How to make your website understandable to Google and AI agents

What changes when an AI writes the answer?

The selection criteria change. A classic search engine ranks documents and lets you choose; an assistant picks a handful of sources, extracts specific claims and combines them into prose. To make that cut, your content has to be easy to read without context: every claim must stand on its own, away from the page that contains it.

That is the real shift. A clever headline that needs three paragraphs to explain itself performs adequately on Google and badly in a generated answer, because there is nothing to extract.

Note that last risk, which almost nobody looks at. If your website does not clearly state what you do, a model will infer it from context. And inference is exactly where the errors come from.

Start by being explicit about who you are

The first screen of your website should let anyone answer four questions without interpreting anything: what you do, who for, where you operate and what makes you different. Not as a slogan — as a declarative sentence.

Compare these two versions:

“We power brands that refuse to settle.”

It sounds good and says nothing. A human reader can keep scrolling; a system that has to summarise your business in one sentence cannot.

“We are an agency for custom software development, web design and SEO. We work remotely with companies in any country, in English and Spanish.”

This second sentence is extractable, verifiable and comparable. You can keep the slogan as an emotional headline as long as the declarative sentence appears right beneath it, as plain text in the HTML. That is the pattern: hook on top, definition immediately after.

Structure communicates as much as the words

A real heading hierarchy is not a style preference. It is the index that crawlers and models use to segment your content.

• One H1 per page, matching the main topic.
• H2s shaped as questions or concrete claims. “How much a project costs” beats “Investment”.
• The first sentence of each section answers its heading. Context comes after, not before.
• Lists and tables wherever items are comparable. A paragraph with five criteria jumbled together extracts far worse than five bullets.
• Nothing critical hidden inside JavaScript-only accordions that does not exist in the initial HTML.

Structured data: the language that reads without ambiguity

Structured data with schema.org turns interpretation into explicit fact. Instead of hoping a system deduces that you are a company, you tell it using a vocabulary it already understands.

The minimum set we recommend for a service business:

• Organization with name, logo, description, email, languages and official profiles in sameAs. This is your business identity card.
• Service for each service, linked to the Organization through provider.
• FAQPage on your FAQ. Even with Google's restrictions on the rich result, it remains a clean way to expose question-answer pairs.
• Article or BlogPosting on every article, with datePublished, dateModified and an author.
• BreadcrumbList to make the site hierarchy explicit.

Two rules when implementing. First: the markup must match what the user sees. Declaring an FAQ that is not on the page earns a penalty, not an advantage. Second: use @id to link entities to each other instead of repeating the same data on every page. A coherent graph is worth more than five loose cards that contradict each other.

What makes a paragraph citable

Citable content has four properties. All four are about writing, not technology:

1. It is specific. “A corporate website usually takes 3 to 6 weeks” is citable. “Timelines vary by project” is not.
2. It is self-contained. It makes sense without the preceding paragraph. Avoid opening sentences with “this”, “the above” or “as we said”.
3. It is attributable. It has an author, a date and an entity behind it. Unsigned, undated text is harder to use as a source.
4. It is consistent. If your home page says one thing and your blog another, the conflict reduces confidence in both.

Consistency is the most neglected part. It is worth periodically checking that the company name, description, services and contact details are identical across every page and in the markup.

The technical basics still rule

None of the above matters if the content cannot be crawled. Before working on semantics, check the boring things:

• Content exists in the initial HTML. If text only appears after JavaScript runs, some crawlers will never see it.
• robots.txt and sitemap.xml exist and are correct. It sounds basic and it is the most frequent failure we find in audits.
• Every page has its canonical, and language versions point at each other with reciprocal hreflang. A language that only changes via JavaScript does not exist for a search engine.
• Reasonable performance and Core Web Vitals. You do not need a perfect score; you need performance to stop being an obstacle.
• Accessible HTML. Headings in order, alt text on meaningful images, sufficient contrast and keyboard navigation. What helps a screen reader helps a crawler.

What about llms.txt?

It is a community-proposed convention: a text file at the domain root describing the site and linking its main content for AI systems. It is worth being honest about its status: it is not currently an officially adopted standard among the major model providers, and there is no public evidence that it improves rankings.

That said, it takes fifteen minutes and does no harm. Our position is to publish it, keep it current, and expect nothing miraculous. The work that actually moves the needle is in the sections above.

Review checklist

If you only have one afternoon, this is the order by impact:

1. Add the declarative sentence about what your company does to the first screen of the home page.
2. Check that every page has a single H1 and headings that answer real questions.
3. Implement Organization, Service and BreadcrumbList linked with @id.
4. Verify robots.txt, sitemap.xml and the canonical on every page.
5. Make sure language versions are distinct URLs with reciprocal hreflang.
6. Rewrite the first sentence of each section so it answers its heading directly.
7. Publish an FAQ using the questions people actually email you, marked up with FAQPage.

None of these is a trick. They are the same old hygiene applied with a new criterion: that anyone — person or system — can understand your business from a single screen.