How a real estate agency gets found inside ChatGPT

These are two separate problems, and confusing them is the origin of nearly all the bad advice circulating on this subject since February 2026. The rest of this page untangles them, names its sources with their dates, and says at each step what is not documented.

The closest source to the issuer that we could reach is the page "Recherche immobilière avec ChatGPT" published by leboncoin on its own professional site, which credits at the foot "Communiqué de presse leboncoin, 10 février 2026". The press release itself is not reachable. The group's corporate site, which now redirects to leboncoincorporate.com, carries no news item on the subject as of 6 October 2026. So we are working from a leboncoin publication summarising its own release, plus the trade press.

The figures come from that trade press. Immo Matin, on 18 February 2026, under the byline of Christian Capitaine, gives the volume of around one million real estate listings and leboncoin's own formulation, the first real estate portal to offer an app in the ChatGPT App Store. Immobilier 2.0, on 10 February 2026 under the byline of Vincent Lecamus, gives the 89 million listings across all categories. MySweetImmo covered the launch the same day.

On what this changes for a professional, the two most useful quotes are in Immo Matin of 18 February 2026. Sophie Bourg, marketing director, describes the app as a new acquisition channel and a new traffic crossroads towards professionals. Julien Jouhault, chief technology officer, says he hopes for millions of additional visits. Those are statements of intent, made by the party with the most to gain, and they should be read as such.

On the mechanics, the most concrete piece is an interview published by Immobil-IA on 4 March 2026, in which Sophie Bourg says that information quality drives performance and that a listing that is complete, precise and properly filled in on its location and technical criteria will have a better chance of being visible and relevant. She adds that AI can help write or enrich content, but that the underlying data remains the deciding factor. That is, as far as we can tell, the only public indication of what counts, and it comes from the portal itself.

Technically, leboncoin has not published its app's architecture. What is documented, on OpenAI's side, is the general framework: the Apps SDK documentation describes publishing "plugins with skills, MCP servers, and optional UI", meaning servers that expose tools to the assistant. If that layer matters for your own systems, we cover it in our guide to the Model Context Protocol in the enterprise.

The detail matters, because the difference between a conversational search on the portal's own site and an app inside ChatGPT is not cosmetic. In the first case the user is still at the portal. In the second, they never leave their assistant.

Two notes on Omny, because it is the most interesting case for an agency. The company was founded by Franck Le Tendre, former VP Operations of the SeLoger and Meilleurs Agents group, and tells MySweetImmo on 7 September 2026 that it holds close to a million professional listings, new-build and existing, for sale and to rent. Its flow runs all the way to booking an appointment with a professional without leaving ChatGPT.

And on 29 September 2026, again according to MySweetImmo, iad connected its advisers' listings to Omny. Franck Le Tendre puts the thesis of this whole movement better than any vendor: the challenge for professionals is no longer only to make their listings visible, but to make sure their properties are understood and recommended by artificial intelligence. Hold on to the word "understood". It describes a data problem, and a data problem can be worked on.

This is the point the press coverage was read past. No reachable source, leboncoin's own publications included, states that the agency name is displayed in the app's results, nor how a listing is picked out of a million. The Immobilier 2.0 article of 10 February 2026 addresses neither displayed seller identity, nor ranking, nor the private-versus-professional distinction. So we do not know, and we say so.

The practical consequence is unwelcome for anyone looking for a shortcut. Polishing your leboncoin listings serves the first column and only the first column. And a property found that way reaches you through the portal, with one more layer of intermediation. We devoted a whole guide to that dependence and to building an asset that belongs to you: why an estate agency should own its data.

The exact wording is on Google Search Central's "AI features and your website" page, last updated 10 December 2025: "Both AI Overviews and AI Mode may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources, to develop a response." Those are Google's own words, in the documentation it addresses to site owners.

For an agency this translates very concretely. Somebody asking whether now is a good time to sell their flat in Nantes triggers searches on selling times, on observed prices, on tax treatment, on the choice of mandate. A page that properly answers one of those sub-questions, with its figures and its edge cases, has a chance of being read. A page titled "Estate agency Nantes" made of three generic paragraphs has none, because it answers none of the sub-questions.

On the ChatGPT side, the documented lever sits elsewhere: in the crawlers. OpenAI publishes the list of its bots and what each is for. OAI-SearchBot is, per that documentation, the one "used to surface websites in search results in ChatGPT's search features". GPTBot is the one "used to crawl content that may be used in training our generative AI foundation models". Two distinct crawlers, two distinct purposes.

Between 2023 and 2025 many sites blocked anything that looked like an AI crawler, often with one broad rule set at CDN level. The result is that they also shut the door on the crawler that displays them. If you want to be found by ChatGPT, the first check costs nothing: your robots.txt and the bot rules at your host.

1. Complete listings at the portal. This is the only documented parameter. Sophie Bourg, in Immobil-IA of 4 March 2026, stresses location and technical criteria: floor area, number of rooms, heating type, condition. A listing whose description runs to two lines and whose fields are half empty cannot be "understood" by a model that has only that text to read. Fill the input fields the way you would fill a database.

2. An agency page readable as text. Trading name, registered name, address, telephone, professional licence number, areas covered, team. As text in the HTML, not inside an image, and not only inside a component that appears after a click. A crawler that cannot read your address cannot place your agency, and a model that does not know where you are will not put you forward for a local question.

3. A name and address identical everywhere. Your site, Google Business Profile, trade directories, the French company register, the portals, social accounts. Google's documentation on local business structured data, updated 8 September 2026, lists only two required properties: name and address. That minimalism is instructive: if those two facts vary from one corner of the web to another, you are not one clear entity with a few variants, you are several weak ones.

4. Content that answers the questions people really ask an assistant. How long it takes to sell in a given town, which documents a preliminary sale agreement needs, who pays the agency fee, how long a given survey stays valid. Long, conversational, local questions. That is exactly what query fan-out goes looking for, and it is the kind of page where a small agency can beat a national portal, because it knows its market while the portal is speaking about the whole country.

5. Being present in the sources assistants already read. The portals, the local press, public registers, professional directories, your partners' sites. One mention in a local news article carries more weight than one more page on your own site, for the simple reason that you did not write it. No platform manufactures those mentions, and the offers that claim to produce citations in directories nobody reads.

6. Not blocking the crawlers that feed the answers. Check that OAI-SearchBot is not disallowed in your robots.txt, and that your host or CDN does not reject it upstream. A ten-minute check that, if it fails, cancels out all the other work.

7. An assistant over your own data, for questions about your stock. A visitor who asks your site whether you have anything under 300,000 euros with outdoor space should get an answer drawn from your real portfolio. That means connecting an assistant to your data rather than to fixed text, which we set out on our Claude integration for the enterprise page and, on the commercial side, in the intelligent CRM for real estate.

The sentence sits on the same 10 December 2025 page as query fan-out: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", followed by "There's also no special schema.org structured data that you need to add." It is worth keeping that quote to hand: it settles most of the commercial pitches circulating about markup for AI engines.

That does not make markup useless. It has another, more modest and quite real job. RealEstateAgent, a subtype of LocalBusiness at schema.org, lets you declare your agency's name, address, telephone, service area and opening hours without ambiguity. Google documents the local business feature and its two required properties. Add that markup because it makes your entity legible, and expect its effect to be that of a properly filled-in identity card.

Two traps to avoid. Never mark up an average rating or reviews you have not genuinely collected: that is a false statement, and it is the kind of mistake that costs for a long time. And do not confuse markup with a declarative file dropped at the root of the site: no Google or OpenAI documentation makes one a condition for appearing.

There is no position to measure. Two close phrasings of the same question can return different sources, for the same user, on the same day. A report announcing your average position in ChatGPT is describing the average of a prompt sample chosen by whoever sells the report. Track verifiable things instead: are your listings complete, is your name identical everywhere, are your crawlers allowed, do your pages answer a whole question.

The selection mechanism is not public. This is the most important point on this page. Nobody has published how a listing is picked out of the million in leboncoin's real estate catalogue. The only public indications come from the portal and concern how complete the data is. Everything beyond that is a guess, including when it is stated with confidence.

No figure measures the channel yet. As of 6 October 2026 we found no publication giving the number of contacts generated for agencies by the leboncoin app inside ChatGPT. The hope of millions of additional visits voiced by leboncoin's CTO in February 2026 is an intention, and eight months later it has not been quantified publicly. Do not commit budget against a figure that does not exist.

The channel deepens your dependence. A buyer who finds your property through a portal's app reaches you through that portal, on its terms, with one more intermediary in the chain. Sophie Bourg says it herself, accurately: a new acquisition channel, a new traffic crossroads. A crossroads belongs to somebody. This work is worth doing, and it does not replace building a contact file and an audience that belong to you.

And the least pleasant limit. If your agency is mentioned nowhere but on its own site, no amount of optimisation will manufacture the missing material. A model recommends what it has read something about. The lever then sits outside your site: the local press, partners, public registers, real reviews from your clients. That is slower and less saleable than a technical audit, and it is what works.

• Step 1, check the crawlers. Your robots.txt and your CDN rules do not disallow OAI-SearchBot. Ten minutes, binary effect.
• Step 2, align name and address. List every place your agency is named on the web and make those two fields identical. Half a day, and it is the foundation for everything else.
• Step 3, audit ten listings at random. Count the empty fields and the descriptions under five lines. That is your measurable baseline on the only documented parameter.
• Step 4, rebuild the agency page. All the identity information as text, the RealEstateAgent markup, the areas covered named one by one.
• Step 5, write three local question pages. One question per page, answered fully, with your figures and your edge cases. Not three keyword pages.
• Step 6, go and earn two outside mentions. Local press, a trade association, a partner. The slowest step and the one that decides the most.
• Step 7, and only then, add tooling. An assistant wired to your real portfolio, once the six steps above hold.

If you are looking for where this work sits among your other AI projects, our sector guide to AI in real estate gives the overview, the five automations that hold up in an agency cover the internal work, and our comparison of AI providers for real estate says what each one charges. Rental management is handled separately in automating rental management. And if the question you actually have is where everyone else stands, what the AI adoption studies in real estate really measure takes the circulating figures apart.