How to write ChatGPT Ads context hints that actually get your ad into the conversation

 

Most advertisers will treat context hints like keywords. That is the first mistake, and it costs them the whole channel.

 

Context hints are not exact match controls, not audience rules, and not instructions to show your ad in specific chats. They describe what your product offers, who it helps, and when it is useful. The system uses that to understand the broader need behind a conversation.

 

Read that again. You are not bidding on words. You are briefing a model on when your product becomes the right answer.

 

Everything below follows from that one shift.

 

 

Stop writing keywords. Start writing situations.

 

In Google Ads the unit of relevance is the query. In ChatGPT Ads the unit of relevance is the moment inside a conversation where a human has a problem your product solves. Nobody types "cushioned running shoes" into ChatGPT. They say something like: "I signed up for a 5K in October, I have never run before, my knees hurt after ten minutes, what am I doing wrong?" The shoes are two questions away.

 

OpenAI's own example makes the point. "Running shoes" is a bad hint. "Cushioned everyday running shoes for beginners training for their first 5K" is a good one. The difference isn't length; the difference is that the second one contains a person, a stage of life, and a reason to buy right now.

 

So the test for every hint you write is three words: what, who, when. What does the product actually do beyond the ad copy? Who is the person in the chat? When, in their life or their week, does this become useful? If a hint answers only one of the three, it's a keyword wearing a longer coat.

 

The four jobs a context hint can do

 

OpenAI lists four things hints help with, and each one maps to a different kind of hint you should be writing. Don't write ten variations of the same job.

 

Job one is covering related use cases. One product, several reasons people reach for it. A project management tool gets used by a freelancer juggling five clients and by a marketing lead planning a product launch. Same product, different conversations. Write one hint per use case.

 

Job two is narrowing a broad ad message. Your ad copy says "accounting software for small businesses." Fine, but which small business, in which panic? "Sole trader who just received a VAT registration letter and has no idea what to file" is a conversation that happens. Write the specific need the copy can't fit.

 

Job three is language variation. People describe the same problem in wildly different ways, and the model needs to know they're the same problem. "My website is slow," "my Shopify store takes forever to load on mobile," and "customers keep abandoning checkout" can all be the same hosting problem. You don't need every phrasing, because the system generalizes; you need enough variety that the model understands the shape of the need rather than a single sentence.

 

Job four is defining conversation types. This one is underrated. Some products are relevant during planning conversations, some during comparison conversations, some during troubleshooting. A travel insurance product belongs in "I'm planning a two week trip to Japan with my parents in their seventies" and not in "cheapest flights to Tokyo." Say which kind of chat you belong in.

 

A format that works

 

Here is the pattern I'd use, and I'd use it consistently across every ad group so the quality stays even:

 

[product attribute] for [specific person] who is [situation or trigger], especially when [conversation context].

 

Applied to a hypothetical POAS tool for Shopify merchants:

 

"Profit based bidding integration for Shopify store owners who see rising Google Ads revenue but shrinking margins, especially when they are asking how to calculate real profit after ad spend, shipping and returns."

 

Compare that to "POAS bidding tool." One of these tells the model something. The other one tells it nothing it couldn't read off the landing page.

 

Not every hint needs all four slots. But if you find yourself dropping the situation slot every time, you've drifted back to keywords.

 

Where advertisers will get this wrong

 

Padding with audience labels. "Small business owners" is a demographic, not a need. OpenAI says this explicitly: describe needs or situations rather than broad audience labels. A label tells the model who; it says nothing about when.

 

Stuffing unrelated use cases into one ad group. The documentation gives a clean rule for splitting: if two hints would need different ad messaging or a different landing page, they belong in different ad groups. That's your structure test. Not "same product," but "same message and same destination."

 

Writing hints that just restate the ad. The system already reads your title, your copy, and your landing destination. A hint that repeats them adds zero information. Every hint should pass the question: what does this tell the model that it couldn't already infer?

 

Fake use cases to widen reach. This is the keyword broadening instinct and it will backfire, because you're not buying impressions on words, you're asking a model to place you in conversations where you make sense. Put a tax tool into cooking conversations and the best case is irrelevance; the worse case is the system learning your product isn't what you claim. Include only real, honest use cases. The doc says it; I'd say it twice.

 

Disconnected terms. "shoes, running, marathon, beginner, cushion" is a comma soup. Write sentences. Natural phrases, the way a customer would explain their problem to a friend.

 

How I'd build an ad group from scratch

 

Start from the conversations, not from the product. Pull twenty real support tickets, sales calls, or review sentences where a customer explained their problem in their own words before they knew your product name. Those are your raw hints. They already contain the who and the when.

 

Cluster them by the message you'd want to show and the page you'd send them to. Each cluster is an ad group. If you end up with one giant cluster, you haven't looked hard enough at the differences between your customers.

 

Inside each group, write hints across the four jobs: two or three use case hints, one or two narrowing hints, a few language variations, and one hint that names the conversation type. Ten well built hints beat forty lazy ones.

 

Then run multiple ads inside each group. OpenAI recommends this and it matters more here than on search, because you're testing message fit against conversational context, not against a query string. Different framings will win in different conversation types and you won't know which without variations.