How to Get Cited by ChatGPT, Perplexity and Google AI Mode

You have a page sitting on the first page of Google for a term you care about. You ask ChatGPT the same question and it cites three other sites, none of them yours. Nothing is broken and no amount of link building will fix it, because the two systems are not doing the same job. Here is how to get cited by AI search in one sentence: publish self-contained passages that completely answer one specific question, on a site whose crawlers are not blocked and whose identity is consistent enough to be recognised. Ranking gets you into the candidate pool. Passage quality decides who gets quoted.

That difference drives every tactic in this article. If you only take one idea away, take this one: the unit of competition in AI search is the paragraph, not the URL.

How do AI assistants decide what to cite?

Retrieval-augmented answering is the process where a system searches an index for relevant text, feeds the retrieved text to a language model, and asks the model to write an answer grounded in that text. ChatGPT search, Perplexity, Google AI Mode and Gemini all work this way in broad outline, and the differences between them matter less than the shape they share.

Three consequences follow immediately. First, one answer typically cites several domains because several sub-queries each had a different winner, so you are not trying to beat everyone, you are trying to own one sub-query. Second, your page enters the shortlist through a chunk, and the chunk is judged without the rest of your page around it. Third, an answer that cannot be traced back to a specific passage usually does not get a citation at all, which is why vague, hedge-heavy writing gets read and discarded.

Passage independence is the whole skill

Passage independence is the property of a paragraph that still makes complete sense, and still says something specific and attributable, when it is lifted out of the page and shown to a reader who has seen none of the surrounding content. It is the single characteristic that separates pages that get cited from pages that merely rank.

Test it the cheap way. Copy any paragraph from your page into a blank document. Read it as a stranger. Does it name its subject, or does it start with "this", "it", "the second option", "as we saw above"? Does it contain a claim someone could quote, or is it a transition sentence doing structural work? Most marketing pages fail this on eight paragraphs out of ten, and the failures are almost always pronouns and back-references rather than bad thinking.

Notice what the right-hand column does. Each line names its subject, states one claim, and carries enough context to be quoted by a stranger. None of them are longer than a normal sentence or two. Passage independence is not about writing more, it is about not leaning on the paragraph above.

Which writing patterns actually get quoted

Some structures get extracted far more often than others, and they are the same structures that make a page easier for a human in a hurry. That is not a coincidence. Retrieval systems reward text that is already shaped like an answer.

One warning about tables. They work when the cells are short and self-describing. A table where the first column is "Option 1, Option 2, Option 3" and the meaning lives in the paragraph above it is worse than no table, because it gets retrieved and then makes no sense.

Owning one fact beats covering one topic

The most common content strategy in 2026 is still "write the most thorough page on the topic". For AI citations it is close to the worst possible use of a week. A 4,000-word ultimate guide competes against every other ultimate guide, and because it covers everything at the same depth, it usually contains no passage that is uniquely the best answer to anything. Thoroughness is a page-level virtue in a passage-level game.

What wins citations is being the primary source of one specific, checkable thing. If your page is the only place that states a particular fact clearly, every sub-query that needs that fact has exactly one candidate. That is a much easier competition to win than "best guide to insurance affiliate marketing".

• Publish your own measurements. Numbers you generated from your own operations are, by definition, unavailable anywhere else. Say how you measured them, over what period, and what the sample was. A small honest sample described precisely beats a large vague one.
• Name a framework. A named checklist or model gives retrieval a specific entity to match on. "The lift test" in this article is a named thing, and named things get quoted with attribution far more readily than unnamed ideas.
• Answer the question everyone asks on sales calls. The questions people are too embarrassed to search are exactly the ones with thin coverage and high intent. Write down the five you answer most often and publish the answers verbatim.
• Be the honest one about numbers. Most pages in a competitive result copy the same figure from each other with no source. A page that gives a range, explains what moves it, and tells the reader how to verify it themselves becomes the one worth citing when the underlying number is genuinely uncertain, which is how our own auto insurance lead cost breakdown is written.
• Date your claims inside the passage. Write "as of September 2026" in the sentence, not just in the byline, because the passage travels without the byline attached. Anything tied to a price, a regulation or a platform behaviour needs an in-text date and a real review schedule.

Freshness cuts both ways. Assistants prefer recent material for anything that changes, so an undated 2023 page about platform behaviour is a liability. But republishing with a new date and no changed content is a transparent trick, and it costs you the trust of readers who notice. Change the content or leave the date alone.

Can an AI system tell who you are?

An entity, in retrieval terms, is a thing the system can recognise and hold facts about, rather than a string of characters it happened to match. Brands that are entities get cited by name, described accurately and recommended in answers to "who does X". Brands that are only strings get their content used without attribution, or not retrieved at all.

The work here is dull and mostly one-off. Use one name for the company everywhere, spelled identically, including in third-party profiles and directory listings. Have an about page whose first sentence states plainly what the company is and does, in the form a machine can extract: "DL Minds is a digital agency based in India that builds websites, apps and AI systems, and runs insurance lead generation for clients." Put Organization schema on the site with a sameAs array pointing at your verified profiles, which we cover in detail in the companion piece on schema markup for AI search citations.

Third-party description matters too. What other sites say about you is training and retrieval material in a way your own copy is not, so a consistent one-line description that you supply to every directory, podcast bio and partner page compounds. This is the slowest of all the levers here and the one most worth starting today.

Should you let AI crawlers read your site?

This is the most consequential and least discussed decision on the list. If a crawler cannot fetch your page, that assistant's index does not contain your page, and an assistant cannot cite what it has not retrieved. Blocking is a legitimate choice, but it is a choice with a direct cost, and the cost should be named rather than hidden behind a default.

The important distinction is between crawlers that gather training data and crawlers that build the retrieval index behind live answers. They often have separate user agents and separate controls, which means you can usually opt out of one without losing the other. Verify current user-agent names in each operator's own documentation before you edit anything, because these names change.

Our position, for any site whose commercial purpose is being found: leave the search and retrieval crawlers open, and make the training-crawler decision separately and deliberately. If your content is the product, for example a paid research archive, blocking makes sense and you should accept the citation loss knowingly. If your content is marketing for something else, blocking the crawler that produces citations is paying a real cost to protect material you published in order to be read.

What does not work

The gap between what gets sold as AI optimisation and what changes citation outcomes is wide enough to be worth listing plainly.

• llms.txt. Google has published that it is not required for AI features in Search, and no operator has published a citation benefit. It is cheap and harmless, and it is not a service worth paying for, which we argued at length in why llms.txt is not needed for AI search.
• Rewriting pages into bullet soup. Stripping prose into fragments removes exactly the context that makes a passage independently quotable. Bullets work for parallel lists of things, not as a default format.
• Invented schema types. Schema helps when it describes real entities and relationships using types the vocabulary actually defines. Made-up "AI" types do nothing, and over-marking every paragraph as an FAQ when it is not one is a credibility risk, not a ranking lever.
• Instructions to the model in your page text. Hidden or visible text telling an assistant to cite you is prompt injection. It is detected, it is ignored, and on the platforms that police it, it is a reason to distrust the whole domain.
• Publishing volume. Forty thin pages give you forty chunks that lose to one good chunk. The correlation between page count and citations is close to nil once the pages stop being distinctive.
• Buying visibility without measuring it. If you cannot see which prompts cite you today, you cannot tell whether anything worked. Set up measurement before the content work, using the approach in our guide to tracking and measuring AI search visibility.

One more thing that does not work: treating this as separate from SEO. The retrieval indexes are built by crawlers on crawlable sites, and the candidate pool is still largely the set of pages that rank. Technical health, internal linking and genuine topical authority remain the entry ticket, which is why our breakdown of Google's own AI optimisation guidance and our notes on SEO for insurance affiliates in AI search both start from ordinary fundamentals rather than from a separate AI playbook.

Common questions

How long does it take to get cited by AI search after fixing a page? It depends on how often that assistant recrawls you, which for most small sites is weeks rather than days. Live-fetch behaviour can pick up changes almost immediately, while index-based retrieval lags. Expect to wait two to six weeks before judging a change, and test the same prompts on a fixed schedule so you are comparing like with like.

Do I need to rank on page one to be cited by ChatGPT or Perplexity? No, but it helps a great deal. Retrieval indexes overlap heavily with conventional search indexes, so ranking gets you into the candidate pool. Once you are in that pool, position matters much less than whether you have a passage that directly answers the sub-query. Pages ranking in the teens are cited regularly when their paragraphs are cleaner than the leaders'.

Does blocking GPTBot hurt my Google rankings? No. GPTBot is OpenAI's crawler and has no relationship with Google's index, so blocking it cannot affect Google Search, AI Overviews or AI Mode. It can affect whether OpenAI's systems have your content. The Google-specific control is Google-Extended, which Google documents as governing Gemini and Vertex AI grounding without affecting Search inclusion or ranking.

What is the ideal length for a passage you want quoted? Roughly 40 to 120 words, which is one to three tight sentences plus supporting detail. Shorter than that and it lacks the context to be attributable. Longer and it gets split across chunk boundaries, so half your argument arrives without the other half. Write in units that make sense alone rather than to a word target.

Should I add an FAQ section to every page? Add one where real questions exist and answer them properly in 40 to 80 words each. FAQ blocks work because each answer is already an isolated chunk mapped to one question. They stop working when they are padded with invented questions nobody asks, which produces low-value chunks that compete with your own good ones.

How do I know whether I am actually being cited? Test a fixed list of prompts across ChatGPT, Perplexity, Google AI Mode and Gemini on a regular schedule and log which domains appear. Referral traffic from assistant domains in your analytics is a second signal, though it undercounts badly because most people read the answer without clicking. Set the baseline before you start changing pages.