What we write for you
Every other page here answers what is wrong with a site. This one answers the other half: three of the files an agent looks for get written from what the site already publishes, and the product says plainly which of them you should publish today and which would make your site worse.
The three files
Generated per site from what discovery already found: the pages, their templates, the links that resolve. Nothing here reaches the network to work out what a site is; it reads what was already measured.
/llms.txt
Publish it todayThe index an agent reads before it starts guessing at URLs. Every line of a generated one is true the moment it is written, because every link came from a page discovery found and the link checker resolved.
text/markdown · Improves LLMs.txt
/.well-known/agent-card.json
Needs a service firstThe A2A card, at the path the spec registered in v0.3.0. A 1.0 card declares where it can be reached in supportedInterfaces rather than in a top-level url, which is what lets one agent offer two protocols at two addresses.
application/json · Improves Agent Card (A2A)
/.well-known/mcp/server-card.json
Needs a service firstThe MCP server card. Not the /.well-known/mcp.json with a tools array we used to write: that path was never in the specification and a static tool list goes stale on the first deploy. The card describes the server; the tools come from a live tools/list.
application/mcp-server-card+json · Improves MCP (Model Context Protocol)
Two of them would make your site worse
An llms.txt describes content you already have. The two cards describe a service: an A2A endpoint that answers, an MCP server a client can connect to. Publishing a card for something you have not built raises your score and lowers the truth. An agent reads it, tries the address, and gets a 404 from a file that promised otherwise, which is a worse outcome than never having advertised.
So both are generated in full and both are marked. Everything derivable from your site is filled in truthfully, and every field that is a promise you have yet to keep is listed beside the file as a claim. A platform that will not tell you when publishing something makes you worse is a score generator, not an optimisation one.
Grounded, not invented
The links are pages discovery found and the checker resolved. The skills are your own sections, described by what is in them and how many pages each holds. Nothing names a page that does not exist, and a section of one page is never described as a set of many.
A model writes the prose: the one sentence saying what the site is, and the description of each section. It is given the site's own shape and nothing else, and the file records which model wrote it. Where no key is configured the deterministic renderer runs alone; the file reads flatter and is still valid, which is a better outcome than no file.
The button in the report
On a failing finding for a check that has a file, the report offers to generate it. It reads the pages you already publish, works out what the file should say, and hands it back ready to copy or download. Nothing is published on your behalf: the file is yours to put where it belongs.
3 of the 26 checks have a file to hand over. The rest are changes to the site itself (a heading that is missing, a server that does not compress, a crawler blocked in robots.txt), and for those the guide says what to do rather than a button pretending it can do it.
The industry answers
The industry plugins ask 47 questions a site of a particular kind owes an agent. 21 of them can be answered from what a page already publishes, and those we write too. The rest are decisions about a business, a returns window or a delivery promise, and the product's position is that it will not invent one.