We scanned 15 AI companies. Can AI agents actually use their websites?

If anyone should have agent-accessible websites, it is the companies building AI agents. They write the papers, ship the protocols and sell the assistants that browse the web on your behalf. So in July 2026 we scanned 15 of the most prominent AI companies with AgentVisible and asked the same question we ask of any site: can an AI agent discover this website, read it cleanly, and find out what it is allowed to do?

Mostly, the answer is no.

The results

Every score below links to a live, evidence-backed profile. Each check shows what was found, where, and when. Scans used scanner v0.3.0; the methodology is public.

Two companies, Synthesia and Lovable, were scanned but their inferred site classifications are under review, so we are not publishing numbers for them until we stand behind the classification. They will appear in the next monthly cycle.

Three front doors, closed

OpenAI, Perplexity and Stability AI could not be scanned at all: their sites refuse a connection from an identified automated agent. We report that as exactly what it is, unreachable, rather than pretending it is a zero.

It is worth sitting with the irony for a moment. OpenAI and Perplexity both sell products whose whole job is to visit websites on a person's behalf. An agent that identifies itself honestly, the way good bots have done since robots.txt was invented, cannot reach their homepages. The agentic web these companies are building only works if sites let agents in; theirs do not.

The MCP company with no MCP card

Anthropic created the Model Context Protocol, the standard that lets agents discover and use tools. Their own site publishes no MCP server card, no llms.txt, no agent-readable version of its content. It scores 19 out of 100. LangChain, the company whose framework popularised agent tooling, scores 10.

None of this means these companies' products are bad. It means their websites, like almost everyone's, were built for human eyes only. The point of measuring it is that these are the organisations best placed to lead by example, and the fixes are small.

What good looks like

ElevenLabs tops the cohort at 80 with a genuinely machine-readable surface: discoverable API documentation and structured resources an agent can find without scraping. And outside the AI cohort, the single best score in our whole index belongs to Zapier, a perfect 100. This is entirely achievable today.

The cohort median is 50, which happens to match the median of the general SaaS and developer-tools companies we scanned in the same window. The AI industry is no better at agent accessibility than the software industry at large.

The fixes are a morning's work

Most of the gap between 19 and 80 is not an engineering programme. It is publishing a handful of files: an llms.txt that tells agents where the useful resources live, a Markdown version of key pages served on request, an API catalogue at a well-known path, and a note about what agents are allowed to do. Our seven signals checklist walks through each one.

See where your own site stands: run a free scan. No sign-up, just a URL.

Fair game, open ledger

Every claim above is reproducible from a stored scan snapshot with a visible timestamp and evidence URL. Scores cannot be bought, and no company in this report has any commercial relationship with us. If you work at one of these companies and want your profile corrected or rescanned, the correction route is on your profile page; improvements get credited in our next monthly cycle. AgentVisible is built by Pixelshed, and yes, our own site sits in the same index, scanned by the identical pipeline.