What an AI visibility tool reads on your site, and what it keeps

A read-only AI visibility scan reads only what an AI crawler already reads: your home page and a handful of public files on your domain. Biencité keeps the scanned address and the result of the scan, hosts that data in the European Union, and installs nothing on your site. Everything beyond that, your email, an account, a deployment, is a separate decision you make later, and one you can decline entirely.

What does a read-only scan actually read?

A Biencité scan reads the home page of the site plus a few public files at the root of the domain: robots.txt, the sitemap, llms.txt, and the files meant for AI agents. Those are addresses anyone can open in a browser, with no login and no permission. The scanner never enters a customer area, never reads a database, never submits a form. It changes nothing, installs nothing, and does not ask you to create an account.

That scope is deliberately narrow, and the narrowness costs something real: a scan does not see your product pages, your service pages or your articles. Our article "Our scan reads one page of your site. Here is why, and what that costs." sets out what the limit leaves behind, and why we defend it rather than quietly widen it.

Why does that scope match what an AI actually sees?

AI crawlers read the HTML your server returns, and do not execute the page's JavaScript. The analysis Vercel published with MERJ on December 17, 2024, based on traffic across their hosting network, found that none of the major AI crawlers render JavaScript: GPTBot and ClaudeBot do fetch JavaScript files, but never run them, so they cannot read content that is assembled in the browser. Over the month observed, that analysis counted 569 million GPTBot requests and 370 million from Anthropic's crawler.

That measurement dates from December 2024 and covers one hosting network rather than the whole web, so treat it as an order of magnitude rather than a law. The same analysis flags one notable exception, Google, whose Gemini rides on Googlebot's infrastructure and therefore does render pages. For everyone else, reading the raw server response is reading what an AI receives. A tool that scored you on a page rendered in a full browser would hand you a flattering number that AI assistants never see.

What gets stored, and at which point?

How much data is kept depends entirely on where you stop. Three steps, three different footprints, all published in the Biencité privacy policy.

• Scan without an email: the address of the scanned site and the result of the scan, nothing else.
• Scan with an email: your email, the site address, the captured result and your IP address, the last one to curb abuse of volume.
• With an account: your email, your sites and the history of their scans.

None of that data is sold on. You can request access, correction, export or erasure, including for data submitted before any account existed, and the request is handled within thirty days. Deleting an account erases the sites, the scans, the generated files and the citation measurements: no working copy survives somewhere in a backlog.

Where is the data hosted, and who else sees it?

The data is hosted in the European Union, and the list of sub-processors is published rather than summarised. Database and hosting run on Supabase, on an EU project, and on Upstash, in an EU region. Payments run through Stripe. Text generation runs through Anthropic and Mistral AI, which receive only public information from the scanned site, never your account data. Access is partitioned account by account, traffic is encrypted in transit, and API keys stay server side.

A data processing agreement is available on request, and it is published on the site. Its page carries a line most vendors would rather bury: it is a draft, pending review by a lawyer before signature. Better to write that down than to let a buyer discover it at signature. If your legal team requires a reviewed and counter-signed agreement, ask for it before you buy, from Biencité as from anyone else.

What, by contrast, writes to your site?

Deployment is the only stage where write access to your site comes into play, and it is separate from the scan. It runs through a dedicated WordPress plugin, a Shopify integration, a file export you publish yourself on a hand-coded site, or an MCP server for Claude Code. You pick the channel, and the implementation stays under your control.

The question does come up eventually, because a diagnosis changes nothing about your visibility until it is published. Our article "An AI visibility report changes nothing until it is deployed: what happens after the scan" describes what each channel actually writes. The order matters: measure without touching anything, read the result, and only then decide whether to open a door or copy files across by hand.

What should you demand from any tool in this market?

• A read-only trial, with no account to create and no password to your site.
• The named list of sub-processors, and the country where the database lives.
• Deletion that genuinely deletes, generated files and measurements included.
• A written commitment that your data is not sold on.
• A measurement scope stated in plain words, and published signals rather than an opaque score.
• No promise of a citation: a tool measures eligibility, never an outcome.

That list is not a sales pitch, it is a buying grid. Biencité publishes its nineteen signals on its method page for exactly this reason, and deliberately keeps the weight of each one server side, so that a score is compared against itself over time instead of being gamed. Run those six points on Biencité as readily as on its competitors, and walk away from any vendor that will not answer on one of them.

When is this the wrong question to be asking?

If your site is a public brochure, with no customer area and no personal data, the data question is close to empty, and dwelling on it becomes a comfortable way to postpone the real problem. In that case your blocker is not privacy: it is that nothing on the site is quotable, or that nobody is searching for your company name yet. A measurement tool creates neither.

If you work in a regulated sector, with a data protection officer and a formal procurement process, a draft processing agreement will not satisfy your legal team, and that is a fair call. Finally, if your platform locks the root of your domain, some of the fixes will stay out of reach whichever tool you choose: a first budget is then better spent on content than on finer measurement.