How do I get cited by AI assistants?

Getting cited starts before anything about writing quality: a page has to be retrievable at all. Most AI assistants either crawl the web directly, similar to a search engine, or call a live search API and read back a shortlist of results. If a page is blocked by robots.txt, returns errors, loads its real content only after heavy client-side JavaScript execution that a crawler does not run, or simply is not indexed anywhere, it cannot be retrieved, and none of the content-level advice that follows matters until that is fixed. This is the same technical floor traditional SEO already requires, and it is the most common reason a business sees zero AI citations despite writing decent content.

Once a page is retrievable, the next filter is chunking. Retrieval systems break pages into pieces, commonly by heading and paragraph boundaries or by a fixed token window, and each chunk is embedded and matched against a query independently of the rest of the page. A chunk that contains a full, self-contained thought, a clear subject, and a specific claim is far more useful to a retriever than a chunk that only makes sense in the context of the paragraph before it. Practically, this means writing paragraphs that could be read in isolation and still make sense: state who or what you are talking about by name rather than with a pronoun, state the claim, and support it in the same short block.

Directness matters more than cleverness here. A model deciding what to cite is effectively pattern-matching for a statement that answers the query with minimal ambiguity. A paragraph that opens with the actual answer, in plain language, followed by supporting detail, gives the retriever and the generation model an easy, low-risk quote. A paragraph that opens with a rhetorical question, an anecdote, or marketing framing before getting to the point buries the useful chunk under material the model has to work harder to use, and it often skips it in favor of a competitor's more direct phrasing.

Structured data is a second, independent channel that reinforces the prose rather than replacing it. Schema markup such as Organization, FAQPage, HowTo, and Article types gives a machine-readable statement of facts that sits alongside the visible text. When the schema and the prose agree, a model gets two corroborating signals instead of one, which increases confidence in citing the claim. Schema that contradicts the visible page, or schema that is present purely for the sake of having markup without matching real content, does not help and can actively confuse a system trying to reconcile the two.

Trust signals beyond the page itself also factor into citation selection. Retrieval and generation systems appear to weight source credibility using signals like clear authorship, visible publication and update dates, consistency with other sources on the same topic across the web, and the general reputation of the domain. A single isolated page making an unusual claim with no supporting context anywhere else is a riskier citation than a claim that is corroborated across multiple credible sources, including the business's own site, its schema, and third-party mentions. This is part of why entity clarity, meaning a consistent, unambiguous description of who a business is across its site, its Google Business Profile, and other properties, feeds into AI citation as much as it feeds into traditional search.

It also helps to write for the actual phrasing people use when they talk to an assistant. People tend to ask AI systems full questions in natural language, closer to how they would ask a person, rather than the fragment-style queries typed into a search box. Headings and section titles that mirror that phrasing, framed as direct questions, make the semantic match between query and chunk stronger, which improves the odds of retrieval before citation is even a question.

None of this produces a guarantee, and it is worth being blunt about that. Retrieval and citation logic inside ChatGPT, Perplexity, Google AI Overviews, and other systems is proprietary, opaque, and changes without notice. A page can be well structured, technically sound, and well marked up and still not get cited on a given query, because the model chose a different source, or because the index simply had not picked up a recent change yet. What this approach does reliably is remove the avoidable reasons a page gets skipped, which is the realistic target.

The last piece is patience paired with verification. Because there is no stable rank to check, improvement shows up as a trend across repeated testing over weeks, not a single before-and-after snapshot. Running the same handful of representative queries against assistants on a regular cadence, and watching for the business to start appearing, being named correctly, or being linked as a source, is the only honest way to know whether the underlying changes are working.