In February, OpenAI began testing ads in ChatGPT with a hand-picked group of American advertisers. On 5 May a beta self-serve Ads Manager opened at ads.openai.com, where any US business with a card and a budget could buy its way in, and on 18 August the company announced 31 European markets, live from the 24th, joining the nine countries (including the UK) where the platform already operates. Six months from closed pilot to most of the Western world is fast by any standard, and the speed tells you where the priorities sit: the biggest marketing story of 2026 is a billing system, not a model release.

For the best part of two years, a recommendation from an AI assistant has been the last genuinely free discovery surface in marketing. When ChatGPT named your firm to someone asking for a supplier, that mention cost you nothing — no auction sat behind it, and it never appeared on a media plan. Most of us have been enjoying that traffic without ever calling it a channel. I wrote at the end of 2024 about the shift from a list of links you browse to a single answer you’re handed, and for a while that answer had no commercial layer at all, which made it feel like early Google all over again — a place where good content and a bit of craft could win you visibility that money couldn’t buy. That free run was a phase, though, and every discovery platform of the last twenty-five years has been through the same one.

The introductory offer always ends

Facebook taught the lesson most brutally. In 2012, a post from a brand page reached around 16% of the people who had chosen to follow that page. By February 2014 the figure was 6.5%, and closer to 2% for large pages — a collapse steep enough that Ogilvy titled its research note “Facebook Zero” and advised clients to plan for the day organic reach stopped altogether. The platform had simply matured: the reach it once gave away to attract businesses became inventory to sell back to them, leaving anyone who had built their marketing on free distribution standing on someone else’s land.

Google ran the same arc more slowly and more politely, with ad slots multiplying above the organic results and the labels growing subtler with each redesign, but the direction was identical and so was the reason. Once a discovery surface has a margin to defend, it stops giving discovery away. And when the change came, complaining achieved nothing. The businesses that came through it best were the ones that had treated free reach as a windfall rather than a plan — building their email lists and their ad accounts while their rivals were still counting likes.

OpenAI has been explicit about where its lines sit. Today’s format is a labelled sponsored card that appears beneath ChatGPT’s reply, on the free tiers only; ads are barred from influencing what the assistant actually says. I take those principles as sincerely meant, but Facebook’s early partners heard warm assurances too, and principles have a way of being renegotiated when the numbers demand it. A company committing hundreds of billions to data centres needs revenue on an extraordinary scale, and it has just spent six months building bidding, conversion tracking, custom audiences and geo-targeting — the complete machinery of a serious ad platform. Nobody assembles that machinery planning to keep it small.

AEO splits into two disciplines

Answer engine optimisation has until now been an entirely earned game: structure your content well, earn citations, hope the models pick you up — and, as I argued in June, struggle to measure whether any of it worked. The ads launch splits the discipline in two, much as search split into SEO and PPC a generation ago. Being worth citing remains the only way into the answer itself; what money now buys is the labelled slot beside it. That distinction matters, and it cuts both ways: a paid card can’t make ChatGPT recommend you, and an earned mention can’t be scheduled, budgeted or scaled. If the history of search is any guide, the paid lane will develop faster — it has a vendor behind it with a revenue target and a product team, while the earned lane has to be reverse-engineered from the outside by everyone else.

The martech industry has already noticed. Scott Brinker’s 2026 landscape barely grew for the first time since he began charting it in 2011, yet AEO and GEO tools sit among the few categories still accelerating. When a stagnant market keeps investing in one corner, that corner is where it expects the next budgets to go.

What to do while it’s cheap

Another paid channel is the last thing most small marketing budgets wanted to hear about this year, which is exactly why the sensible first move is small. OpenAI hasn’t published rate cards, and click prices will vary by category. But young ad markets have few advertisers and thin auctions, and platforms hungry for success stories tend to make the early years the cheap years. The way to find out what the channel costs in your category is a capped test, run now rather than next year, when the case studies are published and your competitors have read them. If you’re in the UK you can register directly at ads.openai.com; in the new European markets access starts through agencies and partners, with self-serve to follow. Three things are enough:

  1. Run a small, deliberate test. Pick one high-value intent — the question a ready-to-buy customer brings to an assistant, in the fuller, messier way people describe a problem in conversation — put a modest budget behind it, and set a firm ceiling before you start. The real purchase here is knowledge of how the channel behaves: which intents trigger your ads, what a click costs, and what someone who arrives mid-conversation is worth to you. That knowledge compounds, and when the crowd arrives you’ll have a year of it while they’re starting from the rate card.
  2. Instrument from day one. Install the pixel or the Conversions API before you scale anything, and decide which single conversion actually matters, because a test you can’t measure teaches you nothing at any price. Keep tracking your earned visibility alongside the paid (the share-of-model checks I described in June still apply), so you can tell whether paid placement is winning you new ground or merely replacing recommendations you were already getting for free.
  3. Keep the earned lane running. Paid placement doesn’t make citability obsolete, any more than PPC made SEO obsolete. The businesses that did best out of paid search were largely those with strong organic foundations underneath, because the two feed each other — and your content still determines what the model says about you in every conversation nobody has paid for.

Will ads in a conversation even work?

Whether ads inside a conversation actually work is still an open question; the format is simply too new to say. The case for is obvious: nowhere else can an advertiser reach someone mid-decision, moments after they’ve explained their situation, their constraints and their budget in their own words. The case against is just as real — a conversation is a more intimate space than a results page, and an ad that lands clumsily there reads as an interruption by a stranger.

It’s also possible OpenAI holds its lines better than its predecessors did, the paid tiers keep growing, and the ad load stays light for years, in which case the free window stays open longer than the Facebook precedent suggests. I don’t know which way it breaks — but finding out costs a few hundred pounds today, and it will cost a great deal more once the results are common knowledge.

Free AI recommendations haven’t stopped, and some businesses will spend the next two years insisting they never will. Everything we know about how these platforms mature says otherwise. The bill is being printed, and the cheapest time to learn a paid channel is before it arrives at your table.