Agent readiness
Software that AI can read correctly.
AI assistants answer questions about products and companies before anyone opens a website. When an assistant cannot reach the real information, it does not stop and say so. It fills the gap with something plausible.
So the risk is not being invisible. It is being described incorrectly, confidently, to someone who will never check. That is a data quality problem, and it is fixable at the source.
What PX7 publishes
The facts, in a form a machine can take literally.
Every page has a Markdown twin. Ask for text/markdown and you get the text; ask for HTML and you get the page. An assistant reads what was written instead of guessing at a layout.
A map of the site. llms.txt lists what exists and what each thing is for, so nothing has to be inferred from navigation.
Structured data that states facts rather than decorating a page: what a product is, who made it, where it runs, what it costs.
A catalog where there is an API. Published as a linkset, so an agent can discover what it may call without a human reading documentation first.
Next
Reading is only half of it.
An assistant that can only read is limited to describing things. The more useful case is an agent that can act: book the thing, fetch the record, hand a result to another agent and carry on without a person in the middle.
That needs somewhere to act, not just something to scrape. PX7 is building towards the Model Context Protocol for this, so an agent works with a site through a declared interface rather than by imitating a browser.
Measuring it
AICanary
None of this is worth doing blind. AICanary checks whether it actually works on a given site: what assistants can reach, what they get right, what they invent, and which sources they lean on instead.
AICanary →