that fails validation leaves the current one running.

6. Associate the uploaded model with the desired Campaign or Line by following: Model Association

The Planning Tool allows users to quickly create a media plan with the help of LLM-powered research and recommendations. Users can provide information about their brand, campaign goals, target audience, geography, budget, and other relevant details, along with optional supporting documents.

Based on these inputs, the tool researches the brand, identifies potential audience and content alignment, and recommends inventory. It also uses Pontiac’s sample bid request data to match the campaign’s goals and targeting strategy with available inventory and expected volume, helping ensure that recommendations are grounded in real marketplace availability.

Users can review and refine the proposed plan before using it to launch campaigns directly in the platform.

Describe the campaign you want to run. The system will research the brand, extract a target audience, and pick taxonomy categories you can review and edit before generating the plan.

Instructions for generating a Media Plan. Follow the steps below:

1. Name: Enter a name for the Planning report. For example: McDonald’s – Q3 National Push.
2. Brand Site: Enter the brand’s site. This is an optional field, but if provided, the LLM may fetch it for additional context.
3. Inventory Types: Select at least one inventory type: CTV, Display / Banner, or Audio. Budget is optional for each selected inventory type. Users can also enter a target number of impressions per month.
4. Description: Enter a description of the campaign, including relevant details such as the brand, goals, media types, geography, budget, and desired scale.
   • Example: CTV plan for McDonald’s targeting the U.S. nationally with a $500K monthly budget. Research the brand to identify the target demographic and relevant content alignment.
   • Users can enter free-text instructions and request tool use, such as “Search the internet for…” or “Summarize the brand site.”
5. Supporting documents: Upload optional supporting files in a text-readable format, including .txt, .md, or .csv. Files can be up to 200 KB each. A short description is required for each upload so the LLM can understand the document’s purpose and how it should be weighted when generating the plan.
6. Click Submit.
7. Once the report is submitted, it will process and the status will change to Draft.
8. Move onto the next set of steps: New Plan Draft

Continue building your Media Plan in the Draft Section:

1. Audience Identified: Provides a summary of the audience the plan is designed to reach, based on the campaign inputs, brand research, and any supporting context provided. This may include the brand or product positioning, primary and secondary audience segments, demographic characteristics, income range, interests, behaviors, lifestyle traits, and other relevant audience signals.
   • Example fields may include:
     • Demo: Primary and secondary audience descriptions.
     • Age: Recommended age range.
     • HHI: Recommended household income range.
     • Lifestyle: Relevant interests, activities, behaviors, and affinities associated with the target audience.

2. Inventory Plan: For each selected inventory type, the Planning Tool generates recommended content veins based on the campaign goals, audience, brand context, and available inventory. Users can review and edit these recommendations before building the final plan. Users are not required to proceed with every inventory type originally selected during setup. For example, if both CTV and Audio were included initially, a user can remove Audio and continue with only the CTV plan.
3. Content Veins: Content veins represent distinct audience or inventory strategies within an inventory type. Each vein includes:
   • Name: A descriptive label for the strategy.
   • Tier: Indicates the role of the strategy within the plan, such as premium endemic, value endemic, or premium reach.
   • Description: Explains the intended audience, content environment, and strategic purpose of the vein.
   • Taxonomy Picks: Structured targeting categories selected to support the strategy. Each category includes a weight indicating its relative importance to the content vein. Users can remove existing categories or add additional categories.
   • LLM Free-Form Hints: Additional contextual signals generated by the LLM to help identify relevant inventory that may not be fully represented by the structured taxonomy.
   • Users can review and modify the recommended content veins, by changing the weights, or removing taxonomy picks.

4. Buy Constraints: These constraints are used when matching the plan to available inventory and expected volume.
   • Target Impressions per Inventory Type: Set the desired monthly impression volume for each selected inventory type. Values may be pre-populated based on the budgets or impression goals entered during setup and can be adjusted before the plan is built.
   • Line Breakout: Users can choose how recommended inventory is organized into campaign lines.
     • Fan Out by Inventory Axis (Recommended): Creates separate lines for each content or inventory axis, such as channels, series, networks, apps, or domains. This provides more granular visibility and pacing control but may create more lines.
     • Single Line per Tier (Multifield): Creates one line per tier and combines multiple content values into a broader multi-field target. This reduces the number of lines to manage but provides less granular pacing and reporting by individual inventory axis.
     • Add an Open Exchange Line: Adds a lower-weight open exchange line using the broader campaign targeting, such as geography and device, to help capture additional scale outside of the curated inventory recommendations.
   • Geo Constraints: Users can further define where the plan should run by entering:
     • Countries using ISO country codes
     • State or region codes Nielsen DMA codes
     • ZIP codes (Max 1,000)
5. Click the Save Draft to save the contents and come back to it later.
6. Click the Build Plan button when ready to active and generate the plan.
7. Move onto the following steps: Plan Results
   

Once the plan is built, users can review the generated strategy, campaign structure, inventory recommendations, projected scale, and downloadable deployment files. The Results section is split into three subtabs:

• Executive Summary: Provides a high-level overview of the generated media plan, including the campaign brief, brand and audience insights, planning constraints, projected scale, room to scale, and a summary of the generated output files.
• Campaign: Provides an editable view of the proposed campaign structure. Users can review inventory types, lines, bids, budgets, targeting, recommended deals, and availability estimates, and can click into each line to review or modify its settings.
• Downloads: Provides access to the complete set of generated planning and deployment files, including human-readable plans, spreadsheets, Hawking import files, raw CSVs, and structured JSON outputs.

Executive Summary

The Executive Summary provides a high-level overview of the generated media plan, including:

• Brief & Brand: Summarizes the campaign brief, brand context, objectives, competitive landscape, flight dates, budget, and intended media strategy.

• Audience Identified: Describes the recommended audience, including demographics, lifestyle attributes, interests, and how those signals map to the proposed inventory strategy.

• Constraints: Summarizes the geo, inventory types, budget, flight dates, and other planning requirements.

• Scale Delivered: Shows projected impressions, spend, blended CPM, and line count by inventory type.

• Room to Scale: Evaluates available inventory and capture rates to identify where additional scale may be available and where supply may be more limited.

• Output Bundle: Describes the files generated with the plan and how they can be used for review, analysis, or campaign deployment.

Campaign

The Campaign tab provides an editable view of the proposed campaign structure.

The Inventory Summary shows the number of lines, projected monthly impressions, projected monthly spend, and share of spend by inventory type.

Users can select an inventory type to view its recommended lines. Each line displays:

• Line name
• Bid CPM
• Monthly impression goal
• Estimated monthly spend

Selecting a line opens the full Line Details, where users can review and edit settings such as:

• Bid CPM
• Monthly impression goal
• Estimated spend per month
• Budget allocation weight
• Device targeting
• Frequency cap
• Geo targeting: Countries, Regions / States, DMAs, ZIPs
• Inventory targeting: Channels, Series, Networks, Apps, and Domains

• Sample availability and supply estimates

• Deals that the inventory will be bought through
• Expected capture rate
• Average deal floor

The plan uses sample bid request data to connect the recommended content and inventory strategy with observed marketplace supply. This helps show not only what inventory is contextually aligned with the campaign goal, but also how much relevant inventory may be available.

Users can click through each proposed line to review and refine the settings before saving or launching the campaign.

Downloads

The Downloads tab provides the complete output bundle generated for the plan. These files are helpful for sharing and audit, but follow the documentation here to implement the Media Plan: Plan to Action

Recommended

• plan.md: Human-readable version of the media plan

• plan.xlsx: Spreadsheet version of the plan for review and sharing

• plan_bundle.zip: Complete bundle of generated planning files

Hawking Import

• hawking_import.json: Structured Hawking API JSON containing the campaigns, lines, and inventory lists required to deploy the plan

Raw CSVs (Hawking-shaped)

• campaigns.csv : Campaign records to create
• lines.csv: Line records to create
• inventory_lists.csv: Site, app, and deal list records
• inventory_list_items.csv: Individual items contained within each inventory list
• deals.csv: Deal-level details and QA information

Other

• plan.json: Full structured version of the generated plan
• brief_text.txt: Original campaign brief used to generate the plan

These files allow users to review the strategy at different levels of detail, share the plan across teams, conduct additional analysis, or use the structured outputs to deploy the recommended campaign setup.

Once a Planning report has been reviewed and is ready to use, it can be imported directly into a campaign. Follows these steps to launch your media plan:

1. Import the Plan at the Campaign Level

• Navigate to the campaign level and click Import.

• In the pop-up window, find and select the desired Planning report by name.

• Click the Planning report name.

2. Review and QA the generated campaign to ensure all settings, targeting, budgets, and line configurations are as desired before launch.

The following campaign-level settings are selected by default when a Planning report is imported:

• Manage Budgets
• Auto Pace
• Manage Creatives
• Manage Dayparts
• Manage Pixels

All lines included in the Planning report are also automatically populated as part of the import.

3. Associate your Creatives.

4. Associate Audience Segments if desired.

5. Associate your Pixels.

6. Once the campaign has been fully reviewed and all desired creatives, audiences, and pixels have been associated, click Submit.

The Audience model uses demographic regression and persona clustering to understand who is most likely to convert. It identifies which census-based features are predictive of performance and groups users into distinct audience segments. All modeling is based on aggregated, privacy-safe data at the ZIP code level, with no use of cookies or personally identifiable information (PII).

The Dashboard tab highlights which census features predict conversions, along with scored ZIP codes and lookalike expansion targets. The Personas tab segments audiences into clusters with shared characteristics and geographic distribution. These insights are currently available for the United States only, leveraging standardized data from the United States Census Bureau to ensure consistency and accuracy.

Use Cases: Understand, Activate, Scale

1. Understand Your Audience: Define who your customers are and what drives them to convert.
   • Understand Your Audience Before You Launch (Audience Mode)
     • Use site visitor data to uncover who your audience is, where they’re located, and what defines them—so you can build smarter strategy before spending media dollars.
   • Find Where Your Best Customers Are (ZIP List Mode)
     • Identify high-performing ZIP codes weighted by revenue or LTV to understand what makes them valuable—so you can prioritize the geographies that matter most.
   • Turn Data into Clear, Actionable Audiences (All Modes)
     • Segment users into distinct personas with shared traits and geographic patterns—so you can translate data into clear targeting and messaging.
2. Activate with Confidence: Turn insights into smarter targeting and campaign execution.
   • Plan Smarter Campaigns with Data-Backed Insights
     • Use demographic and geographic signals to guide targeting, messaging, and channel strategy—so every campaign is built on what actually drives performance.
   • Connect Personas to Performance (Campaign / ZIP List Modes)
     • Link high-performing ZIP codes to persona segments—so you can activate and optimize at the audience level, not just the campaign level.
3. Scale What Works: Grow efficiently while maintaining performance.
   • Measure, Learn, and Optimize in Real Time (Campaign Mode)
     • Analyze past and live campaigns to identify what’s driving conversions—then optimize targeting and spend while campaigns are running.
   • Scale What Works with Lookalike Expansion (Campaign / ZIP List Modes)
     • Use top-performing audiences or geographies to find similar, high-potential segments—so you can expand reach without sacrificing efficiency.

Choose how you want to analyze your audience, whether through campaign performance, site visitors, or custom ZIPs, then configure dates, optimization type, and inputs to generate the report results and model.

Instructions for generating a Audience Report and Model. Follow the steps for initial setup below:

1. Navigate to the Audience Tab.
2. Click New Report button.
3. Give report a name.
4. Select the Advertiser to run the analysis on.
5. Select the Analysis Mode:
   • Campaign – analyze ad performance based on campaign delivery.
     • Requires campaigns and lines with media served.
     • Identifies which audiences and demographics drive conversions.
   • Audience – profile site visitors using pixel data only.
     • Requires pixel fires on the advertiser’s site.
     • No campaign or line selection required.
     • Tip: Can be used for pre-campaign analysis to understand your audience before launch.
   • ZIP List – profile a custom set of ZIP codes with weights.
     • Model learns which demographics correlate with higher-weighted ZIPs.
     • Tip: Useful for profiling ZIPs based on sales, revenue, or LTV.
6. Choose whether the report should run once or repeat on a schedule.
   • Run once builds a single report.
   • Schedule re-runs it on the chosen cadence — each run retrains the model, replaces the live model in S3 when Replace the live model each run is enabled, and is saved as a new run grouped under this report.
     • Rerun Interval:
       • Weekly
       • Every 2 Weeks
       • Monthly
     • Rerun Day of Week:
       • Monday
       • Tuesday
       • Wednesday
       • Thursday
       • Friday
       • Saturday
       • Sunday
     • Replace the live model each run checkbox
       • On: each run’s new model overwrites the live model so it goes live automatically.
       • Off: runs still produce reports and archived models, but the live model is left untouched.
7. Click Save Report button.

Campaign Mode Example

After selecting Campaign Mode, complete the following:

1. Select the Campaign(s) within the chosen Advertiser.
   • If no campaigns are selected, all campaigns will be included.
2. Select the Line(s) within the chosen Campaign(s).
   • If no lines are selected, all lines will be included.
3. Optimization Type:
   • Conversions: Choose the conversion pixel to be used for optimization.
4. Select the analysis Start Date.
5. Select the analysis End Date.
6. Enter the desired lookback window (in days).
   • Default is 30 days.
   • This defines how far back the model will attribute conversions or clicks to ad exposure.
7. Click the Save Report button.

Audience Mode Example

After selecting Audience Mode, complete the following:

1. Select the Optimization Type:
   • Conversions
     • If selected, choose the conversion pixel to be used for optimization.
   • Clicks
2. Select the analysis Start Date.
3. Select the analysis End Date.
4. Enter the desired lookback window (in days).
   • Default is 30 days.
   • This defines how far back the model will attribute conversions or clicks to ad exposure.
5. Click the Save Report button.

ZIP List Mode

After selecting ZIP List Mode, completed the following:

1. Enter ZIP codes with weights.
   • One per line or comma-separated. Format: ZIP:weight (weight defaults to 1 if omitted).
   • Example input: 11215:2, 33138:8
2. Click the Save Report button.

The report results include the pipeline used, date created, when the report started, and when the report completed. The output is separated into 7 sections:

1. Summary
2. Dashboard
3. Lookalike
4. Personas
5. Explorer
6. Charts
7. Downloads

This information provides a complete audit trail and transparency into the model, allowing users to understand how insights are generated and tie them back to audience behavior and performance, rather than relying on a black-box approach.

Summary

The Summary section provides a high-level overview of the analysis, including the Report ID, Advertiser ID, and Date Range, along with an automatically generated Executive Summary.

This summary explains:

• Key demographic drivers of conversions
• Audience composition and persona breakdown
• Geographic performance and distribution
• Lookalike expansion opportunities (if applicable)
• Strategic recommendations and action items

Tip: You can input the Executive Summary into your preferred LLM (e.g., ChatGPT or Claude) to quickly generate a presentation deck or case study based on the results.

Dashboard

The Dashboard section provides a visual overview of audience composition and demographic drivers of performance. It highlights which census-based features influence conversions, how audiences are distributed geographically, and how different demographic segments perform.

Dashboard Sections:

• Tile Metrics
• Most Important Demographic Categories
• Feature Impact Distribution — SHAP Beeswarm
• Demographic Sub-Feature Impact
  • Shows the top three most important features based on your audience dynamically.
• Geographic Performance — Click a State to See ZIP Codes
• Cost per Conversion by Demographic Category

Tile Metrics

• Analysis Model: Campaign Mode — results are based on ad-attributed conversions. Impressions served by your lines that led to pixel fires within the lookback window. The model learns which demographics respond to your ad.
• Optimization: Conversions or Clicks
  • Conversions: Optimizing for conversions (pixel fires). The model predicts which demographics drive conversion events.
• R² (Coefficient of Determination): Measures how much of the conversion rate variance is explained by demographics. 5–15% is typical. Demographics are just one signal among many.
• RMSE (Root Mean Squared Error): Lower is better. Measures the average prediction error in the same units as the target (conversion rate).
• MAE (Mean Absolute Error): Average absolute difference between predicted and actual conversion rates. Less sensitive to outliers than RMSE.
• Scored ZIPs: Total ZIP codes scored with demographic affinity. Example: 1,850 positive, 25,483 negative or zero (this figure changes dynamically).
• Total Spend: Total ad spend across all campaign ZIPs in this analysis period.
• Avg Cost/Conv: Average cost per conversion across all campaign ZIPs. Lower is more efficient.
• Samples: Number of ZIP codes included in the analysis after filtering for (>300 events) and outlier removal.
• Features: Number of census demographic features used by the model.
• Reliability: Pipeline quality assessment. Passed 3/4 checks — the results change dynamically.

Most Important Demographic Categories

This section shows the total SHAP importance aggregated by demographic category, helping identify which broad demographic dimensions have the greatest influence on the model’s conversion rate predictions. Use this to understand which broad demographic dimensions matter the most.

Chart View

What It Shows:

• Each bar represents a demographic category (e.g., Household Income, Age, Housing, Transportation)
• Bar height reflects total SHAP importance across all features within that category

Interpretation:

• Taller bars indicate categories that have a greater impact on the model’s predictions
• Categories at the top represent the strongest drivers of conversion likelihood
• Lower bars indicate categories with less influence on performance

Table View

Provides a detailed breakdown of individual features within each demographic category, including:

• Category: High-level demographic group
• Feature: Specific sub-feature (e.g., income bracket, age range)
• Mean Abs SHAP: Overall importance of the feature
• Mean SHAP: Direction of impact (positive or negative influence)

Users can download this table as a CSV file for further analysis.

Feature Impact Distribution — SHAP Beeswarm

This section shows how each demographic feature impacts conversion predictions across all ZIP codes.

What It Shows:

• Each dot represents a single ZIP code prediction
• X-axis shows SHAP value (impact on predicted conversion rate)
• Color represents the feature value:
  • Red = higher values
  • Blue = lower values
• Features are ordered by importance (top = most important)

Interpretation:

• Points to the right increase predicted conversion rate
• Points to the left decrease predicted conversion rate
• Color helps identify whether higher or lower values drive performance
• Clusters indicate consistent impact patterns across ZIP codes
• Wide spread shows variable impact, while tight clusters show consistent behavior

Example Interpretation: This audience skews toward non-married individuals and people aged 30 to 40, who are more likely to convert, while higher-density housing areas tend to underperform.

• Households Marital Status: Never Married
  • Takeaway: Prioritize ZIPs with higher concentrations of never-married populations as this is a strong positive signal for conversion.
    • Many points extend to the right, indicating this feature often increases predicted conversion rate
    • Higher values (red) are more concentrated on the right, showing that higher concentrations of never-married populations drive performance
    • The distribution is relatively wide, meaning the impact varies across ZIP codes
    • Visible clustering suggests consistent positive patterns in certain regions
• Housing Type 10 to 19
  • Takeaway: Deprioritize areas with high concentrations of this housing type (mid- to high-density housing such as apartments or condos). Lower concentrations tend to perform more consistently.
    • Points are distributed on both sides of zero, indicating this feature can both increase and decrease predicted conversion rate
    • Red values (higher concentrations) extend more to the left, showing that higher values tend to decrease conversion likelihood
    • Blue values (lower concentrations) are more clustered on the right, indicating lower values tend to increase conversion likelihood
    • The spread is moderately wide, suggesting variable impact across ZIP codes
    • Clusters near the center indicate more neutral or mixed performance overall
• Age 30 to 40
  • Takeaway: This is a stable, reliable positive signal and a good baseline demographic to include in targeting.
    • Points lean slightly to the right, indicating a generally positive impact on conversion rate
    • Higher values (red) tend to appear more on the right, suggesting higher concentrations in this age group improve performance
    • The distribution is more tightly clustered, indicating a more consistent effect across ZIP codes
    • Less spread means this feature behaves more predictably compared to others

Demographic Sub-Feature Impact

This section breaks down a top demographic category into its individual sub-features to show how each one influences conversion predictions.

Household Income Example

This chart is a SHAP Beeswarm for the Household Income category.

What It Shows:

• Each dot represents a ZIP code prediction
• Each row represents a specific sub-feature
• X-axis shows SHAP value (impact on predicted conversion rate)
• Color represents the feature value:
  • Red = higher values
  • Blue = lower values
• Features are ordered by importance (top = most important)

Interpretation:

• Points to the right increase predicted conversion rate
• Points to the left decrease predicted conversion rate
• Color helps identify whether higher or lower values drive performance
• Clusters indicate consistent impact patterns across ZIP codes
• Wide spread shows variable impact, while tight clusters show consistent behavior

Example Interpretation: This audience skews toward low to mid-income households, which are more likely to convert than higher-income segments.

• Household Income $125,000 to $149,999
  • Takeaway: Higher concentrations of this income group tend to decrease conversion likelihood, while lower presence performs better.
    • Red points (higher values) are concentrated left of zero, indicating a negative impact
    • Blue points extend more to the right, showing lower concentrations improve performance
    • Wide spread indicates variable impact across ZIP codes
• Household Income $25,000 to $29,999
  • Takeaway: This income bracket is a strong positive signal when present at higher levels.
    • Red points cluster to the right, indicating higher values increase predicted conversion rate
    • Blue points appear more on the left, showing lower values reduce performance
    • Moderate spread suggests some variability, but generally consistent direction
• Household Income $50,000 to $59,999
  • Takeaway: A strong and reliable positive segment worth prioritizing.
    • Red points are clearly right-skewed, showing higher concentrations drive conversions
    • Blue points are more left or neutral, indicating weaker performance when absent
    • Wider spread shows impact varies, but direction is consistently positive

Geographic Performance — Click a State to See ZIP Codes

This section provides an interactive map of conversion performance by geography, helping identify where your audience is performing best.

What It Shows:

• States and ZIP codes colored by conversion rate
• Visual distribution of performance across regions and local markets
• Ability to analyze performance at both state and ZIP-level
• Toggle between Heatmap and Boundaries view:
  • Boundaries: Clearly outlines geographic regions
  • Heatmap: Highlights performance intensity across regions

Boundaries View

Heatmap View

Interactions:

• Hover over a state to view:
  • State Name
  • Conversion Rate
  • Impressions
  • Attributed Impressions
  • Unique Conversions
  • Unique Users
• Click a state to drill down into ZIP level performance – Zip code view:
  • Zip code
  • Zip code name
  • Conversions
  • Impressions
  • Attributed Impressions
  • Unique Conversions
  • Unique Users
  • Predicted Rate

Zip Code View

Table View

This table contains the underlying data used to power the Geographic Performance map, providing detailed metrics at the ZIP code level. Users can download this table as a CSV file for further analysis.

Each row represents a ZIP code and its associated performance, which is aggregated to render state-level views in the map.

What It Includes:

• Geographic identifiers
  • Geo Zip
  • Geo Region Name
  • City
  • State ID
• Delivery Metrics
  • Impressions: Total number of times ads were served
  • Attributed Impressions: Impressions tied to users who later converted within the lookback window
  • Unique Users: Number of distinct users exposed to ads
• Performance Metrics
  • Conversions: Total number of conversion events attributed to the campaign
  • Conversion Rate: Attributed Impressions / Impressions
  • Attributed Conversion Rate: Conversions / Impressions

Example insight:

• 07103 (NJ) shows a Conversion Rate (2.20%) and Attributed Conversion Rate of 0.30%, indicating performance well above average (0.077% conversion rate from Executive Summary for this data).

Cost per Conversion by Demographic Category

This section shows how cost efficiency changes based on how strongly a demographic category is represented in a ZIP, helping identify which types of areas deliver the best ROI.

What It Shows:

• Each demographic category (e.g., Income, Age, Marital Status) is broken into quartiles based on its most important sub-feature
• Each bar represents a quartile segment (Low to High)
• Bar height represents cost per conversion (CPA)

Interpretation:

• Lower bars = cheaper conversions (more efficient)
• Higher bars = more expensive conversions (less efficient)
• Compare across quartiles to see how performance changes as the concentration of the demographic increases or decreases within a ZIP
• Differences across categories show which demographic dimensions drive efficiency vs. cost

Example Interpretation: Balance scale and efficiency by prioritizing ZIPs with strong marital status and age signals, while reducing spend in areas where income signals increase CPA.

• Households Marital Status
  • Takeaway: ZIPs where this category has a stronger presence are more cost-efficient, while weaker presence is expensive.
    • The Low quartile has the highest CPA ~$4.05, indicating poor efficiency
    • CPA decreases steadily across quartiles, with High segments being much cheaper to ~$1.21
    • Indicates areas where marital status signals are stronger perform more efficiently
• Household Income
  • Takeaway: ZIPs where income-related signals are stronger (based on the model’s top feature) tend to be less cost-efficient.
    • CPA increases from Low to High quartiles (bins) (~$1.2 to ~$3.15)
    • Indicates stronger income signal areas are more expensive to convert
    • Lower-signal areas are more efficient
• Age
  • Takeaway: ZIPs with stronger age-related signals are more cost-efficient.
    • CPA decreases from Low to High quartiles
    • Indicates areas where age is more predictive deliver cheaper conversions

Table View

The table provides the underlying data behind the chart, showing how cost and performance vary across demographic category bins. Users can download this table as a CSV file for further analysis.

What It Shows:

• Each row represents a demographic category and its top contributing sub-feature
• Data is broken into quartile (bins): Low to High based on how strongly that feature is represented within each ZIP
• Columns included:
  • Category: High-level demographic group (e.g., Household Income, Age, Marital Status)
  • Top Feature: The most important sub-feature within that category driving the segmentation (e.g., a specific income bracket or age group)
  • Bin: Quartile grouping (Low, Medium-Low, Medium-High, High)
    • Represents relative concentration of that feature within a ZIP, not the value itself.
    • Bins are calculated independently for each category based on that category’s top feature.
  • ZIP Count: Number of ZIP codes in that bin
  • Impressions: Total impressions delivered across those ZIPs
  • Conversions: Total conversions generated in those ZIPs
  • Total Spend: Total media spend across those ZIPs

Lookalike

Lookalike targeting uses the trained demographic model to predict conversion rates for all ~33,000 US ZIP codes not in your campaign. ZIPs with demographics similar to your high-performing areas are identified as expansion targets. Confidence tiers reflect prediction reliability — high confidence ZIPs are within the model’s training distribution with consistent predictions across all decision trees.

This section is only available when your campaign is not running nationally and is not already serving all ZIP codes, as lookalike modeling requires out-of-sample areas for expansion.

Tile Metrics

• Lookalike ZIPs: Total ZIP codes in the U.S. not in your campaign that were scored for demographic similarity.
• High Confidence: ZIPs within the training distribution, with low prediction variance and a positive score. Most reliable for targeting expansion.
• Medium Confidence: ZIPs within the training distribution but with higher prediction uncertainty or neutral scores. Consider for testing.
• Low Confidence: ZIPs that are demographically different from your campaign/analysis ZIPs (out-of-distribution). Predictions are unreliable. Shown on the map in gray for context, but not recommended for targeting.
• States Covered: Number of U.S. states with at least one lookalike ZIP.
• Est. Additional Reach: Estimated additional unique IPs reachable in high-confidence lookalike ZIPs. Based on average IPs per campaign ZIP — treat as a rough estimate.
• Model Confidence: Model quality weight (e.g., 3.4%). All lookalike scores are dampened by this factor to reflect model uncertainty.

Lookalike Expansion Map — Click a State to See ZIP Codes

This section provides an interactive map of lookalike opportunities, showing where to expand based on demographic similarity to your best-performing ZIPs.

What It Shows:

• ZIP codes colored by confidence tier:
  • High (green): Strongest expansion opportunities
  • Medium (yellow): Test-and-learn opportunities
  • Low (gray): Out-of-distribution, not recommended
• Geographic distribution of lookalike audiences across the U.S.

Toggle Tiers

Show or hide High, Medium, ad Low confidence ZIPs for easy investigation and comparison.

Toggle Map View

Toggle views:

• Boundaries: Clearly outlines ZIP/state regions
• Heatmap: Highlights density and intensity of lookalike opportunities

Boundaries View

Heatmap View

Zip Code View

For each state, click onto it so see how the ZIP codes within that specific state map to the confidence tiers.

Lookalike ZIP Details

Full table of lookalike ZIPs sorted by score. Filter by confidence tier, search by city or ZIP. High confidence ZIPs are the safest expansion targets. Users can download this table as a CSV file for further analysis.

Columns:

• Geo ZIP: ZIP code being evaluated
• Predicted Rate: Model-predicted conversion rate for that ZIP
• Prediction Std: Standard deviation of predictions across decision trees
  • Lower = more consistent predictions
  • Higher = more uncertainty
• Confidence: Model confidence score (0–1) based on prediction stability and similarity to training data
• Score: Final model score used for ranking ZIPs
  • Combines predicted performance and confidence
  • Higher = better expansion opportunity
• Lift vs Avg: Predicted performance relative to campaign average
  • Positive = above average
  • Negative = below average
• Unique IPs: Estimated number of unique users in that ZIP (if available)
• Confidence Tier:
  • High: Reliable, within training distribution
  • Medium: Moderate confidence, test recommended
  • Low: Out-of-distribution, not recommended
• Observed Impressions: Number of impressions seen in this ZIP (if any historical data exists)
• Needs More Data:
  • TRUE: Limited data, predictions less stable
  • FALSE: Sufficient data for more reliable estimates
• City: Associated city name
• State ID: State abbreviation

Interpretation:

• Prioritize:
  • High confidence, high score, and positive lift are the best expansion targets
• Be cautious with:
  • High score but low confidence should be tested before scaling
• Use:
  • Prediction Std and Needs More Data to gauge reliability
  • Lift vs Avg to benchmark against current campaign performance

Personas

The Personas section groups your audience into distinct segments using K-means clustering based on shared demographics. It highlights how each cluster performs, what defines it, and where it is concentrated, helping identify the personas that drive the most value.

Persona Sections:

• Tile Metrics
• Persona Performance — Top 10% vs Next 30% vs Bottom 60%
• Audience Clusters — PCA Projection
• Geographic Distribution — Click a State to See ZIP Codes
• Most Important Characteristics Across Personas
• What Differentiates Each Audience — Category Importance
• Persona Comparison (Dynamic Top 4 Features)
  • Shows the most important differentiating features across personas based on your audience’s personas
• Per-Audience Deep Dive

Tile Metrics

• Audiences: Number of distinct audience personas identified by K-means clustering. Each persona represents a group of ZIP codes with similar demographic profiles.
• Total ZIPs: Total number of converting ZIP codes that were clustered into audience personas.
• Avg. Audience Size: Average number of ZIPs per audience persona.
  • Large variance may indicate a mix of broad (dominant) and niche (specialized) personas
• Largest Audience: Identifies the largest persona segment and its share of total clustered ZIPs.
  • Includes the top defining trait for that persona
  • Helps quickly understand the most dominant audience type in your campaign
• Persona Tiles: Each tile represents an audience segment with:
  • A descriptive name
  • A short explanation of its key distinguishing features
  • ZIP count and percent of total ZIPs to indicate size
  • Conversion rate and relative lift vs. average to indicate performance

Persona Performance — Top 10% vs Next 30% vs Bottom 60%

This section shows how performance is distributed within each persona by ranking ZIP codes based on conversion rate, conversion volume, or unique users, and splitting them into three tiers.

What It Shows:

• Each persona’s ZIPs are ranked based on the selected metric:
  • Conversion Rate: Efficiency of each tier
  • Conversion Volume: Total conversions generated per tier
  • Unique Users: Audience size within each tier
• ZIPs are split into three tiers:
  • Top 10%
  • Next 30%
  • Bottom 60%
• Bar charts display performance for each persona across these tiers
• The gray horizontal dotted line indicates the average conversion rate

Interpretation:

• A tall Top 10% bar indicates that a small portion of ZIPs drives a large share of performance
• A more even distribution across tiers suggests consistent performance across the persona
• The Bottom 60% highlights lower-performing areas or opportunities for optimization
• Identifies whether performance is concentrated or evenly distributed within a persona
• Helps determine if a persona should be:
  • Scaled selectively (focus on top-tier ZIPs)
  • Scaled broadly (consistent performance across tiers)
• Enables ZIP-level optimization within each persona

Audience Clusters — PCA Projection

Each dot represents one ZIP code projected into 2D space using Principal Component Analysis (PCA). PCA compresses the 130 demographic features into two dimensions that capture the most variance in the data.

• PCA 1 (x-axis) is the single direction that best separates the data. It typically captures the strongest demographic contrast (e.g., urban vs. suburban).
• PCA 2 (y-axis) captures the next most important contrast, orthogonal to PCA 1.

Colors represent audience persona assignments. Well-separated groups indicate distinct personas, while overlapping areas suggest shared demographic characteristics.

Chart View

• Each point represents a ZIP code
• Position reflects similarity in demographic composition
• Color indicates persona cluster assignment
• Clusters show how clearly personas are differentiated:
  • Tightly grouped clusters are more consistent personas
  • Overlapping clusters have more shared characteristics across audiences

Table View

Provides the underlying data for each plotted ZIP code. Users can download this table as a CSV file for further analysis.

Columns:

• Geo ZIP: ZIP Code
• Persona: Assigned audience persona name
• Performance Tier: Top 10%, Next 30%, or Bottom 60%
• PCA 1: X-axis coordinate from PCA projection
• PCA 2: Y-axis coordinate from PCA projection

Tip: Download the CSV to easily create ZIP code lists by persona and performance tier, enabling direct activation for targeting, exclusions, or testing strategies.

Geographic Distribution — Click a State to See ZIP Codes

This section provides an interactive map showing where each audience persona’s users are located, helping you understand geographic concentration and regional patterns.

What It Shows:

• Geographic distribution of users across all combined audience personas and each persona individually
• Visual representation of where each persona is most concentrated
• Ability to compare geographic footprints across personas

Interactions:

• Filter by Audience:
  • View All Audiences or isolate a specific persona
• Toggle Views:
  • Boundaries: Clearly outlines geographic regions
  • Heatmap: Highlights user density and concentration
• Hover to view ZIP-level details
  • Conversion Rate
  • Impressions
  • Attributed Impressions
  • Unique Conversions
  • Unique Users
• Click a state to drill down into ZIP-level data

Toggle by Audience

View All Audiences or isolate a specific persona.

Boundaries View

Heatmap View

Zip Code View

Table View

Provides detailed data for each ZIP code. Users can download this table as a CSV file for further analysis.

Columns:

• Geo ZIP: ZIP code
• Cluster: Numeric assigned audience persona
• Lon: Longitude (geographic coordinate)
• State ID: State abbreviation
• Demographic Score: Relative strength of demographic alignment for that persona

Most Important Characteristics Across Personas

This section compares the top demographic categories across all personas, showing which features are most important in defining each audience segment.

Chart View

What It Shows:

• Top demographic categories ranked by overall importance
• Each category is displayed with one bar per persona
• Bar height represents the importance of that category for that specific persona

Interpretation:

• Taller bars indicate that a category is more important in defining that persona
• Compare across personas to see which demographics:
  • Are shared across multiple personas
  • Are unique drivers for specific audiences
• Categories with consistently high bars across personas represent broad drivers, while variation highlights key differences between segments

Table View

Provides the underlying data for each persona and its most important demographic categories. Users can download this table as a CSV file for further analysis.

Columns:

• Cluster: Persona ID (corresponds to each audience segment)
• Category: Demographic category (e.g., Household Income, Housing Type, Education Level)
• Importance: Relative importance score of that category for the persona
  • Higher values indicate stronger influence in defining that persona
  • Values are comparable within and across personas

This provides a clear view of what drives each persona at a category level, supporting persona naming, messaging, and targeting strategies while validating insights from the chart with precise importance values.

What Differentiates Each Audience — Category Importance

This section highlights which demographic categories (e.g., income, education, age, housing) are most important for distinguishing each audience persona and how their importance differs across personas.

The analysis is based on SHAP values from per-audience Random Forest models, showing which features most strongly define membership in each persona. This presents the same underlying data as the category importance views, but in a heatmap-style, comparative format designed to emphasize how personas differ from one another.

Chart View

Table View

Provides the underlying data behind the heatmap, showing the importance of each demographic category for each audience persona.

Persona Comparison

This section compares all audience personas across the top 4 most differentiating features, helping highlight how each audience over- or under-indexes relative to the national average.

Household Income Example

Example Interpretation: Household Income

• Affluent Homeowners: Strongly over-indexes in higher income brackets.
  • Positive values for $150K+ and $200K+ indicate this audience skews toward high-income households
  • Under-indexes in lower income ranges, reinforcing the affluent profile
• Urban Renters: Shows a polarized income profile typical of urban markets.
  • Over-indexes in both $200K+ and less than $10K brackets
  • Under-indexes across many middle income ranges
  • Reflects a mix of high-income urban professionals and lower-income renter populations, a common pattern in dense urban areas
  • Additionally, housing type and house size are more influential for this persona, aligning with an urban renter profile where living structure plays a key role
• Budget Households: Skews toward mid-to-lower income ranges.
  • Positive values for sub-$50K ranges indicate strong presence in lower-income households
  • Negative values for higher income brackets highlight clear differentiation from affluent audiences

Per-Audience Deep Dive

This section allows you to explore each persona individually, showing what defines the audience and where it is located.

Select a persona to view its feature importance, category-level drivers, SHAP impact, and geographic footprint. Toggle between personas to dynamically update all charts and tables, enabling you to analyze each audience independently and understand its key traits and distribution.

Feature Importance

This subsection shows the top demographic features that distinguish the selected persona from all other audiences. Higher importance means the feature is more useful for identifying members of this persona.

Chart View

Interpretation:

For this persona, housing-related features are the most important drivers, including:

• Single, Detached
• 1 Bedroom
• 3 Bedrooms

Important Note: These features are the most useful for identifying members of this audience relative to others. Feature importance reflects how useful a feature is for distinguishing the persona, not whether it increases or decreases likelihood of membership. Directional impact (positive vs. negative influence) is shown in the SHAP Impact section.

Table View

Category Importance

This section aggregates feature importance at the demographic category level for the selected persona, showing which broad dimensions are most defining.

Chart View

What It Shows:

• Feature importance grouped into categories such as:
  • Income
  • Education
  • Age
  • Housing
  • Transportation, etc.
• Each category reflects the combined importance of its underlying features
• Available in both chart and table views

Interpretation:

• Higher values indicate that a category plays a larger role in defining the persona
• Compare categories to understand which broad demographic dimensions matter most Helps simplify detailed feature-level insights into high-level driversImportant

Important Note: Category importance reflects how useful a category is for identifying the persona, not whether it increases or decreases likelihood of membership. Directional impact is shown in the SHAP Impact section.

SHAP Impact

Displays the top 3 most impactful features using SHAP Beeswarm charts.

• Each dot represents one observation
• Color:
  • Red = high feature value
  • Blue = low feature value
• Position (x-axis):
  • Right = pushes toward this persona
  • Left = pushes away

Helps explain how specific features drive membership into the persona.

House Size Example

Example Interpretation: House Size

From the Feature Importance chart, 3-bedroom homes appear as one of the top features defining this persona. However, SHAP analysis reveals a more nuanced story.

• 1 Bedroom: Higher concentrations increase likelihood of belonging to this persona.
  • Red points extend to the right, showing positive impact on membership
  • Blue points cluster more to the left, indicating lower values are less aligned
• 3 Bedrooms: While this feature is important, higher values actually decrease likelihood of belonging to this persona.
  • Red points (higher values) are concentrated on the left, indicating they push away from this persona
  • Blue points (lower values) are more centered or slightly right, suggesting lower presence is more aligned with the audience

Feature importance tells you what matters, while SHAP shows how it matters. In this case, although 3-bedroom homes are an important differentiator for Urban Renters, higher concentrations actually reduce the likelihood of belonging to this persona, reinforcing a profile centered around smaller, renter-aligned housing such as apartments and studios.

Geographic Footprint

This section shows where the selected persona is geographically concentrated, helping connect demographic insights to real-world locations. Hover over states to see the Conversion Rate, Impressions, and Unique Conversions. Click into each state to drill down.

What It Shows:

• States where the persona is present
• User density by location, indicating where the audience is most concentrated
• Visual distribution of the persona across states and regions

Interactions:

• Toggle Views:
  • Boundaries: Clearly outlines ZIP and state regions
  • Heatmap: Highlights density and concentration of users
• Click a state to drill down into ZIP-level details
  • ZIP Code and Name
  • Conversion Rate
  • Impressions
  • Attributed Impressions: Impressions that led to a conversion
  • Unique Conversions
  • Unique Users

Zip Code View

Explorer

The Explorer tab allows you to browse and interact with all available datasets used in the model. Select a dataset to view its chart and table. Use the metric buttons above charts to switch between available metrics.

What It Does:

• Lists all datasets in the left-hand panel
• Displays each selection as a chart and table

This section is intentionally extensive and exploratory. Users are encouraged to navigate different datasets and metrics to uncover additional insights.

Charts

The Charts tab provides access to all visualizations generated by the model, organized into categories for easier navigation and analysis.

This section is intentionally extensive and exploratory. Users are encouraged to explore different chart visualizations to uncover additional insights and include in presentations.

Downloads

The Downloads section provides full access to all model outputs, enabling deeper analysis, reporting, and sharing across teams. These outputs can be used for investigation and validation, but more importantly, the model can be applied directly to campaigns or line items, allowing insights to seamlessly translate into real-time bidding and optimization.

Available Downloads (JSON & CSV):

• Executive Summary: High-level narrative of key findings and recommendations
• Scored Campaign ZIPs (Targeting): ZIP-level performance and model scores for areas included in the campaign
• Lookalike ZIPs (Expansion): Scored ZIPs outside the campaign identified as expansion opportunities
• Audience Persona Profiles: Clustered audience segments with demographic and geographic characteristics
• Category Comparison: Performance comparison across key demographic categories
• All Combined: Full dataset including all outputs for comprehensive analysis

The Targeting model analyzes campaign performance to identify which variables and combinations are most strongly driving conversions. It surfaces patterns across inventory, geography, device, and time to help optimize how and where ads are delivered.

Model outputs can be directly applied to lines to improve real-time bidding and delivery efficiency, enabling smarter, more automated optimization.

Use Case Examples

Improve Performance While Campaigns Are Live

Continuously learn from campaign performance and adjust delivery in real time. Improve results without manually breaking out dozens of lines or constantly adjusting targeting.

Reduce Waste and Focus Spend on What Works

Automatically prioritize high-performing impressions and avoid low-value ones, ensuring budget is spent more efficiently over time.

Note:

• Optimize towards Conversions or Clicks.
• Edit the report/model associate to your Line and click the Save & Rerun button to easily refresh the model

Instructions for generating a Targeting Report/Model. Follow the steps below:

1. Navigate to the Targeting Tab.
2. Click New Report button.
3. Give the report/model a name.
4. Select the Advertiser to run the analysis on.
5. Select the Campaign(s) within the chosen Advertiser.
   • If no campaigns are selected, all campaigns will be included.
6. Select the Line(s) within the chosen Campaign(s).
   • If no lines are selected, all lines will be included.
7. Select the Optimization Type:
   • Conversions
     • If selected, choose the conversion pixel to be used for optimization.
   • Clicks
8. Select the analysis Start Date.
9. Select the analysis End Date.
10. Enter the desired lookback window (in days).
    • Default is 30 days.
    • This defines how far back the model will attribute conversions or clicks to ad exposure.
11. Select what Bid Request Fields are used.
    • Only the checked fields are considered when the model is built. Unchecking one removes it from the regression entirely, so it can never influence a bid.
    • All are checked by default, but users can make it customizable.
    • Fields list:
      • Geography:
        • Region / State
        • ZIP Code
      • Inventory:
        • Publisher
        • App Bundle
        • Deal ID
        • Site Domain
        • Exchange
      • Content
        • Channel
        • Series
        • Title
        • Genre
        • Network
        • Livestream
      • Device
        • Device Type
      • Time
        • Day of Week
        • Hour
      • Operating System
      • Browser
12. Select what the Inventory Scope is.
    • This looks at which impressions the model learns from. This is different from the fields above: unchecking a field removes it from the model, while unchecking inventory here removes those impressions from the analysis entirely — they influence no score, and the campaign average every score is compared against is calculated without them. No filter = the model trains on all inventory.
    • Exchanges
      • Display / App
        • Index
        • Microsoft
        • DV+
      • Video / CTV
        • Magnite
        • Freewheel
      • Audio
        • Triton
        • Adswizz
    • Device Types
      • Desktop
        • Personal Computer
      • Mobile
        • Phone
        • Mobile / Tablet (generic)
        • Tablet
      • TV
        • Connected TV
        • Set Top Box
      • Other
        • Connected Device
        • Unknown / Not Provided
13. Choose whether the report should run once or repeat on a schedule.
    • Run once builds a single report.
    • Schedule re-runs it on the chosen cadence — each run retrains the model, replaces the live model in S3 when Replace the live model each run is enabled, and is saved as a new run grouped under this report.
      • Rerun Interval:
        • Weekly
        • Every 2 Weeks
        • Monthly
      • Rerun Day of Week:
        • Monday
        • Tuesday
        • Wednesday
        • Thursday
        • Friday
        • Saturday
        • Sunday
      • Replace the live model each run checkbox
        • On: each run’s new model overwrites the live model so it goes live automatically.
        • Off: runs still produce reports and archived models, but the live model is left untouched.
14. Click the Save Report button.

Additional Notes

• Ensure selected campaigns and lines have sufficient data for reliable analysis.
  • Recommended: ~2,000 conversions for stable model performance.
• The lookback window should align with typical user conversion behavior.
• Once the report and model are generated, view the results. Users can also edit the report and models setup and rerun the report.

The Reports results include the pipeline used, date created, when the report started, and when the reported completed. The output is separated into 6 sections:

1. Summary
2. Dashboard
3. Multi Variable
4. Explorer
5. Charts
6. Downloads

This information provides a complete audit trail and transparency into the model, allowing users to understand how results are generated and directly tie insights back to campaign performance, rather than relying on a black-box approach.

Summary

The Summary section provides a high-level overview of the analysis, including the Report ID, Advertiser ID, and Date Range, along with an automatically generated Executive Summary that explains what is driving performance and helps inform optimization decisions using model-based predictions.

Tip: You can input the Executive Summary into your preferred LLM (e.g., ChatGPT or Claude) to quickly generate a presentation deck or case study based on the results.

Dashboard

The Dashboard section provides a visual overview of campaign performance. It shows which features drive conversions, geographic performance with drill-down maps, temporal patterns, CPA predictions, and the model’s top and bottom performers across all dimensions.

Dashboard Sections:

• Tile Metrics
• What Drives Conversions – Feature Importance (SHAP)
• Feature Impact Distribution
• Geographic Performance – Click a State to See ZIP Codes
• Geo Zip – Best vs Worst Performers
• Device Type – Best vs Worst Performers
• Content Genre – Best vs Worst Performers
• Conversion Rate by Day of Week
• Conversion Rate by Hour
• Time to Conversion
• Impression Frequency & Conversion Rate
• Cumulative eCPA Over Time
• Cost Efficiency by Feature
  • Content Genre
  • Content Livestream
  • Day
  • Device Type
  • Geo Region
  • Geo Zip
  • Hour

Tile Metrics

• Total Impressions: Total Ad impressions served across all geographies.
• Unique Users: Unique IP addresses reached.
  • Note that IP does not equal a person. Shared IPs (Households, offices, etc.) mean actual reach may differ.
• Attributed Impressions: Impressions that were part of a conversion path. Multiple impressions can contribute to the same conversion.
• Total Spend: Total media cost for the reports date range.
• ECPM: Effective cost per thousand impressions.

What Drives Conversions – Feature Importance (SHAP)

At a high level, this section shows which factors matter most, which helps you quickly focus on the biggest drivers of performance. If certain features are not shown, it means they were not included in the report and did not have a statistically meaningful impact on the model.

Chart View

This view ranks the factors that have the greatest impact on conversions based on the model’s analysis. It uses SHAP (Shapley values) to quantify how much each feature contributes to performance.

In the above example, features such as Geo Zip, Device Type, Content Genre, Day, Geo Region, Hour, and Content Livestream are ranked by their impact on conversions, where higher values indicate a stronger influence on outcomes. Hover over the bar chart to see the features mean absolute SHAP value.

Example insight from the above model:

• Geo Zip is the strongest driver
• Followed by App Bundle, then Publisher Name, and Geo Region

Table View

Provides detailed statistics for each feature:

• Mean Abs SHAP: Overall importance (primary ranking metric)
• Mean SHAP: Direction of impact (positive or negative influence)
• STD SHAP: Variability across observations
• Max / Min SHAP: Range of impact
• Confidence Intervals (CI Lower / Upper): Stability and reliability of the importance score

Users can download this table as a CSV file for further analysis.

Feature Impact Distribution

At a high level, this section shows how each feature influences conversions directionally across all predictions, not just how important it is.

Chart View

Each violin represents the distribution of SHAP values for a feature:

• Wide Spread: more predictions fall at that impact level
• Right of zero: pushes toward conversion
• Left of zero: pushes away from conversion

The box inside each violin shows:

• Median: Center line
• Quartiles: Interquartile range of typical values

How to Interpret:

• Wide spread: Feature impact varies significantly across observations
• Centered Near Zero: Limited average influence on predictions
• Skewed Right: Generally contributes positively to conversion
• Skewed Left: Generally contributes negatively to conversion

Example insights from the above model:

• Geo Zip has the widest spread, ranging roughly from -3 to +3, with density on both sides of zero.
  • This means geography can both strongly increase and decrease conversion likelihood depending on the ZIP—it’s highly impactful but varies significantly across users.
• Device Type is mostly concentrated between -0.5 and +1.5, with more density on the positive side.
  • This indicates device generally has a positive influence on conversions, but with moderate variability.
• Content Genre is tightly clustered around 0 to +0.5, slightly right-skewed.
  • This suggests a consistent but smaller positive effect on conversion likelihood.
• Day is centered very close to zero with a narrow spread (roughly -0.3 to +0.3).
  • This indicates limited overall impact, with only minor variation by day.

These insights reflect how the model predicts conversion likelihood and help explain what is driving those predictions. The resulting prediction scores are then used to inform bidding decisions, where impressions above or below defined thresholds influence whether the system chooses to bid in real time.

Table View

This table contains the same data as the What Drives Conversions — Feature Importance (SHAP) section and is available for download for further analysis.

Geographic Performance – Click a State to See ZIP Codes

This section provides an interactive map view of campaign performance by geography, allowing you to quickly identify high- and low-performing regions.

What It Shows:

• Performance by State, with the ability to drill down into ZIP code level data
• Toggle between key metric views:
  • Conversion Rate
  • Conversions
  • Impressions
• Toggle between Heatmap and Boundaries view
  • Boundaries: Clearly outlines geographic regions
  • Heatmap: Highlights performance intensity across regions

Boundaries View

Heatmap View

Interactions:

• Hover over a state to view:
  • State Name
  • Conversion Rate
  • Impressions
  • Attributed Impressions
  • Unique Conversions
  • Unique Users
  • Predicted Rate
• Click a state to drill down into ZIP level performance – Zip code view:
  • Zip code
  • Zip code name
  • Conversions
  • Impressions
  • Attributed Impressions
  • Unique Conversions
  • Unique Users
  • Predicted Rate

Zip Code View

Metric Views

The map can be toggled between three key performance views:

• Conversion Rate: Shows efficiency by geography (conversions ÷ impressions).
  • Best for identifying high-performing, efficient areas to scale.
• Conversions: Shows total conversion volume by geography. Use Conversions to validate impact and volume.
  • Best for understanding where results are coming from at scale.
• Impressions: Shows delivery volume by geography.
  • Best for identifying where ads are being served and any gaps in delivery.

Table View

This table contains the underlying data used to power the Geographic Performance map, providing detailed metrics at the ZIP code level.

Each row represents a ZIP code and its associated performance, which is aggregated to render state-level views in the map.

What It Includes:

• Geographic identifiers
  • Geo Zip
  • Geo Region Name
  • City
  • State ID
• Delivery Metrics
  • Impressions: Total number of times ads were served
  • Attributed Impressions: Impressions tied to users who later converted within the lookback window
  • Unique Users: Number of distinct users exposed to ads
• Performance Metrics
  • Conversions: Total number of conversion events attributed to the campaign
  • Conversion Rate: Attributed Impressions / Impressions
  • Attributed Conversion Rate: Conversions / Impressions
• Model-derived Metrics
  • Conversion Rate Z-Score: Predicted / Normalized performance
  • Anomaly Flag: Identifies outliers (0 = normal, 1 = anomaly)

Example insight:

• 80513 (CO) shows a high Conversion Rate (17.53%) and Attributed Conversion Rate of 0.53% with a strong z-score (~7.9), indicating performance well above average.

Best vs Worst Performers

This section highlights the highest and lowest performing values for a given dimension based on the model’s predictions and observed conversion rates.

Features include:

• Geo Zip
• Device Type
• Content Genre

At a high level, it shows which values are driving strong positive or negative performance, along with how consistently they perform.

Chart View

Geo Zip Example:

• Displays the distribution of SHAP values for top and bottom performing ZIP codes
• Green violins = top performers (high predicted contribution to conversions)
• Red violins = bottom performers (negative impact on conversion likelihood)

How to Read:

• Above Zero (Green): Increases likelihood of conversion.
• Below Zero (Red): Decreases likelihood of conversion.
• Narrow Shape: Consistent performance across users.
• Wide Shape: Variable performance across users.
• Box Plot (inside the violin) – Shows the median (center line) and quartiles (typical range of values)

Table View

Provides detailed metrics for each feature:

• Value: Feature dependent examples
  • Geo Zip: 85029 – Phoenix, AZ
  • Device Type: 3 – Connected TV (CTV)
  • Content Genre: food & cookiing
• Rank: Top or Bottom performer
• Conversion Rate : Observed conversion rate
• Mean SHAP: Average contribution to model predictions
• Samples: Number of observations (data volume)

Example Insight

• 64128 (Kansas City, MO) shows a strong conversion rate (87.50%) with a high positive mean SHAP (~1.67) and a solid sample size (96), indicating a reliable, high-performing market with both scale and consistency

Conversion Rate by Day of Week

This section breaks down conversion rate by the day the ad was served, helping identify which days drive the strongest performance and can be used as a day parting optimization. These insights can be applied manually within the platform. If the model is applied to the campaign or line, day-of-week performance is automatically incorporated into real-time bidding decisions.

What It Shows:

• Conversion rate for each day of the week
• Performance trends across Monday–Sunday
• Relative differences in efficiency by day

How to Use:

• Identify high-performing days to increase spend or prioritize delivery
• Identify underperforming days to reduce spend or adjust bidding
• Inform dayparting strategies to improve overall efficiency

Conversion Rate by Hour

This section breaks down conversion rate by the hour the ad was served (0–23), helping identify peak performance windows throughout the day. These insights can be applied manually within the platform. If the model is applied to the campaign or line, day-of-week performance is automatically incorporated into real-time bidding decisions.

What It Shows:

• Conversion rate for each hour of the day (0–23)
• Performance trends across morning, afternoon, evening, and overnight
• Relative differences in efficiency by hour

How to Use:

• Identify peak hours to increase bids or prioritize delivery
• Identify low-performing hours to reduce spend or adjust bidding
• Inform hour-of-day bid adjustments to improve efficiency

Time to Conversion

This section shows the distribution of time between when an ad was served and when a user converted, helping you understand how long the typical conversion path takes. Short times suggest direct response; long times suggest consideration-based purchase behavior.

Chart View

What It Shows:

• Time between impression to conversion, grouped into time buckets:
  • <1 hr
  • 1–6 hrs
  • 6-24 hrs
  • 1–3 days
  • 3-7 days
  • 7-14 days
  • 14+ days
• Volume of conversions occurring within each time window (hover to see the value).
• Overall shape of the conversion lag distribution

Interpretation:

• Short time to conversion indicates more direct response behavior
• Longer time to conversion indicates more consideration-based or delayed decision-making
• Peaks in specific time ranges highlight when users are most likely to convert after exposure

Table View

Example Insights:

• Conversions are highest within the first few hours (0–3 hrs), indicating strong immediate response behavior
• There is a secondary concentration in the 1–7 day range, suggesting some users convert after additional consideration
• Very long conversion windows (14+ days) show lower volume, indicating diminishing impact over time

This distribution helps inform lookback window selection, attribution settings, and frequency strategy.

• If conversions happen quickly (short lag), higher frequency over a shorter window can be effective.
• If conversions take longer (delayed lag), sustained frequency over time is needed to stay top of mind.
• Helps avoid overexposing users too early or underexposing during longer consideration periods.

Impression Frequency & Conversion Rate

This section shows how conversion rate changes as users are exposed to more impressions, helping identify the optimal frequency for performance. Higher-frequency users typically had more time in the campaign, so results reflect correlation, not causation. Frequency should be interpreted alongside time-to-conversion and campaign duration.

Chart View

What It Shows:

• Bars: Conversion rate at each impression frequency
• Line: Total users reached at each frequency level
• Green arrow annotation: Peak marginal return, the point where each additional impression drives the most incremental lift

How to Interpret:

• Rising conversion rate: additional impressions are improving performance
• Peak point (green marker): optimal frequency where incremental lift is highest
• After peak: diminishing returns, where additional impressions add less value
• User curve (line) : shows how many users are exposed at each frequency level

Table View

Provides detailed performance metrics at each impression frequency level:

• Frequency Bucket: Number of impressions served per user
• Total Users: Number of users reached at that frequency
• Converting Users: Number of users who converted at that frequency
• Conversion Rate: Conversion rate at that frequency level

Interpretation:

• Find the optimal frequency
  • Identify the point where conversion rate is highest before diminishing returns.
• Avoid overexposure
  • If performance drops at higher frequencies, you’re wasting impressions.
• Balance scale vs efficiency
  • Higher frequency may increase conversion rate but reach fewer users.
• Inform frequency caps and bidding strategy
  • Helps determine how often to show ads per user

Additional Notes:

• Higher frequency users often had more time in the campaign, so results are correlated, not causal
• Should be used alongside:
  • User volume (Total Users)
  • Time to conversion

Cumulative eCPA Over Time

This section shows how spend and conversions accumulate over time, providing a complete view of attribution across the campaign and its lookback window.

Chart View

What It Shows:

• Blue line: Cumulative eCPA (efficiency over time)
• Green dotted line: Cumulative spend
• Blue dotted line: Cumulative conversions
• Shaded region: Actual study period (active campaign dates)

Interpretation:

• Spend begins accumulating before the study period due to the lookback window
• onversions continue after impressions are served, as users convert over time
• Early in the timeline, eCPA may appear inflated or volatile because not all conversions have occurred yet
• As time progresses, the lookback “tail” fills in, and eCPA stabilizes

Table View

Provides a daily breakdown of spend, conversions, and efficiency metrics used to build the cumulative chart.

Columns:

• Imp Date: Date impressions were served.
• Impressions: Total impressions delivered on that date.
• Unique Users: Number of distinct users reached.
• Total Cost: Spend for that day.
• Unique Conversions by Imp Date: Conversions attributed back to the date the impression occurred.
• Unique Conversions by Conv Date: Conversions counted on the date the conversion actually happened.
• Conversion Rate: Conversions relative to impressions for that day.
• eCPA: Cost per acquisition or conversion. Calculated as: Total cost / Conversions.
• eCPM: Cost per thousand impressions. Calculated as: Total cost / Impressions.
• Cumulative Spend: Running total of spend over time.
• Cumulative Conversions by Imp Date: Running total of conversions attributed back to impression dates.
• Cumulative Conversions by Conv Date: Running total of conversions based on when conversions actually occurred.
• Cumulative eCPA: Running cost per conversion over time, calculated from cumulative spend and cumulative conversions.

Cost Efficiency by Feature

This section compares cost and performance metrics across values of a selected feature, helping identify the most efficient segments to prioritize.

Select from the following features:

• App Bundle
• Content Genre
• Content Series
• Day
• Device Type
• Geo Region
• Geo Zip
• Hour
• Publisher Name
• Site Domain

Content Genre Example

eCPA ($): Cost per acquisition or conversion. Calculated as: Total cost / Conversions.

• Shows cost per acquisition by feature value
• Bars are sorted from lowest to highest eCPA
• Color intensity reflects efficiency (darker = higher cost / less efficient)

Interpretation:

• Look for shorter bars and lighter color. These are the most efficient segments.
• These represent lowest cost per conversion (best ROI).
• This is the primary view for optimization decisions

eCPM ($): Cost per thousand impressions. Calculated as: Total cost / Impressions.

• Shows cost per thousand impressions by feature value
• Bars are sorted from highest to lowest cost
• Color intensity reflects relative cost (darker = more expensive)

Interpretation:

• Look for shorter bars and lighter color. This is lower-cost inventory.
• Use this to understand where you’re paying more or less to reach users.
• Low cost doesn’t always mean good performance.

Conversion Rate: Efficiency of converting impressions

• Shows conversion rate by feature value
• Bars are sorted from highest to lowest conversion rate
• Color reflects relative performance (darker = lower performance)

Interpretation:

• Look for taller bars and lighter color. These are higher-performing segments.
• These indicate where users are most likely to convert
• Use this to identify strong audiences or content

Multi Variable

This section explores how multiple features interact together to influence conversion performance, rather than looking at each feature in isolation.

It helps uncover combinations of variables (e.g., Geo + Device + Content) that drive stronger or weaker outcomes.

Subsections:

• Waterfalls
• High Single Variables
• Low Single Variables
• High 2-Way
• Low 2-Way
• High 3-Way
• Low 3-Way

Waterfalls

This view explains how the model builds individual predictions. Each chart starts from a base rate and shows how each feature pushes the prediction up (green) or down (red) to reach the final score, making the model transparent rather than a black box.

High Confidence Example

Low Confidence Example

What It Shows:

• Each chart starts from a base conversion rate (baseline prediction) Individual features are then added step-by-step
• Each feature pushes the prediction up or down:
  • Green bars increase likelihood of conversion
  • Red bars decrease likelihood of conversion
• The final value represents the model’s predicted conversion score

Interpretation:

• Start at the base rate (average performance)
• Follow each step to see how features contribute to the final prediction
• Larger bars = stronger impact on the prediction
• The final value shows how likely that specific combination is to convert

High Single Variables

This section identifies individual feature values that appear significantly more often in top-converting predictions than in the overall population. High lift means strong positive signal for targeting.\

Chart View

What It Shows:

• Each point represents a single feature value (e.g., a specific ZIP, device type, or content genre)
• Compares:
  • Overall frequency (how often it appears in the dataset)
  • Top-converting frequency (how often it appears in high-performing predictions)
• Color indicates lift strength (darker = stronger signal)

Interpretation:

• Points above the baseline appear more often in top-performing outcomes
• Points further right are more common overall
• Points higher on the chart show stronger positive signal
• Darker color indicates higher lift and a stronger targeting signal

Table View

Provides detailed metrics for each high-performing single feature value identified by the model. Users can download the table as a CSV for further analysis.

Columns:

• Segment: Which classification it falls under
• Feature: The dimension being analyzed (e.g., Geo Zip, Device Type, Content Genre)
• Value: The specific value within that feature
• Segment Frequency: Share of this value within top-performing predictions
• Baseline Frequency: Share of this value across the overall dataset
• Lift: Ratio of Segment Frequency to Baseline Frequency showing how much more often this value appears in top-performing outcomes vs. normal.
• Segment Count: Number of occurrences of this value within top-performing predictions
• Unique IPs: Number of unique users associated with this value
• Score: Model-derived strength of the signal. The higher the value the stronger the posititive contribution to conversion likelihood.

Interpretation:

• High lift with high segment count indicates a strong and scalable opportunity
• High lift with low segment count indicates a niche but promising segment that should be tested before scaling
• Score reflects how impactful the value is within the model and helps prioritize which signals matter most

Low Single Variables

This section identifies individual feature values that appear significantly more often in low-converting predictions than in the overall population. Low lift indicates a negative signal and these segments may be deprioritized or excluded from targeting.

Chart View

What It Shows:

• Each point represents a single feature value (e.g., a specific ZIP, device type, or content genre)
• Compares:
  • Overall frequency (how often it appears in the dataset)
  • Low-converting frequency (how often it appears in low-performing predictions)
• Color indicates lift strength (darker = stronger negative signal)

Interpretation:

• Points above the baseline appear more often in low-performing outcomes
• Points further right are more common overall
• Points higher on the chart show stronger negative signal
• Darker color indicates lower lift and a stronger signal to avoid

Table View

Provides detailed metrics for each low-performing single feature value identified by the model. Users can download the table as a CSV for further analysis.

Columns:

• Segment: Which classification it falls under
• Feature: The dimension being analyzed (e.g., Geo Zip, Device Type, Content Genre)
• Value: The specific value within that feature
• Segment Frequency: Share of this value within top-performing predictions
• Baseline Frequency: Share of this value across the overall dataset
• Lift: Ratio of Segment Frequency to Baseline Frequency showing how much more often this value appears in top-performing outcomes vs. normal.
• Segment Count: Number of occurrences of this value within top-performing predictions
• Unique IPs: Number of unique users associated with this value
• Score: Model-derived strength of the signal. The higher the value the stronger the posititive contribution to conversion likelihood.

Interpretation:

• Low lift with high segment count indicates a consistently underperforming segment that may be worth reducing or excluding
• Low lift with low segment count indicates a weaker signal with limited impact
• Score reflects how negatively the value impacts performance and helps prioritize what to deprioritize

High 2-Way and High 3-Way

These sections identify combinations of feature values that together produce conversion rates above the campaign average.

• 2-Way: Combinations of two features (e.g., Geo Zip + Device)
• 3-Way: Combinations of three features (e.g., Geo Zip + Device + Day)

These represent layered signals, where performance improves when features are used together.

Chart View

What It Shows:

• Each point represents a combination of feature values (e.g., 35209 × 1)
• X-axis shows total count / volume (log scale)
• Y-axis shows conversion rate
• Color indicates lift vs campaign average (darker = stronger performance)

Interpretation:

• Points higher on the chart have higher conversion rate
• Points further right have more volume (more scalable)
• Top-right quadrant contains the best combinations (high performance + scale)
• Top-left quadrant contains high performance but low volume (niche opportunities)
• Bottom-right quadrant contains high volume but lower performance (less efficient)
• Color intensity shows stronger lift vs average (darker = better signal)

Example Insight:

• 35209 × 1 shows:
  • Very high conversion rate (~0.94)
  • Strong lift (~2.08x vs average)
  • Moderate volume (53 samples)
• This indicates a high-performing combination with meaningful scale, making it a strong candidate for targeting.

Table View

Provides detailed metrics for each high-performing combination. Users can download the table as a CSV for further analysis.

Columns:

• Feature 1: First dimension in the combination
• Feature 2: Second dimension (and Feature 3 for 3-way analysis)
• Combination: Combined feature values (e.g., 29220 × sc)
• Conversion Rate: Conversion rate for this specific combination
• Conversions: Total conversions generated
• Total Count: Number of observations for this combination
• Unique IPs: Number of unique users associated with this combination
• Lift vs Avg: Performance relative to campaign average
  (>1 = above average performance)
• Score: Model-derived strength of the signal
  (higher = stronger positive contribution to conversion likelihood)
• p-value: Statistical significance of the result
• p-adjusted: Adjusted p-value accounting for multiple comparisons
• Significant: Indicates whether the result is statistically significant (TRUE/FALSE)

Interpretation:

• High lift with high total count indicates a strong and scalable combination
• High lift with low count indicates a promising but niche combination
• Statistically significant results provide higher confidence in the signal
• Score helps prioritize which combinations have the strongest impact

Low 2-Way and Low 3-Way

These sections identify combinations of feature values that together produce conversion rates below the campaign average.

• 2-Way = combinations of two features
• 3-Way = combinations of three features

These represent negative interaction signals, where performance declines when features are combined.

Chart View

What It Shows:

• Each point represents a combination of feature values
• X-axis shows total count / volume (log scale)
• Y-axis shows conversion rate
• Color indicates lift vs campaign average (darker blue = more negative performance)

Interpretation:

• Points higher on the chart have higher conversion rate
• Points further right have more volume (more scalable)
• Bottom-right quadrant contains the worst combinations (low performance + high volume)
• Bottom-left quadrant contains low performance but low volume (limited impact)
• Top-right quadrant contains higher volume with moderate performance (mixed efficiency)
• Top-left quadrant contains stronger performance but low volume (niche and less impactful)
• Color intensity shows stronger negative lift vs average (darker = worse signal)

Table View

Provides detailed metrics for each low-performing combination of feature values. Users can download the table as a CSV for further analysis.

Columns:

• Feature 1: First dimension in the combination
• Feature 2: Second dimension (and Feature 3 for 3-way analysis)
• Combination: Combined feature values
• Conversion Rate: Conversion rate for this combination
• Conversions: Total conversions generated
• Total Count: Number of observations for this combination
• Unique IPs: Number of unique users associated with this combination
• Lift vs Avg: Performance relative to campaign average
  (<1 = below average performance)
• Score: Model-derived strength of the signal
  (lower = stronger negative impact on conversion likelihood)
• p-value: Statistical significance of the result
• p-adjusted: Adjusted p-value accounting for multiple comparisons
• Significant: Indicates whether the result is statistically significant (TRUE/FALSE)

Interpretation:

• Low lift with high total count indicates a consistently underperforming combination that should be reduced or excluded
• Low lift with low total count indicates a weaker signal with limited impact
• Statistically significant results provide higher confidence in deprioritization decisions
• Score reflects how strongly the combination negatively impacts performance and helps prioritize what to avoid

Explorer

The Explorer tab allows you to browse and interact with all available datasets used in the model. Select a dataset to view its chart and table. Use the metric buttons above charts to switch between available metrics.

What It Does:

• Lists all datasets in the left-hand panel
• Displays each selection as a chart and table
• Allows switching between multiple metrics:
  • SHAP Value
  • Count
  • Conversions
  • Conversion Rate

This section is intentionally extensive and exploratory. Users are encouraged to navigate different datasets and metrics to uncover additional insights.

Charts

The Charts tab provides access to all visualizations generated by the model, organized into categories for easier navigation and analysis.

Charts are grouped into the following sections:

• Categorical Features: Visual breakdowns of performance across individual feature dimensions
• Waterfall Analysis: Shows how the model builds predictions step-by-step
• SHAP Summary: Provides a comprehensive view of feature importance and impact across the model using multiple visualization types:
  • Heatmap: Shows how feature values impact predictions across observations
  • Decision Plot: Visualizes how features combine step-by-step to form predictions
  • Beeswarm Plot: Displays the distribution and direction of feature impact across all data
  • Bar Chart: Ranks features by overall importance
• Feature Violin: Distribution of feature impact across all predictions

Downloads

The Downloads section provides full access to all model outputs, enabling deeper analysis, reporting, and sharing across teams. These outputs can be used for investigation and validation, but more importantly, the model can be applied directly to campaigns or line items, allowing insights to seamlessly translate into real-time bidding and optimization.

Available Downloads (JSON & CSV):

• Executive Summary: High-level summary of findings, insights, and key drivers
• High Confidence Single Variables: Top-performing individual feature values with strong positive signals
• Low Confidence Single Variables: Underperforming individual feature values to deprioritize
• High Performing 2-Way Combinations: Top-performing pairs of feature values
• Low Performing 2-Way Combinations: Underperforming pairs of feature values
• High Performing 3-Way Combinations: Top-performing combinations of three features
• Low Performing 3-Way Combinations: Underperforming combinations of three features
• CPA Predictions by Segment: Model-predicted cost efficiency across segments
• All Combined: Complete dataset including all outputs in a single export

Understanding and Using Model Outputs

The downloadable files provide the underlying model results used to identify high- and low-performing variables and combinations. Depending on the file selected, results represent:

• A single variable, such as device_type = 4
• A 2-way combination, such as geo_zip = 33147 and device_type = 4
• A 3-way combination, such as geo_zip = 60602, publisher_name = iheartmedia, and device_type = 6

The All Combined file includes these outputs together in a single export.

Single-Variable Results

Single-variable outputs identify individual feature values that are more strongly associated with high- or low-performing segments.

Example:

{
  "feature": "device_type",
  "value": "4",
  "segment_frequency": 0.81578652,
  "baseline_frequency": 0.5408660915,
  "lift": 0.5082966614,
  "segment_count": 29073,
  "unique_ips": 6523,
  "score": 0.122608
}

This result identifies Device Type 4 as a high-performing signal in the analyzed dataset.

A negative score indicates a feature value associated with lower-performing segments. For example:

{
  "feature": "site_domain",
  "value": "www.iheart.com",
  "score": -0.469354
}

This identifies www.iheart.com as a negative signal within the analyzed data.

2-Way Combinations

2-way combinations identify specific pairs of feature values that perform differently when evaluated together.

Example:

{
  "feature1": "geo_zip",
  "feature2": "device_type",
  "combination": "33147 × 4",
  "conv_rate": 0.7123287671,
  "conversions": 208,
  "total_count": 292,
  "unique_ips": 154,
  "lift_vs_avg": 2.1878006942,
  "score": 0.478138,
  "significant": true
}

This result represents:

ZIP Code 33147 + Device Type 4

Within the analyzed dataset, this exact combination had:

• 208 conversions out of 292 total records
• A 71.2% conversion rate
• 154 unique IPs
• A conversion rate approximately 2.19 times the overall average
• A positive model score of 0.478138
• A statistically significant result

Low-performing combination files use the same structure but identify combinations associated with weaker performance.

3-Way Combinations

3-way combinations identify patterns involving three feature values.

Example:

{
  "feature1": "geo_zip",
  "feature2": "publisher_name",
  "feature3": "device_type",
  "combination": "60602 × iheartmedia × 6",
  "conv_rate": 0.6818181818,
  "conversions": 180,
  "total_count": 264,
  "unique_ips": 141,
  "lift_vs_avg": 2.0512602798,
  "score": 0.456544,
  "significant": true
}

This result represents:

ZIP Code 60602 + Publisher iHeartMedia + Device Type 6

The model identified the exact combination of these three values as a high-performing pattern in the analyzed dataset.

Creating Your Own Model

The model generated outputs are easy to apply to a campaign with a click of a button, but users are not limited to the exact combinations provided by the model. Users can also create their own custom JSON file using the same general structure as the model outputs.

This allows users to define variables and combinations based on their own campaign learnings, business rules, historical performance, or optimization goals. For example, a user may already know that a certain geography, publisher, device type, content genre, or day of week is strategically important and can build those values into a custom model.

For detailed information, see here: User Defined Models

The Incrementality report measures the true impact of advertising by comparing users who were exposed to ads against a matched control group who were not.

It identifies what actually works by removing organic conversions and calculating incremental lift, then uses these results to train a model that classifies ZIP codes based on incremental performance.

This model segments markets into actionable groups (e.g., Scale, Observe, Ignore), enabling direct application to targeting and real-time bidding strategies.

Use Case Examples

Understand the True Impact of Your Campaigns

Measure the incremental lift driven by your campaigns—so you can separate what your media is actually contributing from what could have happened anyway.

Prove the Value of Your Media Investment

Quantify how much of your performance is driven by advertising—so you can demonstrate CPA, ROI, and justify spend with confidence.

Optimize Budget and Reduce Waste with Incrementality Insights

Identify which campaigns, audiences, or strategies are truly driving incremental results—so you can shift budget toward what delivers real impact.

Identify underperforming or non-incremental ZIP codes and segments to further investigate—so you can refine messaging, creative, and targeting to improve efficiency and drive greater incremental lift.

Test and Validate Your Strategy

Evaluate different targeting strategies, creatives, or channels to understand what actually drives lift—so you can make smarter, data-backed decisions moving forward.

Choose how you want to analyze campaign impact, then configure campaign inputs, dates, and optimization settings to generate the report results and incremental model.

Instructions for generating a Incrementality Report and Model. Follow the steps for initial setup below:

1. Navigate to the Incrementality Tab
2. Click the New Report button
3. Give the report a name
4. Select the Advertiser to run the analysis on
5. Select the Campaign(s) within the chosen Advertiser.
   • If no campaigns are selected, all campaigns will be included.
6. Select the Line(s) within the chosen Campaign(s).
   • If no lines are selected, all lines will be included.
7. Optimization Type:
   • Conversions: Choose the conversion pixel to be used for optimization.
8. Select the analysis Start Date.
9. Select the analysis End Date.
10. Enter the desired lookback window (in days).
    • Default is 30 days.
    • This defines how far back the model will attribute conversions or clicks to ad exposure.
11. Select the Analysis Mode:
    • Fast: Runs quickly using a smaller sample size for directional insights with lower statistical confidence.
    • Balanced: Provides a balance between speed and statistical reliability, suitable for most analyses.
    • Thorough: Uses the largest sample and most rigorous evaluation for maximum accuracy and highest statistical confidence.
12. Choose whether the report should run once or repeat on a schedule.
    • Run once builds a single report.
    • Schedule re-runs it on the chosen cadence — each run retrains the model, replaces the live model in S3 when Replace the live model each run is enabled, and is saved as a new run grouped under this report.
      • Rerun Interval:
        • Weekly
        • Every 2 Weeks
        • Monthly
      • Rerun Day of Week:
        • Monday
        • Tuesday
        • Wednesday
        • Thursday
        • Friday
        • Saturday
        • Sunday
      • Replace the live model each run checkbox
        • On: each run’s new model overwrites the live model so it goes live automatically.
        • Off: runs still produce reports and archived models, but the live model is left untouched.
13. Click the Save Report button.

Key Notes:

• Requires campaign delivery and pixel fires to measure incrementality
• Uses exposed vs. control methodology to train a model that classifies ZIP-level performance
• Outputs are directly usable for targeting, budget allocation, and real-time bidding strategies

The Incrementality Reports results include the pipeline used, date created, when the report started, and when the reported completed. The output is separated into 5 sections:

1. Summary
2. Dashboard
3. Explorer
4. Charts
5. Downloads

This information provides a complete audit trail and transparency into the model, allowing users to understand how results are generated and directly tie insights back to campaign performance, rather than relying on a black-box approach.

Summary

The Summary section provides a high-level overview of the analysis, including the Report ID, Advertiser ID, and Date Range, along with an automatically generated Executive Summary. The Executive Summary explains what is driving incremental performance and highlights where advertising is creating true lift, helping inform optimization decisions using model-based predictions.

Tip: You can input the Executive Summary into your preferred LLM (e.g., ChatGPT or Claude) to quickly generate a presentation deck or case study based on the results.

Dashboard

The Dashboard provides a visual overview of incremental lift and true campaign impact, comparing exposed users (who saw ads) against a matched control group (who did not).

• Exposed Group: Unique IPs served ads, matched to pixel fires where the ad was shown before conversion within the lookback window
• Control Group: 1% random sample of bid stream IPs in the same ZIPs and deal IDs, not served ads, matched to the same pixel fires

ZIP codes are filtered to within 1 standard deviation of mean impression volume to ensure reliable comparisons.

The model classifies ZIPs into:

• Scale: Strong positive lift with high confidence
• Observe: Moderate or emerging signal
• Ignore: No measurable incremental impact
• Block: Negative lift

Dashboard Sections:

• Tile Metrics
• ZIP Classification — Scale / Observe / Ignore / Block
• Geographic Analysis
• Top Scale ZIPs — Incremental Conversions
• Lift Distribution Across ZIPs
• Exposed vs Control — Conversion Rate by ZIP
• Ad Impact by ZIP — Control vs Exposed Conversion Rate

Tile Metrics

• Overall Lift: Percentage increase in conversion rate for users who saw ads versus those who did not. Positive = ads are driving incremental impact. Range in brackets is the 95% bootstrap confidence interval.
• Incremental Conversions: Estimated conversions that would NOT have happened without advertising. Range in brackets is the 95% bootstrap confidence interval.
• ZIPs Analyzed: ZIP codes with impression volume within 1 standard deviation of the mean – statistically meaningful sample sizes.
• Scale: ZIPs with high positive lift and high confidence – increase budget here.
• Block: ZIPs with negative or zero lift – ads are not working here, stop spending.
• Exposed Users: Unique IPs served ads by this campaign.
• Exposed Conversions: Unique exposed IPs that fired the pixel AFTER being served an ad (within lookback window). Matched via pixel server logs.
• Exposed CR: Conversion rate for the exposed group (Exposed Conversions / Exposed Users).
• Control Users: Unique IPs from 1% sample bid steam in the same ZIPs and deal IDs, NOT served ads by this campaign.
• Control Conversions: Unique control IPs that fired the pixel AFTER their sample request (within lookback window). Same pixel server log source as exposed – apples to apples comparison.
• Control CR: Conversion rate for the control group (Control Conversions / Control Users). This is the baseline organic conversion rate.
• FDR Significant: ZIPs that pass Benjamini-Hochberg FDR correction at 10% – these lift estimates are statistically reliable after correcting for multiple comparisons.
• Underpowered ZIPs: ZIPs where sample size is too small to detect the observed lift reliably (lift < minimum detectable effect). These are classified as Observe regardless of lift magnitude.
• Incremental CPA: True cost per conversion caused by advertising: (Spend / Incremental Conversions. This is the honest acquisition cost – excludes conversions that may have happened anyway.
• Standard CPA: Traditional cost per conversion: (Spend / All Conversions). Lower than incremental CPA because it counts conversions that would have happened possibly without ads.
• CPA Inflation: How much higher the true (incremental) CPA is vs standard CPA. A factor of 2x means that the real cost of acquiring a customer is twice what standard reporting suggests.
• Total Spend: The total media spend across al ZIPs (full volume, no down sampling).

ZIP Classification — Scale / Observe / Ignore / Block

This section categorizes ZIP codes based on incremental lift and statistical confidence, helping identify where advertising is driving real impact and where it is not. Each ZIP is classified based on its lift magnitude, statistical confidence, and sample quality. Scale indicates high lift with strong confidence, while Block reflects negative lift or no measurable effect.

Chart View

What It Shows

• Each ZIP is classified into one of four groups based on:
  • Lift magnitude (positive or negative impact)
  • Statistical confidence (reliability of the result)

Classification Definitions

• Scale: ZIPs with strong positive lift and high statistical confidence
  • Proven to drive incremental conversions
  • Best candidates for increased budget and expansion
• Observe: ZIPs with moderate or emerging signals
  • Potential to perform well but require more data
  • Suitable for continued testing and monitoring
• Ignore: ZIPs with no measurable incremental impact
  • Conversions likely occurring organically
  • Candidates for reduced spend or deprioritize
• Block: ZIPs with negative lift
  • Advertising may be harming performance
  • Should be excluded from targeting

Table View

This table provides the full ZIP-level output of the incrementality model, combining exposure data, lift measurement, statistical confidence, and cost efficiency. Users can download this table as a CSV file for further analysis.

Columns:

• Geo ZIP: ZIP code being evaluated
• Unique Users Control: Number of users not served ads
• Conversions Control: Total conversions from control group
• Unique Converting IPs Control: Unique converters in control group
• Diversity Ratio Control: Distribution of conversions across users in control group
• Unique Users Exposed: Number of users served ads
• Conversions Exposed: Total conversions from exposed users
• Unique Converting IPs Exposed: Unique converters in exposed group
• Diversity Ratio Exposed: Distribution of conversions across users in exposed group
• CR Control Raw: Raw conversion rate for control group
• CR Exposed Raw: Raw conversion rate for exposed users
• Raw Lift: Difference between exposed and control conversion rates
• Diversity Valid Exposed: Whether exposed data passes diversity validation checks
• Diversity Confidence: Confidence in diversity-based data reliability
• CR Control Bayes: Bayesian-adjusted conversion rate for control group
• CR Exposed Bayes: Bayesian-adjusted conversion rate for exposed users
• CR Control CI Low: Lower bound of control conversion rate confidence interval
• CR Control CI High: Upper bound of control conversion rate confidence interval
• CR Exposed CI Low: Lower bound of exposed conversion rate confidence interval
• CR Exposed CI High: Upper bound of exposed conversion rate confidence interval
• Lift Bayes: Modeled incremental lift (preferred lift metric)
• Incremental Conversions: Estimated conversions driven by advertising
• Lift Confidence: Confidence level of the lift estimate
• Relative Lift: Percent increase in conversion rate vs control
• Probability Positive: Probability that lift is positive
• Probability Large Positive: Probability of strong positive lift
• Probability Negative: Probability that lift is negative
• Expected Lift: Average expected lift value
• Lift Std: Standard deviation of lift
• Lift 5th Percentile: Lower bound of lift estimate range
• Lift 95th Percentile: Upper bound of lift estimate range
• Expected Value: Expected economic value generated by lift
• Value Std: Standard deviation of expected value
• Probability Profitable: Probability that the ZIP is profitable
• Value 5th Percentile: Lower bound of expected value
• Value 95th Percentile: Upper bound of expected value
• Sample Adequacy Score: Score indicating if sample size is sufficient
• Incremental Conv CI Low: Lower bound of incremental conversions
• Incremental Conv CI High: Upper bound of incremental conversions
• Classification: Final label (Scale, Observe, Ignore, Block)
• P Adjusted: P-value adjusted for multiple comparisons
• FDR Significant: Whether result is statistically significant after correction
• Min Detectable Lift: Minimum lift detectable with current sample size
• Lift vs Avg: Lift relative to campaign average
• Unique IPs: Total unique users observed
• Score: Model-derived ranking score for prioritization
• City: City name
• State ID: State abbreviation
• Total Spend: Total spend in the ZIP
• CPM: Cost per thousand impressions
• Total Impressions: Total impressions served
• Incremental CPA: Cost per incremental conversion
• Standard CPA: Traditional cost per conversion

Geographic Analysis

This section visualizes ZIP-level performance, showing how incremental impact varies across locations.

Each dot represents an analyzed ZIP code, allowing you to quickly identify where advertising is working and where it is not.

Toggle Views:

• Classification: Colors ZIPs by Scale, Observe, Ignore, Block
  • Quickly identify high- and low-performing markets
• Impressions Served: Dot size represents user volume / impressions
  • Larger dots = more reach and spend
• Lift Magnitude: Color represents direction and strength of lift
  • Green lift (positive) = stronger incremental impact
  • Red lift (negative) = underperformance

Classification

What It Shows:

• Quickly identify where to scale (Scale) and where to monitor (Observe and Ignore)
• Look for clusters of Scale ZIPs to find high-performing regions
• Use alongside other views to understand both performance and scale

Hover Data:

• ZIP code, name and state
• Lift: Lift relative to the analysis’s average
• Users: Unique users exposed to in that ZIP
• Incremental Conversions: The number of additional conversions driven by the ad
• FDR Significance: Indicates whether the lift is statistically reliable

Impressions Served

What It Shows:

• Dot size represents impressions served
• Color scale reflects Lift vs. Average intensity:
  • Light blue = below average / negative lift
  • Red = above average / positive lift

Hover Data:

• ZIP code, name and state
• Lift: Lift relative to the analysis’s average
• Users: Unique users exposed to in that ZIP
• Incremental Conversions: The number of additional conversions driven by the ad
• FDR Significance: Indicates whether the lift is statistically reliable

Lift Magnitude

What It Shows:

• Color represents direction and strength of lift
  • Green (Positive) lift = ads are driving incremental conversions
  • Red (Negative) lift = ads are underperforming

Hover Data:

• ZIP code, name and state
• Lift: Lift relative to the analysis’s average
• Users: Unique users exposed to in that ZIP
• Incremental Conversions: The number of additional conversions driven by the ad
• FDR Significance: Indicates whether the lift is statistically reliable

Table View

This is the same table as the previous section, ZIP Classification — Scale / Observe / Ignore / Block. Refer to the above section for detailed column definitions and interpretation.

Top Scale ZIPs — Incremental Conversions

This section highlights ZIPs classified as Scale (high lift, high confidence), ranked by estimated incremental conversions. Error bars show the 90% credible interval from Monte Carlo simulation. These are the proven winners – increase budget here.

Chart View

What It Shows:

• Only ZIPs classified as Scale Ranked by incremental conversions (highest at the bottom)
• Bars represent the estimated number of conversions driven by ads
• Error bars show the 90% credible interval from Monte Carlo simulation
  • Reflect the range of likely outcomes
  • Narrower intervals = more confidence
  • Wider intervals = greater uncertainty
• Hover to see the Incremental Conversions

Identify proven high-impact ZIPs backed by statistical confidence.

Table View

This is the same dataset as the ZIP Classification table, filtered to Scale ZIPs only. Refer to the ZIP Classification — Scale / Observe / Ignore / Block section for full column definitions.

Lift Distribution Across ZIPs

This chart shows the distribution of relative lift (%) across all analyzed ZIP codes, helping you understand how incremental performance is spread across your campaign.

Chart View

What It Shows:

• Histogram of relative lift (%) for all ZIPs
• Each bar represents the number of ZIPs within a lift range
• The vertical dashed line at 0 represents no incremental impact

Interpretation:

• Right of zero (positive lift): ZIPs where advertising is driving incremental conversions
• Left of zero (negative lift): ZIPs where advertising is not effective or may be harming performance
• Cluster around zero: ZIPs with little to no measurable incremental impact
• Long right tail: A small number of ZIPs driving very high incremental lift

Why It’s Useful:

• Shows whether performance is broadly distributed or concentrated
• Helps assess overall campaign effectiveness
• Identifies if results are driven by:
  • Many moderately positive ZIPs
  • Or a few high-impact outliers

Table View

This is the same dataset as the ZIP Classification table. Refer to the ZIP Classification — Scale / Observe / Ignore / Block section for full column definitions.

Exposed vs Control — Conversion Rate by ZIP

This chart compares conversion rates between users who saw ads (exposed) and those who did not (control) across each ZIP code, showing where advertising is driving true incremental impact.

Chart View

What It Shows:

• Each do represents a ZIP code
• X-axis (Control CR): Conversion rate of users who were eligible but not served ads (baseline)
• Y-axis (Exposed CR): Conversion rate of users who were served ads
• The diagonal line represents equal performance between exposed and control

Interpretation:

• Above the diagonal: Exposed CR > Control CR and ads are driving incremental conversions
• Below the diagonal: Exposed CR < Control CR and ads may be ineffective or counterproductive
• Near the line: Minimal difference and/or little to no measurable incremental impact
• Farther from the line: Greater difference and stronger positive or negative impact

Table View

This is the same dataset as the ZIP Classification table. Refer to the ZIP Classification — Scale / Observe / Ignore / Block section for full column definitions.

Ad Impact by ZIP — Control vs Exposed Conversion Rate

This chart shows the direct impact of advertising at the ZIP level by comparing conversion rates between control (no ads) and exposed (with ads) users and visualizing the lift.

Chart View

What It Shows:

• Each row represents a ZIP code
• Blue dot: Control conversion rate (baseline, no ads)
• Green dot: Exposed conversion rate (with ads)
• The line connecting the dots represents the incremental lift
  • Longer lines = larger impact
• Whiskers on each dot show the 95% credible intervals

Hover Data:

• ZIP Code and name
• Conversion Rate
• FDR-adjusted p-value: Indicates statistical confidence after correcting for multiple comparisons
  • Lower values = higher confidence
  • Helps confirm whether observed lift is reliable

Table View

This is the same dataset as the ZIP Classification table. Refer to the ZIP Classification — Scale / Observe / Ignore / Block section for full column definitions.

Explorer

The Explorer tab allows you to browse and interact with all available datasets used in the model. Select a dataset to view its chart and table. Use the metric buttons above charts to switch between available metrics.

What It Does:

• Lists all datasets in the left-hand panel
• Displays each selection as a chart and table

This section is intentionally extensive and exploratory. Users are encouraged to navigate different datasets and metrics to uncover additional insights.

Charts

The Charts tab provides access to all visualizations generated by the model, organized into categories for easier navigation and analysis.

Charts are grouped into the following sections:

• Decision Support
• Distributions
• Confidence
• Geographic

Downloads

The Downloads section provides full access to all model outputs, enabling deeper analysis, reporting, and sharing across teams. These outputs can be used for investigation and validation, but more importantly, the model can be applied directly to campaigns or line items, allowing insights to seamlessly translate into real-time bidding and optimization.

Available Downloads (JSON & CSV):

• Executive Summary: High-level summary of findings, insights, and key drivers
• ZIP Classification (Scale/Observe/Ignore/Block)
• High-Lift ZIPs (Scale Targets)
• Full ZIP-Level Lift Data
• Campaign Summary Metrics
• Incremental CPA by ZIP
• Targeting Rules (JSON)
• All Combined