# ChatGPT Ads performance study protocol

Status: Methodology only. No campaign benchmark or ROAS result has been published.
Advarde · 11 September 2026

## Research question

Question:
Business decision the study supports:
Observation period:
Eligible account population:
Inclusion rules:
Exclusion rules and reasons:
Publication permissions:

## Data dictionary

| Field | Definition / source |
| --- | --- |
| account_key | Redacted stable account identifier |
| campaign_key | Stable campaign identifier, redacted where needed |
| cohort_start / cohort_end | Complete dates with account timezone |
| currency | Preserve separate currencies unless a documented conversion method is used |
| geography | Actual included campaign markets |
| offer | Product / service and destination version |
| paid_spend | Media spend from a verified source |
| other_acquisition_costs | Explicit included costs; unknown stays unknown |
| impressions / clicks | Platform observations and refresh timestamp |
| conversion_event | Exact business action and trigger |
| attributed_conversions | Platform attribution definition and window |
| qualified_outcomes | First-party qualification rule and source |
| customers / orders | Confirmed business records |
| revenue / refunds | Financial source and completeness |
| contribution | Defined revenue less included variable costs |
| follow_up_complete | Whether the cohort has matured enough for the outcome |
| exclusions | Missing, incompatible, duplicate, or otherwise ineligible records |

## Calculation rules

CPC = paid spend / clicks; undefined when clicks are zero.
Cost per qualified outcome = included acquisition cost / qualified outcomes.
ROAS = attributed revenue / paid spend; label attribution and treatment of refunds.
Contribution after acquisition = contribution less included acquisition costs.
Recompute rates from totals. Do not average row-level rates without a justified weighting method.

## Analysis requirements

- Separate paid campaigns from organic ChatGPT referrals.
- Preserve account, currency, attribution, and observation-window boundaries.
- Include eligible unsuccessful tests, not only selected successes.
- Show sample counts, distributions, missingness, and incomplete follow-up.
- Mark creative, destination, budget, and tracking changes.
- Distinguish descriptive comparisons from causal incrementality claims.
- Report uncertainty appropriate to the sample and design.

## Publication gate

- [ ] Source totals reconciled.
- [ ] Exclusions and missingness documented.
- [ ] Metrics reproducible from a redacted dataset or inspectable calculation record.
- [ ] Claims reviewed against the actual evidence.
- [ ] Permission obtained for identifiable account examples.
- [ ] No credentials or personal customer data in the publication.
- [ ] Limitations and selection process included beside the findings.

Reference: https://advarde.com/research/chatgpt-ads-performance-study