How AI assistants read your website

A growing share of my traffic doesn't come from someone clicking a blue link. It comes from an AI assistant that read the site, decided it answered a question, and cited it. That changes what "being findable" means. The good news: the work that helps AI assistants read your site is mostly the same work that always made good SEO.

Assistants read the page, not the vibe

An AI assistant fetching your site sees HTML, not your design. If the answer to a question is locked inside a JavaScript widget, an image of text, or a wall of marketing copy, it may as well not exist. Three things make a page legible to a machine:

Semantic structure. One h1, headings in order, lists marked up as lists, and the main content in main. This has been the advice for twenty years. Now it decides whether an assistant can extract your answer at all.

Answers near the top. Assistants quote the passage that answers the question, not the page as a whole. If your post takes four paragraphs of throat-clearing to reach the point, the quotable passage is buried. Lead with the answer, then expand.

Server-rendered content. Most crawlers that feed AI systems don't execute JavaScript reliably. Static or server-rendered pages are read completely; client-rendered pages are a gamble. This is one more reason static generation should be your default.

Structured data: saying it in the machine's language

JSON-LD tells crawlers what a page is, not just what it says. Every post on this site ships an Article block:

<script
  type="application/ld+json"
  dangerouslySetInnerHTML={{
    __html: JSON.stringify({
      "@context": "https://schema.org",
      "@type": "Article",
      headline: post.title,
      description: post.description,
      datePublished: post.date,
      author: { "@type": "Person", name: "Diego Paganelli" },
    }),
  }}
/>

The types worth implementing for a small business site are few: Organization or Person on the homepage, Article on posts, Service for what you sell, and FAQPage where you genuinely answer questions. Don't mark up what isn't on the page; assistants cross-check, and mismatches cost trust.

What gets cited

Watching which of my pages assistants actually cite, a pattern is clear. It's the pages that commit to an answer. A pricing post with real numbers. A comparison that ends in a recommendation. A how-to with the exact steps.

Hedged content ("it depends, contact us to learn more") gives an assistant nothing to quote. Specific content gets lifted verbatim, with a link. Being quotable now beats being comprehensive.

Check what the machines see

Two quick tests tell you most of what you need to know. Fetch your page without JavaScript and see if the content is there:

curl -s https://crea-tech.uk/blog/some-post | grep -c "the answer you expect"

Then paste your URL into Google's Rich Results Test to confirm the structured data parses. Five minutes, and you know whether you're legible.

The decision in one question

If a machine read only your HTML, would it find a direct answer?

If yes, you're set for both kinds of search. If no, the fix is rarely exotic: render on the server, structure the headings, put the answer first, and describe the page in JSON-LD. The sites that win AI search are the ones that were built properly for the old search.