Guide
How ads in ChatGPT work: the sponsored slot, the auction, and the answer above it
Founder of Epovest
Ads have been part of ChatGPT since 9 February 2026, first in the United States, then across thirty one European markets from late August and in Southeast Asia and Taiwan from 23 September: OpenAI now counts more than sixty countries. They serve the Free and Go plans. This guide describes the mechanism as OpenAI documents it: where the ad sits, how one is chosen, what it costs, what it is not allowed to touch, and what the person on the other side of the screen can turn off. Then it reports what advertisers have published about running it. Facts verified on 8 October 2026 against the OpenAI help centre and the ad policies.
Where ads run, and who sees them
The pilot opened in the United States on 9 February 2026. Eight further markets followed over the next six months, then OpenAI opened thirty one European countries on 18 August 2026, among them Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria. On 23 September came Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, after Australia, New Zealand, Japan, South Korea and India, and OpenAI now counts more than sixty countries.
The advertiser side follows its own rule, and it is worth keeping apart. To open a self-serve account, the legal entity that advertises and is billed has to be based in an open country: there were sixty three of them on 8 October 2026 on the list OpenAI keeps, eleven more than a month earlier, covering North America, most of Europe, Brazil, Mexico, India, Japan, Korea, Southeast Asia, Taiwan, Australia, New Zealand and markets across the Middle East and North Africa. A new self-serve account starts in its home country, and becomes eligible to advertise in the other supported countries once its identity is verified and it has reached the required level of spend at home.
An ad can appear below the end of a response. It is labelled as sponsored and visually separated from the response itself. It serves accounts on the Free and Go plans, and ads also appear to people using ChatGPT without signing in, where they are kept suitable for all audiences. Plus, Pro, Business, Enterprise and Edu accounts do not see ads, and neither do accounts identified as belonging to people under 18, based on account age information and age prediction. Temporary Chats carry none, and during this test neither does the ChatGPT Atlas browser. That eligibility rule decides more than it looks like it does, and the last section comes back to it.
The ad unit
An ad carries six elements: the advertiser name, a favicon, a title, a line of copy, a landing page and an image. One response may carry one or more units, which can feature several items or several advertisers, and length of conversation is taken into account. Campaigns can also be built from a product feed, which is how a Shopify catalogue reaches the auction.
Since 16 September 2026 the unit can adapt to the thread. When the advertiser switches the option on, its existing titles and descriptions are rewritten to the context of the conversation and translated into the person's preferred language. The language of an ad is chosen from the person's location, language settings and conversation, and it can differ from the interface language.
Two new formats sit beside it, both in the United States for now. Sponsored Agents, tested since the same day with a selection of advertisers, let a person who has seen an ad open a labelled conversation with an agent run for the business, separate from ChatGPT's own answers and from the thread where the ad appeared. And a visual format with larger images, announced on 5 October, is to be tested during image generation from later in October with a first group of advertisers, kept distinct from the image being created.
How one ad is chosen rather than another
Selection weighs the context and intent of the current conversation, the ad's own landing page, title and copy, the context hints and targeting the advertiser sets, and, when ads personalisation is on, signals from that person's broader ChatGPT experience.
That last branch does not apply everywhere. Personalisation is not open in the European Economic Area or Switzerland at launch: in those markets an ad is chosen on the current thread and on what the advertiser set, and on nothing else.
Context hints are written at ad group level. They describe the conversations, topics or situations where a product may be relevant, and OpenAI is explicit that they are not exact match keywords and guarantee delivery in no particular conversation. A buyer arriving from search advertising meets the difference in the first hour: there is no term to bid on, and the reporting does not return the prompts that produced a click.
The ranking, the objectives and the price
Among eligible ads, ranking runs a relevance weighted second price auction. Buying happens by objective, and there are three:
- Views, billed per thousand impressions (CPM), for reach and awareness;
- Clicks, billed per valid click (CPC), for traffic;
- Conversions, which steer delivery towards one standard conversion event and never bill per conversion: either per valid click (oCPC), or per thousand impressions (oCPM), the newer option, in which conversions that follow a view without a click also count towards optimisation.
Since August, eligible new ad groups run on automatic bidding by default: the Maximize results strategy sets and adjusts bids to get as many results as possible from the budget, and guarantees no cost per click, cost per acquisition or return target. An advertiser who needs a ceiling switches back to a manual maximum bid, for which OpenAI suggests three to five dollars per click as a starting point, and Ads Manager can indicate whether a bid is likely to be competitive. On a conversion campaign the cap is set per conversion, and OpenAI gives no recommended amount.
Budgets are set as a campaign total or per day, with a daily minimum that depends on the billing currency: fifteen euros, fifteen pounds, twenty five dollars. A daily budget is an average over a Sunday to Saturday week: a single day can spend up to twice the amount, the week never more than seven times, and a campaign that uses up its week early stops until the next one while still showing as active.
A campaign also carries its targeting: the countries it aims at and, where available, a region, city, market or postal code; the surfaces, among the iOS and Android apps, mobile web and desktop web; custom audiences to include or exclude, inclusion requiring at least twenty five thousand matched people and being best avoided in the European Economic Area and Switzerland while personalisation is closed there; and, for eligible advertisers since 5 October, phrases to exclude.
What the purchase does not touch
Ads run on systems separate from the chat model, and an advertiser has no ability to shape, rank or alter a response. Seeing an ad does not mean the product is endorsed or recommended, and OpenAI states that separation as a principle rather than a setting.
That boundary is the reason this guide exists. Two competitions run one above the other on the same screen, and they are won in two different ways. The card below is bought at auction. The sources cited inside the answer are not for sale: they are what the engine retrieved and chose to credit. We measure that second competition continuously, and on ChatGPT it is narrow. The engine cites at least one source in roughly three quarters of its answers, and credits about six distinct domains when it does cite, two figures we publish under the names AI Coverage and AI Density and whose spread across engines we measured in a study of citation chances. The next wave re-measured it one month on: ChatGPT opened its answer just as often, and credited five distinct domains instead of six. And the sources it credits are largely its own: asked the same question on the same day, ChatGPT shares at most 3.5% of its cited sources with each of the other engines we measure, as we measured over August and September.
The sponsored agents sit on the same side of that line. They open from the ad, carry the advertiser's label, and stay separate from ChatGPT's independent answers and from the thread the person started.
Where an ad is not allowed to go
Ads are not eligible near sensitive or regulated content, including personal health, mental health and politics, and political advertising is not accepted. During the initial test, ads are concentrated in consumer verticals: lifestyle and household goods, local services, travel and experiences, digital products and education. Since 5 October, DoubleVerify and Integral Ad Science have been running pilots that assess those safeguards in a controlled test environment, without access to real conversations.
Regulated categories follow a map of their own, and it is not the map of open markets, which is the thing to check before building a plan. Advertisers in financial services, healthcare and legal services can be approved case by case, after manual review and on proof of a licence to practise, but in the United States only: outside the United States, advertising for those three families of services remains prohibited. The rules were sharpened three times in two months: version 1.4 on housing and employment and version 1.5 on American legal services in August, then version 1.6 in September, which spells out OpenAI's right to refuse an ad that conflicts with its advertising principles, its commercial interests or its competitive position. An ad cannot carry an individual job opening or a specific property to rent or buy; linking to a platform that lists them stays possible as long as neither the creative nor the landing page points at one particular listing.
What the person on the other side controls
Ad controls live in the ChatGPT settings for Free and Go accounts in the regions where ads run. From there a person can turn personalisation off, hide an ad, report it, see why it was shown, and clear the data used to select ads, which is retained for up to thirty days before removal.
With personalisation off, ads still match the current thread, but other threads, ad history and topics are no longer used. If memory is on, memories and recent chats may inform which ad is selected. Interactions with ads, such as hiding or clicking one, are not written into memory. By default ChatGPT cannot see the ads shown to a person: sharing one into the conversation is a deliberate gesture, and it does not let the advertiser influence what ChatGPT then says about it. On the Free plan, an ads free experience is available in exchange for lower usage limits and no access to some tools, image generation and deep research among them, and the switch back and forth lives in those same settings.
What comes back to the advertiser
Reporting covers impressions, clicks, spend, click through rate, average cost per click, average cost per thousand impressions and conversions. Conversations, chat history, memories, names, email addresses, IP addresses and precise location never reach the advertiser: what returns is aggregated. If someone messages an advertiser through an ad, the advertiser sees only the messages sent to them. Impressions, clicks and click through rate refresh about every fifteen minutes; spend can lag by seven to eight hours.
Post click measurement gained instruments over the summer, and again in October. To the static tracking parameters, UTMs among them, which survive an ad click, OpenAI added a pixel, a conversions API and integrations with third party measurement providers in August. On 5 October it extended that list: Hightouch, Tealium and LiveRamp to send conversion data from existing systems, attribution providers for web and apps such as AppsFlyer, Adjust, Branch, Northbeam, Triple Whale and DV Rockerbox, and first geographic experiments on incrementality with Haus, Measured and WorkMagic. The channel is therefore identifiable in your own analytics, by the same means as any other paid source, and a conversion can be traced back to the campaign that produced it.
Five doors into the same inventory
Buying happens through the self-serve Ads Manager, through the agency groups that placed the first campaigns, or through ad technology partners. Criteo was the first such partner to integrate with the pilot, followed by others including Adobe, Kargo, Pacvue and StackAdapt. Since 16 September two more doors open from tools businesses already use: HubSpot, the first CRM partner, where an account can create ads and follow up leads, and a Shopify app for merchants, open in the United States first and in the other ad markets from 23 September. A plugin also lets an advertiser create and analyse campaigns from inside ChatGPT. Every door reaches the same auction; they differ in who builds the campaign and in what reporting sits on top. OpenAI said in late August that its ads had reached a billion dollars in annualised revenue in under two hundred days, with tens of thousands of advertisers running campaigns.
What advertisers report
Published first hand accounts are still few, so we set an entry rule and kept it: a named author, a stated spend, a stated period. Anything reporting benchmarks without a sample or a date was left out, and that removed more articles than it kept; this month again, a press release without a dated period and a test without a stated spend stayed out. Thirteen accounts qualify.
| Author | Period | Spend | Clicks | CTR | CPC | Conversions |
|---|---|---|---|---|---|---|
| Murat Bock, adlibrary.com | 2 weeks | $200 | ~40 | ~0.8% | ~$5 | 0 |
| Out of the Box Advisors | published 24 July | $500 | ~140 | not stated | ~$3.50 | 0 |
| Shawn Walker, Symphonic Digital | 1 month, published 3 August | ~$500 | 53 | not stated | ~$9 | 0 |
| Austin LeClear, Grow My Ads | ~2 weeks, published 3 August | ~$1,200 | 92 | 1% | ~$13 | 0 |
| Alex Thilén, Opascope | 15 days | ~$60,000 | ~34,900 | not stated | ~$1.72 | 2.35% conversion rate, 1.49x return on ad spend |
| Q1Media | published 24 July, aggregated | not stated | not stated | 0.5% to 2.5% | ~$5 | ~2%, cost per acquisition around $500 |
| Conversion Logix | 30 days in June, published 19 August | $4,000 | not stated | 0.45% to 1.34% | not stated | 16 key events across five campaigns |
| Andy Brice, Successful Software | 5 to 28 August | £289.52 | 2,988 | not stated | £0.10 | 14 installs, 0.46% |
| Hüseyin Ograk, Hostinger | several weeks, published 18 May | ~$70,000 | not stated | not stated | not stated (CPM above $65) | purchases, not quantified |
| Joel Horwitz, Synter | 2 to 10 June | $4,428.84 | 448 | 0.35% | $9.89 | not reported |
| Common Thread Collective | 34 days from late May, published 1 July | $9,620 | ~2,180 | 0.94% | $4.41 | 3.3x to 6.8x return on ad spend, by attribution model |
| Amanda Shomo, Windmill Strategy | 22 June to 21 July | $1,470 | 259 | 0.66% | $5.68 | 0 |
| Floyd Blaikie | a few weeks, published 27 July | ~C$7,000 | ~750 | 0.7% | C$9.29 | 0 |
Two campaigns at $200 and $500 carry no average, and the authors of the small tests say as much themselves: thin creative written around features rather than needs, a budget spent in about an hour, a first run its own author calls a smoke test rather than a test. Read the table for its axis rather than its levels.
The axis is visible, and it held with five more accounts. The six that report no conversion all sell to professionals, and ads serve the Free and Go plans, so a large share of the buyers they are looking for sit outside the inventory by construction. The two that report a positive return both sell to consumers: Opascope, about $60,000 over fifteen days at a cost per click five to seven times lower, with daily figures ranging from 0.2 to 2.9 times spend, and Common Thread Collective, for a product sold above a thousand dollars, between 3.3 and 6.8 times depending on the attribution model.
Floyd Blaikie measured that axis directly, which no other account had done. A visitor identification tool on the landing page put names on part of the paid traffic: 336 clicks, 146 organisations identified, five of them in the profile the campaign was aimed at. Ads Manager does not return that reading; the team had described its target in the context hints, and planned to test the custom audiences opened that same month.
The newer entries add two variables the first six did not show. The first is the bid. Andy Brice bid between seven and fifteen pence where the manager suggested two pounds, took close to three thousand clicks at ten pence each, and ended with fourteen installs. Windmill Strategy set a maximum of twenty dollars and paid between five and six, which is what a second price auction does. Conversion Logix compared the two buying modes across five property campaigns and paid less than half as much per session buying on clicks as on impressions, $4.63 against $10.06, with a better click through rate on top. The second is the market. Across the five countries Synter ran in over nine days, the cost per thousand impressions hardly moved, but people in the United Kingdom clicked at 0.72% against 0.33% in the United States, so a click cost $5.10 there against $10.62. A low bid buys volume, then, and a market that clicks more buys it cheaper; what that volume is worth is the question left standing.
The picture from the ad technology side is different in scale and in kind. Criteo reported in June 2026 that more than two thousand brands were live through its integration, with click through rates two to three times those of comparable formats elsewhere and more than eighty per cent of ad driven traffic coming from new customers, apparel, home furnishings, consumer electronics, automotive and beauty leading the categories. These are Criteo's own figures on its own client base, in curated verticals, and independent verification at that scale does not exist yet. OpenAI published a second set on 5 October, figures from its measurement partners on named brands: a cost per acquisition 15.3% below WeightWatchers' search benchmark according to DV Rockerbox, 67% of incremental purchases coming from new customers for the wellness brand Dose according to WorkMagic, and 93% of new visitors for Portland Leather according to Triple Whale. Neither set states the spend or the period behind it, which is why both sit beside the table rather than in it.
The brands named in the pilot say something else again, and their silence is itself informative: Target, Williams-Sonoma, Albertsons and The Knot Worldwide have all confirmed taking part, and none has published performance. The Knot's chief marketing officer described the pilot as carrying no hard performance targets, framed as a way to observe behaviour and traffic while the format is young.
The measurement gap, and what moved it
Four of the first six accounts described the same thing without citing one another: the platform's click count and the site's analytics did not agree. Symphonic Digital recorded 53 clicks in the platform against 35 sessions in analytics in July. Out of the Box Advisors found no matching traffic in their own tools. Grow My Ads hit a tracking bug, rebuilt the campaign without UTM parameters, and watched the traffic land as direct. Opascope moved to first party tracking, reporting that the native dashboard was under counting conversions.
The August and September readings describe something else, and that is the most useful update on this page. Andy Brice began by testing that same gap and ruled it out: the clicks OpenAI reported matched the ones his analytics recorded. Conversion Logix does not talk about counting either, but about qualifying: the native conversion metric did not let the agency judge quality, so it tracked its own key events instead. In between, the tooling announced on 18 August added a pixel, a conversions API and third party integrations to tracking parameters that already survived the click.
The gap has moved, then, from counting to value. Counting clicks is a problem with instruments that exist; knowing what a click is worth stays the advertiser's work, and it is prepared before the budget starts running. Andy Brice had to instrument his site with a third party to establish that most of his clicks came from humans, and he established it. What was missing after that was not a counter, it was the intent behind the click.
The accounts added this month confirm that move, and show how wide the value question is. Common Thread Collective's return doubles depending on the attribution model, from 3.3 to 6.8 times, on the same spend and the same sales. Hüseyin Ograk notes that a return read on the first purchase undervalues customers who come back and buy again. Floyd Blaikie had to identify the companies behind the clicks to learn what they were worth.
OpenAI's October announcements answer on that ground rather than on counting. A conversion campaign billed on impressions counts conversions that follow a view without a click, and the incrementality experiments with Haus, Measured and WorkMagic aim at what an ad caused rather than at what merely followed it; OpenAI describes that work as being at an early stage. The question has moved one step further, from what a click is worth to what the ad changed.
Where this leaves a publisher
Three things follow from the mechanism itself. The eligible audience is the Free and Go plans and signed out use, which sets the ceiling of who the channel can reach. The sponsored card, the sponsored agent and the cited sources are decided by separate systems, so a campaign changes nothing about how the answer above it is composed. And an ad click is identifiable on arrival, provided the parameters are in place before the budget starts running.
The corpus offers one comparison worth reading twice, because it comes from the same site, over the same period, in the hands of the same author. Across the twenty three days of August that Andy Brice measured, clicks bought inside ChatGPT converted at 0.46%, and visits arriving from ChatGPT answers without a purchase, at 4.2%. It is one site, one period and one author, and it is the only measurement in this corpus where both competitions on that screen are counted side by side.
If the goal is to appear inside the answer rather than below it, that is the other competition. It has its own mechanics, and we describe them in what GEO is and in why ChatGPT recommends your competitors. The budget that aims at the answer itself follows another map: the sources the engines read, ranked by measured weight.