How to get your brand cited by ChatGPT and other AI assistants
Being cited by an assistant is not the same as ranking. A 2026 factorial study across six models and 252,000 trials found what actually drives the first citation, and it is not formatting.
Key takeaways
- Relevance and position in retrieved context are the primary drivers of citation, ahead of most content rewrites.
- Assistants cannot cite what they cannot fetch: allow AI crawlers and render content server-side.
- Open with a self-contained forty to sixty word answer with your brand named inside the claim.
- Citing authoritative sources inline measurably increases how often you are cited, most of all for pages not already ranking first.
To be cited by an AI assistant your content must be retrievable, relevant to the exact question, and easy to lift as a self-contained claim. A 2026 factorial experiment covering 252,000 trials across six large language models and eighteen variables found that relevance and position in the retrieved context are the primary determinants of the first citation. Moving a source higher in what the model retrieves matters more than most rewrites of the source itself.
- 252,000
- trials in the 2026 factorial study of generative citation across six models
- 115%
- citation visibility lift measured for sites ranked fifth that added inline authoritative sources
- 18.5%
- share of AI referrals attributed to Claude by mid-2026, now the second largest source
- 63%
- ChatGPT's share of AI referrals, down from 89% eight months earlier
Retrieval comes before everything
An assistant cannot cite what it cannot fetch. That makes the unglamorous work decisive: your robots.txt must allow the AI crawler user agents you want citing you, your content must be present in the server-rendered HTML rather than appearing only after hydration, and the page must respond quickly enough not to time out. Most sites that believe they have a generative visibility problem in fact have an access problem, and it can be diagnosed in ten minutes by requesting the page without JavaScript and reading what comes back.
Classic search visibility still feeds this. Several assistants ground their answers in live search results, so ranking well for the query remains one of the most reliable routes into the retrieved set. Generative optimisation does not replace search optimisation; it sits on top of it.
Write claims that can be lifted whole
A model quoting you will take a sentence or two, not a page. If your key claim depends on three paragraphs of build-up, it cannot be extracted without losing its meaning, and it will be passed over for a source that states the same thing cleanly. Open each page with a direct answer of forty to sixty words that would still be true and useful with no surrounding context.
- Attach your brand to the claim inside the sentence, not only in the byline. Models reproduce text; a name inside the quoted sentence travels with it.
- Give figures with a unit, a date and a named source. Unattributed numbers are unusable to a system that has to stand behind what it says.
- State prices explicitly where you can. The 2026 study found measurable effects for explicit prices and recent dates, while formatting changes alone had weak effects.
- Answer the question in the heading, then explain. A heading phrased as a question with the answer immediately beneath it matches how retrieval segments a page.
Cite other people to be cited yourself
One of the more counter-intuitive findings is that pages linking out to authoritative sources inline are themselves cited more often, and the effect is largest for pages that are not already dominant. Sites ranked around fifth saw the greatest visibility gains from adding inline citations. The mechanism is plausible: a page that shows its working reads as more reliable both to a human evaluator and to a model trained on human judgements.
Consistency across the web decides who you are
Assistants assemble a picture of your organisation from many sources. If your description, location, service list and name differ across your site, your profiles and third-party listings, the model has to pick, and it may pick wrong. Publish one canonical description of the business, mark it up as a single Organization entity with a stable identifier, link to the profiles you control, and make the same statements everywhere.
Then check. Ask each assistant directly what your company does, once a month, with the same wording. What comes back is the most honest available measure of whether the work has landed, and the errors it makes will tell you exactly which source needs correcting.
A model quotes sentences, not pages. If your best claim cannot survive being lifted alone, it will not be lifted.
Sources