LLM SEO: how language models choose citations, and how to earn one

An assistant does not remember your page. When ChatGPT, Perplexity or Gemini answers a question about the live web, it works from search: the question is rewritten into narrower queries, results are fetched, and the answer is composed from passages in the pages it fetched. If your page is not in that fetched set, nothing else on this list matters.

Three things follow. First, the queries an assistant issues are usually narrower than the question a person typed, so a page that ranks for the head term but not for the specific sub-questions gets skipped. Second, retrieval works on passages rather than whole pages: systems of this kind split a document into chunks and score each chunk, so a section with its own heading and a complete answer inside it travels better than prose that depends on the paragraph above. Third, the crawler has to be allowed in. The vendors publish the user-agent names their fetchers use, GPTBot, OAI-SearchBot, PerplexityBot and ClaudeBot among them, and honour robots.txt for those agents, so a blanket block written years ago for another reason still applies. The AI Crawler Check reads your robots.txt and reports which assistants you have shut out.

Models are trained on a snapshot and then asked about a moving world, so recency carries weight in the retrieval layer. For any question where the answer can change, prices, versions, regulations, product availability, assistants lean towards recently updated pages, and they often show the date of the source they used.

Freshness is read from signals a crawler can parse, not from your intention. A visible published date and a visible updated date near the top of the article, a matching dateModified in the Article schema, and a sitemap lastmod that moves only when the content moves are the three that matter. A lastmod that changes nightly while the copy stands still is an unreliable signal, and unreliable signals get discounted rather than rewarded.

The practical consequence is that content has a half-life. An article cited in March and untouched since is competing against pages covering the same ground with a July date on them, and it tends to lose. IT Master's nightly loop tracks which published articles are losing impressions in Search Console and queues them for a refresh, which costs less than writing a replacement and keeps the URL that has already earned its links.

Look at how an assistant cites. It quotes or closely paraphrases a short passage from a source, then attaches the link. Pages that get cited tend to contain sentences that survive being lifted out of context: a subject, a verb, a claim, and no pronoun pointing back to the previous paragraph.

The pattern that works is answer-first. Open each section with a sentence that states the conclusion in plain terms, then spend the rest of the section on evidence and caveats. Define every term the first time it appears, in a single sentence under 40 words. Put numbers in the sentence that explains them, not in a chart the crawler cannot read. Give each section its own heading that repeats the question in the reader's words, because that heading travels with the chunk.

Callout boxes like the ones on this page exist for this reason. They isolate the one sentence we would want an assistant to quote, mark it up so it is unambiguous, and leave the surrounding prose as detailed as the topic needs. Marketing copy fails this test almost every time: adjectives are not quotable, and a model asked for facts skips them.

Assistants are tuned to prefer sources they can attribute. A page with a named author, a stated organisation, a date and links to the primary material it drew from is easier for a model to justify citing than an anonymous post asserting the same thing. This is E-E-A-T applied by a machine, and the machine is literal: it looks for the byline, the about page, the author's other work, and whether other pages on the web name the same person for the same subject.

Primary data is the strongest version of this. A measurement you took, a specification from your own datasheet, a figure from your own support tickets: these are not available elsewhere, so a model that wants to state that fact has to cite you. IT Master articles are grounded in the customer's own first-party data where it exists, product data, datasheets and support Q&A, for exactly this reason. The page becomes the origin of a fact rather than a restatement of one.

Cite outward as well. Linking to the standard, the regulation or the manufacturer page you relied on costs you nothing and shows the model your claim has a source it can check.

JSON-LD does not make a page rank, and it does not make an assistant cite it. What it does is remove ambiguity for the systems that decide both. Article schema with headline, author, datePublished and dateModified tells a parser who wrote what and when, without inferring it from layout. FAQPage schema pairs each question with its answer in a form that indexes cleanly as a self-contained chunk, which is the same shape the quotable-definition rule produces in prose. Organization and Person schema connect the byline to an entity that exists elsewhere on the web.

The failure mode is schema that disagrees with the page. A dateModified that says June while the visible date says January, an author in the markup who never appears on the page, or FAQ schema for questions the copy never answers: each one is a contradiction, and a parser resolves contradictions by trusting the page less.

Every article IT Master publishes ships with JSON-LD schema and an FAQ block alongside the prose the reader sees. If you write your own markup, validate that it describes the page in front of you and not an earlier draft of it.

A model rarely reads one page of your site in isolation. During a research query it may fetch three or four and try to reconcile them. If your pricing page says one figure, a blog post from last year says another and a product page says a third, the model has no way to tell which is current. It may pick one, and it may decline to state the fact at all and take it from a source whose pages agree with each other.

Consistency is a maintenance discipline rather than a writing one. Keep a single source for every fact that appears in more than one place: names, prices, specifications, dates, the spelling of your own product. Use canonical URLs so near-duplicate pages do not present as separate, slightly different sources. Interlink related pages so that when the model fetches one, the others are a hop away and confirm it.

An llms.txt file at your root is worth an hour on one condition: treat it as a proposed convention, not a documented ranking input. No major search or assistant vendor lists it as one, and Google has said Search does not use it. The llms.txt entry sets out what it is good for.

Tools in this category do four jobs. Visibility checkers ask assistants a set of questions and record whether your domain is cited in the answer. Crawler checkers read robots.txt and server responses to confirm the assistants' bots can fetch your pages. Schema validators confirm that your JSON-LD parses and matches the page. Log analysers show which AI user agents actually visited and what they fetched.

All four are diagnostic. None of them changes a citation on its own, and visibility checkers in particular are noisy: the same question asked twice can return different sources, so a single run tells you little and a weekly series tells you a lot. Ask which assistants a tool queries, whether it uses the assistant's web-search mode or its bare model, because a bare model measures brand fame rather than your pages, and how many prompts make up a score.

IT Master publishes three of these free. The AI Visibility Checker records whether assistants cite your domain for the questions you choose. The AEO Checker reads a page for how quotable it is. The AI Crawler Check reads your robots.txt and tells you which assistants you have blocked.

A citation is a measurable event rather than an impression, so a service selling them should be able to show you the mechanism. Before you buy, ask four questions. Which assistants do you measure, and how often? Does the same model write and check the content, or does a different one check it? Where do the facts come from, my data or the model's memory? And what happens to the pages that stop being cited?

IT Master's answer to each is on the record. The engine finds topics from real search demand and Search Console, researches competitor pages, forums and the customer's own first-party data, writes with a Claude model, then puts the draft through a validation gauntlet run by Gemini and GPT judges: fact-check, novelty against the competing pages, an E-E-A-T critique and AI-tell detection, with up to three revision rounds. Across 186 Standard-tier runs, 51% of drafts passed every check first time; the rest were revised or rejected.

Articles that pass publish to your site with schema, internal links and an FAQ. You pay per published article, see pricing; the full flow is on how it works.

Rankings have a position; citations have a rate. Pick the twenty questions a buyer asks before choosing you, put each to every assistant once a week in its web-search mode, and record whether your domain appears among the cited sources. The share that do is your citation rate, and it moves slowly enough that weekly is the right resolution.

Pair it with Search Console. Both depend on the same retrieval, so a page losing impressions is a page at risk of losing its citations. When a page has lost impressions for a month, refresh it before the citation goes with them.

IT Master's nightly loop does this measurement, Search Console alongside GPT and Perplexity citations, but the loop is not proprietary. Start with the free AI Visibility Checker, run it on a fixed day each week, and keep the results in one sheet.