The barrier is gone. OpenAI's self-serve ads manager dropped its $50,000 pilot minimum, moved to cost-per-click billing, and any business in an open market can now run ads inside ChatGPT. So the question is no longer whether you can — it's whether you should. In July 2026 someone finally measured it, across 50,006 commercial prompts. The headline number is this: 96% of the time, the advertiser paying for the sponsored card is not cited anywhere in the answer sitting directly above it. Paying and being recommended are two different things, and only one of them is for sale.
Where these ads actually run
Start with the map, because a lot of what circulates on this topic manufactures urgency that doesn't exist yet.
| Surface | Ads? | Since |
|---|---|---|
| ChatGPT — United States | Yes | February 2026 |
| ChatGPT — United Kingdom | Yes | June 2026 |
| ChatGPT — Canada, Australia, New Zealand | Yes | 2026 |
| ChatGPT — France, Germany, Ireland, Singapore | Not yet | Ad teams hiring, no official date |
| Google AI Mode — United States | Yes (Shopping formats) | February 2026 |
| Google AI Overviews & AI Mode — France | No | Explicitly excluded from the 22 July 2026 launch |
The format is the same everywhere: a single "Sponsored" card, placed underneath the generated answer. No block of four competing ads the way a Google results page opens. One slot, one advertiser, below the text the user actually came to read. Hold on to that word — below. It carries most of the argument.
What the first 50,000 prompts show
In July 2026, SE Ranking analysed 50,006 commercial prompts across 20 niches in the US. As far as I know it is the first serious measurement of this new inventory, and it deserves reading past the headline.
Volume first. 25.94% of commercial prompts trigger an ad in ChatGPT. Google AI Mode sits at a comparable 29.45%. Roughly one commercial query in four — not marginal at all, and confirmation that the channel is real.
Relevance next, and this is where it gets uncomfortable. 14.35% of ads had no meaningful topical connection to the prompt. The average hides brutal spreads: 2.6% mismatch in pets, north of 50% in the more conversational categories. Targeting doesn't run on exact keywords but on plain-language context hints — more flexible, considerably less precise.
Concentration last, and it is the figure almost nobody quotes. Just 1,159 advertisers share the entire inventory. One financial comparison site accounts for 13.56% of all ads observed. In travel, Booking.com holds 62.57% of placements; in food, a single player takes 79.92%. This isn't an advertising market yet — it's a handful of national buyers absorbing the space. A local business isn't outbid there. It isn't in the room.
The number that should stop you
Here is the central finding, and the one I have seen reported least.
When an advertiser appears in the sponsored card under an answer, it is cited as a source of that answer 3.63% of the time. Which means 96.37% of the time, the text the user just read makes no mention of the brand that paid to sit right beneath it. The exact promoted URL shows up in the citations in 0.09% of cases. The brand is mentioned somewhere in the answer in 4.44%.
Translate that for a business owner. Your customer asks a question, reads an answer that recommends three places, then sees a labelled ad below it. Those three recommendations and that ad do not come from the same place. The recommendation is what the model retained from the web. The ad is a slot that was sold.
Why this isn't a teething problem. It would be tempting to read it as a setting that will improve. It's structural. The answer is assembled by a source-selection process — the one I describe in how a single question is split into eight to twelve parallel searches. The ad is allocated by an auction. Two separate pipes that don't talk to each other. Buying into the second has no reason to place you in the first, and it would be a serious trust problem if it did. AI advertising is not a shortcut to citation: it is a parallel channel that never touches it.
It is the same distinction I drew about agentic booking: there are several doors, they don't open with the same key, and none of them excuses the others. Getting cited by an AI is earned; a slot is bought; and the purchase does not produce the citation.
Three more reasons, if you're a local business
Beyond that number, three properties of the channel make it a poor fit for a neighbourhood business today. These aren't opinions — they're documented product limits.
Geo targeting stops at the country. This is the most disqualifying of the three by a wide margin. At launch, the only geographic granularity available was national. A bakery serving four streets would be paying for clicks from the other end of the country. On Google Ads you set a three-kilometre radius around your door; here you buy the whole map. The cost per click — a few dollars — isn't the problem. The problem is that the overwhelming majority of those clicks cannot become a customer.
Your most valuable users don't see ads. Ads are served to free and entry-tier accounts. Higher paid subscribers see none. The inventory therefore excludes, by design, a share of the most intensive users of the tool — precisely the people who delegate the most decisions to their assistant.
The reporting is blind. Advertisers get no visibility into the prompts or conversations that triggered their ads. You see impressions, clicks, spend. You don't see on what. Diagnosing bad targeting becomes impossible, and optimising without a diagnosis is an act of faith. It's a remarkable step backwards from thirty years of search advertising habits.
One last data point to settle the picture: the team behind the study also bought its own campaign as a test. Over 97,000 impressions, 1,263 clicks — a 1.30% click-through rate — and almost no sign-ups at the end of it. A young channel, still rough, that hasn't found its conversion mechanics yet.
What this does not mean
I'm not telling you advertising in AI answers is stupid. That would be as lazy as the opposite claim.
It makes sense for national players with a broad catalogue and high customer value, able to absorb coarse targeting because their catchment area is the entire country. That is exactly the profile of the advertisers holding the inventory today. And it will make sense for far more businesses the day local targeting exists — which it will, because that's where the budgets that keep an ad network alive actually come from.
My claim is narrower and easier to check: in 2026, for a shop or a restaurant, this channel cannot be targeted at neighbourhood scale, cannot be diagnosed, and cannot produce the thing you're actually missing — being the answer rather than the box underneath it. When local targeting ships, the right question won't be "should I advertise?" but "can I target my city, and what does it cost me per customer who actually walked in?". Two questions you answer with your own numbers, not with an article.
What I'd do with the time this buys you
Whether ads are already live in your market or still months away, the work below is the part that isn't for sale — and it stays valuable either way.
Measure where you stand first. The Generative AI performance report in Search Console rolled out to most properties on 11 August. It gives you your impressions inside AI Overviews and AI Mode — not your clicks, and that limit matters before you draw conclusions from it: I've covered what the report tells you and what it hides. Open it now, even if the number is unflattering. Without a starting point you'll never know whether anything worked.
Work on what decides citations. Exact primary category on your Google Business Profile, correct hours including holiday closures, recent and detailed reviews, identical contact details everywhere. Same fundamentals as two years ago — they haven't changed, they've just become considerably more profitable. That's what I lay out in the AI Overviews checklist for local businesses.
Take back control of your answers. A questions page on your own site, written in natural language, marked up as FAQPage, covering the questions customers actually ask on the phone. It is now one of the few places where you still write, yourself, the words the engines will repeat on your behalf — and it costs nothing.
Check that AI can read you at all. A slow, JavaScript-heavy or thin site gives nothing to quote. If you don't know where you stand on that, why a local business doesn't show up in ChatGPT is a good starting diagnosis.
And write your budget rule while you're calm. Decide today what you would accept paying for a customer who genuinely walks through your door. Once the ad network opens in your market you will be sold to, and you will be sold to with urgency. A rule written before the pressure beats an arbitration made during it.
The checklist
- Don't budget for AI advertising on the strength of the hype. Check first whether the channel is even open in your market, and whether it can target your catchment area.
- Open the Generative AI performance report in Search Console and write down this month's impressions. That's your baseline.
- Fix your primary category on your Google Business Profile — the most specific one that genuinely describes what you do.
- Update your hours, holiday closures included, and get in the habit of entering them ahead of time.
- Keep reviews fresh so you always have detailed feedback from the last six months.
- Publish a marked-up FAQ on your site using your customers' real questions, in their words.
- Write your cost-per-acquired-customer rule before anyone calls you, and keep it.
Frequently asked questions
Where are ChatGPT ads actually available?
In the US since February 2026, the UK since June 2026, and in Canada, Australia and New Zealand. France, Germany, Ireland and Singapore are in preparation with no official date. Google, separately, excluded advertising entirely from the French rollout of AI Overviews and AI Mode on 22 July 2026, while Shopping formats have run inside US AI Mode since February.
Does buying an ad get you cited in the answer?
No. Across 50,006 prompts, the advertiser appeared as a source of the answer displayed above its own ad in 3.63% of placements, and the promoted URL in 0.09%. The answer and the sponsored card come from two separate systems: one selects sources, the other sells a slot. Buying the second does not put you in the first.
Can I target my city or my neighbourhood?
Not today. Geo targeting in OpenAI's ads manager stops at country level. For a local business that's more disqualifying than the price: you'd pay for clicks from everywhere with no way to restrict delivery to your catchment area. It's the first thing to re-check whenever the product updates.
Who actually sees these ads?
Free and entry-tier users. Higher paid subscribers see none. The inventory structurally excludes a share of the most intensive users of the tool — often the ones delegating the most purchase decisions to their assistant.
What if my competitors start advertising before me?
Look at who actually holds the space: 1,159 advertisers for the entire US inventory, one of them taking 13.56% of all ads and another 62.57% of the travel category. Those aren't the businesses down your street — they're national players. Your real competitive risk isn't a rival buying an ad. It's a rival being cited for free inside the answer while you aren't there at all.
Want to know whether AI currently cites you, and what's missing if it doesn't? Drop me a line: I'll look at your profile, your site and your AI impressions, and tell you what to fix first — without selling you an ad budget on a channel that may not even be open where you are. That's the core of my work on AI search optimization, extended by AI for local business and, when the foundations are missing, by a site that finally gives machines something to read.