ChatGPT ads work as sponsored units shown below ChatGPT’s answer on the Free and Go plans, and to visitors who are not logged in. Advertisers buy them in OpenAI’s Ads Manager through an auction, and OpenAI’s system picks the ad from the conversation.
The ad never changes the answer. For B2B demand generation, the part that matters most is who is left out: a buyer on a company Business or Enterprise seat never sees an ad, while the same buyer on a free personal account can.
How ChatGPT picks which ad to show
An ad is not triggered by a single search phrase. In OpenAI’s account of how it built the ad system, published in October 2026, the ad service remembers context across the turns of a conversation and across conversations, and runs apart from the model writing the answer.
The selection is run by an AI agent of its own, a language model OpenAI calls the orchestrator. Each time there is a chance to show an ad, the system works through these stages:
- It turns the conversation into a brief summary, then into embeddings that capture its meaning.
- It retrieves candidate ads by that meaning, not by matching keywords.
- It drops candidates that fail on budget, the user’s settings or other eligibility checks.
- It scores the rest for quality and relevance, then ranks them.
- An auction admits only ads above a minimum of relevance and value, paces each budget, and sets the price.
- The winner appears below the answer, clearly labelled, if it arrives in time. If it arrives too late, no ad shows that turn.
The advertiser supplies the landing page, a title, copy, an image and context hints: plain descriptions of who the offer helps and when. OpenAI says hints “are not exact-match keywords” and do not guarantee delivery in any given conversation.
None of it touches the answer. OpenAI says ads “run on separate systems from our chat model”, so money buys the slot below the answer, never the brands ChatGPT names inside it.
Its engineers also call returning no ad “an ordinary, successful outcome”, which is why a campaign can spend less than its budget.
Who sees ChatGPT ads, and who never does
OpenAI’s help page on ads in ChatGPT sets out who sees them. Testing began in the United States on 9 February 2026, and ads are rolling out region by region:
| Who | Detail | Sees ads? |
|---|---|---|
| Free plan | Unless the user switches to Ads-Free, which lowers message limits | Yes |
| Go plan | Ads-Free is a Free plan option only | Yes |
| Plus, Pro, Business, Enterprise and Edu | Business and Enterprise are the company seats | No |
| Accounts identified as under 18 | Judged from account age information and age prediction | No |
| Temporary Chats | On any plan | No |
| Visitors not logged in | Only ads “appropriate for all audiences” | Yes |
Ads are never placed near conversations about personal health, mental health or politics, and political advertising is not allowed at all.
For a B2B advertiser, the question is how many buyers sit on a company seat. OpenAI counted more than 9 million paying business users in February 2026, at a time when more than 900 million people used ChatGPT each week.
Many people bring work to a personal account. Epoch AI and Ipsos surveyed 1,103 employed US adults in July 2026, and among the 469 who had used AI for work in the past week, free plans were the most common way in for most occupations, at 45% to 61% of workers.
Technical work is the exception: 67% of computer, engineering and science workers used an employer-provided subscription at least some of the time, against 41% in management, business and arts. The survey asked about AI tools in general, not ChatGPT alone.
So whether a buyer sees your ad depends less on their job title than on the account they are signed into, and engineers and scientists are the hardest buyers to reach this way.
How to buy ChatGPT ads, and what they cost
Advertisers buy through Ads Manager, OpenAI’s self-serve tool, in beta since May 2026. It is open in 89 countries; the company that pays must be based in one of them, and some new accounts can advertise only in their home country at first. Larger advertisers can ask OpenAI for sales support.
The campaign objective decides what you pay for, and it cannot be changed once the campaign exists:
| Objective | Best for | You pay |
|---|---|---|
| Views | As many people as possible seeing the ad | Per 1,000 impressions (CPM) |
| Clicks | Visits to your site | Per valid click (CPC) |
| Conversions | A conversion event you choose, such as a lead | Per valid click or per 1,000 impressions, never per conversion |
Bidding changed on 9 November 2026. New click campaigns can set a Target CPC and new conversion campaigns a Target CPA, or leave bids to Intelligent bidding within the budget. The older Max Bid and Max CPC controls retire for existing campaigns on 1 January 2027.
There is no list price. “The final winning auction bid sets the price you pay”, and a target is an average the system works toward, so a single click or conversion can cost more or less. OpenAI’s own basics page suggests starting click bids at $3 to $5.
Its advertiser FAQ also says the platform “does not yet have performance benchmarks across advertisers, industries, or campaign types”, so any cost per click quoted elsewhere was measured on someone else’s campaigns.
The minimum spend in Ads Manager is a daily budget per campaign: $25 in the US, 15 euros, 15 pounds, or 25 Canadian or Australian dollars. One day can spend up to twice the daily budget, and seven days never more than seven times it.
Billing is postpay. OpenAI charges after ads are delivered, starting at a low payment threshold that can rise with a history of successful payments.
A worksheet for your first test budget
With no benchmarks to borrow, the test budget comes from your own numbers, taken from paid search or paid social on a similar landing page. Work down the lines in order:
| Line | How to work it out |
|---|---|
| Conversions needed to judge the test | Your team’s call, made before launch |
| Landing page conversion rate | From your own paid search or paid social on a similar page |
| Clicks needed | Conversions needed divided by the conversion rate |
| Target cost per click | The Target CPC you will set |
| Test spend | Clicks needed times the target cost per click |
| Days to finish | Test spend divided by the daily budget, at least $25 in the US |
| Cost per opportunity | Spend divided by the opportunities your CRM credits to the campaign, set against your other paid channels |
If the days to finish run past a quarter, the test will not reach a verdict in time. Raise the daily budget, judge the test on a conversion that happens more often, or wait until the category or the audience fits better.
What you can target, and what you cannot
A campaign targets countries, then regions, cities, markets and postal codes where available, and up to five platforms: the Android and iOS apps, and the web on Android, iOS and desktop. To target only the United States, it has to be included explicitly; excluding other countries is not enough.
There is no setting for job title, company, industry or seniority, the fields a LinkedIn campaign is built on. Beyond location and platform, the conversation decides who sees the ad, and OpenAI is plain that context hints are “not targeting rules or delivery instructions”.
The one list-based option is a custom audience: emails, phone numbers or Google Advertising IDs, plain or hashed. Including an audience, or adjusting bids for it, needs at least 25,000 matched users, and OpenAI recommends 100,000. Smaller audiences can only be excluded.
That rules out targeting the account list account-based marketing runs on, unless it holds 25,000 matched people. A shorter list still earns its place as an exclusion, keeping current customers and open deals out of the audience.
Ad formats, landing pages and measurement
The standard ad is one unit below the answer: advertiser name, favicon, title, copy, an image and a landing page. A multi-image format of 4 to 6 square images opened in Ads Manager as a beta on 26 October 2026, and OpenAI is testing a multi-image ad shown while ChatGPT generates an image.
Sponsored Agents, which let a person chat with an AI representative of a business inside ChatGPT, are “available only to selected advertisers in a limited alpha test”.
Landing pages must not block OAI-AdsBot, OpenAI’s ad crawler, and OpenAI recommends allowing OAI-SearchBot as well, the crawler that decides whether ChatGPT search can show your pages. Every ad is reviewed before it serves.
Conversions reach Ads Manager through the OpenAI Pixel, the Conversions API or both. Reporting windows run 7, 14 or 30 days after a click, and 0 or 1 day after a view, so a deal that closes months later never shows up there.
If your sales cycle runs longer than a month, tag every ad with UTM parameters, capture them on the form, and let the CRM carry the first touch through to the opportunity, the only place a long B2B deal can be credited to the ad.
How ChatGPT ads differ from search ads
A search ad answers the words someone typed; a ChatGPT ad answers a whole conversation, matched by meaning. In a search campaign the advertiser chooses keywords. Here the advertiser describes the offer, and the system decides which conversations it fits.
Google runs ads in its own AI answers too, inside AI Overviews and AI Mode, which makes a paid slot beside an AI answer something both platforms now sell.
When a ChatGPT ads test makes sense in B2B
Start with the category. OpenAI’s ad policy, last updated in September 2026, says that during the initial test period ads are “primarily limited to consumer verticals”, such as household goods, local services, travel, and digital products or education.
Financial, healthcare and legal advertisers may be approved case by case. Software can fit: OpenAI’s own example of a context hint is bookkeeping software for growing retailers. A B2B service outside those groups should confirm it is eligible before planning a budget.
Worth a test when
- Your category is clearly allowed under the ad policy.
- Your buyers work in roles where a free plan is the usual way to use AI.
- The problem you solve is one a buyer would describe in a conversation.
- Leads from the ad can be traced to pipeline in your CRM.
- The worksheet’s test finishes inside a quarter.
Wait when
- Your buyers are mostly engineers or scientists on company seats.
- Your category is restricted, or allowed only case by case.
- The only result you can measure is a form fill.
● About the author
Mark Zides is the founder and CEO of Katama. He founded it to do the work rather than sell it. Meet the team.