On 23 July, Axios detailed a deal between Yelp and OpenAI: the American directory licenses 330 million cumulative reviews and more than 8 million business listings to ChatGPT. Seen from Paris, the reflex is to shrug — Yelp is irrelevant in France. That is true, and it is precisely why the story is worth stopping on. What it reveals is not an American news item but a method: assistants do not explore your street shop by shop, they plug in suppliers of local data. And in France, a supplier has already signed. Since 22 June, PagesJaunes has been available as an app inside ChatGPT.
What the Yelp deal actually says
On the substance, the deal is ordinary: a content publisher sells access to its database. Yelp already did this for Apple Maps and for Yahoo. Financial terms were not disclosed, and there is no exclusivity — Yelp remains free to sign with other AI companies. What changes is the visible trade-off for the user: when OpenAI uses that content, Yelp branding and links accompany the answer, even though OpenAI controls how it is presented.
The detail that matters sits elsewhere. Three weeks later, on 10 August, Yelp pushed its reservations and waitlists into the conversation for the US and Canada; Resy did the same that day, and OpenTable, whose coverage is global, makes the feature usable from France. In other words, in three weeks we moved from "AI reads a database" to "AI acts inside a database". That is the direct extension of what I described on the Google side in agentic booking inside AI Mode: the answer no longer merely cites, it executes.
One note of caution, because it is instructive. A spot check run in San Francisco on 4 August, across ten local queries, surfaced no visible Yelp citations at all: Angi, Thumbtack and Tripadvisor were the ones showing up. Signing a deal and actually being cited in answers are two distinct stages, separated by weeks or more. Keep that in mind: announcing a partnership tells you nothing about the real weight of that source in the answers your customers see.
Three doors, not one
People talk about AI crawlers as if they were the only route by which a machine learns a business exists. It is the most visible door, it is not the only one, and on local it is probably not the main one. There are three, and they do not follow the same rules.
| Door | How it works | Dated example | Your room to act |
|---|---|---|---|
| Data licensing | The AI company pays for wholesale access to an entire database | Yelp → OpenAI (23 July 2026) | Indirect: be present and accurate at the supplier |
| Connected app | A partner exposes its data; the model calls it when the question warrants | PagesJaunes in ChatGPT (22 June 2026), Tripadvisor | Indirect: the quality of your listing at that partner |
| Web crawling | The bot reads accessible public pages | Your site, local press, blogs, guides | Direct: the only door you own |
Two honest caveats on the second row. An app inside ChatGPT is not invoked on every question: it is a connector the model may call, not a mandatory step. And its availability does not guarantee that your listing in particular will be selected. What it does guarantee is that a missing or wrong listing can never be used.
The practical consequence fits in one sentence: you control one door out of three, but you feed all three. The first two run on data you entered somewhere, once, and may well have stopped updating.
Who has signed, market by market
This is the useful question, and its answer is national rather than global. Here is the picture in France as of early September 2026, separating what has been officially announced from what is observation.
| Supplier | Status inside assistants | Real weight in France |
|---|---|---|
| Buys nothing: owns Maps, profiles and reviews | Dominant, feeds AI Overviews, AI Mode and Gemini | |
| PagesJaunes (Solocal) | App inside ChatGPT since 22 June 2026 | General-purpose database, all trades, national coverage |
| Tripadvisor | App inside ChatGPT since late 2025 | Restaurants, hotels, tourism |
| OpenTable | Booking inside ChatGPT since 10 August 2026, global coverage | Restaurants, minority share against local players |
| TheFork | No official app announced to date | Booking leader in France, large web footprint |
| Yelp, Resy | Deals signed with OpenAI, features limited to the US and Canada | Close to zero |
Two lessons come out of that table. First: the supplier map is national, not global. Copying the American list means optimising for directories your customers will never open. Second, and more uncomfortable: the player that signed in France for the general-purpose database is the one many shop owners had decided, for perfectly good commercial reasons, to walk away from. Solocal has announced €130 million of investment over four years in AI and its local database. You can enjoy the irony; the listing is still free, and the subscription question is a separate matter.
Do not confuse the listing with the subscription. What an assistant consumes is listing data: name, address, phone, hours, category, reviews, photos. It exists in the free tier. A subscription buys promotion inside the publisher's own product, not a citation in an AI answer — and no publisher can seriously promise you the latter. Check the listing; judge the subscription on its own commercial merits.
Why directories outrank your site in citations
Two pieces of 2026 research measure the same thing from different angles and land in the same place.
The first, by Acromatico, ran 100 "best [vertical] in [city]" queries across 10 verticals and 25 US cities, logging the sources the engine actually used. The result: directory and ranking sites are cited in 78% of answers, and the most frequent domains are directories, "best of" lists and review sites — never the recommended business's own website.
The second, published by Formative Digital in May 2026, ran 176 queries across 9 Ontario cities, 5 trades and 4 assistants (ChatGPT, Claude, Gemini, Perplexity), for 1,732 citations spread over 326 domains. Two numbers stand out. 83.7% of sources were cited by a single engine only: four assistants, four different webs. And individual business websites appear once, or not at all.
That last point deserves to be said plainly, because it contradicts a very common sales pitch. There is no single ranking to conquer, therefore no single method to sell. What the four engines share is not the list of sources: it is the type of source they favour. Aggregators. Pages other people write about you.
Necessary caveat: both studies are North American and vertical-specific. The domains they surface have no audience in Europe. What transfers is the structure of the behaviour, not the list.
And yet your site is still the top factor
An attentive reader should baulk here. A few days ago I wrote about an audit of 4,776 venues whose central finding was that owning your own website is the strongest factor for appearing in AI recommendations (×1.92), ahead of every reputation signal. And today I am writing that business websites are almost never cited as sources. Both are true, and the contradiction is only apparent.
It dissolves as soon as you separate two things everyday language conflates:
- Being named in the answer: your business is one of the three or five the assistant mentions. That is what the 4,776-venue audit measured.
- Being the source displayed under the answer: the URL a user can click to verify. That is what the citation studies measure.
Your site works on the first plane, not the second. It makes your facts verifiable, consistent and cross-checkable: an address, opening hours, services, prices, a menu in text. That material is what makes you eligible, and it is also, very concretely, what directories copy and what local guides quote. The full mechanism is in what an AI citation actually is.
Two planes, one thing in common: your facts. The directory is the source the machine displays; your site is the source everyone copies from, the directory included. You do not choose to be the cited URL. You choose whether the facts inside the cited sources are correct — and that decision is made on your site, then propagates.
Immediate corollary: an inconsistency between your site, your Google profile and your national directory listing does not merely confuse one engine. It replicates across three doors at once. That is the whole point of NAP consistency, a topic once filed under local SEO housekeeping that has become a condition of existing.
What I take from this for a local business
One. "How do I show up in ChatGPT?" is the wrong question. The right one is: where does ChatGPT get its local information in my country and my trade? It has verifiable answers, it changes every six months, and you can check it in a handful of queries. I listed the usual causes of absence in why your business doesn't appear in ChatGPT; the Yelp deal adds one nobody suspected — being absent from the database the AI company just bought.
Two. The work that pays is neither glamorous nor new. It is accuracy, repeated across a small number of sources that matter. Three correct listings beat thirty registrations. And accuracy compounds: it serves all three doors at once, whereas a trick specific to one engine serves a quarter of the market, for six months.
Three. Google is a separate case and should be treated as one. It does not buy local data: it owns it. Your business profile and your reviews feed AI Overviews, AI Mode and what Gemini assembles inside Maps, as I detailed about Ask Maps. A flawless Google profile does not make you exist at its competitors. These are two projects, not one.
The checklist, in order
- Inventory your real suppliers. Five questions your customers would genuinely ask, run on ChatGPT, Gemini, Perplexity and Copilot. Record the domains shown as sources, not the names mentioned. That gives you your list — the one for your trade and your city.
- Fix consistency before adding anything. Exact name, identically formatted address, same number, same hours everywhere. A discrepancy costs more than a missing listing: it creates a doubt the machine resolves by leaving you out.
- Claim and complete your national directory listing. In France that means PagesJaunes — free, a few minutes, and now wired into ChatGPT. Without subscribing to anything.
- Handle the sector platform that matters in your trade. Restaurants: the booking platform you already use, plus OpenTable if your clientele is international, since that is what books inside ChatGPT from France. Hotels and tourism: Tripadvisor.
- Make your site the canonical source. Hours, services, prices and menu in text — no PDF, no image, no "contact us for pricing". That is what directories copy and what guides quote.
- Check that AI bots can read your site. Five minutes in your
robots.txt. An unreadable site cancels the previous point. - Measure, then repeat quarterly. The Search Console report on generative surfaces gives you a numeric baseline: how to measure your AI visibility. Licensing deals change; so does your supplier list.
Frequently asked questions
Do I need a PagesJaunes listing to show up in ChatGPT?
It is not a requirement, but it is an entry point you should not leave empty if you trade in France. Since 22 June 2026, PagesJaunes has been available as an app inside ChatGPT: when a question calls for it, the assistant can query that database and pull your contact details, hours, reviews and photos. If your listing is missing or wrong, you lose on that channel without ever seeing it. The basic listing is free and takes minutes to claim. What I would not do is confuse that check with buying a visibility subscription: that is a separate commercial decision, not a technical condition. The same logic applies in any country to whichever aggregator has signed there.
If directories get cited and my site doesn't, what is my site for?
Being named in an answer and being the URL shown as a source are two different things. Citation studies measure the second, and directories dominate there: 78% of answers in the Acromatico study. The audit published in August 2026 measured the first, and owning your own website came out as the strongest factor for appearing at all, ahead of every reputation signal. The two complement each other: your site makes your facts verifiable and cross-checkable, which gets you into the shortlist; the directory supplies the displayable URL. Without a site, you depend entirely on what third parties write about you. The full reasoning is in why a restaurant still needs a website.
Those studies are North American. Do they apply elsewhere?
The domains, no. The mechanism, yes. Acromatico covers 25 US cities and 10 verticals; Formative Digital covers 9 Ontario cities, 5 trades and 4 engines. The names that surface have no audience in Europe. What transfers is the structure: assistants lean heavily on aggregators rather than on businesses' own websites, and each engine reads a largely different set of sources. In any other market the exercise has to be redone with the local aggregators — Google, the national general directory, and the sector platforms of your trade.
Do I need to pay for a directory subscription to be visible in AI?
No, and this is the main misreading the news invites. What an assistant consumes is listing data: name, address, phone, hours, category, reviews, photos. It exists in the free tier of the major databases. A subscription buys promotion inside the publisher's own product, not a citation in an AI answer, and no publisher can seriously promise you the latter. The useful rule runs the other way: three accurate listings on sources that matter in your market beat thirty paid entries.
Does Google buy local data too?
Google does not need to buy it: it owns it. Maps, business profiles and reviews form the most complete local dataset on the market, and that dataset feeds AI Overviews, AI Mode and Gemini's answers. Hence an asymmetry that is often misread: work on your Google profile serves the Google ecosystem directly, while your visibility inside ChatGPT runs through third-party suppliers Google does not control. You have to work both, and never assume a flawless Google profile is enough to exist elsewhere.
How do I find out which sources speak for my business?
By running the questions yourself, across several assistants, and recording the sources shown rather than the names mentioned. Take five phrasings your customers would actually use, with the city and the need, and run them on ChatGPT, Gemini, Perplexity and Copilot. The domains that recur are your de facto suppliers; check your listing on each. On the Google side, the Search Console report covering generative surfaces gives you a measured starting point. Repeat every quarter: licensing deals change, and the supplier list changes with them. That is the core of GEO.
If you want to know which suppliers your business depends on today, get in touch: I run the queries for your trade and your city, record the sources actually cited, and tell you which one is wrong or empty about you. That is the starting point of my work on AI search optimisation, and when the canonical source is missing, it starts with a website that finally gives machines facts to copy. For a restaurant or café, the restaurant websites page lays out the method.