How AI search reads your website, and how to be legible to it

Somewhere in the last two years, a chunk of research quietly moved. People who would once have typed a query and scanned ten blue links now ask an assistant and read one answer. When someone asks ChatGPT or Claude how much a small business website should cost, or Google's AI results summarise the best way to brief a designer, that answer is assembled from websites. Possibly yours. So the useful question is not "how do I rank in AI". It is plainer than that: when a machine reads my site, can it understand it, and is there anything worth quoting?

How assistants source their answers

Strip away the mystique and there are three routes from your website into an AI answer.

Training data. The model read a snapshot of the public web at some point in the past. If your site was crawlable and clearly written then, some of it may be baked in. You have no control over this after the fact, which is a decent argument for getting pages right early and keeping them stable.

Live retrieval. This is the big one. When an assistant answers a current or specific question, it usually runs a search behind the scenes, fetches a handful of pages, reads them in seconds and writes a synthesis. Your site is competing to be one of that handful, and then to be the page whose sentences survive the summary.

Search engines' own AI layers. Google and Bing generate AI summaries on top of their existing indexes. If you are indexed and well structured, you are already in that pool.

Notice what all three have in common. The raw material is a fetched page and parsed text. There is no separate "AI version" of your website. The version the assistant reads is the same HTML everyone else gets, read quickly and literally.

What actually helps

Semantic HTML

An assistant skims the way a hurried human does: headings first, then the sentences under them. A page with one clear h1, descriptive subheadings and real paragraphs is easy to skim. A page built as anonymous div soup, with the headline baked into an image and the key facts trapped inside a JavaScript widget, is close to invisible. This is the least glamorous item on the list and the most important.

Structured data

JSON-LD is a small block of machine-readable facts in the head of a page: this is a business, here is its name, its location, its services, its FAQ. Schema.org markup has existed for well over a decade and search engines have used it for rich results the whole time. It removes ambiguity, which is exactly what a machine reading at speed needs. It is invisible to visitors, so there is no design cost to shipping it.

Sentences that can stand alone

Assistants lift answers, and an answer has to survive being lifted. "We craft holistic digital experiences" survives nothing. "A bespoke marketing site from an independent studio starts from around £2,500" is a sentence a machine can quote with confidence. Commit to specifics: prices, timelines, process, who the service is for. My post on what a website costs in the UK is written this way deliberately: real ranges, plain claims, no adjectives doing the work of facts. Honest pages were always better pages. Now they are also more quotable pages.

llms.txt, with appropriate scepticism

llms.txt is an emerging convention: a plain text file at yoursite.com/llms.txt that summarises what the site is and links to its key pages, written for language models the way robots.txt is written for crawlers. Some AI companies read it, some ignore it, and nobody should tell you it is a ranking factor, because there is no evidence it is one. But it costs ten minutes, it cannot hurt, and it forces a useful discipline: if you cannot summarise your site in a page of plain text, that is a problem worth discovering.

Letting AI crawlers in, on purpose

AI companies fetch pages with named crawlers, GPTBot and ClaudeBot among them, and your robots.txt file can allow or block each one. Publishers who sell their words have real reasons to block. A small business that wants to be found has the opposite incentive: if the assistant cannot fetch your page, it will answer the question from someone else's. Either position is legitimate. What is not fine is having no position because nobody has looked at the file since launch.

What is snake oil

The moment "AI search" became a phrase, an industry appeared to sell it. Be suspicious of anything shaped like this:

• Promised placement. "We'll get you cited in ChatGPT answers." Nobody outside the AI companies controls what the models say, and the AI companies are not selling placement. This is the old "guaranteed first page of Google" lie in a new jacket.
• Proprietary scores. An "AI visibility score", measured by the same agency that sells the fix for it.
• Secret techniques. There is no secret. The mechanisms are the ones in this post, and they are public.

The test is simple. Ask the seller to explain the mechanism by which their work changes an assistant's output. If the answer is fog, so is the service.

The unfashionable conclusion

Everything that makes a site legible to AI made it a good site ten years ago: crawlable, fast, properly structured, specific, honest. If your site was built well, you are most of the way there already. If it was built as a pretty picture with no structure underneath, no amount of "AI optimisation" sprinkled on top will fix it.

This is why I treat it as foundation, not upsell. Structured data and technical SEO foundations are built into my Conversion and Platform work, and the pricing page sets out exactly what each tier includes. The Key's site, for one, went out with clean URLs, structured data and copy written to be found, in place from day one. The other case studies show how each build ran.

One honest caveat to end on: nobody can guarantee an assistant will quote you, including me. What you control is what happens when the machine arrives. Whether it finds a page it can parse, facts it can trust, and a sentence clear enough to lift. That part is entirely buildable.