Direct answer: OpenAI Ads targeting does not use keywords or audience segments. You write "context hints", natural-language descriptions of the ChatGPT conversations where your product belongs, at the ad-group level, and a relevance-weighted auction matches your ad to live chat topics. Advertiser geo targeting is country-level only (US, CA, AU, NZ today). Because relevance multiplies bid, a tightly specific ad can outrank a higher one.
Key takeaways
- The unit of targeting is a conversation, not a keyword. Context hints are natural-language descriptions of relevant chats, set per ad group; they guide matching but do not guarantee delivery and are not exact-match keywords.
- Relevance multiplies bid. Placement is decided by a relevance-weighted auction across four signals, context hints, landing page, ad title, ad copy, so specific, well-written ads can beat a higher bid.
- Use the Audience-Intent-Topic formula. Name the role/firmographic, the active verb (evaluating, switching, comparing), and the category plus a filtering constraint. One ad group per audience-and-intent combination.
- Geo is country-level. Advertisers can target the US, Canada, Australia, and New Zealand today. There is no documented state, DMA, or ZIP targeting. Where ads are shown is expanding; where you can buy is narrower.
- Targeting is privacy-safe by design. No user-level tracking, no demographic profiles, no chats or PII reach the advertiser, a durable model as third-party cookies disappear.
- Treat it as top/mid-funnel. ChatGPT ads catch research and comparison conversations; pair them with bottom-funnel search ads that capture decided intent.
Why does OpenAI Ads targeting work differently from Google and LinkedIn?
OpenAI Ads targeting breaks the model every paid-media team knows. On Google, you buy a keyword. On LinkedIn, you buy an audience. On ChatGPT, you describe a conversation. That single shift changes how you plan every campaign, structure every ad group, and measure every result. If you import a Google or LinkedIn playbook unchanged, you will mis-target from day one.
When a logged-in adult user on the Free or Go tier has a chat with commercial intent, ChatGPT may show a sponsored card at the bottom of the response, clearly labeled and visually separate from the answer (OpenAI: Testing ads in ChatGPT). The matching engine reads the topics in the current thread and surfaces the most relevant eligible advertiser. There is no keyword box and no audience builder. You define which conversations you belong in.
Takeaway: the durable mental shift is that you are buying relevance to a moment, not a query or a person.
This guide covers targeting specifically. For the ad format and basics, see ChatGPT Ads explained. For bidding and budgets, see OpenAI Ads campaign types and objectives. For writing the card itself, see the ChatGPT ad creative guide.
What does each ad platform actually target?
Most comparisons stop at ChatGPT vs Google. The useful framing is three-way: each platform targets a different layer of the buyer, the query, the person, or the conversation.
What are context hints in OpenAI Ads (and what are they not)?
Context hints are short, natural-language descriptions of the conversations where your product is genuinely relevant, set at the ad-group level. They are the targeting unit that replaces keywords and audiences on ChatGPT. OpenAI is explicit that hints "help guide matching but do not guarantee delivery in specific conversation types" and are "not exact-match keywords," and that ads are selected primarily on relevance to the conversation rather than keyword matching (PPC Land).
The difference from keywords is structural. A keyword is a token to match. A hint is a description of intent and situation. "CRM software" is a keyword. "RevOps leaders at Series A–C US healthtech companies evaluating HIPAA-eligible CRMs with EHR integration" is a context hint, it carries audience, intent, and a constraint that filters to your product.
How is ad placement decided on ChatGPT?
Placement is decided by a relevance-weighted, second-price-style auction that evaluates four signals: your context hints, your landing page content and quality, your ad title, and your ad copy. Ranking is roughly bid × relevance score, and keyword matching is not used (Context Hints: Definitive Guide). The practical consequence: specificity can beat raw spend. A tightly relevant ad at a $3 bid can outrank a generic one at $5.
| Dimension | Google Search Ads | OpenAI / ChatGPT Ads |
|---|---|---|
| Targeting unit | Keyword + match type | Context hint (natural language) |
| Audience segments | Demographic, in-market, remarketing | None, no user-level profiles |
| Auction inputs | Bid × Quality Score | Bid × relevance (hints, LP, title, copy) |
| Geo granularity | Country, region, city, radius | Country-level only (US, CA, AU, NZ) |
| Funnel fit | Bottom-funnel, transactional | Top/mid-funnel, research & comparison |
| Data back to advertiser | Query, device, geo, audience | Aggregated performance only |
| Benchmark CTR (directional) | ~4–6% avg | ~0.68% overall, ~1.0% top quartile |
CTR benchmarks are secondary and directional (WebFX). A lower CTR is expected, these are research conversations, not someone typing a buying query.
How do you write a context hint? The Audience-Intent-Topic formula
The Audience-Intent-Topic formula is the reliable way to write a context hint: combine three parts, never write a keyword.
- Audience, the role, company size, and industry. "RevOps leaders at Series A–C US healthtech companies."
- Intent, an active verb that signals where they are: evaluating, switching, comparing, migrating from.
- Topic, the category plus a constraint that filters to your product: "HIPAA-eligible CRMs with EHR integration."
Put together: "RevOps leaders at Series A–C US healthtech companies evaluating HIPAA-eligible CRMs with EHR integration." That single line tells the matching engine who, in what mindset, about what, with a constraint generic competitors will not match. Vague hints like "marketing software" lose the auction to specific ones because relevance, not volume, drives ranking.
How should I structure ad groups for ChatGPT ads?
Structure one ad group per audience-and-intent combination, and never mix audiences. "Founders and CMOs and developers" in one ad group dilutes the relevance score, because the hints pull in three directions and match none cleanly.
The math is simple. If you sell to three verticals across three funnel stages, that is three audiences × three intents = nine ad groups, each with its own distinct set of hints.
Within each ad group, use roughly 5–15 hints, written as variants of that one audience-intent-topic theme. That gives the matching system enough surface area to find relevant threads without blurring the target.
Skip the manual work
Abmatic AI runs targets, sequences, ads, meetings, and attribution autonomously. One platform replaces 9 tools.
See the demo →Can you target ChatGPT ads by city, state, or ZIP? Geo limits explained
OpenAI Ads geo targeting is country-level only, there is no city, state, DMA, or ZIP option. Today the advertiser footprint is the United States, Canada, Australia, and New Zealand, expanded on March 26, 2026; the self-serve Ads Manager itself is US-only (PPC Land). PPC Land's documentation review found no state, DMA, or ZIP-level sub-national targeting, contradicting some secondary blogs. Do not plan around ZIP or DMA targeting, it is not in the documentation.
Abmatic AI's OpenAI Ads integration supports advertiser geo targeting in the US, CA, AU, and NZ today, the same countries the API allows. No global promise, no sub-national claims.
Where ads are shown vs where you can buy
Where ads are shown and where advertisers can buy are two different maps, and competitors blur them. On May 7, 2026, OpenAI announced expanding the consumer ads pilot beyond US/CA/AU/NZ to the UK, Japan, South Korea, Brazil, and Mexico "in the coming weeks" (Digiday). That is where ads are shown to users. It is not where advertisers can buy and geo-target. Plan delivery geos against the advertiser footprint, not the consumer show map.
What personalization signals does ChatGPT use, and where is the privacy line?
Matching starts from the topics in the user's current chat thread. If the user opts into personalized ads, OpenAI may also use signals such as past chats, memory, prior ad interactions, and basic context like general location or language depending on settings (WebFX).
Here is the line you cannot cross, and it is a feature. There is no user-level tracking and no demographic or cross-platform behavioral segmentation. Chats, names, email addresses, IP addresses, and precise locations stay inside ChatGPT and never reach the advertiser, who receives only aggregated performance data. Advertisers have no access to chats, history, memories, or personal details (OpenAI Help Center: Ads in ChatGPT). And ads do not influence the answers ChatGPT gives, they run on systems separate from the chat model (OpenAI).
For privacy-first B2B teams, this is the durable model. As third-party cookies and behavioral audiences erode, contextual matching against live intent does not depend on tracking anyone. You are not buying a profile; you are buying relevance to a moment.
How do you build an audience strategy when there is no audience tool?
Because there is no demographic audience data to optimize against, you build your own signal. Tag every ChatGPT-ads conversion with the topical and intent context that generated it, the hint theme, the vertical, the funnel stage. Over a few weeks that tagging becomes your conversion-context framework: a de facto audience-optimization layer assembled from which conversations actually convert, not from who the person is.
Refine hints before you raise spend. OpenAI recommends a starting maximum bid of roughly $3–$5 per click, and because ads are selected primarily on relevance, a well-targeted ad group can win delivery at the lower end of that range while generic hints get priced out (PPC Land). Treat your delivered cost as a relevance gauge: if you are paying near your max bid for thin volume, tighten the hints before adding budget. Bid mechanics are covered in the campaign types playbook.
Note the eligible advertiser categories at the US open-up, household and consumer goods, local services, travel, entertainment, digital products, and education, with more to come as safeguards mature (Digiday). Third-party measurement and CPA bidding are in development with no firm timeline. Plan for performance you can attribute today.
The first-party demand signal
Those conversations already convert to real visits, ChatGPT overtook LinkedIn as a referral source in our 2026 data (the full data case).
Pointing hints where you do not already rank
If buyers already reach you organically through ChatGPT, the highest-leverage ads buy where you are absent. Abmatic AI's AI-optimization engine queries the AI engines directly to see where you surface and where you do not for buyer-relevant topics, cross-references your organic rankings to find phrases where you have demand but sit on page three or are missing from the AI answer, and builds context-hint strategy to fill exactly those gaps. The logic is simple: buy ads precisely where you do not show up organically. See autonomous, AI-managed ChatGPT ads for how that runs end to end.
Abmatic AI manages OpenAI Ads in the same console as your LinkedIn and Google campaigns, fully synced both ways, and pipes results into your CRM. See how context-hint targeting works in Abmatic AI.
FAQ
What is the difference between context hints and keywords?
A keyword is a token Google tries to match against a typed query. A context hint is a natural-language description of a conversation where your product belongs, set at the ad-group level. OpenAI states hints guide matching but do not guarantee delivery and are not exact-match keywords. The matching engine reads the topics in the live chat, not a literal phrase.
Can I target ChatGPT ads by city, state, or ZIP code?
No. OpenAI Ads geo targeting is country-level only, currently the US, Canada, Australia, and New Zealand. PPC Land's documentation review found no state, DMA, or ZIP-level targeting, contradicting some secondary blogs. Plan delivery against the country footprint, and do not build campaigns expecting sub-national geo.
Do advertisers see who they are reaching on ChatGPT?
No. There is no user-level tracking and no demographic audiences. Chats, names, emails, IP addresses, and precise locations stay inside ChatGPT. Advertisers receive only aggregated performance data and have no access to chats, history, or memories. That privacy boundary is a deliberate design choice, not a temporary limitation.
How many context hints should I use per ad group?
Roughly 5 to 15, written as variants of one audience-intent-topic theme. That gives the matching system enough surface area to find relevant conversations without blurring the target. Keep one audience-and-intent combination per ad group, mixing audiences dilutes your relevance score in the auction.
Is ChatGPT ad targeting better for top-funnel or bottom-funnel?
Top and mid-funnel. ChatGPT conversations are research and comparison moments, so context-hint targeting reaches buyers while they are still deciding. Pair it with bottom-funnel search ads that capture already-decided transactional intent. The two are complements, not substitutes.
Does the country where ads are shown match where I can target?
Not necessarily. OpenAI is expanding the consumer pilot where ads are shown (to the UK, Japan, South Korea, Brazil, and Mexico, reported May 2026), but advertiser targeting access is separate and currently limited to the US, CA, AU, and NZ. Always plan against the advertiser footprint, not the consumer show map.
How do I improve my OpenAI Ads targeting relevance score?
Tighten the three parts of every hint, audience, intent, topic, so each ad group describes one specific buyer in one mindset about one constrained category. Align your ad title, copy, and landing page to the same theme, since all four feed the relevance auction. OpenAI recommends a starting maximum bid of about $3–$5 per click; if you are paying near the top of that range for little volume, your hints are likely too broad.
