
Context hints in ChatGPT Ads
how to write them
ChatGPT ads became available to Norwegian advertisers in late August 2026. Ad groups contain one field that looks familiar to anyone who has worked with Google Ads, and that behaves completely differently: the context hint. Here is what the field does, how the system uses it and how to write hints that actually put the ad in front of the right conversations.
A context hint is a plain-language description of the conversations, needs and situations your ad group is built to support. The field sits at ad group level, next to the group’s title, and applies to all ads in the group.
OpenAI is unusually clear about what the field is. Context hints describe a situation. The system decides for itself which conversations are similar enough, and the field gives you no exact-match control, no audience targeting rules and no guarantee of showing in specific conversations. You are therefore buying a situation. You describe a situation, and the system judges for itself whether the conversation is similar enough.
The difference is bigger than it sounds. In Google Ads, the keyword is almost a contract: there are match types, negative keywords and a search terms report that shows what you actually paid for. In ChatGPT Ads, the hint is an argument you put forward to a model that already reads the whole conversation. You improve the odds of the right placement, and the model controls the outcome.
Relevance is decided by several things at once: the context and intent of the ongoing conversation, the landing page, the ad title, the ad text and your context hints. When the user has turned on ad personalization, selected signals from broader ChatGPT use count too. The selection itself happens in a relevance-weighted second-price auction.
Two things follow from this.
The hint is one of several inputs, alongside the landing page, the ad title and the ad text. A well-written hint on a thin landing page ends up as weak as the page.
The landing page counts both as a relevance factor and as a destination. It must also be accessible to OAI-AdsBot and OAI-SearchBot, or the ad is rejected no matter how good the text is.
Worth including in your planning: ads are shown only to free users and Go users over 18. Plus, Pro and Business subscribers get an ad-free experience. Your audience in ChatGPT is therefore systematically skewed toward those who pay the least for the tool, and that should affect both the offer and your expectations for conversion value.
A keyword lets you advertise to people you don’t know. A context text makes that impossible. You actually have to know why the customer is there, or you can’t write it.
A usable hint answers four questions in one to two sentences: who, where in the decision process, why now, and what characterizes a good match.
[Who] who [is exploring, comparing or ready to buy] [product or category] because [trigger or need], and who typically mentions [concrete match signals].
Put together, it can look like this:
Small online stores comparing email marketing tools because follow-up after purchase is random, and who need automatic cart reminders and predictable pricing as the list grows.
Compare with “email marketing.” One describes a conversation you can picture. The other describes an industry. Only the first gives the system something to recognize.
One ad group, one need. OpenAI recommends that each group covers one product category, one topic or one intent area. Merge two needs and the hint becomes vague for both, and the ads respond imprecisely to both.
Write full sentences. Natural phrasing beats isolated words. The system reads meaning, and meaning lives in sentences.
Name the problem before the persona. “Marketing managers” says little about when you are relevant. “Marketing managers who need to document campaign effect without an analytics team” says it.
Say where in the buying process the reader is. Some are trying to understand the category, some are comparing two vendors, some are about to order today. That is three different conversations, three different ads and therefore three different groups.
Include concrete match signals. Platform, team size, budget level, geography, deadline, integrations. Include what actually decides whether you fit, and leave the rest of what you know about the customer out.
Skip synonym lists. Matching is semantic. Listing “cheap, affordable, inexpensive” gives the same coverage as one of the words and only makes the hint muddy.
Set boundaries positively. Here you steer with a precise description instead of negative keywords. If you want to avoid the wrong conversations, describe the right situation precisely instead. If you must mark a boundary, do it as a property: “paid tools for businesses, not free resources for students.”
Be specific and brief. One to two sentences are enough. If the hint needs five, you probably have two ad groups waiting to be split.
Keep hint, title, text and landing page in the same story. All four should solve the same need. Conflicting signals are the most common reason a technically approved ad performs poorly.
Test topics in separate groups. Mix two hypotheses in one group and you get a number that points in two directions. Isolate the variable, or you learn nothing from the budget you spent.
Most weak hints share the same flaw: they name a category instead of describing a need.
The same customer asks different questions at different times. A single hint that has to cover all the phases gets diluted, so split them up.
The temptation is to build twenty groups in the first week, one for every nuance you can think of. That rarely produces learning, because each group gets too little volume to say anything. Three to five groups are a better starting point. Create one for the problem phase, one for the comparison phase and one for the purchase-ready, possibly supplemented with one per main audience if they need different documentation or a different landing page.
Then you expand where it works. A group that delivers can be split into two more precise ones. A group that delivers weakly is usually either too broad in the hint or mismatched with the landing page. You see both faster when you have few groups to keep track of.
The Google Ads keyword list pasted straight in. Isolated terms are exactly what OpenAI asks you to avoid.
A hint that repeats the ad text. The field should add something beyond the title, text and landing page.
All audiences in one group. A solution for an individual and a solution for a corporation need different documentation, different pricing and different landing pages.
Persona without a problem. A job title alone says who might be relevant, and little about when you are.
A hint that promises more than the landing page delivers. That means rejection at best and wasted clicks at worst.
Collect 25 to 50 questions customers actually ask. The sources already exist: sales calls, customer service, search data, chat logs.
Group the questions by decision.
Write one hint per group, using the formula above.
Create a separate ad group per topic, with several ads in each so you can test messages.
Check that the landing page answers the same need and is open to the OAI bots.
Evaluate on results. Impressions only tell you the system found you relevant enough to try.
Keep a change log. With a change log, adjustment becomes testing.
Can you picture the concrete conversation the hint describes?
Is the problem there, in addition to the audience?
Is the intent phase clear?
Is there at least one concrete match signal, and are all of them relevant?
Does the hint add anything beyond the title, text and landing page?
Do hint, ad and landing page point to the same need?
Is the hint one to two sentences?
The field is new, and OpenAI lacks an official character limit and recommended length. Industry guides have settled on one to two sentences as a practical norm, based on observation more than documentation.[1] It is unclear whether hints should be written in Norwegian or English for Norwegian conversations. Until someone has tested it properly, the most reasonable approach is to write the hint in the same language as the conversation you want to match. Set up an English variant as its own group that you measure against.
Sales in the Nordics currently go through agency partners, and a self-serve platform has been announced without a date. Expect both the interface and the advice above to change quickly. What is unlikely to change is the premise: you describe a need to a model that reads the conversation, and the one who knows their customer best wins.
Read more

700 meters inside the mountain: Lefdal Mine Data Centers
Read more in the glossary about ChatGPT, AI advertising and keywords.

