A website built for AI

For twenty years a website was built for two readers: the person visiting it and Google, which indexes it. Recently there is a third, and it grows faster than the other two: the language model that reads your site when someone asks ChatGPT, Gemini, Perplexity or Claude about what you sell. That reader does not see your design. It reads your HTML, decides in seconds whether it understands who you are, and cites you or leaves you out.

What a model does not see

Let us start with what gets lost, because that is where most mistakes are:

• Images. The price in a photo, the menu in a scanned PDF, the “from $X” in a banner: they do not exist. A model reads text.
• What JavaScript paints too late. Many modern sites deliver an empty page and fill it with code in the browser. A crawler that does not run that code —or does not wait— sees the empty page.
• Adjectives. “Unique”, “unforgettable”, “world-class” are not information. A model that needs to cite a fact finds none and moves on to the next site.
• What is behind a block. A robots.txt that bans GPTBot or ClaudeBot —often set by the hosting provider without anyone deciding— makes the site invisible to AI even if Google indexes it perfectly.
• What loads slowly. A crawler has a time budget. An eight-second site does not get read in full.

What a model does see, and rewards

1. Headlines that state. The h1 is the first sentence the model reads from you. “30-suite aparthotel with kitchens in the Zona Rosa of San Salvador” is quotable; “Welcome” is not.
2. Semantic structure. Sections with h2s that name what they contain, lists that are lists, a question and its answer together. HTML has tags for all of that, and a model uses them to understand what each thing is.
3. Structured data (schema.org). A JSON-LD block that declares “this is a hotel, here it is, it has these rooms, these policies and these frequently asked questions”. It is the part of the site written directly for machines, and the one that weighs most in an AI describing you well.
4. An llms.txt. The business summary in plain text, at the root of the domain. What you are, where, what you offer, to whom, and the links that matter. It costs an hour.
5. Verifiable facts. Measured distances, schedules, policies, real room or product names. And consistency: the same data on your site, on Google and on your listings. A model trusts what it can confirm in more than one source.
6. Content that answers questions. An AI answers questions; a site that answers them in writing —“can I get there by public transport?”, “do you accept pets?”, “how long does shipping take?”— is the natural source.
7. Open crawling and measured speed. robots.txt reviewed by hand, and a mobile load under two and a half seconds, measured in production and not estimated.

Why this is not “SEO under another name”

It looks alike, and on purpose: what makes Google understand you is almost the same as what makes an AI cite you. The difference is the severity of the outcome. On Google, being in position eight still brings visits. In an AI’s answer there are four or five names and no page two: you are either there or you do not exist. And the AI does not give you a position you can improve little by little; it cites you or not, depending on whether it can understand and trust you.

That is why we say a site must be built for AI and not “optimized afterwards”: structure, real text and consistent data are construction decisions, not a layer added at the end.

What we do

Every site we build ships with complete schema, llms.txt, open crawl permissions, headlines that state, real text and not images with text, speed measured on mobile and a GEO score taken before publishing. And when the site already exists, it gets worked on: at Porto Marina, opening crawling, completing the schema and adding llms.txt took the score from 62 to 80 out of 100 without changing a line of the design.

The previous lesson explains how AI search engines find us. And if you want to know whether your site is blocking AI today, the diagnostic asks you directly.