Barometer · October 2026 edition

Are French websites open to AI?

On October 18, 2026, we read the directives that the top 1,000 .fr sites give to AI assistant crawlers. The result: 89.1% of sites let ChatGPT, Claude and Perplexity in. The remaining 10.9% close the door on at least one of them, in whole or in part — and among them are France's biggest news publications.

Assistant by assistant

Share of the 919 sites whose directive is known. "Blocked": the assistant's search crawler is denied access to the homepage, and the assistant loses access to the site to find and cite it. "Partial": search is allowed, but reading on a user's request is denied.

Perplexity is the most frequently blocked assistant. Google and Copilot go through search engine crawlers, which sites keep open almost everywhere: more than 98% welcome them. Only four sites block all six assistants from their homepage.

Training or citing: two distinct doors

AI companies send out two kinds of crawlers. Some collect pages to train the models; others read a site to answer a question and cite it. A robots.txt file can treat each kind separately, and sites do just that.

The telling figure: 130 sites block GPTBot, the crawler that trains OpenAI's models. Among them, 79 — six out of ten — leave OAI-SearchBot open, the one that powers ChatGPT's search. These sites choose to be cited, while keeping their texts out of training.

The press each draws its own line

The media form the most divided group. Every major publication has picked its assistants, and the choices differ from one newsroom to the next.

Le Monde shuts out Claude and keeps ChatGPT; Radio France does the opposite. The magazines Vogue, GQ and Vanity Fair also keep ChatGPT and shut out Claude. TF1 and LCI open search to all three assistants and close on-demand reading. Le Chat by Mistral, which publishes a single robot, the one for on-demand reading, remains outside this table.

Commerce and public services: open doors

Leboncoin, Vinted, Leroy Merlin, Carrefour, Decathlon, Free, Orange, SFR, La Poste and Doctolib welcome the six assistants. Public services too: Service-public.fr, impots.gouv.fr, ameli.fr, the CAF, France Travail, Légifrance, the CNIL, the BnF, the CNRS, the Senate, the National Assembly and the Élysée. Amazon is the exception: its search remains open to Google and Copilot, and closed to ChatGPT, Claude and Perplexity.

llms.txt: 5.5% of sites

The llms.txt file introduces a site to assistants, much like a table of contents. 51 of the 924 reachable sites publish one, including Crédit Agricole, ManoMano, Veepee, Capital and Trivago. Ouest-France and La Voix du Nord publish one as well, while shutting their door to the robots of ChatGPT, Claude and Perplexity: the table of contents is ready, and the instruction guards the entrance.

The method

• The sites: the top 1,000 .fr domains from the Tranco list (list 8PX8V), the public reference ranking used in research work.
• The reading: on 18 October 2026, for each site, the homepage, the robots.txt and the llms.txt, once each, with an ordinary browser agent. When the bare domain name leads to the site via www., it is the www. address that is read.
• The verdict: the robots.txt, read according to the RFC 9309 standard, for every robot known to the six assistants, on the page where the homepage lands after its redirects. This is the instruction the site writes itself, the most reliable evidence. A firewall can add its own rules: our free test measures them site by site.
• The counts: 924 sites reachable over HTTPS out of 1,000; the other 76 have no site or an invalid certificate. Percentages are based on the 919 sites whose instruction is known (robots.txt read, or absent, which opens everything); 5 sites returned a server error in place of their robots.txt.
• The next edition: January 2027, same method, to track what changes.

Read next

• Getting cited by AI assistants →
• All guides →