· Anvilstead · AI visibility · 4 min read
How we make a website readable by AI answer engines
People increasingly ask an AI assistant instead of scrolling search results. What decides whether your business is the one it names — and the concrete files and markup that make a site quotable.
A growing share of buying decisions now starts with a question typed into an AI assistant rather than a keyword typed into a search box. That changes what a website needs to do: it is no longer enough to rank. The site has to be readable, quotable and attributable by whatever model answers the question.
We build this into every showcase site we ship. Here is the actual list, without the mystique.
1. Say who you are, in one place, in plain sentences
Models assemble an answer from fragments. If your company name appears only inside a logo image or a JavaScript bundle, there is nothing to quote.
Every site we build states, in ordinary text on a real page:
- the legal entity behind the brand, and the relationship between brand and entity
- what the business does, in one sentence, without adjectives
- who it does it for
- where it is registered and where it operates
That paragraph is what an assistant will reuse. Write it once, keep it consistent everywhere.
2. Structured data for the entity and the content
Search engines and assistants both read Schema.org JSON-LD. We emit it per page rather than in one lump:
OrganizationandWebSiteon the home pageBlogPosting, with author and publication date, on every articleBreadcrumbListon nested pagesFAQPagewhere a page genuinely answers questions
Two rules matter more than coverage: only mark up what is visible on the page, and keep the markup consistent with the text. Contradictions are worse than absence.
3. An llms.txt that an AI can actually use
llms.txt is a plain-text summary at the root of the domain, written for models rather than people. Ours contains the entity facts, the three services with timelines, the pricing structure, working hours, and links to machine-readable files.
https://anvilstead.com/llms.txtThe important part is that it is factual and specific. “We are a leading innovative solutions provider” gives a model nothing to work with. “Showcase websites are typically 2–4 weeks from kickoff” is a sentence a model can quote and a reader can verify.
4. Machine-readable companions
Alongside llms.txt we publish small JSON files that answer the obvious questions directly:
| File | What it holds |
|---|---|
/ai/summary.json | entity facts, services, engagement model, pricing structure |
/ai/faq.json | the questions clients ask before engaging, answered in quotable form |
/ai/service.json | service catalogue with deliverables, timelines and fit criteria |
They cost an hour to write and remove the guesswork for anything that reads them.
5. Let the crawlers in — deliberately
Most sites either block AI crawlers by accident or never think about them. In robots.txt we allow the ones we want reading us, by name: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, CCBot, meta-externalagent and others. The only paths disallowed are internal ones with no public value, such as the API endpoint.
If your goal is to be quoted, being read is the prerequisite.
6. Content shaped like answers
This is the part nobody can automate away. An article that an assistant will cite usually:
- answers one question, visibly, near the top
- is dated, and updated with a visible date when it changes
- separates fact from opinion
- includes the specific numbers, limits and conditions — the things that make a claim checkable
- links to the primary source rather than paraphrasing it
Ten thin articles restating each other do less than one that says something only you can say.
7. Verify it yourself
None of this requires trust in a vendor. You can check any site in a few minutes:
- Open
https://<domain>/llms.txt— does it exist, and is it accurate? - Open
https://<domain>/robots.txt— are the AI crawlers allowed? - Paste a page URL into Google’s Rich Results Test — is the structured data valid and visible?
- Ask an assistant a question your customer would ask, then ask it to name its sources. See whether your site appears, and whether the description is right.
Step 4 is the only one that measures the outcome, and it is worth repeating monthly with the same questions so the comparison means something.
None of the six points above is exotic. The reason most sites fail at them is not difficulty — it is that nobody owned the job. On the sites we build, the structure, the markup, the llms.txt and the news section ship as part of the project, not as a later optimisation phase.
If you want to know what an assistant currently says about your business, that is a reasonable first conversation: book a 30-minute call and bring the questions your customers ask.
