Contextual targeting in ChatGPT ads: how context hints work and how to write them

Ads in ChatGPT are targeted by context. An ad group carries context hints, free text that describes what you sell, what it is for and what need it answers, and the platform decides when a conversation fits. There are no keywords, no match types and no query reports, so the quality of the hints and the way ad groups are structured decide what you can learn. This guide covers what a hint is, how to write one, which angles to cover, what happens when you update the list, the open question of language, and how to measure per angle.

What is a context hint?

A context hint is a free-text description of a product, a use case or a need, attached to an ad group, up to 2,000 per group. It is not an exact-match keyword: the platform matches the meaning of a conversation against the hints, and it does not tell you which prompts produced an impression. Hints are set when the ad group is created and can be updated later.

Because matching is by meaning, a hint is addressed to the platform, not to a search engine. Think of it as a brief to a person who has to decide whether your ad belongs in a conversation: describe the situation, the need and the fit, and the platform can match conversations you would never have listed as keywords. The other two targeting settings are locations and platforms; hints carry the rest.

How do you write a useful hint?

As a sentence about a situation, not as a phrase. A hint such as espresso grinder tells the platform only the product category. A hint such as someone in a small flat wants a burr grinder that is quiet enough to use early in the morning, for under 300 euro describes a need, a constraint, a setting and a budget, which is what a conversation actually contains. The second form is what Adsonomy generates and what its hint quality check enforces.

Three tests. Is it specific: does it name the need, the setting or the constraint rather than the category? Is it honest: does it avoid competitor names and claims that would need substantiation? Is it one situation: a hint that bundles three ideas is three hints. Adsonomy's editor writes 20 to 200 hints per narrative as full sentences, with a live count, and flags a hint that reads as a keyword, names a competitor or makes a claim it cannot back. The examples in this guide are ours; the platform documentation states the principle, free text describing products, use cases and needs, and Adsonomy applies it.

Which angles should the hints cover?

One ad group should cover one angle on one offering. Adsonomy calls this a narrative and works with ten angle kinds: use case, problem, occasion, persona, comparison, objection, complement, season, service outcome and local. Each becomes its own ad group and is measured on its own.

One example hint per angle, for a grinder and for a service:

• Use case (what it is for): wants to make flat whites at home with a grinder fine enough for espresso.
• Problem (what it solves): the grinder they own is too loud to use while a child is asleep.
• Occasion (when it is bought): looking for a present for someone who has just bought an espresso machine.
• Persona (who buys it): a first-time home barista who does not want to learn dialling in from scratch.
• Comparison (versus the alternative): deciding between a stepped and a stepless grinder for espresso.
• Objection (why not to worry): worried that a cheap grinder produces uneven grounds and bitter shots.
• Complement (what it goes with): has an espresso machine and needs a grinder that suits its portafilter size.
• Season (the calendar): setting up a coffee corner at home before a winter of working from home.
• Service outcome (the result the client gets): wants a bookkeeper who closes the year without an extension and explains what changed.
• Local (a city or region): needs a bicycle repair shop in Utrecht that can fix an e-bike motor this week.

A grinder sold on quietness and the same grinder sold as a gift are two different conversations, with different rates of clicks and orders. Merged into one group, the numbers average out and neither angle can be improved.

What happens when you update the hints?

The update replaces the list. Sending a shorter list removes the hints that are missing, so an addition must resend the full list plus the new lines. Adsonomy stores every narrative's hints and always resends the complete list; its hint action is an add, never a replacement by accident.

Updates are also worth spacing. A hint change alters which conversations an ad group matches, and it takes time before performance reflects it. Adsonomy's guardrails set a cooldown per action type, 72 hours by default for hints, so a group is not rewritten before its last change has been measured. Creating an ad group or an ad is never automatic: Nomy proposes a narrative and the customer approves it. A changed creative also goes back through review on the platform.

What about language?

The OpenAI documentation has no language targeting field and says nothing about language. Targeting is locations and platforms only. Adsonomy's operating assumption, marked as an assumption in the product and to be confirmed with OpenAI, is that matching is multilingual while a creative is not: a Dutch title and body are read by a Dutch speaker. So Adsonomy places language at the ad group level and keeps one ad group per narrative and language.

In practice a narrative has a language tag, its hints are written in that language, and an English mirror of the hints is kept alongside and can be switched off. The mirror costs nothing and covers the case where the platform matches on English regardless of the conversation language. If OpenAI confirms that matching is independent of hint language, the mirror can be turned off; if it confirms the opposite, the per-language structure is already in place. Either way, the assumption is visible rather than buried.

How do you measure per angle?

Because one ad group is one angle in one language, the platform's ad group insights are the per-angle report: impressions, clicks, spend, CTR, CPC, conversions, cost per acquisition, post-click conversion rate and, with order events, attributed sales and ROAS. Adsonomy refreshes a 7-day and a 30-day performance snapshot on each narrative and shows them in a matrix of offerings by angle by language, with coverage gaps highlighted.

Reporting can be segmented by one dimension at a time: country, device, platform or product. There is no segment by hint and no prompt report, so the measured unit stays the ad group. Daily insights are retained for a year, hourly for about 30 days. Conversions take at least a day to appear, so a narrative is judged on a window, not on a morning. Rules can act on these numbers, for example pausing a narrative with no conversions after a set spend. In Autopilot, Nomy proposes new narratives from the catalogue, service pages, GA4 landing pages and the strategy, and retires the ones that do not earn their spend, within the same guardrails and change log.

Frequently asked questions