ChatGPT ad prices collapsed before OpenAI went global

I.The expansion that reads as a warning

OpenAI expanded its ChatGPT advertising test on 11 August 2026 to the UK, Mexico, Brazil, Japan and South Korea. The pilot began in the United States on 9 February 2026 and reached Canada, Australia and New Zealand on 26 March. Ads appear only to logged-in adults on the Free and Go tiers; Plus, Pro, Business, Enterprise and Education accounts see none. They are labelled as sponsored, kept away from health, mental-health and political queries, and — OpenAI states — do not influence the answers themselves. Advertisers receive aggregate views and clicks, not conversations.

Five days earlier, on 6 August, OpenAI gave Free and Go users unlimited text chats and made GPT-5.6 Luna the default. Read together, the two moves describe one strategy: give away far more inference, then monetise the attention it buys. The tell is in the fine print of the ads page — a Free user may switch ads off, but only in exchange for a lower daily message cap. OpenAI is now explicitly pricing attention against tokens, which means it has a number for what a user's eyeballs are worth relative to compute. That number is falling.

II.What the price actually did

ChatGPT ads launched at $60 per thousand impressions, reported by The Information in January 2026 as roughly triple Meta's average. By 17 April, buyers were seeing $25 to $45 depending on inventory mix, with one executive quoting $15. The minimum spend fell from $200,000–$250,000 at launch to $50,000 by 13 April, then to $10,000 through Criteo in June, alongside a shift of selected inventory from impressions to cost-per-click. Falling floors and a move to performance pricing are what a seller does when reach alone will not clear.

The counter-reading is that this was deliberate. Adthena's chief marketing officer told Digiday the cuts looked like preparation for wider auction access ahead of a global rollout — and a global rollout duly arrived four months later. That reading is plausible and worth holding. It is also unfalsifiable in the direction its proponents want, because a price cut ahead of expansion and a price cut caused by weak demand look identical until the price recovers. Four months on, it has not.

III.The bottleneck is ad load, not demand

Roughly 85% of ChatGPT users are ad-eligible, yet fewer than one in five encounter an ad on a given day. That gap is the whole problem. A search results page can carry four ads above the fold without breaking; a conversational answer can carry one sponsored unit before it stops feeling like an answer. eMarketer's June forecast names this directly as the first of three structural constraints, alongside CPM compression and the fact that most consumer AI interaction still happens inside Google Search rather than a standalone chatbot.

The demand side is genuinely more encouraging than the price suggests. Criteo data across 500 US retailers found LLM-referred users converting at 1.5 times other referral channels; almost 20% of shoppers who touch a sponsored prompt keep talking about that brand, and around 70% of post-click purchasers are new to it. High intent, low volume. The missing piece is verification: as one marketing consultant put it to Digiday, industry measurement is the real moat, and until independent attribution exists, advertisers are taking OpenAI's word for what an impression was worth.

Then there is the audience. eMarketer's survey work finds 63% of US adults saying chatbot ads reduce their trust in the AI's output and 57% saying they reduce trust in the advertised brand; only 24% regard chatbots as unbiased, against 41% for Reddit. Sam Altman had called advertising a last resort and warned against a trust-destroying moment. The trust cost is not hypothetical, and it is borne by the product that generates the subscription revenue.

IV.The arithmetic that does not close

OpenAI disclosed a $100m annualised ad run-rate on 26 March 2026, seven weeks into the pilot. Its internal target for the year is $2.5bn, rising to roughly $100bn globally by 2030. eMarketer's US AI advertising forecast, published 4 June 2026, projects that the entire US standalone chatbot ad market — ChatGPT, Copilot, Gemini and Alexa combined — generates under $1bn this year and $5.41bn by 2030, inside a US AI ad market growing from $32.03bn to $68.25bn. Most of that growth accrues to ads placed beside AI-generated search results, which is Google's inventory, not OpenAI's.

For scale, Alphabet booked $63.3bn in search and other advertising revenue in Q2 2026 alone, up 17%, with total Google advertising at $81.6bn — after a full year of AI Overviews. The thesis that AI assistants would strand search advertising has not shown up in the accounts. What has shown up is capex: Alphabet raised full-year 2026 guidance to $195–205bn. Meanwhile OpenAI is forecast to lose about $14bn in 2026 on roughly $25bn of revenue. Ads at a few hundred million dollars do not move that line; they are a rounding error against the compute bill.

V.Who pays instead

Trace the channels. Prices: if the ad tier cannot subsidise free inference, the subsidy has to come from paid tiers, so the pressure runs toward higher Plus and Pro pricing or tighter free limits — the opposite of the 6 August giveaway, and a reason to read that giveaway as customer acquisition rather than sustainable unit economics. Capital flows: a business that cannot monetise a billion weekly users through advertising is a business that keeps returning to primary markets, which is why funding rounds and IPO timing, not CPMs, remain the binding constraint. Market structure: a chatbot ad market worth $5bn in 2030 does not disrupt a search ad market worth over $250bn a year; it concentrates advertising further with the incumbent that owns both the surface and the measurement. Labour: the agency work being created here is small and mostly experimental, and the measurement gap means the buying side cannot yet reallocate budget with confidence.

The second-order effect matters more than the first. If conversational advertising yields poorly, the commercial logic of consumer AI shifts from attention to transactions — agentic checkout, affiliate take-rates, retail media inside the assistant. That is a materially different industry: it puts AI labs into payments and merchandising, it makes retailers the counterparty rather than brand advertisers, and it changes what regulators need to look at from ad disclosure to conflicted recommendation.

VI.The European hole in the map

One large advertising bloc is missing from every phase of the rollout. OpenAI's own documentation states that personalised ads are not initially available in the European Economic Area or Switzerland. The UK is in; the EU is not. Separately, on 2 August 2026 the European Commission began enforcing the AI Act's transparency obligations, which require disclosure that a user is interacting with AI and machine-readable marking of synthetic content, with fines up to €15m or 3% of worldwide turnover and an extended marking deadline of 2 December 2026 for existing systems.

Attributing the EEA exclusion to the AI Act would be over-reading; GDPR and the Digital Services Act already constrain profiling-based advertising, and OpenAI has not published a reason. The judgment worth recording is narrower: the ad product ships in a two-speed form, with personalisation available where consent rules are looser, and the European variant remains unbuilt in the market with the most demanding disclosure regime. If personalisation is what makes the CPM, a permanently unpersonalised Europe is a permanent haircut on the addressable market.

VII.How it plays out: three scenarios

Probabilities below are analytical judgment, not forecasts drawn from a model. Horizons are 12 to 36 months from August 2026.