The question changed. A plant manager who once typed keywords into Google now asks an agent: what should we use to plan production from our ERP exports? The agent fetches a handful of pages, reads what it can, and gives one answer. Not three blue links. One answer. If your page cannot be read by the thing doing the fetching, you are not in the answer — no matter how good the page looks to a person.
What we found on our own site
We audited algorilla.org the way an agent reads it. Some of it held up: our home page carries fourteen questions and answers in plain text, and the production evidence block — the morning brief, the scoreboard, the pilot shape — is real text a fetcher can quote. Some of it did not. Our llms.txt was a bare list of links. The richest proof on the site, forty-one demo scenarios, played inside a JavaScript chat widget that a fetching agent never runs. Our product pages had few of the question-shaped answers the home page had. And asking the site for a Markdown page returned the home page's HTML, which is worse than returning nothing: it teaches the reader a wrong fact with confidence.
What we shipped
Every page on this site is now published as Markdown, generated from the same source files as the HTML — one source, so the two can never drift apart. Request any page with an Accept: text/markdown header and you get the Markdown directly. The llms.txt index now carries the substance, and llms-full.txt holds the whole site as one text file. The demo scenarios are written out as a text page. Products and Valet OS gained the questions-and-answers treatment the home page already had. There are comparison pages — against the spreadsheet morning rebuild, against general chat AI, against generic automation — and a page for each major platform we connect to, saying exactly what we read from it and what we return. There is a page for agents themselves: how they connect to Canopy, what they may do, and the rules they work under.
The rule underneath
None of this is a trick, and that is the point. The discipline is the same one we sell: spell out in plain text anything a widget shows, keep one source of truth, and answer the questions people actually ask in the words they ask them. An agent that reads this site will describe us accurately — including the things we chose to state plainly rather than leave to guessing. We do not publish pricing, for example, and now the machine-readable site says so in so many words, so a model can state that fact instead of inventing a number or calling the offer obscure.
Designing for agents turns out to be designing for readers who cannot see your design. What survives that test is the substance: the facts, the evidence, the answers. Which is what was supposed to carry the site all along.