ChatGPT served ads on 28.69% of healthcare prompts and 15.21% of finance prompts in a 50,006-prompt study run on July 23, 2026. Same policy update, nearly double the inventory in one vertical over the other. That gap is the actual story, and the headlines saying “OpenAI opens ads to health and finance” missed it.
Here is what happened. OpenAI published its ad policies in March 2026 and has revised them four times since. The current page was last updated on August 10, 2026. The health and finance expansion did not land in that August edit. It landed in July, when OpenAI added a dedicated advertiser policies section and spelled out which financial and health categories are eligible, and where.
Four revisions in five months
| Version | Timing | What changed |
|---|---|---|
| v1.0 | March 2026 | Initial publication of the OpenAI Ads Policies |
| v1.1 | April 2026 | Medical, legal and financial advice contexts were no longer categorically blocked from ads by default |
| v1.2 | May 2026 | Added a section on what standards apply, how they’re implemented, and what happens when ads fail the safety bar |
| v1.3 | July 2026 | Added advertiser policies; clarified eligible financial and health categories and markets |
| v1.4 | August 2026 | Clarified the policy stance on housing and job listings |
Four rewrites in five months is not a rounding error. It is the operating condition.
Who can buy now, and who still can’t
Financial services, US only. Approved advertisers can run auto loans and leasing, credit cards, credit monitoring, deposit accounts, financial planning, insurance, investment services and brokerages, mortgages, personal loans, and payment services.
Still prohibited: credit repair, debt settlement, debt assistance, and alternative investments such as bullion. Advertisers may be asked for proof of licensure. Financial services advertising outside the US is generally prohibited.
Healthcare and medicine, US only. Eligible categories include consumer medical devices, wearables and apps; dental services; dietary supplements; disease awareness campaigns that don’t advocate for public policy; health insurance; hospitals and urgent care; medical testing and diagnostic products and services; minimally invasive cosmetic procedures; and vision products and services.
Still prohibited: unsafe or unapproved medical products, unsupported treatment claims, experimental or investigational therapies, invasive or high-risk procedures, and adult-oriented sexual health services. Same licensure requirement. Same non-US ban.
Legal services are still out. This is the part the April changelog entry quietly obscures. v1.1 said medical, legal and financial advice contexts were no longer categorically blocked. That’s about where ads can appear, not who can buy them. Ads for legal advice, court-related work and similar services remain prohibited. General legal education or media may be allowed where no legal services are offered.
If you run marketing for a legaltech company and read the April note as an opening, it wasn’t one.
Approval to advertise is not permission to appear
OpenAI separates three gates: advertiser eligibility, ad content, and placement. Pass one, you can still fail another.
Most review runs on automated systems with human oversight. Restricted categories get additional safeguards, either enhanced advertiser verification or manual review, and some advertisers have to produce licensure documentation.
Then placement narrows it again. Ads don’t run near emotionally reliant contexts, mental or personal health conversations, or sensitive user journeys, and they’re kept out of brand-unsafe contexts including child safety, cyber abuse, graphic violence, hate and harassment, misinformation, terrorism and weapons. Ads also aren’t shown to users under 18.
So a health insurer can be an approved advertiser and still never appear beside a personal health conversation. That is the design, not a bug in the rollout.
What the inventory data actually shows
SE Ranking analyzed 50,006 commercially oriented prompts across 20 niches in the US, collected on July 23, 2026. Ads appeared on 25.94% of them.
Healthcare came in at 28.69%. Finance at 15.21%.
Two more numbers from that study are worth more than the headline rate. About 14.35% of ads had no meaningful topical connection to the query they ran against. In categories like Relationships and News & Politics, mismatch topped 50%. And in 96.37% of placements, the advertiser was not cited in the answer itself.
Read together: the inventory is real, relevance is uneven, and appearing as an ad is not the same as being the source ChatGPT recommends. For anyone treating ChatGPT Ads as an AI-visibility play, those are two different budgets solving two different problems. Our breakdown of Generative Engine Optimisation covers the organic side of that split.
One caveat the original coverage skipped: this sample was built from commercial prompts in the US, where the ad test has been live longest. It is not the ad rate across all ChatGPT usage, and nobody should quote it as one.
The measurement problem hasn’t been solved
OpenAI began rolling out conversion-optimized campaigns in early June 2026, adding CPA bidding on top of the earlier CPM and CPC models. To use it you need a tracking pixel or the Conversion API feeding events back.
That makes the conversion event you select a delivery input, not just a reporting line. Pick “form fill” and the system will find you form fills. For B2B SaaS, a form fill is not pipeline.
Meanwhile the context asymmetry runs the wrong way for advertisers. Ad selection can consider the intent of the current conversation, the landing page, ad title and copy, advertiser-provided context hints, and, when personalization is on, broader signals from the user’s ChatGPT experience. Advertisers get impressions, clicks, spend, CTR, average CPC, average CPM and conversions. No conversation-level visibility.
The platform knows more about the query than you do. If your server-side event pipeline is thin, you are asking a black box to optimize against a signal you barely feed it. Worth reading alongside how to track AI and LLM chatbot traffic in GA4 before you switch a campaign to CPA.
The housing and jobs clarification matters more than it sounds
The August v1.4 edit reads like housekeeping. It isn’t, if you sell HR tech or proptech.
Ads for individual job listings or housing rentals and sales are prohibited. But advertisers may link to platforms that carry those listings, as long as neither the creative nor the landing page references a specific listing.
That’s a creative and landing-page constraint, not a category ban. A jobs marketplace can advertise the marketplace. It cannot advertise the job. Anyone running dynamic listing-level creative on other channels will need a separate asset set here.
What to do about a rulebook that moves this fast
A normal media plan assumes platform rules will still be recognizable on launch day. Four revisions in five months makes that assumption hard to defend. A category rule can change while a campaign is still in build.
Three practical moves:
- Move the policy check from planning into launch. Re-verify eligibility before creative approval, before any new geography goes live, and before you scale spend. Not once at kickoff.
- Don’t copy a US test into another market. Both financial services and health services advertising remain generally prohibited outside the US, even though the ad test itself now runs in more countries. OpenAI confirmed on August 11 that the UK, Mexico, Brazil, Japan and South Korea had launched, adding to the US, Canada, Australia and New Zealand. Market availability and category eligibility are not the same list.
- Treat the landing page as ad copy. Review covers creative and landing pages both. An ad can be rejected because the destination introduces a prohibited category or a claim you can’t support. Good reason to audit what your landing pages actually promise before submission, not after rejection.
The OneMetrik Takeaway
Health and finance are not “open” on ChatGPT. What changed is that OpenAI now has a process for admitting some regulated advertisers, one at a time, while keeping sensitive contexts off limits entirely.
That’s a real test window for a US healthtech or fintech company. It is not a channel you can plan two quarters out.
We’d run it as a controlled test with a compliance checkpoint attached: verify eligibility at launch, document every claim you got approved, feed server-side conversion events, and track qualified pipeline rather than leads. Then assume the rulebook changes again before the quarter closes, because on current form, it will.
If you’re weighing whether ChatGPT Ads deserves budget next quarter, we’re happy to look at your setup.