In Brief: As of the end of August 2026, ads in ChatGPT are available. They are not purchased based on keywords. The system matches them to the meaning of the entire conversation, the content of the landing page, and the description of the situation in which the client needs assistance. Ads appear directly below the model's response, meaning their effectiveness also depends on what ChatGPT said about the brand just above. Brand Semantics research across eleven industries, from construction and manufacturing to hotels, fashion, and legal services, shows that this background is rarely neutral. Therefore, this is a channel for prepared companies. A brand without organised entities, without a clear landing page, and without knowledge of what models say about it today will not gain an advantage here – it will only buy clicks to a conversation it lost just above. This shifts the skill set required for running campaigns: from managing bids and phrases to managing meaning, context, and brand representation in language models. Below, I explain how this channel works, why it could change the rules of the game in conversion, what it requires from a company wishing to leverage it, and how to check if you are ready for it.
1. What Actually Happened
OpenAI announced ads in ChatGPT on February 9, 2026, and tested them in the United States for six months. On August 24, 2026, the channel was launched in 31 European countries. The Ads Manager is currently in beta. In Europe, the first campaigns were launched through partners, with self-service to be gradually rolled out.
Here are some key facts to start with, as they clarify the rest:
Ads appear below ChatGPT's response, marked as sponsored and visually separated from the model's text. They are not embedded within the response.
One format: brand name with a favicon, a headline of up to 30 characters, a description of up to 60 characters, a square image of 256×256 px, and a link to the landing page. No video, no carousel.
Currently, they are visible to users on free plans (Free, and in markets where available – also Go). OpenAI states that it does not display ads on Plus, Pro, or Business plans, nor to users under 18. Ads do not appear in temporary chats or for users who are not logged in. This is important to consider when planning reach. It is also worth monitoring whether this policy changes.
Advertisers have no influence over the model's response and receive no user data: neither conversation content, history, nor email addresses.
Billing: CPM (cost per thousand impressions), CPC (cost per click; OpenAI recommends rates of $3–5), and – from August – oCPC, which optimises for conversions. The previous minimum budget of $50,000 has been lifted.
Measurement: impressions, clicks, CTR, average CPC and CPM, spend, conversions. Campaigns can be tagged with UTM parameters, and conversions can be measured with a pixel and through the Conversions API. OpenAI notes that the channel currently lacks benchmarks for effectiveness across industries.
2. How ChatGPT Selects Ads. And Why This Is Not “New Google Ads”
In search engines, the mechanics have been known for nearly a quarter of a century: a user enters a phrase, an advertiser bids on that phrase, and the one with the best combination of bid and quality score wins. The phrase is a common denominator. It can be bought, excluded, and measured.
In ChatGPT, there is no common denominator. According to OpenAI's documentation, the system selects ads based on four factors simultaneously:
the context and intent of the entire conversation, not just the last sentence;
the content of the landing page of the ad, which OpenAI reads with its own bot (OAI-AdsBot) and checks for compliance with advertising policy;
the title and description of the ad;
contextual hints provided by the advertiser – and, if the user has enabled personalisation, selected signals from their previous conversations.
Contextual hints are not keywords. OpenAI explicitly states that they do not function like exact match and do not guarantee display in a specific conversation. They describe the situation of need: who is asking, about what, at what decision-making moment. Instead of the phrase “CRM system,” the advertiser describes “a sales team comparing reporting automation tools and wanting to reduce manual work.” The final decision on ad placement is made by OpenAI's relevance model.
Additionally, there is an auction. The bid is just one input, and the relevance of the ad and landing page to the conversation jointly determines who wins and how much they pay. In other words: if the landing page clearly describes what the product is, for whom, in what situations, and how it differs from alternatives, the system has a basis for matching, and the ad wins at a lower cost.
Element | Search Engine Ad | ChatGPT Ad |
|---|---|---|
Purchase Unit | keyword | need situation described in natural language |
Matching Signal | query | meaning of the entire conversation + landing page + creative + hints |
Advertiser Control | high (phrases, exclusions, matches) | limited (hints without guarantees, geography, platform, own lists) |
User Sees Above Ad | list of links | the model's complete response, often with a recommendation |
Auction | bid × quality score | second price weighted by relevance |
Formats | text, graphics, video, shopping | one card: headline, description, image, link |
User Data for Advertiser | limited | none |
The conclusion is straightforward: in this channel, you cannot “set a campaign” on phrases. You need to understand how people actually talk to the model about a given category, at what decision-making moment this occurs, and what the model responds before displaying the ad.
3. Why Conversion May Look Different Than in Search Engines
To be fair: OpenAI does not publish benchmarks, and the market is just beginning to gather initial data. However, there are three independent reasons that help understand why this channel has a different conversion potential than traditional media.
First, the moment. A ChatGPT user is not “searching”; they are already conversing. They asked a question, received an answer, and often clarified further. The ad appears at the moment when the need is already articulated, and the selection criteria discussed. This is the closest to the decision moment that advertising has ever been.
Second, the quality of traffic. Shopify reported that in Q1 2026, traffic from GenAI assistants converted 49% better on product pages than traffic from search engines, and visits from this source increased eightfold year-on-year. Criteo, in a sample of 500 sellers, measured that traffic referred by language models converts about 1.5 times better than other referral channels. This data pertains to organic traffic, not paid, but it shows the type of user we are dealing with: someone who has already gone through the comparison stage in conversation.
Third, initial paid tests. A two-week test by SE Ranking in English-speaking markets yielded a 1.30% CTR on over 97,000 impressions. Another test published by a Canadian agency yielded a 0.65% CTR at a cost per click of around 4.4–4.9 CAD. The variance is significant and should be treated as a signal that the channel works, not as a benchmark.
There is also a downside that needs to be clearly stated. Ads currently reach users on free plans, not Plus, Pro, or Business subscribers. Two published B2B tests suggest that business decision-makers use paid plans more frequently than the average user, meaning that ads are currently not reaching them. A click costs $3–5. Inventory is limited, there is one format, and optimisation tools are still in their infancy. Every mismatched ad in this channel is costly. This is precisely why contextual precision ceases to be a matter of aesthetics and becomes a condition for profitability.
4. One Window, Two Layers: What the Model Says Above Your Ad
Here we arrive at something that is not covered in any guide to ChatGPT Ads, but which – in my opinion – determines the outcome of the campaign.
The ad appears below the model's response. The user first reads what ChatGPT thinks about their problem, what solutions it recommends, which brands it mentions, and how it describes them. Only then do they see the sponsored card. This means that in one window, there are two layers of brand messaging: the organic representation of the brand in the model's response and the paid card. The advertiser only controls the latter.
Let’s consider four typical situations we observe in model response studies conducted with our research tool Semantio:
Situation A: the model recommends a competitor. The user asks for a solution in a category, the model lists three brands and recommends one of them. Below appears an ad for a fourth brand – yours. The paid card competes with the recommendation the user just read and trusted. A click is possible, but the ad starts from a position the model has already lost for it.
Situation B: the model provides incorrect information about the brand. Your brand appears in the response, but with the wrong price, a non-existent certificate, incorrect location, or a feature attributed to a competitor. The ad below leads to a page that states otherwise. The user receives conflicting messages and does not know whom to trust.
Situation C: the model does not know the brand. The response is about the category, but your brand is not mentioned. The ad is the first contact – and the only argument. This is the most expensive scenario, as all persuasive work rests on 30 characters of the headline and 60 characters of the description.
There is also Situation D: the model recommends your brand, and the ad confirms it. In this case, the two layers work together: the organic response builds trust, while the paid card provides the shortest path to action. This is the only arrangement where the cost per click has a chance to return with a profit.
What We See in Research: Eleven Industries, the Same Mechanism
The following examples come from brand representation studies we conducted in 2026 at Semantio for Brand Semantics clients and as part of research by the AI Business Center Foundation using Semantio data. We provide the industries and scale of the study because they demonstrate that the mechanism is independent of the sector. Each result is a dated snapshot from one day of research – models change their responses over time, which is why measurements are repeated.
Industry (Client) | Scale of Study | What the Model Does in Response Above Which the Ad Would Appear | Situation |
|---|---|---|---|
Gate and Door Manufacturer (Premium, Construction) | 150 responses, 4 systems | Sentiment 8.1/10, but 0% direct recommendations at the decision phase; in general questions, models indicate mass-market and budget brands 88% of the time. Additionally, models “grant” the brand a lifetime warranty that the manufacturer does not offer. | A + B |
Polish ERP System Manufacturer (B2B Software) | 225 responses, 5 systems | Visibility 9.3% in general questions, 49–51% in comparisons and decisions. 11 out of 225 responses contain inaccuracies – the most concerning being the attribution of the system to competing manufacturers. | C, then B |
Regional Sausage Producer with Online Store | 120 responses, 4 systems | 0 out of 40 responses to general questions about delivery purchases in the region mention the brand; models indicate national chains. When the name is mentioned: 24 out of 120 responses contain inaccuracies (no bread in the offer, no same-day delivery, delivery to Warsaw) – all at the consideration and decision stage. The model bases its response on the brand's website only 17% of the time. | C + B |
Lighting Manufacturer (Home and B2B) | 199 responses, 4 systems | 95% visibility after the name, 40% without the name. 18 responses with an error alert or representation issue – 16 of them with high recommendation. The model recommends, but with incorrect information. | D with risk of B |
Fire Ventilation Manufacturer (Industry, B2B) | 130 responses, 5 systems | 0 out of 30 responses to category questions mention the brand; 82 out of 100 responses with the name recommend it “highly.” 14 responses contain inaccuracies, including four attributions of the Polish company to foreign capital groups. In one out of three responses: “no public data, ask the manufacturer.” | C, D, B – depending on the stage |
Legal and Consulting Firm (Professional Services) | 240 responses, 4 systems | Without name hints, the brand appears in 1 out of 92 responses – in eighth position on the list. In 7 responses, the model confuses the founder's name. In 51.5% of questions about proof of competence, it directs the client “for further verification.” | C |
Premium Hotel (Kraków, Tourism and MICE) | 225 responses, 5 systems | Visibility 47%, sentiment 8/10, but the model reduces the venue to “wellness hotel” and omits it in conference scenarios despite extensive facilities. The same location is an asset in some responses and a drawback in others. | D in wellness, A in MICE |
Boutique Hotel (Kraków; parallel study for the Norwegian market) | 5 systems, two markets | One of the systems interpreted the hotel's proper name as a category of property and on that basis directly discouraged it as a place for tourists from Norway. | B (entity error) |
Shoe Brand (Fashion, E-commerce) | 200 responses, 5 systems | Visibility 55%, good recommendation at the consideration stage, but at the decision stage, models repeat mentions of durability issues and complaints. The ad below such a response leads the customer who just read a warning. | A at the decision stage |
Plastic Profiles Manufacturer (Industry, B2B, part of an international group) | 180 responses, 6 systems | 0 out of 60 awareness scenarios; 82% recommendations when the brand is mentioned. Only 7% of responses cite the Polish domain – models attribute references and certificates to the foreign parent company. | C + B |
Building Centre (Retail, Local Market) | site audit + Semantio | The bots reading the internet see a “blank page” on the company's site – content only renders in the client's browser. For the models, the offer does not exist; recommendations are gathered by competitors. | C – and landing page without relevance |
Three things from this table have direct implications for ChatGPT Ads.
First, “no brand in category questions” is the norm, not the exception. In eight out of eleven studies, the brand is almost absent when the client does not yet know its name, precisely in those conversations where the ad would ideally enter. An ad in such a context operates in Situation C: it is the only argument and pays the full price for every click.
Second, high recommendation does not mean a safe background. At the lighting manufacturer, 16 out of 18 errors appeared in responses where the model recommended the brand. At the ventilation manufacturer, models recommended the company while simultaneously attributing it to a foreign capital group. An ad below such a response reinforces the recommendation and the error together.
Third, the same question in different systems yields different backgrounds. For the regional sausage producer, the recommendation score (on a scale of 0–100) ranged from 16 to 51 depending on the system; for the ventilation manufacturer, four systems saw the brand in 20 out of 20 questions, while the fifth saw it in 13 out of 20. A campaign in ChatGPT concerns one system, but the client compares across several – and the organic background must be known in each of them.
The difference between Situation A and D is not a matter of bid or creative. It is a matter of how the model represents the brand before the campaign even starts. And this is not visible in any advertising panel. To see this, one must ask the models the same questions clients ask – many times, in various forms, across several systems – and measure the responses. This is precisely the subject of Generative Engine Optimization (GEO) and the focus of measurement in Semantio.pro.
5. Four Competencies Required by This Channel
Since ChatGPT does not match ads to phrases but to meanings, the question is: who can work on meanings? From our experience, four competencies are needed. It is worth checking which of these your campaign team – internal or external – possesses.
5.1. Entity Map and Assertion Map – To Ensure the Landing Page Is Understandable to the Model
The OpenAI system reads the landing page and assesses the relevance of the ad to the conversation based on it. If the page describes the company as an “innovative technology partner,” the model has no basis to infer in which conversations the ad makes sense. If the page clearly states: what product, for whom, for what problem, in which category, with what evidence, and how it differs from alternatives – the model has a basis for matching.
Two examples from our audits show how often this condition is not met. The local building centre's page rendered all content only in the client's browser – the bots reading the internet saw a “blank page”; the only thing they could read was two words. Such a page as a landing page for ads in ChatGPT has no chance of winning the relevance auction, and the OAI-AdsBot has nothing to check. Conversely, a Polish industrial manufacturer operating in an international group had 699 out of 729 addresses in its site map leading to the foreign parent company's domain; models cited the Polish domain in 7% of responses and attributed references and certificates to another company. An ad for such a brand would lead to a page that the model does not associate with the entity it discusses in the response above.
We call this semantic clarity. In practice, it means separating two things: entity map (what objects create the brand's world: products, recipients, problems, competitors, evidence) and assertion map (what statements about these objects should be true, source-verified, and retrievable by the model). A well-prepared landing page is simultaneously a better page in search engines, a better source for language models, and a cheaper landing in the weighted relevance auction. This is one job, three effects.
5.2. Intent Scenarios – To Ensure Contextual Hints Are Data-Driven, Not Brainstormed
A contextual hint is a description of the situation of need: who, what need, what context, what decision moment. To write it well, one must know how customers actually ask models about a given category – at the discovery stage (“how to solve problem X”), comparison stage (“how does A differ from B”), and decision stage (“where to buy, whom to commission”).
At Semantio, we work precisely on such units: scenarios, which are full questions embedded in a specific persona, product, and decision stage, directed to selected systems. A company that has been studying purchase scenarios in the client's category for months writes contextual hints based on measurement: it knows which situations end in recommendations, in which the model mentions competitors, and in which the brand is absent. A team that lacks such measurement writes hints from intuition – and only tests them with the budget.
5.3. Measurement of Brand Representation – Before, During, and After the Campaign
This is a competency that distinguishes running a campaign in ChatGPT from running a campaign in any other medium. Because there is an organic response from the model above the ad, one must know:
before the start – in what percentage of purchase scenarios the model mentions the brand, how often it recommends it, what its Share of Voice is against competitors, what tone it uses to describe it, and whether it provides any false information about it;
during – whether the organic response supports the ad (Situation D), or competes with it (A, B, C); whether changes on the page and in sources shift the model's responses; what sources the model cites when discussing the category;
after the campaign – what exactly the model said about the brand on the day of airing, in a form that can be shown to management or clients.
In practice, this measurement changes campaign decisions. For the gate and door manufacturer, the organic background at the decision stage was neutral (0% direct recommendations with a sentiment of 8.1/10) – the ad only makes sense there together with evidence on the landing page that allows the model to justify a higher price. For the shoe brand, the background at the decision stage was directly unfavourable (mentions of complaints) – a campaign without prior work on sources would pay for clicks from customers who just read a warning. For the premium hotel, the background was excellent in wellness scenarios and poor in conference ones – contextual hints and landing pages must differ for each of these segments.
Semantio measures these dimensions on live responses from ChatGPT, Gemini, Claude, Perplexity, Google AI Overview, Microsoft Copilot, Grok, and DeepSeek: Visibility Rate, Share of Voice, recommendation level, sentiment, source impact, and detection of statements contradicting verified facts about the brand. Each result is a dated measurement from a specific study, not a forecast based on signals from the site. This way, a campaign in ChatGPT is not run in the dark: we know what organic background the ad is displayed against and whether that background works for or against it.
5.4. Evidence – To Ensure the Results Can Be Defended
In a channel without benchmarks and with a 24–48-hour delay in conversion reporting, it is particularly important that diagnoses and conclusions are auditable. At Semantio, every model response is logged along with metadata, classification, and cited sources, and selected responses can be preserved as a independently verifiable, tamper-proof record of what a given system said on a given day.
6. How We Run Campaigns in ChatGPT
At Brand Semantics, we manage ChatGPT Ads campaigns from diagnosis to reporting – with direct access to the Ads Manager and our own measurement tool. The process looks like this:
Diagnosis of brand representation in models. Research in Semantio: real purchase scenarios in the client's category, in market language, across several systems. Result: where the brand is recommended, where it loses to competitors, where models make mistakes, what sources shape the response.
Entity map and preparation of the landing page. Organising what the model needs to understand about the brand and translating it into the landing page: clear, source-based, allowing OpenAI bots (OAI-AdsBot and OAI-SearchBot – blocking either means rejection of the ad or disappearance from organic responses).
Scenarios → contextual hints and creatives. Measurement yields need situations where the ad makes sense, and variants of headlines and descriptions tailored to the decision stage.
Campaign in Ads Manager. Goal (reach, clicks, or conversions), geography, platform, daily budget, conversion measurement via pixel and through the Conversions API configured before launch.
Monitoring two layers. Simultaneously: campaign results from Ads Manager and periodic measurement of organic model responses in Semantio. If the organic background works against the ad – GEO actions on sources and content, not just on bids.
Report with evidence. What the model said, what the ad did, what changed between measurements – with logs and verifiable records of responses.
7. Who Should Enter Now, and Who Should Wait
It is worth testing now if the brand operates in a category where the client compares, asks, and returns with further questions: e-commerce with considered products, electronics, personal finance (within OpenAI's advertising policy), telecommunications, automotive, travel, education, services chosen after research, as well as home products, building and finishing products, or children's items. It is also worthwhile if the company wants to gain experience in the channel before competitors do – in a young channel, the cost of learning is lower than the cost of delay.
It is worth waiting if:
clients are mainly B2B decision-makers using paid ChatGPT plans – ads will not reach them for now; in this case, the organic representation of the brand in responses, which all users see regardless of plan, is more important. In the study of the fire ventilation manufacturer, we termed this “a tender lost before the tender”: the designer asks the model what to include in the specification, receives a list without the brand – and the brand only enters as a substitute at the execution stage, competing on price. A sponsored card will not fix this, only presence in the response itself;
the product is an FMCG or local service with a low basket value – with CPC at $3–5, the economics do not add up;
the industry is subject to advertising policy restrictions (alcohol, gambling, political content are prohibited; health and financial services are limited, and mostly not allowed outside the USA);
the company does not have a landing page that the model could understand – in this case, the first step is not a campaign, but organising brand representation.
Regardless of whether a company enters paid advertising, the organic layer always operates: a user on a paid plan will not see the ad, but will see what the model says about the brand. For many B2B companies, this is now a more important channel than any sponsored card.
8. ChatGPT Ads Readiness Checklist
Before launching a campaign, it is worth answering ten questions:
Do we know what percentage of real purchase questions in our category ChatGPT mentions our brand at all?
Do we know whom it recommends instead of us – and why (what sources)?
Does the model provide false information about us that the ad would reinforce instead of correcting?
Does the landing page clearly describe the product, recipient, problem, and evidence – in a language the model can adopt?
Does the page allow OAI-AdsBot and OAI-SearchBot?
Do we have intent scenarios for three decision stages, from which contextual hints arise?
Do we have several variants of the headline and description, each showing a different benefit, rather than the same message?
Is conversion measurement (pixel, Conversions API, UTM) configured before launch, not after?
Do our clients use the free ChatGPT plan?
Do we have a way to show management what the model said about the brand on the day of airing – in a verifiable form?
If the answers to questions 1–3 are “we don’t know,” the campaign will launch in the dark – regardless of who leads it.
9. Glossary of Terms
ChatGPT Ads – OpenAI's advertising channel where sponsored cards appear below ChatGPT responses for users on free plans; matching is based on the context of the conversation, landing page, creative, and contextual hints from the advertiser, not on keywords.
Contextual Hints – descriptions of the situation of need (who, what need, what context, what decision moment) that the advertiser provides in Ads Manager; they are not exact matches and do not guarantee placement.
Generative Engine Optimization (GEO) – a discipline following SEO: instead of ranking on a list of links, the goal is how language models describe, compare, and recommend the brand in a finished response. It includes organising entities and assertions about the brand, building sources that models cite, and measuring responses.
Brand Representation in Language Model Responses – the observable way in which a brand is present, described, compared, recommended, and sourced in responses from GenAI systems in a specific study. Visibility is just one of its dimensions.
Share of Voice (SoV) – the brand's share of “airtime” in the model's responses compared to named competitors in a given set of scenarios.
Brand Hallucination – a model's assertion about a brand that contradicts verified facts: incorrect price, non-existent certificate, wrong location, feature attributed to a competitor.
Entity Map / Assertion Map – tools for organising brand representation: a list of objects that create the brand's world (products, recipients, problems, competitors, evidence) and a list of statements about these objects that should be true, source-verified, and retrievable by the model.
Semantio – Brand Semantics' research platform for analysing and monitoring brand representation in responses from ChatGPT, Gemini, Claude, Perplexity, Google AI Overview, Microsoft Copilot, Grok, and DeepSeek, based on live responses, in market language, with logs and verifiable evidence.
10. Frequently Asked Questions
Can advertising in ChatGPT change what the model says about the brand? No. OpenAI separates the advertising system from the model. The advertiser cannot influence the content, order, or tone of the responses. Changing the organic representation of the brand is only possible through work on sources, content, and information structure – that is, through GEO.
Can I buy an ad for the phrase “best X in Warsaw”? No. There are no keywords. You can describe the situation of need in which the ad makes sense and ensure that the landing page and creative are relevant to it. The decision on placement is made by OpenAI's relevance model.
How much does a click cost? OpenAI recommends CPC rates of $3–5. The first published foreign tests fall within this range. There are no Polish benchmarks yet.
Will my B2B clients see this? If they are on a free plan – yes. Ads are not yet available on Plus, Pro, and Business plans; however, their users do see the organic responses from the model about the brand, and these determine whether the brand makes it to the shortlist.
Where to start? With measurement: what do ChatGPT and other models say about the brand in real purchase questions today. Without this, it is unclear whether the ad will reinforce the recommendation or compete with it.
