ChatGPT Ads in 2026: How It Works and Where It's Going

This is a field guide to what ChatGPT Ads is today, how the buying mechanics work, what advertisers can and can't control, what the measurement stack actually looks like, and where the platform stands relative to the discovery channels planners already know. For the strategic framing of what this shift means for agencies, see our companion piece, ChatGPT Ads Are Live: The AI Answer Is the New Ad Unit.

A snapshot of where things stand

What ChatGPT Ads actually is

When eligible users on the Free or Go tiers ask ChatGPT a question (planning a dinner party, comparing software, researching a trip), they may now see a sponsored card below the answer. The card shows an advertiser name, favicon, headline, short description, image, and a link to a destination page. It is clearly labelled as sponsored, and OpenAI's policy is explicit that ads cannot influence the assistant's answer.

Inventory is narrower than headline ChatGPT user numbers suggest. Ads do not appear for paid Plus, Pro, Business, Enterprise, or Education subscribers, and they are excluded from accounts identified as under 18. OpenAI also offers eligible users an "Ads-Free" option on the Free tier in exchange for lower usage limits, which trims available impressions further. A company spokesperson told reporters earlier this year that fewer than 20% of eligible users were being shown ads daily during the pilot, which is useful context when sizing the channel.

The single ad format documented today is called a chat_card. It carries a 3–50 character title, up to 100 characters of body copy, an image, a favicon, and a target URL. There is no publicly documented self-serve video, carousel, or interactive format yet, although OpenAI has signalled that more formats will follow.

How buying works

Two campaign objectives are available: a CPM-priced awareness objective (called either "Views" or "Reach" depending on which OpenAI page you read), and a CPC-priced "Clicks" objective. Both run through a relevance-weighted, second-price auction, which means bids are only one input; ad and landing-page relevance to the user's conversation also factor into ranking. Budgets, dates, and country targeting are set at the campaign level; bids and context hints are set at the ad-group level.

Public guidance on pricing is high-level. The table below summarises what is currently published and what external reporting suggests about real-world cost.

OpenAI has not published average CPC, CPM, CTR, CPA, or ROAS benchmarks by vertical, so any media plan today should treat efficiency assumptions as provisional. One welcome change: trade reporting indicates the previous US$50,000 pilot spend threshold has been removed for the self-serve beta, making meaningful tests accessible to mid-market and SMB advertisers, not just enterprise pilots.

Targeting and creative

This is the area most likely to surprise planners coming from Google Ads or Meta. The most important distinction to keep straight is between signals that shape delivery internally at OpenAI and controls that advertisers can explicitly configure.

Context hints are explicitly described as broad thematic guidance, not exact-match keywords, and OpenAI recommends writing them as descriptions of the questions, needs, or situations users bring to ChatGPT. Advertisers also do not receive any user data: chats, names, emails, IP addresses, and precise locations stay inside ChatGPT.

The creative implication is straightforward: this channel rewards specific, intent-matched offers far more than generic brand messaging. OpenAI's own guidance pushes advertisers to build many distinct creative variations, write benefit-led titles and descriptions, and link to the most relevant page rather than a homepage. This is squarely conversion optimization territory.

The campaign workflow

The end-to-end setup process for a new advertiser, drawn from OpenAI's onboarding and quickstart material:

1. Create advertiser account
2. Complete verification
3. Set account info and favicon
4. Add billing profile and payment method
5. Invite team members and generate API keys if needed
6. Create campaign
7. Choose objective, dates, budget, countries
8. Create ad groups
9. Add context hints and bids
10. Upload creative asset
11. Create ads with title, body, image, target URL
12. Submit for review
13. Serve if approved
14. Monitor in tables, charts, CSV, or Insights API
15. Optimise bids, context hints, ads, and landing pages

Operationally, the major friction points today are verification, policy review, and beta feature gaps. Ads Manager supports guided creation, bulk upload, performance monitoring, in-line and bulk edits, member permissions, billing, API keys, and change logs; public docs stop short of detailing review SLAs or the full bulk-upload schema.

Measurement and the pixel

For a beta platform, the measurement stack is more developed than the buying controls. Ads Manager exposes impressions, clicks, spend, CTR, average CPC, average CPM, and conversions, with table views, charts, and CSV export. An Insights API provides programmatic access to performance data at the ad account, campaign, ad group, and ad level. The capability matrix below summarises what is and isn't there today.

For conversion tracking, OpenAI offers two complementary mechanisms.

The JavaScript pixel (distributed as a small SDK called OAIQ, currently version 0.1.3) handles browser-side event collection. When a user clicks a ChatGPT ad and lands on the advertiser's site, the SDK sets a first-party cookie called __oppref on the merchant's domain with a 30-day (720-hour) lifetime, then posts subsequent event data to OpenAI's measurement endpoint. Because the cookie is first-party (set on your domain, not OpenAI's), it isn't subject to the same browser restrictions that have hollowed out third-party cookie attribution over the last few years. Independent reverse-engineering analyses describe the attribution payload as built around encrypted tokens (oppref, olref, an ad_data_token, and a server-side spam-integrity payload), but planners can treat that as plumbing rather than something they need to operate.

A practical quirk: OpenAI generates the pixel for the advertiser based on what they need to track, rather than letting advertisers freely write their own measurement code. This keeps event taxonomy consistent across the platform but also means changes to your tracking go through OpenAI's tooling, not yours.

The Conversions API is the server-to-server alternative. OpenAI explicitly describes it as more reliable than the pixel alone, and it supports deduplication when you reuse a common event ID across browser and server events. Out of the box, it accepts events from web, mobile_app, offline, physical_store, phone_call, and email action sources, meaning CRM and offline-conversion use cases are technically possible, though OpenAI has not yet published native first-party connectors for major CRM, MMP, or CDP platforms. The pragmatic posture for serious testers is to run pixel and Conversions API together, with a shared event ID on key conversions, and let deduplication handle the overlap. If your team isn't confident the underlying tracking is clean before the test starts, see Why Your Conversion Tracking Probably Breaks Once a Quarter.

What is not yet publicly documented: attribution windows, view-through methodology, lift studies, incrementality tooling, and most of the advanced attribution language Google and Meta planners are used to citing in business cases.

Crawlers and landing pages

OpenAI's documented expectation is that ChatGPT ads landing pages must be valid and must not block two specific user-agents: OAI-AdsBot and OAI-SearchBot. This is worth understanding because aggressive bot management at the edge (Cloudflare, Akamai, AWS WAF) can quietly break ad approvals before any human at your agency notices.

OpenAI now publicly documents four distinct crawlers, each with a separate purpose:

• OAI-AdsBot is the validation crawler. It only visits pages submitted as ChatGPT ads, checks them against OpenAI's ad policies, and uses the page content to assess relevance for ad serving. Its user-agent string is Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; OAI-AdsBot/1.0; +https://openai.com/adsbot. OpenAI states data collected by OAI-AdsBot is not used to train its generative models. Unlike the other three crawlers, OpenAI has not yet published a JSON IP range list at openai.com/adsbot.json, which means edge-level filtering currently has to rely on user-agent matching rather than IP verification.
• OAI-SearchBot powers ChatGPT's search and citation features. Allowing this bot is what lets your pages surface inside ChatGPT's organic answers. Blocking it removes you from those results.
• GPTBot is the training crawler. This is the one to block if you do not want your content used in training future OpenAI models; blocking it has no effect on whether your ads run or whether you appear in ChatGPT search.
• ChatGPT-User fetches pages on demand when a logged-in ChatGPT user asks the assistant to look at a specific URL. Because visits are user-initiated, robots.txt may not always apply.

The practical takeaway is that these are independent tokens. Blocking one does not block the others, and allowing one does not allow the others. A baseline robots.txt that lets ads work while keeping training opt-out intact looks like:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: OAI-AdsBot
Allow: /

User-agent: ChatGPT-User
Allow: /

If you front your site with a CDN or WAF, also confirm that OAI-AdsBot is not being challenged or rate-limited by managed bot rules. Until OpenAI publishes an IP range file, the safest approach is to allow the user-agent string explicitly in your bot-management ruleset and monitor server logs for visits.

Eligibility, geography, and policy

Two eligibility lenses matter here, and they don't fully agree.

On the user side, OpenAI's Help Centre states that eligible users may currently see ads in the United States, Canada, Australia, and New Zealand.

On the advertiser side, OpenAI's campaign setup guidance still describes advertiser delivery as US-only, even though API examples include country arrays like ["US","CA"]. The safe planning assumption is that non-US advertiser access and cross-market delivery are pilot-controlled until OpenAI confirms otherwise inside your account. Canadian advertisers, in particular, should expect a partner-mediated path or a wait, even though Canadian users are seeing ads.

Policy is the most underrated planning constraint. At launch, allowed categories are concentrated in lifestyle and household goods, local services, travel and experiences, and digital products or education. Disallowed launch categories include dating and sexual content, health claims, alcohol and drugs, healthcare, financial services, legal services, gambling, and political content. Baseline rules also prohibit misleading claims, offensive language, discrimination, and creative that imitates the ChatGPT interface.

OpenAI has signalled that some categories (including medical, legal, and financial advice contexts) are moving from blanket exclusion toward more selective contextual rules, but the operational thresholds aren't public. For regulated verticals, expect approvals to be conservative and policy to keep evolving.

OpenAI also bars ads from sensitive conversation contexts entirely, including child safety, self-harm, hate, weapons, terrorism, mental health, emotionally reliant interactions, and politics, which puts a real ceiling on inventory in a way feed-based platforms don't have.

Early performance signals

Hard, comparable performance benchmarks from advertisers are not yet public, but a few directional data points have emerged.

OpenAI says trust metrics in the user experience have not been negatively affected by ads and that ad dismissals are low. Reuters has reported that the US pilot crossed roughly US$100 million in annualised revenue within six weeks and that more than 600 advertisers participated. Criteo, a launch tech partner, has said that across a sample of 500 retailers it observed, LLM-referred traffic converted at roughly 1.5x the rate of other referral channels, consistent with the platform's intent-rich positioning. Individual brand pilots, including VistaPrint's, have stayed under wraps.

These are encouraging signals, but they are channel-level and partner-level, not standardised campaign benchmarks. A planner should treat them as evidence that the channel is working for someone, not as a forecast of what their own performance will look like.

Where it fits, by goal

The table below maps each common media goal to the platform's current capability and the kind of test scope that is realistic given today's constraints. Suggested budget ranges are estimates of what is needed to generate statistically useful early signal at current bid guidance; they are not OpenAI-published recommendations.

The pattern: ChatGPT Ads is best when the user is already in a research, comparison, or decision mindset and your offer can be expressed clearly in one short card pointing to a tightly relevant page. It is weakest when scale, granular audience control, or mature attribution are the primary requirements.

Where the gaps are

Several capabilities planners reasonably expect from a mature ad platform are not yet there: more than one self-serve creative format; demographic, audience, customer-list, lookalike, and retargeting controls; documented attribution windows and view-through methodology; native CRM, MMP, and CDP connectors; published vertical benchmarks; and clearer geographic eligibility for non-US advertisers.

OpenAI has indicated that more formats, objectives, and buying models are on the roadmap, and the partner ecosystem (agency holdcos Dentsu, Omnicom, Publicis, and WPP, plus tech partners Adobe, Criteo, Kargo, Pacvue, and StackAdapt) is filling some of the operational gaps in the meantime.

The bottom line for planners in May 2026

ChatGPT Ads in its current form is best understood as an assistant-adjacent contextual channel that sits closer to search-intent media than to social or programmatic display. It rewards specific, well-targeted offers in categories where users actively research and compare. It is not yet a substitute for platforms that depend on rich audience controls, mature attribution, or large creative formats.

The practical posture is straightforward: treat ChatGPT Ads as an experimental line in the media plan rather than a core channel, scope budgets and KPIs accordingly, wire up the OAIQ pixel and Conversions API from day one if you do test, confirm your robots.txt and edge bot-management rules allow OAI-AdsBot and OAI-SearchBot, and watch for OpenAI's next round of updates, particularly around new ad formats, expanded geography, and more granular targeting.

The platform is still being built in public. The honest read in May 2026 is that the foundation is now real enough to plan around, but the ceiling and the rules are both still in motion.