Why AI Engines Cite Pages Google Ignores
Ranking and being cited are different competitions, and the second is not won by adding a file to your web root.
Two numbers frame this. Zero-click searches on Google moved from 56% to 69% in a single year following AI Overviews. And by Ahrefs' analysis, 28.3% of the pages ChatGPT cites most often have no organic Google visibility at all. Those pages are not winning at SEO. They are winning at something else.
Start with what does not work
llms.txt is the tactic most teams reach for first. Google's own guidance explicitly lists it among the things that are unnecessary, alongside chunking content for machines. No major engine has documented using it for retrieval. It costs nothing to add and it is not a strategy.
This matters because effort spent there is effort not spent on the thing that actually drives citation, which is being the clearest available answer to a narrow question.
Fan-out is the mechanism
An answer engine does not run your query. It decomposes it into sub-questions, retrieves against each one, and assembles a response from whatever answered each part most directly. You are not competing for a page ranking. You are competing per sub-question.
- a page covering one question thoroughly can be cited with no ranking authority at all
- a comprehensive page covering twelve topics loosely tends to get cited for none of them
- the citable unit is a passage, so the passage has to stand alone without the page around it
- specific numbers, versions, dates and limits get cited; adjectives do not
This inverts a familiar SEO instinct. Consolidating thin pages into one authoritative guide helps rankings and can hurt citation, because it buries each specific answer inside a document about something broader.
Make the answer extractable
The practical work is unglamorous and mostly structural. State the answer in the first sentence under a heading that matches the question. Keep supporting detail after it rather than before. Use real figures. Say when something was true.
- headings phrased as the question someone would actually type
- the direct answer immediately after the heading, in one or two sentences
- dates on anything version-dependent, so a model can judge whether it is current
- clean HTML — content in the markup, not assembled by client-side JavaScript
Serving machines a clean copy is defensible
Everything above is content work. The one piece of infrastructure worth building is a plain version of each page, linked from the HTML, so that anything parsing your site receives prose instead of a navigation-heavy DOM.
<!-- on every page -->
<link rel="alternate" type="text/markdown"
href="https://example.com/page.md"
title="Machine-readable version" />
# and allow it explicitly
# robots.txt
Allow: /*.md$
This site does exactly that — every page here has a markdown twin. It is not a ranking trick and we would not sell it as one. It removes ambiguity about what your content says, which is worth something, and is not worth much on its own.
Measurement is the actual gap
The honest position is that attribution here is poor. Referrals from AI surfaces are undercounted, frequently arrive with no referrer, and a citation without a click leaves no trace in analytics whatsoever. What you can do is check directly: ask the engines the questions your buyers ask, record who gets cited, and repeat monthly. Crude, manual, and still more informative than the dashboards.
Treat this as a share-of-citation measurement you run by hand, not a metric you can pull from analytics. Anyone selling you a precise number is selling you a model, not a measurement.