ChatGPT Ads: how it works and what changes vs Meta

ChatGPT Ads is OpenAI's advertising system: labelled ads that appear below ChatGPT's answer, selected mainly from the context of the conversation in progress rather than from the person's profile. Only people on the free and Go plans see them. Since late August the channel has been open across 31 European markets, and you can buy it yourself from OpenAI's panel without going through an agency.

The panel opens, you add a card, you build your first campaign in a single sitting.

The trouble starts afterwards, when you go looking for the settings you have used for years on Meta. The panel gives you three levers and nothing more: where (country, and where supported also region, city, metro area and postcode), which platform, and which custom audiences to include or exclude. Interests, age brackets and lookalikes built from a list are not among the options.

In their place there is a field called context hints, and it asks you for something you are not used to giving: not who you want to reach, but when your product is useful.

Download the free ChatGPT Ads guide

Ten pages: campaign structure, context hints, ad specs, minimum daily budgets by currency, a tracking checklist with the UTM string ready to paste, and a thirty-day test plan.

Download the PDF →

Who sees these ads, and who never will

Ads appear to people using ChatGPT on the free and Go plans. Nobody on Plus, Pro or a business plan sees them, and neither do accounts identified as belonging to under 18s, based on account age information and age prediction where available. Ads are labelled, they sit separately from the answer and they do not influence it; advertisers receive no chats, no memory and no personal data from the people writing them.

There is also a context rule worth knowing before you plan anything: OpenAI states that it keeps ads away from sensitive conversations and brand-unsafe environments, following its own advertising policies.

Before any other consideration, ask yourself where your audience sits. If you sell to people who work with AI all day, many of them are subscribers and will never see you. If you sell to someone choosing bathroom tiles, the odds change considerably.

Nobody will put it to you that plainly while explaining how promising the channel is.

How ChatGPT picks which ad to show you

It reads the conversation in progress. OpenAI says so in its European privacy policy: the advertising system uses the context of the current chat, together with other available information, to choose the ads most relevant to a given request.

For people who have consented to personalisation, past conversations and previously engaged ads come into play too. For those who have not, what remains is generic advertising chosen from whatever is left: current chat context, location, time of day.

On Meta you describe a person and hope they are looking at their phone. Here you describe nobody: you wait for someone to say out loud the problem you solve. The moment has taken the profile's place.

Context hints, which is the genuinely new thing to learn

The advertiser documentation defines them with disarming simplicity: context hints give additional information about when you think your ads might be useful, and help guide when they appear. You provide a list of descriptions or words for the cases where your product or service might be relevant.

The official example is a two-item list: productivity, team collaboration.

Three things to keep in mind, and the third is the one that does damage.

They are not exact-match keywords. You are not buying the query "field service software for plumbers", you are describing a territory of conversations and the system decides.

They do not guarantee delivery on a specific question. No promise, ever.

And above all they are not your Meta interests in a new box. If you write "furniture, design, home" you are describing product categories, which is the old way of thinking. If you write "redoing the kitchen in a rented flat", "choosing between two quotes from a carpenter", "working out whether bespoke is worth it", you are describing moments. That is where the system recognises you.

The rest of the levers are geographic: country, and where supported also region, city, metro area and postcode. Plus the platforms you let the campaign run on, iOS app, Android app and web, where "web" covers desktop and mobile together. And the custom audiences you upload yourself, for eligible accounts. End of list.

One detail worth knowing before you read your first report: you buy on three platforms and the data comes back on two. Insights groups devices into Mobile and Desktop, and mobile web lands under Mobile alongside both apps. If you want to know what an Android app user cost you, today you cannot: split campaigns by platform or give up the number.

Meta stops you, ChatGPT answers you

Here is the real difference, and it is not technical.

On Meta your ad gets in the way. Someone is scrolling the feed to switch off and you interrupt them. Everything you know how to do on that platform comes from this: the first frame that stops the thumb, the hook thrown in immediately, the contrast that breaks the scroll. It is interruption advertising, and it is a hard craft that you learned.

Inside ChatGPT there is nothing to interrupt. The person wrote a question and is waiting for an answer. If you land on them with a shouted hook you do not look sharper than the rest, you look out of place: your ad lives next to a calm piece of text explaining something in order, and striking a wrong note there costs trust instead of earning attention.

	Meta	ChatGPT
Where it appears	in the feed, while you scroll	inside a conversation you started
What it interrupts	entertainment	nothing: you asked the question
What you describe	the person: interests, age, behaviour	the moment: when your product is useful
Who picks the recipient	your targeting and the algorithm	the chat context, plus your context hints and, when personalisation is on, other account signals
Who sees it	anyone	only free and Go plans
How they arrive	distracted	actively evaluating

Two different crafts going by the same name.

How a campaign is built, and where the bid goes wrong

Three levels, as you would expect: campaign, ad group, ad. The campaign holds objective, budget, dates, locations and platforms. The ad group holds theme, context hints and bidding strategy. The ad holds brand, title, copy, image and landing page.

There are three objectives, but four combinations. CPM if you pay per thousand impressions and you need visibility. CPC if you pay per valid click and you need traffic. And then the Conversions objective, which you can buy two ways: oCPC, where you pay per valid click and the system optimises towards conversions that follow a click, and oCPM, in beta, where you pay for impressions served and optimisation also accounts for conversions that follow an ad view.

One thing holds across all of them: objective, billing model and conversion event cannot be changed after creation. If you got it wrong, you rebuild the campaign.

Budget has two controls that are not the same thing. The daily one is a delivery target, and a single day's spend can land above or below it. The campaign total is a ceiling, and it is the rigid control. For a first test, decide the ceiling first and let everything else work inside that perimeter.

Then there is the trap, the one the API documents and the panel does not shout at you. In conversion campaigns the value you type into the maximum bid is not a cost per click: it is what you are willing to pay for a conversion. In the panel it is called Bid Cap, and OpenAI is explicit that it is neither the price of a conversion nor a guaranteed CPA: it is what the system uses to compete in the auction, and what you actually pay is settled by the auction.

Through the API that value is expressed in millionths of the currency unit: in the official example a hundred dollars is written 100000000. One order of magnitude too many in that box tells the system something other than what you meant.

Ad specs, which are tighter than you think

Element	Recommended	Limit
Title	16-24 characters	50 maximum
Copy	32-48 characters	100 maximum
Image	square, PNG or JPG	1200 × 1200 maximum, publicly reachable
Brand icon	your mark, legible small	at least 128 × 128
Landing page	consistent with the ad	valid, reachable, not blocked to OpenAI's crawlers

The advertiser name and the icon complete the unit. No video, no carousel: the standard format is called a chat card. There is a second one, the product ad template, which pulls image and URL from a feed and can also show a price: it is for catalogues, and you do not write it out ad by ad.

A hundred characters is less than this sentence plus the one before it. There is no room for a preamble, for the founder's story, for the hook followed by the promise. There is room for one thing said well.

And there is a requirement that goes unnoticed and counts double if you also work on visibility: the landing page must be valid, reachable and must not block OpenAI's user agents. Which ones exactly is written in the documentation: OAI-AdsBot and OAI-SearchBot.

This is the easiest thing to get wrong, because OpenAI runs three bots and they do different jobs. OAI-AdsBot is the advertising one, OAI-SearchBot is what gets you into ChatGPT's answers, GPTBot is for model training. You can refuse GPTBot without touching your campaigns or your organic visibility. But if your robots.txt carries a rule that blocks anything starting with OAI, you have just switched off your own advertising.

How a conversion gets measured, and where it hurts

There are three tools and you want at least two of them.

The pixel is a snippet you put on the site and call when something happens: registration_completed for a sign-up, lead_created for an enquiry, order_created for an order. Standard events have fixed names, but it is not a closed list: you can send your own by declaring it as custom and giving it a name of 1 to 64 characters, as long as it does not match a standard one. Custom events cannot become a campaign's optimisation goal, though: for that you need the standard ones.

The Conversions API sends the same event from your server, and it exists because the pixel is fragile. Ad blockers, refused consent, browsers that cut scripts: every conversion the browser fails to report is a conversion that, as far as the platform is concerned, never happened.

Then there is deduplication, which holds the two legs together. If you send the same conversion from both sides with the same event identifier, OpenAI keeps the first to arrive and drops the second. Without that shared identifier you count everything twice and optimise on inflated numbers, which is worse than not measuring at all.

One detail that costs money to people arriving from Meta: here too the value goes in minor units, as a whole number. Two thousand euros is written 200000. Send it in euros and you would be declaring twenty-euro leads, and the platform would learn to find you twenty-euro customers.

How long you have to attribute a conversion

You choose the click window when you create the conversion. The view window you do not: it is fixed at one day after the impression, and it is reported as a separate campaign-level metric, outside the conversion total. When a conversion is eligible for both, the click takes precedence.

Here, though, two OpenAI pages do not say the same thing, and it pays to know it. The metrics documentation is blunt: view-through conversions are for reporting only, and cost per acquisition, conversion rate, bidding, billing and optimisation stay click-based. The newer page on conversion-optimised campaigns says that with oCPM the optimisation also accounts for conversions that follow an ad view.

The reasonable reading is that the first page describes the world before oCPM, which is in beta. Either way the operating rule holds: if you buy on clicks, look at click numbers.

If you come from Meta's windows, where the view counts and carries weight, this setup looks timid. For the buyer it is a gift: the numbers the platform optimises on are largely the same ones you can reason with, and you do not spend three months working out why the panel says double what your CRM says.

How the ad you write changes

Flip the order. In the feed you put the hook first and the useful information after, because you have to earn the glance first. In a conversation you already have the glance: open with the useful thing.

Lower your voice. The ad's register has to be able to sit on that page without looking like it arrived from another one.

And take your material from the right place. Real commercial questions, objections, comparisons between alternatives, the criteria people actually decide on: they are in the calls, the emails, the CRM. They are worth more than a translation of your best Meta ad, which was built to stop someone who was not looking for you.

Who should try it now

It is worth it if you sell something people evaluate before buying. Software, professional services, travel, training, furniture, anything that produces the question "which one should I get?". That is exactly what people ask an assistant, with a frankness the search bar never got.

It is worth less if your audience pays for a subscription, for the reason above.

One figure to weigh the decision without leaning on enthusiasm: OpenAI has claimed a billion dollars in annualised revenue in under two hundred days from launch, and says tens of thousands of marketers have now advertised on ChatGPT. They are the company's own numbers, so take them for what they are, but they suggest this is not an experiment that closes at Christmas.

What is worth most today is not the conversion you take home this week. It is understanding how the channel behaves while your competitors are still reading other people's ads.

Frequently asked questions

Is ChatGPT Ads available in Europe?

Yes. The expansion announced on 18 August 2026 covers 31 European markets, including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria, and it went live the following week. Self-service through Ads Manager has been open since 31 August. The country list changes without announcement, so the only reliable reference is OpenAI's Ads Manager availability page on the day you check it.

How much does it cost to start with ChatGPT Ads?

There is a published floor: the minimum daily budget for an account billed in euros is 15 EUR, and it varies by currency. On bidding, for click campaigns OpenAI recommends starting with a maximum of 3-5 USD per click. The rest is your call.

Do you need the pixel to run ChatGPT Ads?

To show ads, no. To optimise, yes: without the pixel or the Conversions API the platform has no event to push towards, and in the panel you only read impressions, clicks, spend, CTR, CPC and CPM. You can still see conversions from outside, because UTM parameters survive the click and land in your analytics and CRM. What you lose is optimisation, not measurement.

Are context hints keywords?

No. They are not exact match and they do not guarantee delivery on a specific question. They describe when your product is useful, not what the person types.

Can I add ChatGPT conversions to my Meta conversions?

No. Different windows, different models. Keep them apart and look at the total of real enquiries reaching you, which is the only number that pays the bills.

Where exactly do the ads appear inside ChatGPT?

Below the answer, as a labelled card separate from the generated text: advertiser name, icon, title, copy, image and landing page. There is no video, no carousel and no full-screen format. If you have a catalogue there is a second format, the product ad template, which pulls image and URL from the feed and can show a price.

Do ads change ChatGPT's answers?

No. OpenAI states that the answer is generated on its own and that advertising does not influence it: the ad sits beside the answer, not inside it. It is the first of the advertising principles OpenAI published, and it holds up commercially too: an answer engine that sells its answers loses the reason people use it.

Who does not see the ads?

Anyone paying for Plus, Pro or a business plan, and accounts identified as under 18. If your audience is made of professionals using ChatGPT all day, a sizeable share are subscribers and this channel does not reach them.

The field guide

This piece explains how the platform thinks. If you have to actually build the thing, the guide at the top is the one: ten pages to keep open while you set up the first campaign.

ChatGPT Ads: the field guide

Ten pages to keep open while you build the first campaign.