How to Optimise a ChatGPT Ads Campaign With No Search Terms Report
By Ansh Khandelwal 9 min read
In brief
Optimize ChatGPT Ads without query data by testing context-hint containers, reading controlled baselines, and using qualified-lead outcomes.
Last verified: 12 September 2026 | Version: 1.0 | Next scheduled review: 12 October 2026
You optimise by changing the container, not the traffic. ChatGPT Ads reports no queries, no conversations and no per-hint data, so every weekly decision has to be made from the four things that are readable: ad group splits, creative variants, the landing page, and where the budget sits.
That is the whole method. The rest of this article is why the familiar method does not transfer, and what a working week looks like once you accept that.
Why the Google optimisation loop does not exist here
On Google Ads, account management is a closed loop with a data source at the top of it. You read the search terms report, you see the queries you actually paid for, you add negatives against the bad ones, you raise bids or budgets against the good ones, and next week the report tells you whether that worked. The query is both the diagnosis and the treatment.
ChatGPT Ads removes the entire top of that loop. There is no search terms report. There is no query. The ad is matched against the meaning of a conversation, and OpenAI's Ads Manager reports impressions, clicks, spend, CTR, average CPC, average CPM and conversions at campaign, ad group and ad level, and nothing below that. There are no negative keywords to add and no match types to tighten. Two of the three steps in the Google loop have no equivalent action available.
The mistake advertisers make in month one is to keep running the loop anyway. They stare at an ad group's CTR trying to infer which conversations it came from, they rewrite hints in response to a number that cannot be attributed to any hint, and they conclude the channel is unreadable. It is readable. It is just readable at a different resolution.
Container optimisation: the four readable units
Call the replacement method container optimisation. You cannot see or filter what arrives, so you change the shape of what it arrives into and read the difference in the aggregate numbers. Four units are genuinely readable, in descending order of how much they tell you.
The ad group. This is the smallest unit the platform will ever attribute a number to. An ad group is a hint theme plus a bid plus a default destination. If an ad group returns a number, that number belongs to its theme. This is why one theme per ad group is not tidiness, it is the only way you get an interpretable result at all.
The creative variant. Ads report at ad level. Two ads in the same group share the same hints, the same bid and the same matching surface, which makes creative the one true controlled comparison available on this channel. Everything upstream is held constant by the platform's own structure.
The landing page. Not reported by OpenAI, but reported by your own analytics and your CRM, at the resolution you choose. This is the only layer where you can see more than the platform shows you.
Budget allocation between groups. Moving money from a group with a bad cost per qualified lead to one with a good cost per qualified lead is the single action that most resembles adding a negative keyword, because it does the same job: less money into the wrong intent, more into the right one. It just costs you one round of spend to learn it, instead of being preventive.
What replaces each Google habit
| Google habit | Why it fails here | What you do instead |
|---|---|---|
| Read search terms, add negatives | No query data of any kind exists | Split the suspect theme into its own ad group, then pause or defund it on its own numbers |
| Tighten match types | No match types exist | Add a constraint inside the hint, before the impression is bought |
| Bid down on poor-converting keywords | No keyword-level bid exists | Bid at ad group level, since that is where the bid lives |
| A/B test with a query-level holdout | No query layer to hold out | Test creative inside one ad group, where hints and bid are already held constant |
| Add high performers to a dedicated campaign | Nothing to promote at query level | Promote a hint theme into its own ad group with its own budget |
| Watch quality score | No visible quality score | Watch CTR and average CPC together, since the auction is relevance-weighted |
The weekly loop that does work
The cadence matters as much as the actions, because the sample sizes on this channel are small and the temptation to act on noise is high.
Once a week, read four numbers per ad group. Spend, CTR, average CPC, and cost per qualified lead from your own system. Not conversions as OpenAI counts them, if you have anything better. Independent agency testing reported to MediaPost in August 2026 found click counts from OpenAI that did not reconcile with Google Analytics sessions, in one case 57 reported clicks against fewer than 20 recorded visits, and not always erring in the same direction. Treat that as reported, single source, and treat your own analytics as the tiebreak rather than the platform's number.
Change one layer at a time. Because your only comparison is before and after, two simultaneous changes destroy the reading. If you rewrite the hints and swap the creative in the same week, you have one number and two candidate explanations.
Move budget before you rewrite hints. Budget reallocation is reversible, immediate and does not reset any learning. Hint rewrites are slow to read and you will never know which hint did it. Exhaust the cheap lever first.
Split before you optimise. When an ad group's numbers are mediocre and you cannot explain why, the correct move is usually not a fix. It is a split. Two groups produce two readable results from the same spend, where one group produced one uninterpretable result.
The reckoning: splitting costs you speed
Every split halves the spend flowing into each resulting group, which doubles the time until either group has enough volume to read. At a 25 USD daily minimum per campaign, and with reported CPC guidance sitting in the three to five dollar range, an ad group running at roughly 12 USD a day is buying you something like three clicks a day. Split it and you are reading two units at under two clicks a day each. That is not a decision-grade sample in a fortnight, and it may not be one in a month.
So the structural advice and the statistical advice are in direct tension, and anyone who tells you to simply split everything has not done the arithmetic. The resolution is a sequence rather than a rule. Start with fewer ad groups than you want, at budgets high enough that each one produces a number you can believe, and split only when a group has accumulated enough spend that halving it still leaves both halves readable. Legibility you cannot afford to read is not legibility.
What we cannot tell you
- Which conversations your ads appeared in. Absent by design. No search terms report, no adjacency reporting, no conversation data.
- Which context hint drove any result. Not reported at hint level, and no workaround exists in Ads Manager.
- How long an ad group needs to produce a stable CTR. No advertiser has published a variance study for this channel, so every recommended waiting period in circulation, including ours, is inference.
- Whether OpenAI's click count is accurate. Agency reports in August 2026 describe discrepancies against site analytics in both directions. OpenAI has not published a reconciliation method.
- What a good cost per qualified lead is here. OpenAI publishes no cross-advertiser benchmarks, so the comparison has to be internal or against your other channels.
Quick answers
Does ChatGPT Ads have a search terms report? No. There is no query layer, no search terms report, no conversation data and no per-hint reporting. Impressions, clicks, spend, CTR, average CPC, average CPM and conversions are reported at campaign, ad group and ad level only.
What do I optimise if I cannot see queries? Ad group structure, creative variants inside a group, the landing page, and where budget sits between ad groups. Those four are the only units the platform or your own analytics will attribute a number to.
How often should I change something? Weekly at most, and one layer at a time. With no query data, before-and-after is your only comparison, and two simultaneous changes leave you with one number and two explanations.
Is rewriting context hints an optimisation lever? It is the slowest and least readable one. Nothing is reported at hint level, so a hint rewrite produces a change you cannot attribute. Reallocate budget between ad groups first.
How do I stop wasting spend on bad-fit conversations? You cannot prevent it, only respond to it. Separate the suspect theme into its own ad group so it becomes pausable, and put a disqualifying condition in the ad copy so the wrong reader does not click.
Should I trust OpenAI's click numbers or my analytics? Reconcile both and investigate the gap. Multiple agencies reported mismatches in August 2026. For spend decisions, your own qualified-lead data is the more defensible input.
Sources
| Claim | Source | Tier |
|---|---|---|
| Reporting covers impressions, clicks, spend, CTR, avg CPC, avg CPM, conversions at campaign, ad group and ad level | OpenAI Ads Manager documentation, 2026 | Confirmed, primary |
| No search terms report, no query data, no per-hint reporting, no conversation data | OpenAI Ads Manager documentation, by absence | Absent |
| Context hints are set at ad group level and have no negative or exclusion equivalent | OpenAI Help Center, Create Ad Groups for ChatGPT Ads, August 2026 | Confirmed, primary |
| Auction is relevance-weighted | OpenAI documentation, 2026 | Confirmed, primary |
| Three to five dollar CPC starting bid guidance | Multiple agency write-ups citing OpenAI guidance, 2026 | Reported |
| 25 USD minimum daily budget | OpenAI Ads Manager, 2026 | Confirmed, primary |
| 57 reported clicks against fewer than 20 analytics visits; discrepancies in both directions | MediaPost, 31 August 2026, citing Nicholas Verity of Cleverly and others | Reported, third party |
| Splitting ad groups halves readable volume per unit | Arithmetic from the 25 USD floor and reported CPC range | Inference, ours |
Related reading
- How to Write a ChatGPT Ads Context Hint
- Can You Exclude Conversations on ChatGPT Ads?
- How to Measure ChatGPT Ads, and What You Cannot Measure Yet
- ChatGPT Ads Troubleshooting: Documented Failure Modes
- How Many Ad Groups Should a ChatGPT Ads Campaign Have?
Changelog
12 September 2026, v1.0. First publication. Establishes container optimisation as the replacement for the Google search terms loop, and states the split-versus-sample-size tension rather than resolving it with a rule.
Field kit
Tools
Site
OpenAI (primary)
Ansh works across GEO strategy, B2B research, and execution. At InPromptAds, he translates new AI advertising products into clear operating advice, tests, and measurement questions for marketing teams.
