Getting your site into AI answers

AI summaries now sit on top of a majority of searches. For the person searching that is convenient. For the person who runs the site the summary was built from, it is a problem: the answer is given, the click never happens.

Studies put the zero-click share at roughly two thirds of searches, and considerably higher where an AI summary appears. In the pure answer engines it is more pronounced still.

Two independent analyses from February 2026 put numbers on it. SISTRIX went through more than 100 million keywords: with an AI summary on top, the click rate for position one falls from about 27 to 11 per cent. Ahrefs, working from 300,000 keywords of Search Console data, lands on the same picture, with position one losing 58 per cent of its clicks. Informational questions take the hardest hit, while brand, local and buying intent hold up far better.

So the question is no longer only how to rank first, but how to end up inside the answer. There is a lot of nonsense circulating about that. Let us start there.

What does not help

llms.txt. The idea sounds sensible: a file at the root that tells language models what is on your site. Hundreds of thousands of sites now have one. And to date there is no documented case of a major model actually using it. Google guidance on AI features says plainly that no machine-readable extra files are needed to appear in generative results, and Google representatives have confirmed their systems do not read the file today.

We ship one at /llms.txt on primero.link anyway. Two reasons, both honestly small: it costs us nothing because our plugin generates it from the page tree, and if the picture changes we are ready. We would not sell it to anyone as a visibility measure.

Keyword density, AI articles by the dozen, boilerplate written for answer engines. Publishing ten generated guides a week produces exactly the thing the models already have too much of. What gets cited is what adds something.

Schema markup for its own sake. Structured data helps, but it is not a spell. A FAQPage marking questions that are not on the page breaks Google's structured data policies and can take the page out of results.

What does help

1. Paragraphs that can be lifted

An answer engine is not looking for a page, it is looking for one paragraph that answers the question. So write so that a single paragraph stands on its own: question as the heading, answer in the first sentence, reasoning after it.

The difference in practice:

• Weak: "In the following we will take a look at the factors that may play a role when choosing a website builder."
• Strong: "A website builder is worth it while you stay under five pages and need no custom integration. Above that it costs more than a bespoke build, because every extra feature carries a monthly fee."

The second one can be quoted. The first says nothing.

2. Numbers and evidence

Models favour claims that can be anchored. One number with a source beats three adjectives. If you have your own figures, publish them. In our post on building an accessible site we wrote down our own contrast values before and after, including the mistakes we had made. Those are the passages that get cited, because nobody else has them.

3. Experience that can be checked

Google groups this under E-E-A-T: experience, expertise, authoritativeness, trustworthiness. In practice:

• A named author, not "the editorial team". With a photo, a role and a profile someone can verify.
• Visible and marked up. The author box under this post is not decoration; it matches the Person markup in the head of the page. Both have to say the same thing.
• Something only you can tell. A project, a measurement, a mistake. Summaries of other people's sources already exist in quantity.

4. Clean delivery

None of the above works if the page cannot be read.

• Content in the HTML, available before JavaScript runs. Many AI crawlers skip rendering.
• Fast pages. A crawler on a time budget gives up.
• A heading outline that holds. Exactly one h1, no skipped level. That is also an accessibility requirement, so one job serves two purposes.
• Structured data that matches the content. Article with author and date, FAQPage for questions genuinely on the page, Organization for the company.
• Consistent naming. Call your product the same thing in every text. Models build entities, and three spellings make three entities.

5. Mentions off your own site

Answer systems pull from more sources than Google does. Directories, trade forums, industry lists, guest pieces. If you exist only on your own domain, you are exactly one source. That is thin for a system that likes to cross-check.

How to measure it

Tools exist, but doing it by hand is more honest and costs nothing:

1. Write down twenty questions your customers actually ask. Not keywords, real questions.
2. Run them monthly through Google, ChatGPT and Perplexity. Note whether you appear and who else gets cited.
3. Read your server logs. Which AI crawlers arrive, how often, on which pages? Look for GPTBot, PerplexityBot, ClaudeBot and Google-Extended.
4. Judge sessions by value. If visits drop and enquiries hold, nothing is broken. The AI is filtering out people who were never going to enquire.

The uncomfortable part

Some of that traffic is not coming back. If your business ran on answering questions a machine answers just as well, those visits are gone, and no measure retrieves them.

What is left is two things. Content a machine cannot generate because it contains first-hand experience. And the rest of the site, where the decision happens when somebody does click. Both were the right work before any of this.

If you want to know where your site stands in the answer engines, drop us a line. We run twenty questions and show you who gets cited instead of you.

Sources

• Google Search Central: AI features and your website
• Google Search Central: structured data general guidelines
• Google Search Central: creating helpful, reliable, people-first content
• smartlemon: study on the impact of AI Overviews on SERPs and traffic in Germany
• Ahrefs: AI Overviews cut position one clicks by 58 per cent
• SISTRIX: 100 million keywords analysed for AI Overviews in German Google
• blckalpaca: llms.txt, 844,000 implementations and no confirmed use
• traffic3: llms.txt, useful GEO signal or hype