Getting cited by AI search: what we changed on our own site

AI assistants do not rank pages. They assemble an answer, and either your page is one of the sources it was assembled from or it is not. That single difference is why a site can hold its rankings and still watch its traffic fall.

Three things decide whether you are in the answer: whether a machine can read your site without guessing, whether your claims are verifiable, and whether a passage can be lifted from your page without being rewritten. In August 2026 we ran a full audit of our own website against exactly those criteria. This is what it found and what we changed.

What actually changed in search

The shift is measurable and it is not subtle:

• Google AI Overviews appeared in roughly 18% of searches in early 2025 and around 25.8% by late 2025, and they show up in about 57% of long-tail queries — the informational questions that used to be the easiest traffic to win.
• Clicks are leaking. A 2026 randomized field experiment by researchers at ISB and Carnegie Mellon measured a 38% drop in organic clicks when an AI Overview was present. Clicks on the links inside the summary happened only about 1% of the time.
• Zero-click is now the default. Around 43% of Google searches end without a click to any external site, rising to roughly 93% when AI Mode is active.
• The engines disagree with each other. Only about 11% of domains are cited by both ChatGPT and Perplexity. Winning one does not win the other.

None of that means SEO is finished. It means the unit of visibility changed from a ranking position to a citation, and citations are earned by different things.

What the audit found on our own site

We build websites for a living, and our own site still failed several of these checks. That is normal, and pretending otherwise would be useless to you. The audit covered every URL in the sitemap.

Every one of those is invisible to a human visitor. All of them are load-bearing for a machine trying to decide what your company is and whether to quote it.

The six changes that mattered, in order

1. Structured data on every single page

Structured data is the difference between a machine reading your page and a machine parsing it. Research into generative engine optimization consistently puts schema markup among the strongest citation signals, and FAQ-formatted structured data has one of the highest citation rates of any format in AI-generated answers.

We added an Organization entity with a logo, Article markup on every case study, Service markup on every service page, and BreadcrumbList across the site so an assistant can see the hierarchy rather than infer it. Thirty-nine pages went from telling a machine nothing to declaring exactly what they are.

2. One page, one claim, one liftable paragraph

This is the change that costs nothing and gets ignored most. If the answer to the question your page targets is spread over four paragraphs and a subheading, no engine will assemble it. If it sits in a single self-contained paragraph near the top, it can be quoted verbatim.

Write the answer first, then earn it. Every page on this site now opens with a direct statement of what it does, before the context.

3. Verifiable numbers instead of adjectives

“Leading agency” is unquotable. “ROI 811% with booked-visit cost held in a 400–600 ₽ corridor” is a fact with a source attached, and a machine can carry it into an answer without risk. We rewrote six service pages so that every figure on them traces back to a named case study on the same site.

This is the same discipline that wins with human buyers, which is the reassuring part: nothing here asks you to write for robots at the expense of readers.

4. Descriptions written once, properly

Thirty-five pages had no meta description, which means the engine wrote its own from whatever text it found first. We wrote 31 by hand for the case studies alone, each naming the client and one verifiable figure, each between 140 and 156 characters, with no duplicates anywhere on the site.

A meta description is not a ranking factor and has not been one for years. It is the sentence a machine reads when it is deciding what your page is about, which is a different job and now a more important one.

5. llms.txt, and the honest caveat about it

An llms.txt file is a plain-markdown summary at your domain root that tells an AI crawler who you are and which pages matter. We publish one at atlasgoat.com/llms.txt, generated by the site itself so it can never drift out of date.

The honest caveat: Google has stated publicly that llms.txt is not required to appear in its generative results, and no engine has confirmed it as a ranking input. We publish it because it costs an afternoon, it forces you to write a one-paragraph description of your business that either survives scrutiny or does not, and because a machine reading it spends less effort guessing. If someone sells you llms.txt as an AI ranking strategy, they are overselling.

6. We removed pages instead of adding them

The least intuitive fix. Four category archives on this site were thin, near-duplicate lists of the same articles. They were set to noindex, follow and the sitemap dropped from 54 URLs to 50. Fewer, denser pages give an engine a clearer picture of what you are authoritative about than a larger number of shallow ones.

What we deliberately did not do

Three tactics are being sold hard right now and we passed on all three.

• Volume publishing. Generating fifty articles to blanket a topic works until the engine notices the pages say nothing. Thin content was a liability under classic search; under citation-based search it is worse, because there is no passage worth lifting.
• Chasing every engine. With only around 11% domain overlap between ChatGPT and Perplexity, optimizing separately for each is an unbounded project. The overlap that does exist is fundamentals, so we invested there.
• Schema on pages that do not deserve it. Marking up a page as an authoritative article does not make it one. Structured data describes a page, it does not upgrade it.

A twenty-minute check on your own site

1. View source on your best page and search for application/ld+json. If it is missing, that page is describing itself to machines in prose only.
2. Open your sitemap and count the URLs you would genuinely defend to a client. Everything else is diluting you.
3. Ask an assistant a question you should own — your service plus your city, or your category plus your specialism. Note who gets cited. That list is your real competitive set now.
4. Read your homepage’s first paragraph aloud. If it does not state what you do and for whom, no engine can quote it, because there is nothing to quote.
5. Check your claims. Every number on your site should be traceable to a page a stranger could open. Unsourced figures are the fastest way to be left out of an answer.

Frequently asked questions

What is generative engine optimization?

It is the practice of structuring content so AI systems cite it when they generate answers. The measurable outcome is not a ranking position but whether your page becomes part of the answer a user reads inside ChatGPT, Claude, Perplexity or Google AI Overviews, whether or not they ever visit your site.

Is GEO different from SEO?

It overlaps heavily. Crawlability, page quality and topical authority still decide most of it, and Google’s own guidance is that its generative experiences rely on the same core ranking systems. What genuinely changes is emphasis: structured data, self-contained answer passages and verifiable claims matter far more than they did when the goal was a blue link.

Do I need an llms.txt file?

No engine requires it and Google has said it is unnecessary for its generative results. It takes an afternoon and can improve how accurately a crawler summarizes you, so it is a reasonable thing to publish, but it is not a strategy.

How do I know whether AI search is sending me anything?

Filter your analytics by referrer for the assistant domains, and watch impressions against clicks in Search Console. The signature of AI-driven loss is stable impressions with falling clicks: you are still being shown, but the answer is being read without you.

Where this leaves you

Nothing in this list is exotic. Structured data, one clear answer per page, numbers you can prove, descriptions written by a human, fewer and better pages. It is the same work good SEO always asked for, with the tolerance for vagueness removed — because the thing reading your site now has no patience for it.

If you want your own site measured against these checks, that is where our SEO work starts, and what our digital strategy engagements are built on.