Test the business action first, then event delivery, duplicate handling and attribution. A script loading successfully does not prove that a completed order is recorded correctly. An accepted event does not, by itself, prove that Ads Manager will attribute it to an ad.
For a founder or marketing team preparing a first campaign, the useful deliverable is a short evidence record: what happened, which event represented it, what your systems received and what remains unverified.
This guide provides a practical test plan. If you are still choosing an implementation, start with our Pixel and Conversions API comparison.
Keep three questions separate
Your team needs to answer three different questions:
- Did the business action happen? Your order, booking or account system supplies the evidence.
- Did measurement receive the right event? Your implementation and event diagnostics supply that evidence.
- Was the event attributed to advertising? Campaign configuration and reporting supply that evidence.
OpenAI says a reported conversion depends on receiving an event from a connected data source, matching the campaign's conversion configuration, satisfying the applicable attribution window and connecting the action to an eligible ad click using available signals. Receiving an event is one part of that process. OpenAI conversion measurement
Keep the advertising boundary clear, too. OpenAI states that ads do not influence ChatGPT's answers and that advertisers do not receive people's private conversations. Conversion testing measures activity on your business's side of the journey; it does not reveal the chat that preceded it. Ads in ChatGPT
1. Write the conversion definition in one sentence
Before opening a tag manager, finish this sentence:
We count a conversion when ______ is confirmed by ______.
For a fictional desk retailer, the answer could be: “We count an order when payment is confirmed by our order system.” For a service business, it might be a successfully stored enquiry. A button click and a saved enquiry are different milestones.
Our recommended test record starts here:
| Field | Fictional example |
|---|---|
| Business action | Confirmed desk order |
| System of record | Store order database |
| Measurement event | order_created |
| Internal test reference | test-order-1042 |
| Responsible owner | Ecommerce developer |
| Evidence | Confirmation record and event diagnostics |
Choose a test environment and procedure that your team can distinguish from genuine customer activity. These examples are a planning worksheet, not evidence of a GPTAd integration or campaign result.
2. Check browser and server delivery separately
The Measurement Pixel sends browser events. Its event options include event_id, used to identify an action also sent from the server. The Pixel captures the OpenAI click reference, oppref, and stores it in a first-party cookie for subsequent events. Measurement Pixel documentation
For the browser test, record the page and exact action that should trigger the event. Then test a failed submission, a successful submission and a revisit to the confirmation page. Compare what actually fires with the conversion definition you wrote.
For server testing, OpenAI provides validate_only: true, which validates events without saving them. The Conversions API must be called from the server. A failed event can cause its entire batch to fail, so inspect the response rather than assuming that sending a request means every event was accepted. Conversions API documentation
A successful validation checks the submitted payload; it is not a complete live attribution test. Record validation and live delivery as separate results.
3. Confirm that two delivery paths represent one action
When the same conversion goes through both methods, OpenAI requires matching identifiers: the Pixel's event_id and the API's id, together with the same Pixel ID and event name. Custom events also need the same custom-event name. The first matching event received is used; later duplicates are ignored. Pixel deduplication guidance
For our fictional order, the intended relationship is:
| Record | Identifier |
|---|---|
| Store order | test-order-1042 |
| Browser event ID | test-order-1042 |
| Server event ID | test-order-1042 |
This is an illustrative identifier map, not a copy-and-paste API payload. Ask the developer to show the actual values from each path. Two different IDs for one order deserve investigation before launch. Two different orders should not accidentally reuse one ID.
Also ask what happens if the server retries. Our recommended acceptance criterion is that retrying delivery does not create a second business action in the measurement record.
4. Test the journey, including the awkward parts
OpenAI recommends preserving oppref through redirects and navigation, and including it in server events when available. It also says conversion data must be shared only when permitted, with the necessary disclosures and consent. Conversion measurement guidance
Use this original QA worksheet to turn those requirements into reviewable checks:
| Scenario | Evidence to collect | Question to resolve |
|---|---|---|
| Landing-page redirect | Destination and parameter handling | Did the redirect preserve the click reference? |
| Successful action | Business record plus event record | Do both describe the same outcome? |
| Failed form or payment | Failure result and event log | Was a success event sent incorrectly? |
| Confirmation-page reload | Event IDs before and after reload | Can this create an extra conversion? |
| Browser and server both send | IDs and event names from both paths | Are the copies eligible for deduplication? |
| Measurement permission denied | Browser and server behaviour | Does collection follow the configured permission decision? |
| Missing click reference | Recorded fields and diagnostics | Is missing data visible rather than invented? |
Mark each row passed, failed or not tested. An untested checkout redirect should remain visible in the launch decision, rather than disappearing into “tracking installed.”
5. Reconcile evidence before reporting performance
Prepare a small reconciliation sheet with three columns: confirmed business actions, received measurement events and attributed conversions. Put the date range and reporting settings beside it.
OpenAI lists attribution settings, time boundaries, browser conditions, deduplication and modeled measurement among the reasons its totals may differ from other analytics tools. A difference alone does not establish a tracking bug. Conversion measurement guidance
Our recommendation is to investigate a specific example before changing the integration. Find one confirmed action, inspect its event and check the relevant reporting configuration. Avoid trying to “fix” a dashboard by making every platform display the same number.
For campaign labels and reporting comparisons, continue with How to Measure Your First ChatGPT Ads Campaign.
The handover your team should keep
A useful launch handover contains the conversion definition, responsible owner, completed worksheet, evidence location and unresolved limitations. It should let another person repeat the test without relying on whoever installed the tag.
GPTAd is an independent product being developed for conversational advertising preparation. This checklist is an editorial resource, not a claim that GPTAd installs the Pixel, sends conversion events or operates your Ads Manager account. Explore the GPTAd product preview.
Sources checked October 5, 2026. Recheck OpenAI's current documentation and your account configuration before implementation.
