OpenAI Publishes Ad Credit Policies Six Months Into Ads Business

OpenAI has been running ads inside ChatGPT for six months. This week it published a credit policy. The paperwork arrived after the money, not before.

The policy itself is thin. Ninety-day expiry, no stacking with other discounts, and a clause letting OpenAI "suspend, revoke or void" credits against fees it never defines. Sensor Tower reports that ads per user per hour more than doubled between April and late July. Inventory is growing. The rules governing that inventory are still being written.

Why big buyers stay out

No Fortune 500 procurement team signs a media agreement where the counterparty can void your credits at its discretion and won't say which fees they apply to. That clause fails legal review at any company large enough to have one. Add no third-party measurement hook, no data residency terms, no audit logs, and no attribution model, and you have inventory a scrappy performance shop might test with house money but that a brand safety officer cannot approve.

The undefined "eligible fees" language is the tell. It suggests OpenAI's internal commercial policy is still being built, which means the disclosure and native-ad-labeling guardrails probably are too. That is regulatory exposure inside a product surface that has no obvious place to put a "sponsored" tag. When ChatGPT recommends a product, users cannot tell whether that is the model's judgment or someone's ad buy. The FTC has not weighed in on persuasion woven into a trusted conversational voice, and OpenAI has not gotten ahead of that question.

The growth stat is weaker than it looks

Ads per user per hour doubling sounds impressive until you notice that the base was near zero and OpenAI controls the denominator. It also measures supply, not whether the ads work. There is no public data on click-through, recall, or lift, and no outside verifier has been let in to check. The Sensor Tower figure tells you how often OpenAI is showing something. It tells you nothing about whether those ads produce results for buyers.

On the cost side: serving a contextually relevant ad inside a generative conversation requires retrieval, ranking, and probably a relevance model running alongside the base model on every ad-bearing turn. Doubling ads per user per hour doubles that compute, stacked on top of generation cost that is already high. The 90-day credit expiry looks less like an advertiser rule and more like cash-flow management on a business where the unit economics are worse than the growth chart implies.

What has to change

Two things are missing and both matter. First, a third-party measurement path so buyers can verify performance instead of trusting a platform-reported headline. Second, a rewrite of the discretionary-revocation and undefined-fee language into something procurement can actually sign. Until both exist, this is a trial balloon, not an ad business.

If your team is evaluating ChatGPT inventory now: run one 90-day credit cycle with your own conversion tracking instrumented on the landing side, since OpenAI will not give you funnel data, and get legal to demand a written definition of "eligible fees." If OpenAI won't define it, that is your answer.

The call: By OpenAI's next DevDay (expected fall 2026), ChatGPT ads will still lack independent third-party measurement or attribution that plugs into standard buyer tooling. Six months in with no measurement signal and policy still unfinished, the direction is clear. Platforms that control their own numbers rarely open them to outside verifiers until buyers force the issue with withheld budget, and no such pressure is visible yet.

OpenAI has run an ads business inside ChatGPT for six months, and this week it finally wrote down the rules for how brands spend ad credits. Ninety-day expiry, no stacking with other discounts, and a clause letting OpenAI void the credits whenever it wants against fees it never bothers to define. Sensor Tower says ads per user per hour more than doubled between April and late July. So: inventory is growing, and the paperwork is arriving after the money.

What's actually being decided here is whether a brand or agency should treat ChatGPT as real ad inventory yet, or keep it in the experimental-budget column. Type 2 decision, easily reversed. Nobody is rebuilding their media stack around this. The forcing function is soft: no deprecation, no rate card cliff, just a growth number and a legal document that reads like it was written by the finance team, not the sales team.

The Skeptic. Ads per user per hour doubling means nothing when the base was roughly zero and OpenAI owns the denominator. Show me a second data point that isn't self-reported. For a PM who's heard of ChatGPT but not the ad business: the impressive-sounding growth stat is a company grading its own homework. There's no third-party verification, no attribution model, no defined fee schedule, and now a credit policy whose main feature is OpenAI's right to take the credits back for reasons it won't state. That verbatim clause, "suspend, revoke or void" against undefined "eligible fees," is the sound of lawyers protecting a platform, not enabling a buyer. Every "AI changes advertising" cycle dies on measurement. This one hasn't reached the measurement question yet.

The Enterprise Buyer. No CMO's procurement team signs a media agreement where the counterparty can void your credits at its discretion and won't say which fees the credits even apply to. That clause fails legal review at any company big enough to have one. Plain version for the PM: the contract lets OpenAI change the deal after you've committed budget, and legal departments exist to kill exactly that. Add no data residency terms, no audit logs, no third-party measurement hook, and you have inventory that a scrappy performance shop might test with house money but that a Fortune 500 brand safety officer can't touch. The credits are a trial lure. The absence of indemnification and eligible-fee definition tells you OpenAI's own commercial policy isn't finished.

The Compute Pragmatist. Serving a contextually relevant ad inside a generative conversation is not free. You need retrieval, ranking, and probably a relevance model running alongside the base model on every ad-bearing turn. So doubling ads per user per hour also doubles ad-serving compute per user per hour, stacked on top of generation cost that already isn't cheap. For the PM: every ad ChatGPT shows costs OpenAI real GPU time to decide what to show, unlike a banner that's basically free to serve. If OpenAI is eating that serving cost to build inventory, the unit economics are worse than the growth chart implies. The 90-day credit expiry looks less like an advertiser rule and more like cash-flow timing.

The Safety Lens. Persuasion inside a conversation is a different animal than persuasion on a banner. ChatGPT's whole value is that it sounds like a trusted advisor, and an ad woven into that voice blurs the line between recommendation and paid placement in a way the FTC has not blessed. For the PM: when the chatbot suggests a product, can the user tell whether that's the model's judgment or someone's ad buy? OpenAI's safety review was built for helpfulness and harm, not commercial intent. The undefined fee language suggests the internal policy architecture is incomplete, which means the disclosure and native-ad-labeling guardrails probably are too. That's regulatory exposure sitting inside a product surface that has no obvious place to put a "sponsored" tag.

The Researcher. The Sensor Tower figure is a proxy for inventory creation, not yield. Ads per user per hour tells you how often OpenAI is showing something, not whether anyone clicked, recalled the brand, or bought. For the PM: it's a measure of supply, not of whether the ads work. There is zero public disclosure on click-through, attention, or lift, and no third-party measurement hookup. Until OpenAI opens a signal that plugs into marketing mix models or lets an outside verifier in, this is unauditable growth. The research and buyer communities cannot validate what they cannot observe, and OpenAI has given them nothing to observe.

Where the council splits. The Compute Pragmatist thinks OpenAI is subsidizing ad-serving to manufacture the growth curve; the Skeptic thinks the growth curve barely exists because the base was near zero. Both can be right, and both point the same direction: the doubling stat is not evidence of a real business. The second fault line is the Enterprise Buyer against everyone assuming this scales. The Safety and Researcher lenses describe a product that isn't measurable or disclosable, and the Buyer says that's precisely why the big money stays out. The bull case needs enterprise brands. The document published this week gives enterprise legal three reasons to say no.

What it hinges on. Two facts. First, does OpenAI open a third-party measurement path, so a buyer can verify performance instead of trusting a Sensor Tower headline. Second, does it rewrite the discretionary-revocation and undefined-fee language into something a procurement team can sign. Both are missing today, and the council leans hard toward "this is a trial balloon with legal cover, not an ad business yet." Before any team commits above house-money levels, run a real campaign for one 90-day credit cycle, instrument your own conversion tracking on the landing side since OpenAI won't give you funnel data, and get legal to demand a definition of "eligible fees" in writing. If OpenAI won't define it, that answers the question.

Prediction: By OpenAI's next DevDay (expected fall 2026), OpenAI will still not offer independent third-party measurement or attribution for ChatGPT ads that plugs into standard buyer tooling.

Confidence: Medium. Six months in with no measurement signal and policy still unfinished.

Why: OpenAI has run ads for six months and this week published a policy that defines expiry and revocation but not a single performance or measurement term, which tells you where its priorities sit. Building verified attribution for a conversational surface is genuinely hard: there's no pixel, no click funnel, no impression standard, and OpenAI would have to expose data it currently keeps entirely inside its own reporting. Platforms that control their own numbers rarely open them to outside verifiers until buyers force the issue with withheld budget, and no such pressure is visible yet. The opposite outcome, a full third-party measurement stack in one release cycle, would require OpenAI to solve conversational attribution and cede data control faster than any walled garden ever has.

Revisit by 2026-11-15: We're right if ChatGPT ads still lack an independent, buyer-side measurement or attribution integration by then. We're wrong if OpenAI ships a verified third-party measurement path (MMM-compatible signal, IAS/DV-style verification, or an open attribution API) before that date.