Context Hints: The Definitive Guide to Targeting in ChatGPT Ads

Tarun Kapoor Updated May 21, 2026 22 min read

Context hints are short descriptions of the conversations, topics, or keywords where your product or service is relevant. They live at the ad group level inside OpenAI's Ads Manager Beta, where they act as the primary targeting signal — replacing the cookie, the device graph, and the literal-match search keyword with a single, conversational instruction the model can reason about. This guide covers what context hints are, how OpenAI's relevance-weighted second-price auction evaluates them, how to write ones that convert, fifty worked examples across ten verticals, the bulk-upload schema, pricing benchmarks of $3-5 CPC and $60 CPM, and the seven mistakes that cause most ad groups to underdeliver.

What context hints are

Inside OpenAI's Ads Manager Beta, the unit of targeting is an ad group. An ad group contains your ads, your bid, your budget pacing — and your context hints. OpenAI's own help-center definition is plain:

"At the ad group level, advertisers can provide context hints that describe the conversations, topics, or keywords where their products or services may be relevant. These hints help guide ad matching, but they are not exact-match keywords and do not guarantee delivery in specific conversations." — OpenAI Help Center, Ads in ChatGPT: The Basics

Three things matter in that definition. First, the hint is a description, not a token. Second, OpenAI explicitly says hints are not exact-match. Third, OpenAI explicitly says delivery is not guaranteed. Each of these is a meaningful break from the way paid-search advertisers have written keyword lists for the last twenty years, and each one changes how you should plan a campaign.

The simplest way to think about a context hint is this: imagine you could whisper a one-paragraph briefing to ChatGPT before every relevant conversation, telling it when your product is worth surfacing. The briefing describes the kind of question being asked, the kind of user asking it, and the kind of decision they are about to make. That briefing is your context hint.

Why context hints replaced keywords

The keyword-as-target model worked because search was a typed query box. The user wrote three words. The advertiser bid on three words. The platform matched the words. That symmetry collapses inside an LLM. A buyer who would have once typed best CRM for healthcare into Google now types a paragraph into ChatGPT:

"We're a Series B healthtech company in Boston, around 90 employees, mostly clinicians and a small RevOps team. We need a CRM that's HIPAA-eligible, integrates with our EHR for patient outreach, and isn't going to take six months to roll out. Salesforce feels heavy, but I don't want a fragile startup tool either. What's the right shortlist for us in 2026?"

That paragraph contains the company stage, the headcount, the geography, the buyer role, the regulatory constraint, the integration requirement, the timeline, the competitive comparison, the budget signal, and the desired outcome. Every one of those is a targeting parameter the old keyword model could not represent. The keyword best CRM for healthcare is in there technically, but if your campaign is built on that string alone, you've thrown away ninety percent of what the user actually told you.

A context hint captures the paragraph instead of the substring. It says something like: founders or RevOps leaders at Series A–C US healthtech companies (50–150 employees) evaluating HIPAA-eligible CRMs with EHR integration who want fast time-to-value and don't want Salesforce. That hint is a fingerprint of the conversation, not a token in it. The auction is then free to evaluate whether your ad belongs in this specific conversation based on how well the conversation fingerprint matches your hint fingerprint.

The five things context hints capture that keywords cannot

• Role and seniority. "VP of Marketing at a B2B SaaS" is a hint. It has never been a viable keyword.
• Company stage and size. "Series B with 50–200 employees" is a hint. The keyword search volume on those exact strings is effectively zero.
• Regulatory or industry constraint. "HIPAA-eligible," "SOC 2," "EU-only data residency" — keywords technically exist, but the user almost never types them as standalone queries.
• Funnel stage. "Evaluating a shortlist," "asking for a comparison," "ready to switch from a competitor" — these are intent states, not strings.
• Competitive context. "Considering switching away from HubSpot to something lighter" is a hint. As a keyword it is too long-tail to bid on profitably.

Where context hints live in Ads Manager Beta

The Ads Manager Beta object model nests three things:

That placement matters strategically. Because hints live at the ad group level, the cleanest way to scale a campaign is one ad group per audience-and-intent combination. If you sell to three different verticals and three different funnel stages, you do not write nine ads — you write one ad per audience and use nine ad groups with nine sets of context hints to point them at the right conversations.

How the relevance-weighted second-price auction scores hints

OpenAI describes the auction as relevance-weighted, second-price. Both halves of that phrase matter.

The second-price half

In a second-price auction the winning bidder pays one cent above the next-highest eligible bid, not their own max. The strategic implication is well-understood from twenty years of Google Search history: you should set your max bid at the true value of a click, because you will almost never pay it. Bidding shy is leaving impressions on the table; bidding aggressively but reasonably is the right move.

The relevance-weighted half

"Relevance-weighted" means the platform multiplies each bidder's bid by a relevance score before ranking them. The relevance score is computed across multiple inputs the help center calls out specifically: context hints, landing page, ad title, and ad copy. An advertiser with a $2.50 bid and a 0.9 relevance score will beat an advertiser with a $4.00 bid and a 0.4 relevance score in the ranked auction, and will likely pay a fraction of their max.

This is the single most important strategic insight in the entire system. You do not win this auction by outspending — you win it by being more relevant. Context hints are the largest lever you have to move the relevance side of the equation, because they are the only input that describes the conversation rather than the offer.

Anatomy of a great context hint

Across the campaigns we have audited at Context Hints, the best-performing hints share a structure. We call it the Audience — Intent — Topic framework. Every great hint names all three layers, in that order.

Layer 1: Audience

Describe the person asking. Role, company size, industry, geography, sometimes seniority. The richer you are, the easier it is for the model to score relevance high. "Founders" is weak. "Solo founders of pre-seed B2B SaaS companies in the US" is strong.

Layer 2: Intent

Describe what they are trying to do right now. "Buying" is too broad — buying what? At what stage? Switching from what? "Evaluating a shortlist of CRMs to switch to after outgrowing HubSpot Starter" is the kind of intent a model can match against. Three useful intent verbs that work across categories: evaluating, switching, comparing.

Layer 3: Topic

The category, sub-category, and any constraints. "CRM" is the category. "Modern CRM with usage-based pricing" is sub-category. "HIPAA-eligible CRM with EHR integration" is sub-category with constraint. The constraint is what disqualifies all the generic CRMs and qualifies yours.

The full template

[Audience: who they are, what they do, where they work]
+ [Intent: what they're trying to accomplish right now]
+ [Topic: the category, sub-category, and constraint that points at your product]

Worked example for a fictional HIPAA-eligible CRM called Cardiff:

And the same hint, weakened by stripping the layers:

The weak hint will match a thousand conversations a day. The strong hint will match thirty. The strong hint will win all thirty because its relevance score is unbeatable; the weak hint will lose most of its thousand because every other CRM advertiser in the auction is also bidding on it.

50 worked examples by vertical

The fifty hints below follow the Audience–Intent–Topic framework. Each is written so you can paste it into an ad group, swap the product reference for your own, and have a usable starting hint within a minute. They are deliberately concrete because the auction rewards specificity.

B2B SaaS (5)

E-commerce and DTC (5)

Fintech (5)

Healthcare and life sciences (5)

Legal and professional services (5)

Education (5)

Local services (5)

Consumer apps (5)

Manufacturing and industrial (5)

Marketing and creative tools (5)

Pricing benchmarks — what $3-5 CPC and $60 CPM actually mean

OpenAI publishes two starting reference points. They recommend a starting CPC max bid of $3-5 USD per click and set the default CPM max bid at $60. These are not floors — they are starting suggestions inside a relevance-weighted auction.

In practice we see three patterns:

For CPM (Reach) campaigns the $60 default is generous. We have seen brand-awareness campaigns clear at $15-25 CPM when the hint is well-scoped to a small but exact audience. Treat the $60 as a ceiling, not a target.

Choosing between CPC and CPM

• Use CPC (Clicks objective) when you have a measurable conversion event and you want to pay per outcome. Most lower-funnel demand-gen campaigns belong here.
• Use CPM (Reach objective) when you are introducing a category or building familiarity ahead of a future evaluation. Most brand campaigns belong here.
• You can run both for the same product. They live in different campaigns with different objectives.

The bulk upload schema

Ads Manager Beta supports bulk creation via CSV upload. The schema mirrors the object model: each row represents an object, and the rows are linked by name references. The high-level shape is:

campaign_name, objective, daily_budget, schedule_start, schedule_end
ad_group_name, parent_campaign, context_hints, bid_strategy, max_bid
ad_name, parent_ad_group, title, copy, landing_page, image_asset, advertiser_name, favicon

A few non-obvious points about working with the CSV:

• Context hints in the CSV are typically pipe- or semicolon-delimited inside a single cell. Verify the exact delimiter in the latest documented schema before importing.
• Ad group names must be unique within a campaign, and ad names unique within an ad group.
• If validation fails on a single row, the platform usually rejects the entire upload — keep your batches small until you trust your template.
• Use the OpenAI-published Bulk Upload Campaign Schema Checklist as your final pre-flight gate. We mirror its structure in the UTM Builder for downstream attribution consistency.

The seven most common mistakes that kill ad-group performance

1. One generic hint per ad group. "Best CRM software." The auction treats you as a commodity bidder. Rewrite using the framework.
2. Hint and ad copy describe different things. The relevance score reads both. If the hint says "for healthcare startups" and the ad says "for everyone," the score is dragged down. Mirror your hint inside the ad copy.
3. Landing page mismatch. The landing page is one of the four named relevance inputs. Linking the hint "modern CRM" to a generic homepage instead of a vertical landing page is a relevance penalty.
4. No competitive context in the hint. "Switching from HubSpot" and "evaluating new CRMs" are different conversations. Include the competitor if your offer is a credible alternative.
5. Hint vocabulary the buyer never uses. If your hint contains your internal product category name and the buyer uses an industry term, you will miss the match. Write hints in the buyer's vocabulary, not yours.
6. Too few hints per ad group. Three hints describing slightly different phrasings of the same audience-intent-topic combination outperform one. Variants compound.
7. Mixing audiences in one ad group. "Founders and CMOs and developers" is three ad groups, not one. The mixed group loses to each focused competitor.

Context hints vs Google search keywords

For paid-search marketers crossing over, the mental model that helps most is this: treat each context hint like a SKAG (Single Keyword Ad Group) — except the keyword is now a paragraph. The discipline of writing one ad group per intent is the same. What changes is that the intent is now a fully-described state rather than a typed string.

Where context hints are going

Three plausible directions in the next twelve months, based on patterns in adjacent platforms:

• Hint quality scoring shown in-product. Expect Ads Manager Beta to surface a relevance estimate per hint, similar to Google's Quality Score, so operators can iterate without launching.
• Audience extension across LLM surfaces. The same hint structure is highly likely to be portable to Perplexity, Gemini and other AI ad inventories, with each platform tuning its own relevance model on top of a common semantic input.
• Negative hints. The ability to exclude a conversation type ("not for enterprise buyers," "not for free-trial-only intent") is the most-requested feature among advertisers we talk to. It's also the most likely to ship next.

Frequently asked questions

Tarun Kapoor

Founder & CEO, Context Hints

Twelve years of media buying across GroupM, WPP, Ogilvy & Mather, and Neil Patel Digital — where he was directly responsible for $3.5M+ added to the bottom line of a $6.5M ARR agency. Has personally owned media for Nestlé, Sage, Qualcomm, Aetna, Weight Watchers, Chubb and Novotel. Now consults Fortune 500s and venture-backed startups as a fractional CMO and operative media buyer.