The ads arrive on schedule

OpenAI updated its ChatGPT advertising page on Wednesday to confirm that the US pilot, quietly running since February, is being extended to five additional markets. The page is unsentimental about it: ads appear to free and Go-tier users; Plus, Pro, Team, and Enterprise remain "ad-free, always." This is the company that spent two years insisting it had no plans to put ads in ChatGPT now publishing a product page for them.

The numbers reported around the pilot are, if accurate, the part worth reading carefully. The Information reported the US test crossed roughly $100 million in annualised revenue within six weeks of launch, with more than 600 advertisers in the system.¹

What was actually published. OpenAI's updated commerce and advertising page says ads are "currently shown to logged-in Free and Go users in the United States" and that the pilot is "expanding to Canada, the United Kingdom, Australia, Germany and France over the coming weeks." Ad placements are described as "sponsored suggestions within the response," with the company committing to clear labelling and a "no targeting based on the contents of your conversations" rule, targeting instead runs off "general signals like region, device, and aggregated interest categories." There is also a paragraph about advertisers not being able to "buy higher ranking in ChatGPT's answers," which is the sentence the lawyers wrote and the sentence the product team will spend the next eighteen months relitigating.

The frame this fits. This is the AI advertising frame arriving on schedule, slightly earlier than I'd have guessed and at a higher initial run-rate than the consensus assumed. The structural question with conversational AI has always been whether ads would be additive (a new surface, incremental revenue, no degradation of the core product) or replacement (the free tier degrades, the paid tier becomes the actual product, the funnel reshapes). OpenAI is publicly betting on additive. The pricing-tier carve-out, ads on Free and Go only, is the explicit promise.

It is the same promise Google made about search ads in 2000, and to be fair, Google kept it for an unusually long time before it didn't.

Why the run-rate matters more than it looks. ChatGPT has, per OpenAI's most recent disclosures, something north of 800 million weekly active users, the overwhelming majority on the free tier. A $100m ARR pilot at six weeks, with 600 advertisers, on a US-only free-tier subset, implies an RPM (revenue per thousand interactions) that is small in absolute terms but interesting in trajectory. Google's search RPM is roughly $40–60 depending on how you count; display is a fraction of that. If ChatGPT ads scale linearly with geography and load, they won't, but as a Fermi exercise, the global free-tier monetisation ceiling is materially larger than the current subscription business.

This is the part where the inference-economics frame intrudes. Free-tier ChatGPT is the most expensive free product in the history of consumer software. Every query runs through a frontier model; every response burns GPU-seconds that cost real money. OpenAI has been subsidising free users out of subscription revenue and investor capital. Ads are the first plausible mechanism by which the free tier pays for itself rather than being a customer-acquisition cost for Plus. Whether the unit economics close depends on whether sponsored-suggestion RPM clears inference COGS per session, and OpenAI has not disclosed either number.²

I would like to see them disclosed. I do not expect to.

The agent-economics complication. The advertising page is careful to say ads appear "within the response" as sponsored suggestions. This is fine when the user is asking ChatGPT what to buy. It becomes structurally weird when the user delegates the buying to an agent. OpenAI is, separately, building Operator and pushing agentic checkout flows; Anthropic and Google are doing the same. If the equilibrium is that consumers ask agents to purchase on their behalf, the sponsored-suggestion model collides with the agent's fiduciary posture toward the user. An agent that recommends the sponsored option rather than the best option is an agent the user fires.

OpenAI knows this. The page's promise that advertisers cannot "buy higher ranking in ChatGPT's answers" is the seam where the conversational ad business and the agent business have to be kept apart. The pilot is a free-tier consumer product; the agent flows will be a paid-tier or API product; the wall between them is the entire business model. I would watch the language on that page like a hawk over the next year. The day "ChatGPT's answers" becomes "ChatGPT's responses" or some softer phrase, the wall has moved.

The day "ChatGPT's answers" becomes "ChatGPT's responses," the wall has moved.

What this is a case of. Two patterns, both familiar.

The first is the standard consumer-internet monetisation arc, compressed. Free product builds scale; scale attracts advertisers; advertisers fund the free product; the free product degrades as ad load rises; paid tiers become the de facto product. Google ran this arc over fifteen years. Meta ran it over ten. OpenAI appears to be attempting it over two. The compression is what the enshittification frame predicts when capital costs are high and patience is low, and ChatGPT's capital costs are extraordinarily high.

The second is the strategic-defensive read. Google's AI Overviews are eating search-ad inventory; Perplexity is selling sponsored answers; Meta is putting its assistant in front of three billion users with an ad stack already attached. If OpenAI does not build an advertising business, it cedes the conversational-ad surface to companies that already know how to run one. The pilot is, in part, OpenAI declaring it intends to compete for that surface rather than retreat into the enterprise API business.

What to watch.

• Whether the five-market expansion produces a roughly proportional revenue bump or materially less (US ad markets are uniquely deep; UK and DE are the next-best; AU, CA, FR are smaller).
• The first disclosed RPM or per-user ARPU figure, which will tell us whether the unit economics close against inference cost.
• Any softening of the "no ranking influence" language on the commerce page.
• Whether the Plus tier price holds. If the free tier monetises well, OpenAI will have an incentive to keep Plus expensive and protect the ad-free promise. If it doesn't, Plus will be where the pressure goes.
• Anthropic's response. Claude.ai is still ad-free and Anthropic has positioned itself, awkwardly, as the safety-coded alternative. Whether that becomes a market position or a margin problem depends on whether OpenAI's ad business works.

The most honest sentence on OpenAI's updated page is the one that says ads are "an early experiment, and we'll keep evolving based on what we learn." Translated: we have built the surface, the surface works, the surface will get larger.

Footnotes

¶