A new campaign owner opens a file called “test final.” It contains a green percentage but no explanation of what changed, why several days are missing or which page visitors actually saw. That result is difficult to use for the next ChatGPT ads campaign. An experiment log should make the decision understandable to someone who was never in the meetings.
The log can live in an internal document, spreadsheet or system the team already uses. The structure below is our recommendation, not a promised feature of ChatGPT Ads or AthillyAds. Its purpose is to let the next owner follow the path from intention to observation and action without guessing.
Give the experiment an identity outside the campaign
Create a durable test identifier, such as EXP-2026-014, distinct from the campaign name. One experiment may involve several campaign objects, while a campaign may continue through successive investigations. A label such as “new headline” will not distinguish those histories several months later.
Map the test identifier to the relevant account, campaign, ad group and ad identifiers. OpenAI recommends durable IDs in its campaign management documentation. Include the destination version and the identifier of any separate testing system. That additional connection must not imply that the advertising platform performed random assignment.
Name the owner of the question and the person authorized internally to decide what happens next. They may be different people. Also identify who can restore the page, change campaign status and explain the analysis. A log with an author but no action owner hands over a document rather than a working process.
Preserve the plan as it existed before launch
The plan entry should describe the specific change, its comparison and the mechanism that could affect the selected outcome. State the population, assignment unit and analysis unit. If regions receive treatments but thousands of impressions are recorded, the log must not make those impressions look like thousands of independent assignments.
Write the primary outcome’s numerator and denominator in words rather than only using an abbreviation such as CVR. “Qualified companies divided by assigned companies” differs from “registrations divided by clicks.” Specify observation duration, exclusions, business threshold, stopping rule and planned analysis. Link to a longer protocol if necessary, while retaining the central definitions in the log.
Give the frozen plan a version number and timestamp. Later changes need new entries with reasons and consequences. Do not replace the original success criterion after examining the outcome. A reader must be able to establish whether a conclusion followed the original plan or arose from later exploratory analysis.
Make every data extract identifiable
For each analysis input, record extraction time, source system, reporting period, timezone, filters, aggregation level and requested fields. For conversion analysis, include the time basis, attribution windows and outcome definition. OpenAI’s reporting documentation is the reference for reporting terminology. Record your actual settings instead of writing “default.”
Keep the source extract alongside the transformed table and calculation version. Record a clear file location or internal link, format and access owner. A screenshot of a total is insufficient if the next analyst needs to investigate a filter error. Equally, avoid collecting personal information irrelevant to the question; an evidence trail can often use aggregated data and controlled internal references.
Consider a hypothetical case where an initial extract contains 48 attributed outcomes and a later extract contains 54 under the same reporting definition. That is six additional outcomes, or 12.5% of the original count. It is a data revision. It does not establish that advertising improved by 12.5% between analysis sessions. Retain both extracts and explain which one supported the decision.
Where refreshed data change the conclusion, create a new decision entry rather than silently replacing the earlier one. Record whether an operational action had already happened. The next owner then sees both what the team knew at the time and what became available later.
Four records that preserve the decision context
- Plan before launch
Freeze question, unit, primary outcome, threshold and analysis version under one test ID.
- First extraction
Snapshot A contains 48 outcomes and the time they were extracted.
- Traceable revision
Snapshot B contains 54: 6 more, or 12.5%. Retain both versions.
- Decision and owner
Link the decision to A or B and identify who executes the next action.
Record deviations while the details remain available
A deviation entry needs a time, affected objects, verified event and possible impact. Examples include an incorrect destination, a changed qualification rule or a discount accidentally offered to control. Keep observation and interpretation in separate fields. “The field was absent between two releases” is more useful than “poor traffic.”
If group experiences overlapped, link to the assessment of test contamination. Preserve incidents that did not change the analysis too, including the reason they were considered immaterial. Otherwise the next analyst may rediscover the same event and repeat an investigation without its original explanation.
For conversion events, note the business action represented by the integration. OpenAI’s conversion tracking documentation provides technical terminology, while the log should clarify whether a lead meant a submitted form or an internally accepted enquiry. That definition determines which conclusions the outcome can support.
Separate the result, the decision and remaining work
The result entry should identify the estimate, uncertainty, population and analysis version. The decision entry should state what the team will do and why. The same uncertain finding can justify different actions when implementation costs and risks differ. Refer to decisions after inconclusive results or handling negative results where appropriate.
Assign an owner, timing and verifiable completion condition to the next action. “Restore page version 3 and check the destination” can be followed up. “Continue optimizing” cannot. Also record the condition for revisiting the idea, preventing the same proposal from returning without a new reason.
Finally, test the handover with a colleague who did not participate. Ask them to locate the original question, the data version used, the main limitation and the agreed action. Anything requiring an oral explanation is still missing from the log. A well-supported rejection or a clearly unresolved question is just as worth preserving as a successful rollout.
Sources and scope
Design an internal experiment register that preserves implementation, data revisions and decisions through a change of campaign owner.
Working methods and examples are editorial suggestions. Check current platform requirements and available features before implementation.