How to Write a ChatGPT Ads Context Hint
By Ansh Khandelwal 9 min read
In brief
A context hint is a brief for a matching system that never reports back, so hints must be written to stay interpretable after the fact, not only to match well.
Last verified: 12 September 2026 | Version: 1.0 | Next scheduled review: 12 October 2026
Every guide to context hints teaches you to phrase one. None of them deal with the thing that decides whether your phrasing was any good, which is that OpenAI will never tell you.
Ads Manager reports impressions, clicks, spend, CTR, average CPC, average CPM and conversions at campaign, ad group and ad level. It reports nothing at hint level. Write eight hints into an ad group, spend two thousand dollars, and you will know precisely what the eight of them did together and nothing about what any one of them did.
That is not a gap that better writing closes. It is a constraint that changes what good hint writing means.
What a context hint actually is
A context hint is an instruction to a matching system, written in plain language, describing the situations in which your product is genuinely useful. It is not a keyword, not an audience definition, and not a query string. OpenAI matches it against the meaning of a live conversation, and it never tells you which hint did the matching.
OpenAI's own documentation, in its Create Ad Groups guide updated August 2026, gives one worked contrast: instead of "running shoes", write "cushioned everyday running shoes for beginners training for their first 5K". Its three stated rules are to use clear natural phrases rather than disconnected terms, include only genuine use cases, and keep each hint inside the ad group's single theme.
That is the whole of the official guidance. Everything below is what you have to work out for yourself.
The three-part template
A hint that matches well contains an audience, an intent and a constraint. Drop any one and it degrades in a predictable way.
Audience is who is in the conversation. Not a demographic. A role and a situation: "an operations lead at a 40-person logistics firm".
Intent is what they are trying to accomplish in that conversation, in the present tense: "comparing route planning tools after outgrowing spreadsheets".
Constraint is the specific condition that makes your product the right answer rather than a plausible one: "needs it to read existing driver schedules without a data migration".
Written out: An operations lead at a 40-person logistics firm comparing route planning tools after outgrowing spreadsheets, who needs the tool to read existing driver schedules without a migration.
Drop the constraint and you match every route planning conversation, including the enterprise ones you cannot serve. Drop the intent and you match people reading about logistics. Drop the audience and you match students writing coursework.
Six worked examples
| Weak hint | What it matches | Rewritten |
|---|---|---|
| project management software | Every PM conversation including free-tool hunting | A team lead at a 15 to 50 person agency whose projects span multiple clients, comparing tools after outgrowing a shared spreadsheet, who needs client-visible timelines without paying per guest seat |
| accounting services | Students, DIY filers, job seekers | A founder of a UK limited company approaching their first year end, deciding whether to file accounts themselves or appoint an accountant, who has payroll and needs someone who handles both |
| CRM for startups | Established teams doing vendor reviews | A two-person sales team at a pre-Series A startup moving off spreadsheets, who want pipeline visibility without an implementation project |
| cyber security | Anyone reading breach news | An IT manager at a mid-size firm being asked by their insurer to evidence endpoint protection before renewal, who needs deployment across 200 laptops without hiring |
| best CRM | Comparison and review-reading traffic, very broad | A revenue operations lead consolidating two CRMs after an acquisition, who needs both histories preserved and reporting unified |
| conveyancing solicitor | Legal study, general property reading | A first-time buyer in England whose offer has just been accepted, who needs a conveyancer to instruct this week and wants a fixed fee quoted upfront |
Two patterns hold across all six. Every rewrite contains a trigger event, the thing that just happened to put this person in this conversation. And every rewrite contains something you could disqualify on, which is what keeps the match narrow.
How specific is too specific
The failure mode at the far end is real but rarer than the advice suggests. A hint written as a single exact sentence does not fail because it is long. It fails because it describes one phrasing of a need rather than the need itself.
The test: write your hint, then write three genuinely different ways someone might arrive at that need. If your hint plausibly covers all three, it is at the right altitude. If it covers only the one you started with, you have written a query, not a hint.
Personas are the opposite failure. Six hundred words describing values, job title, team structure and buying committee gives the matcher more surface to match loosely against, not more precision. Use the shortest description that carries the real need and its important constraint.
The rule nobody writes down
Because nothing is reported at hint level, the ad group is the smallest unit you will ever be able to interpret. Every decision about hint placement is therefore a decision about what you will be able to learn.
Call it legible targeting: structure the account so that every number it produces can be attributed to a cause.
In practice this means one hint theme per ad group, always, even when two themes feel adjacent enough to share. If an ad group holds hints about payroll compliance and hints about expense management, and it returns a 0.4% CTR at a four dollar CPC, you have learned nothing actionable. You cannot tell whether payroll works and expenses drags, or the reverse, or whether both are mediocre. Split them and the same spend produces two readable results.
This is also why the standard advice to "add more hints for reach" is bad advice in the first month. More hints in one group buys you volume at the price of interpretability, on a channel where interpretability is already the scarcest thing you have.
The reckoning: the test that contradicts all of this
In 2026 Search Engine Land ran a controlled test that cuts directly against the house advice of this category, including some of this article. They built one campaign with two ad groups and identical creative in both. One group was filled with a Google-style keyword list. The other was filled with complete natural-language questions of the kind everyone here tells you to write.
The keyword group won. It served far more often, at a slightly higher CTR and a slightly lower average CPC.
That is one test, one advertiser, one category, one point in time, and it is third-party reporting rather than something we have replicated. But it is the only published head-to-head on this question, and it should make you hold the conversational-phrasing orthodoxy loosely. The defensible reading is that OpenAI's matcher is less sensitive to input style than the guidance implies, and that specificity of meaning matters more than the grammar it arrives in. Both the keyword list and the narrative hint can carry specificity. Only one of them reads like advice.
What survives the test is the structural argument, not the stylistic one. Whatever style you use, one theme per ad group still decides whether you can read the result.
What we cannot tell you
- Which of your hints is working. Not reported at hint level, and no workaround exists inside Ads Manager.
- How OpenAI weights a hint against conversation meaning. Undocumented.
- Whether hint count affects delivery. Undocumented. The reach advice circulating in this category is inference, including ours.
- Whether length materially changes matching. Undocumented. OpenAI's example is one sentence long, which is a signal and not a rule.
- Approval rates or match rates by category. No advertiser has published them and OpenAI has not.
Quick answers
What is a context hint in ChatGPT Ads? A plain-language description, set at ad group level, of the situations where your product is genuinely useful. OpenAI matches it against the meaning of live conversations. It is not a keyword and not an audience.
Should I describe my customer or their question? Their situation, which contains both. Audience, what they are trying to do right now, and the constraint that makes you the right answer rather than a plausible one.
How many context hints should one ad group have? Few enough that the ad group's reported numbers still mean something. One theme per group, always. Adding hints for reach costs you the ability to read the result.
Can I use negative keywords or exclusions? No. There is no exclusion layer. You separate low-fit themes into their own ad groups and let creative do the filtering.
How long should a context hint be? One or two sentences carrying an audience, an intent and a constraint. OpenAI's own documented example is a single sentence.
Do context hints guarantee my ad appears in those conversations? No. Hints guide matching in a relevance-weighted auction. They do not reserve placement.
Sources
| Claim | Source | Tier |
|---|---|---|
| Hints are set at ad group level and described in natural language | OpenAI Help Center, Create Ad Groups for ChatGPT Ads, updated August 2026 | Confirmed, primary |
| "Running shoes" versus "cushioned everyday running shoes for beginners training for their first 5K" | OpenAI Help Center, Create Ad Groups for ChatGPT Ads | Confirmed, primary |
| 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 reporting exists at context hint level | OpenAI Ads Manager documentation, by absence | Absent |
| Keyword-style ad group outserved and outperformed a conversational one with identical creative | Search Engine Land, how to run ChatGPT ads, 2026 | Reported, third party, single test |
| Relevance weighting affects what you pay, not only whether you serve | OpenAI documentation plus multiple agency accounts, 2026 | Confirmed, primary for the mechanism; vendor for the magnitude |
Related reading
- ChatGPT Ads Context Hints: What They Are, and What They Are Not
- Context Hints Generator
- How the ChatGPT Ads Auction Works: Objectives, Bids and Budgets
- How to Measure ChatGPT Ads, and What You Cannot Measure Yet
Changelog
12 September 2026, v1.0. First publication. Establishes legible targeting as the design constraint and documents the Search Engine Land keyword-versus-conversational result as an open contradiction in the category's advice.
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.
