OpenAI Ads intent targeting, evaluated: how conversation-stage targeting actually performs

Tarun Kapoor Updated 21 August 2026 12 min read

OpenAI Ads has no keywords, no interest segments, and no lookalikes. It has one targeting primitive: the system's read of the conversation's context and intent, scored against four inputs you control. That is either the most precise intent signal advertising has ever had, or an unaccountable black box — and after months inside the Ads Manager Beta, our honest evaluation is: both, depending entirely on how you structure for it. This guide explains how the targeting actually works, benchmarks it against keyword and audience models, names its failure modes plainly, and shows the ad-group architecture that turns conversational intent into buyable funnel stages.

The short answer

Intent targeting in OpenAI Ads works like this: at the ad-group level you write context hints — short natural-language descriptions of the conversations where your product belongs. When a free-tier user's conversation is eligible for an ad, a relevance-weighted, second-price auction scores every candidate ad's four inputs (hints, landing page, title, copy) against the conversation's context and intent, and the best-scoring ad — not simply the highest bid — renders below the answer. Evaluated as a targeting product: it captures intent that keywords structurally cannot, it punishes lazy setup harder than any platform we have bought on, and its biggest genuine weakness is transparency, not precision.

How marketers target ads on ChatGPT: the complete list of levers

Marketers target ChatGPT ads with context hints, not keywords: plain-language descriptions of the conversations where the ad belongs, written per ad group and scored against the live chat. Every other lever narrows where or to whom an ad can serve; only the hints and creative decide which conversations. The full set available in Ads Manager:

What marketers coming from search and social cannot use: keywords, match types, negative keywords, interest or demographic targeting, and lookalikes. Ads only appear to logged-in Free and Go users; Plus, Pro and Business users never see them. For the translation from a Google account, see context hints vs keywords.

The mechanics: what “intent” means here

In search advertising, intent is inferred from a query — a compressed, keyword-shaped artifact of what someone wants. In a conversation, intent is stated, elaborated across turns, and qualified: “I run a ten-person agency, we've outgrown spreadsheets, I looked at two CRMs and both felt heavy, budget is about $40 a seat.” That single exchange contains category intent, stage (actively comparing), constraints (team size, budget), and even objection data (“felt heavy”). OpenAI's ad system selects primarily on this conversational relevance — its documentation is explicit that selection considers the context and intent of the conversation against context hints, landing page, ad title, and ad copy.

Three properties of the mechanic are worth internalizing before structuring a single campaign:

• Hints are descriptions, not triggers. OpenAI states plainly that context hints are not exact-match keywords and do not guarantee delivery in specific conversations. You are training a matcher, not setting a rule. Writing hints as keyword lists — the most common migration error from Google — feeds the matcher noise.
• All four inputs score together. A precise hint attached to a generic landing page is a contradiction the auction resolves against you. Coherence across hint → title → copy → landing page is itself the targeting. (Full auction mechanics in how ChatGPT ads work.)
• Relevance is priced in. The auction is second-price and relevance-weighted, so a better-matching ad clears at a lower effective price. On this platform, targeting quality is not just a delivery variable — it is your discount.

Against keywords and audiences: an honest comparison

The pattern in that table: OpenAI's model trades transparency for expressiveness. You can say things to this targeting system that Google's cannot hear (“small agency that just churned from a heavyweight tool”), but you cannot audit what it heard. The whole discipline of operating it well follows from that trade. A deeper structural comparison is in context hints vs keywords.

Where intent targeting outperforms

• Mid-funnel evaluation conversations. This is the sweet spot. A user comparing options has stated constraints the auction can match with unusual precision, and an ad that completes the comparison reads as help. Our buyer-evaluation guide covers the organic side of the same moment.
• Constraint-heavy niches. The more qualifiers define your buyer (team size, compliance need, region, stack), the more the conversational signal outruns keywords — those qualifiers are unsearchable but frequently spoken.
• Category creation. When nobody searches your category yet, keyword targeting has nothing to bid on. Conversations about the problem exist long before queries about the solution do, and hints can describe problems.
• Objection-stage capture. Hints like “frustrated with the cost of their current tool” target a state of mind no other self-serve platform sells access to.

Where it fails, and what each failure costs

• No conversation-level transparency. There is no equivalent of the search-terms report. You will never see the conversations you matched — only aggregate impressions, clicks, spend, CTR, average CPC/CPM, and conversions. Mitigation: tight single-intent ad groups, so the ad group itself becomes your query report.
• Broad hints drain silently. A hint like “people interested in productivity” matches an ocean of adjacent-but-wrong conversations, and nothing in the UI flags it. Mitigation: every hint should name a situation, not a topic. If a hint could describe a million conversations, it describes none.
• No true negatives. You cannot exclude conversations. The workaround is repulsion by specificity: qualifiers in creative (“for teams under 25”) that make wrong-fit users self-filter before the click — which CPC billing turns into real savings.
• Reach is structurally capped. Free-tier, 18+, ad-eligible conversations only, with brand-safety carve-outs for sensitive contexts. This is a precision instrument, not a reach platform; budget expectations should follow. (Costs and inventory dynamics in what ChatGPT ads cost.)
• The score is unappealable. When delivery underwhelms, there is no diagnostic beyond your own inputs. Ads Manager's bid-strength guidance tells you whether the bid is competitive; nothing tells you whether the hint is. Mitigation: change one input at a time, weekly, and read CPC as your relevance thermometer — on a relevance-weighted second-price auction, a falling effective CPC at stable bid is the score improving.

Structuring ad groups by funnel stage

Because hints are free text, stage targeting is available to anyone willing to write it. The architecture we deploy for clients — one intent per ad group, stage-split — looks like this for a CRM example:

Rules that make this structure hold: 5–10 hints per group, each describing the same intent from a different phrasing angle (the coverage principle from writing context hints); never mix stages in one group, because mixed stages force one creative to speak to two moments and the relevance score averages down; and give every group its own landing page, since the page is a scored input, not a destination detail. Creative-side craft is in the creative guide.

Tuning a relevance score you can't see

A weekly operating loop that treats observable metrics as proxies for the hidden score:

• CPC trend per ad group (at fixed max bid) is your relevance gauge — second-price mechanics mean rising relevance shows up as falling clearing price before it shows up anywhere else.
• CTR tells you whether matched conversations found the creative useful; high impressions with low CTR usually means the hint matched a wider situation than the creative addresses. Sharpen one or the other, not both at once.
• Conversion rate validates the whole chain; a group with healthy CTR and poor CVR almost always has a landing-page mismatch — the page isn't answering the question the matched conversation was asking.
• UTM-tagged sessions (parameters persist through ad clicks) let you read post-click behavior per ad group in your own analytics, which is the closest available substitute for a matched-conversation report.

The evaluation verdict

Scored as a targeting product: precision 9/10, control 6/10, transparency 4/10 — and the composite depends on your discipline more than the platform's. For advertisers with sharply defined buyers and the patience to structure single-intent ad groups, OpenAI's intent targeting reaches moments of stated, qualified intent that no keyword or audience system has ever sold. For advertisers who port a Google account shape into it — broad themes, one landing page, creative reused from search — it will spend quietly and underdeliver, and the lack of query-level reporting will hide exactly why. The platform rewards the craft it demands. Buy on CPC, structure by stage, change one variable a week, and treat your clearing price as the report OpenAI didn't ship.

Intent extraction is the mechanic; turning it into pipeline is the work. For B2B teams specifically, our ChatGPT Ads service for B2B SaaS covers ICP problem-language mapping, conversion-event selection, and CRM attribution through the Conversions API.

Frequently asked questions

Sources and further reading

• OpenAI Help Center — Ads in ChatGPT: the basics (selection inputs, auction, pricing, measurement).
• OpenAI Help Center — Create ads for ChatGPT (ad-group setup and context hints).
• Context Hints — context hints vs keywords, writing context hints that convert, how ChatGPT ads work, and what ChatGPT ads cost.

Tarun Kapoor

Founder & CEO, Context Hints

Twelve years of media buying across GroupM, WPP, Ogilvy & Mather, and Neil Patel Digital. Has personally owned media for Nestlé, Sage, Qualcomm, Aetna, Weight Watchers, Chubb and Novotel.