llms.txt, schema, and extractability: what actually affects AI citation
Owners ask me to "add llms.txt" the way they used to ask for a meta description fix — as if it's the one setting that unlocks AI citation. It's one of three, and on its own it's usually the least impactful of the three.
On this page
The three layers, defined precisely
Discovery is whether an AI crawler or the search index behind an assistant can find the page at all — crawlability, a sitemap, and increasingly llms.txt as a curated pointer. Structure is whether the page tells a machine what it is — schema.org markup identifying an Article, a Person, an FAQ, a Video, with dates and authorship attached. Extractability is whether an actual fact on the page can be lifted cleanly into an answer — one claim per sentence, the number or definition stated plainly instead of buried three clauses deep in a qualifier. All three have to hold. A perfectly structured page that's unreachable gets nothing; a perfectly reachable page that's vague gets found and skipped.
llms.txt: what it does and doesn't do
llms.txt is a curated index file pointing an AI crawler at your most important content — closer to a sitemap written for a different audience than a ranking signal. Adoption is still uneven across platforms: some lean heavily on their own crawl, some route through a search partnership, and not every AI product reads it the same way yet. Where it earns its keep is content that's hard to discover through ordinary crawling — deep pages, content behind heavier navigation, or a curated "here's what actually matters" list on a large site. It's a supplement to good architecture, not a substitute for it.
Schema.org: which types actually matter
For a services and content site, four types carry most of the weight. Person or Organization establishes entity trust — who's speaking, and whether that identity is consistent and verifiable across every page that mentions them. Article carries provenance — publish date, author, what the piece actually is — which matters for an assistant deciding whether content is current enough to cite. FAQPage is the most directly extractable type that exists: a question and a self-contained answer, structured exactly the way an assistant wants to quote it. VideoObject matters wherever video is genuinely present — it tells a crawler what the video covers without processing the file itself. Schema that doesn't match the visible content is worse than no schema; a mismatch is a trust signal in the wrong direction.
Extractable prose: the writing pattern itself
This is the layer most sites skip, and it's the one with the highest return. A direct-answer paragraph in the first hundred words of a page — not a hook, not a story, an actual answer to the question the title implies — is the single highest-leverage habit for AI citation. Below that: one factual claim per sentence rather than compound sentences that bury the number behind two qualifiers; explicit numbers and named definitions rather than vague comparatives; headers that state the actual question being answered rather than a clever turn of phrase. None of this requires new tooling. It requires writing like you expect to be quoted, because increasingly, you are.
What moves the needle vs what's cargo-culted
Adding llms.txt to a site whose actual content is still hard to crawl accomplishes close to nothing — it's a pointer to content that was already reachable, dressed up as a fix. Schema without content that supports it can actively hurt, flagged as inconsistent rather than ignored. If you can only fix one layer this quarter, fix extractability: rewrite the top of your most important pages into direct-answer paragraphs. It's the cheapest of the three, and in the accounts I've measured it on, the one with the clearest before-and-after in citation rate. The $750 Starter tier of the AI Search audit maps exactly these three layers against your site and tells you which one is actually holding you back.
VERIFIED EXAMPLE · On my flagship account I took AI-search presence from zero to 600+ citations across ChatGPT, AI Overviews, Perplexity and Gemini, alongside ~6.4× organic traffic growth — measured with a fixed query panel, not vibes.
What clients say
"A true professional – a real talent for explaining complex concepts in simple, accessible terms."
Dmitry Kuznetsov ↗ · Product Manager, Vallex
| Layer | What it does | Common mistake |
|---|---|---|
| Discovery | Whether a crawler can find the page at all | Adding llms.txt while the site is still hard to crawl |
| Structure | Whether schema tells a machine what the page is | Schema that doesn't match the visible content |
| Extractability | Whether a fact can be lifted cleanly | Burying the answer three qualifiers deep |
Questions owners ask
Do we need llms.txt if we already have good SEO?
It doesn't hurt, and it's cheap to add — but it won't compensate for pages that are genuinely hard to crawl or thin on extractable content. Treat it as a small supplement, not the fix.
Which schema type should we prioritize first?
FAQPage on your highest-traffic pages, if you don't have it yet — it's the most directly extractable format that exists, and the fastest to implement correctly.
How do we know if our content is "extractable"?
Read your own first paragraph and ask: could someone lift one clean sentence from it and quote it accurately out of context? If the answer requires reading three more sentences to make sense, it's not extractable yet.