Being cited by AI assistants
Answer engines cite pages they can retrieve and trust, so the work is crawlability, clear definitions, sources and consistency across the places that describe you. Track assistants as their own channel and read what the platforms themselves say before trusting a tactic.
17 steps, about 21 hours of work · 10 plays · 2 checked claims
AI search optimization means making a product the one an AI assistant names when a buyer asks it for a recommendation. The assistants search an index, read the pages that rank, and quote passages from them. So it rests on ordinary SEO, plus three things: being fetchable by their bots, writing quotable passages, and appearing on third-party pages.
The concept, explained in full: GEO.
Do it
Every step in the free launch plan that this topic needs, in the order you need them. Each keeps its plan number.
- Can the AI engines actually fetch your pages?6 steps
- How do you write pages an AI will quote?10 steps
- Which third-party pages decide the answer?7 steps
Can the AI engines actually fetch your pages?
Each engine reads a different index. Check the bots and the indexes before writing anything.
1.3.3.1Publish llms.txt
A plain-text map of what you are and which pages state it. Half an hour, then move on: Google says AI Overviews need no AI text file, so it is cheap insurance for the agents and tools that do ask for one, not a ranking lever.
1.3.3.2Put a plain “what this is” paragraph in the first 100 words
Extraction reads the top of the page. Poetry above the definition costs you the citation.
Done when: Read the first 100 words of the homepage body with no page around them, as an extractor would. Quote them. Do they name the product and say plainly what it is?
30mhow to do it
1.3.3.3Record your AI-citation baseline
Ask ChatGPT, Claude, Perplexity and Google AIO “best [your category] tools”. Log the answer verbatim.
Done when: Ask two AI answer engines a question your category owns, and record whether the product was cited. The dated note of what they said is the evidence.
30m
1.3.3.4Confirm the AI search crawlers can actually fetch you, in the logs
Look for the crawlers that decide citations — OAI-SearchBot, Claude-SearchBot, PerplexityBot, Applebot — not GPTBot or ClaudeBot, which gather training data and say nothing about whether an answer can quote you. Cloudflare blocks AI crawlers by default now, so a site can be invisible without anyone choosing it.
Done when: Count the requests from OAI-SearchBot, Claude-SearchBot and PerplexityBot in the server log that returned 2xx. All three must appear, fetching pages rather than only robots.txt. GPTBot and ClaudeBot do not count: they gather training data, not citations.
20m
1.3.3.5Check that Brave has your key pages, and ask it to re-fetch the missing ones
Brave’s index is what Claude’s web search reads. There is no console and no Brave user agent: it crawls only what your robots.txt lets Googlebot crawl. Its submit form re-fetches a page; it does not promise to index a site it has never seen.
Done when: Search site:yourdomain on search.brave.com and list which of the homepage, the pricing page and your three most important pages appear. For each one missing, give the date you submitted it for re-fetching.
1.3.3.6Serve a Markdown version of each content page to agents that ask for it
An agent that asks for Markdown gets the page without menus, scripts or popups, so it reads what you wrote rather than your layout. The same words as the HTML, never extra ones: a version only machines can see is cloaking.
Done when: Request three content pages with the header Accept: text/markdown. For each, give the Content-Type returned, the Vary header, and the canonical URL in its Link header, and say whether the headings match the HTML page. Anything other than text/markdown with Vary: Accept and a canonical pointing at the HTML page fails.
How do you write pages an AI will quote?
An engine lifts a passage, not a page. Write the passage it would want to lift.
3.4.1Publish five pages answering real user questions
Pull the questions verbatim from your support inbox and Phase 1 interviews.
3.4.2Retrofit question-shaped headings on your top 10 pages
Headings that match how people ask are what extraction matches against.
Done when: Count the H2s on your highest-traffic pages and how many are shaped as a question a reader would actually type — ending in a question mark, or opening with what/why/how/when/which. Report both numbers. An extractor matches a question to a heading, so a page of noun phrases answers nothing it was not already ranking for.
3.4.3Publish a definition and stat block per core term
Self-contained, 40–60 words, one citable number. Written to be lifted whole.
Done when: Open a core term's page and read the block under the heading. It must answer 'what is X?' in 40-60 words as a passage that stands alone, and be followed immediately by one dated statistic with its source. Quote the block and count its words — a definition an engine cannot lift whole, or a number with no date and no source, does not count.
3.4.4Earn genuine mentions on Reddit and Hacker News
Both are indexed heavily by assistants. Genuine is doing the work; astroturf gets caught.
3.4.5Write the twenty prompts you will re-ask forever
Prompts are keywords. Six shapes carry the decision: category, use case, comparison, alternative, “is X actually any good”, and the one with the buyer’s constraints in it — must do this, must connect to that, under this price. The last is where a shortlist is actually cut.
Done when: List the prompts. Twenty or more, written as a buyer would type them, and recorded somewhere their wording will not drift — a prompt reworded between runs measures the rewording. The list must include queries you expect to LOSE, or the score can only ever go up.
1h
3.4.6Re-ask the whole portfolio monthly and score the answers
Three clean sessions per prompt, with the model version logged: one answer a month is variance, not movement. Then score position, recommendation, accuracy and the sources behind it.
Done when: Report how many of the twenty prompts named the product and how many CITED it with a link, across the engines, for this month's run. Named and cited are different numbers and both are needed — one is training data bleeding through, the other is the retrieval layer finding a page.
2h
3.4.7Read the prompts you lose as a to-do list
Whoever gets recommended instead has something you do not — a category page, a review count, a comparison, a thread. Name that asset, then go and build it. On a prompt with constraints in it, check the cheaper failure first: often the product qualifies and the fact that proves it is written down nowhere a model can reach.
45m
3.4.8Mine Bing’s grounding queries for the prompts you are already cited on
The AI Performance report counts your citations in Copilot and Bing’s AI summaries, page by page, and lists the grounding queries — what the model actually searched when it used you. It is the only first-party record of an AI engine’s own queries; fold them into the portfolio.
Done when: List the grounding queries the AI Performance report in Bing Webmaster Tools shows for the last three months, each beside the page of yours it cited.
3.4.9Read which pages AI Overviews and AI Mode pull in, in Search Console
The generative AI report counts impressions in AI Overviews and AI Mode by page, country and device — no clicks and no queries, so read it for which pages Google’s answers draw on, not for traffic. Low everywhere can mean the site is opted out: check that setting before blaming the content.
Done when: List the five pages with the most impressions in the Search Console generative AI performance report over the last three months, each with its impression count.
3.4.10Cite a source beside every claim, and quote a named person, on the pages you want quoted
In the Princeton GEO study, citing sources and adding quotations raised a page’s visibility in generative answers as much as adding statistics did. A link at the end of the page does not count: the source goes in the sentence that makes the claim.
Done when: Open your five most important pages. For each, count the factual claims, how many link their source in the same sentence or paragraph, and the quotes attributed to a named person. Report the three numbers per page.
Which third-party pages decide the answer?
Most of the brands an assistant names come from other people’s lists, reviews and threads.
3.10.1List every third-party page that keeps appearing
Run your ten highest-intent searches and ten prompts and write down what comes back: threads, listicles, review categories, videos. The repeats are the map.
2h
3.10.2Rank each source by how often it appears and how much it sways
A thread cited across nine of your queries beats a DR 80 mention nobody buying reads. Work only the handful scoring high on both.
45m
3.10.3Earn a place on the top five, one at a time
Find out who owns each one and how it earns before you spend the afternoon: many ranking listicles are affiliate properties, where the price is a commission and no pitch will move them.
8h
3.10.4Make every source describe you the same way
One category sentence, repeated by your site, your listings, your reviewers and your press. Consensus is what turns a claim into something a model will repeat.
3.2.3.1Create the Crunchbase profile
Heavily weighted in AI training corpora and in “who is this company” lookups.
40mCrunchbaseDR 91 · nofollow
3.2.3.2Create the LinkedIn company page
Required before you can run any LinkedIn distribution in Phase 3.
40mLinkedIn PagesDR 99
3.2.3.3Create the Wikidata item
The structured spine most knowledge panels and assistants resolve entities against.
1hWikidataDR 92
More steps on this topic
Know where signups actually come from
- Read the AI Assistant channel as its own row15m
GA4 files ChatGPT, Gemini, Copilot and Grok referrals under AI Assistant already — but not AI Overviews or AI Mode, which stay inside Organic Search. Judge the row on conversion rate, not sessions: it is the one place on your own property where “the assistant pre-qualified them” is checkable, and it counts the clicks only, never the reads.
Plays
Reusable patterns that explain how: backed by read sources first, then practitioner advice, then our own ideas.
- Build the page an answer engine can quoteBacked · 3 sources
Write the page so any passage still reads correctly once it is lifted out: a plain definition and one dated number at the top, comparisons as HTML text rather than images, a source linked beside each claim, quotes from named people, and a real author in the markup. Then make sure the search indexes behind the engines can fetch it.
- Ship llms.txt: a crib sheet for AI crawlersBacked · 2 sources
Add an llms.txt file and a plain-language facts page stating what the product is, who it serves, pricing, and honest differences from alternatives, written in flat quotable markdown. AI crawlers assemble answers from whatever they can parse; giving them a canonical crib sheet makes the version they repeat the version you wrote.
- Hand agents the Markdown, not the layoutBacked · 2 sources
Serve a Markdown copy of each content page when a request asks for text/markdown, or at the same path plus .md. Keep the words identical to the HTML, send Vary: Accept so caches keep the two apart, and name the HTML page as canonical in the Link header. Agents get a clean read, and nothing competes with the original.
- An FAQ page built from the questions buyers actually askBacked · 1 source
Collect the real questions from sales calls, support tickets, and community threads, then answer each on an FAQ page in self-contained 40-60 word blocks with FAQPage structured data. AI engines answer questions by extracting passages; pages organized as question-plus-liftable-answer are the easiest content for them to quote — with your name attached.
- Retrofit your best pages into question-and-answer shapeBacked · 1 source
Restructure existing high-traffic pages so each section heading is a real buyer question and the first 40-60 words beneath it answer that question completely. AI engines quote passages, not pages; converting prose into question-plus-standalone-answer units makes content you already own extractable without writing anything new.
- Definition + statistic block engineered for AI citationPractitioner · 1 source
Lead a page with a 40-60 word definition that answers "what is [category]?" as a standalone passage, then attach one original, dated statistic with a citation. AI engines extract passages, not pages, and citing sources plus statistics are the two highest-leverage moves for getting quoted in AI answers.
- Author the honest "best [category] tools" list yourselfPractitioner · 1 source
Publish a genuinely fair "best [category] tools" ranking on your own domain, including real competitors, their strengths, and who each option suits, with yourself placed honestly. AI engines and searchers both favor ranked lists; authoring the definitive one means the comparison happens on your turf, framed by your criteria.
- Run the benchmark your category argues aboutOur idea
Build and publish a reproducible benchmark answering the performance question your category argues about — speed, cost, accuracy — with the method, raw data, and a standing update cadence. Journalists and AI engines both reach for concrete comparative numbers, and the benchmark’s maintainer becomes the reference everyone cites, including when competitors win a round.
- Say the same sentence everywhere a machine looksOur idea
Write one sentence naming your category and your buyer, then propagate it verbatim across your homepage, listings, review profiles and community answers. Retrieval systems assemble a picture of you from many sources at once; when they all agree, the claim reads as consensus, and when they contradict each other the model hedges or leaves you out.
- Write the canonical answer to the sub’s weekly questionOur idea
Identify the question your target subreddit asks every week, then write the definitive answer as a structured text post — thorough, current, and free of links to you. Recurring questions guarantee recurring traffic; the canonical answer gets bookmarked, linked by regulars, and quietly makes its author the recognized expert.
What the evidence says
Sources
Founder questions
Do you need llms.txt to appear in AI answers?
No. Google says AI Overviews and AI Mode need no special file, and they read the classic index. An llms.txt takes half an hour and helps some agents and tools, but it is not a ranking lever.
How do you measure AI search visibility?
Bing Webmaster Tools reports your citations in Copilot and the queries the AI searched, and Search Console reports impressions in AI Overviews and AI Mode. For ChatGPT, Claude and Perplexity, ask a fixed set of buyer prompts by hand each month.
Is ranking yourself first in your own “best tools” list a good idea?
No. Assistants and readers both discount a list whose author ranks itself first. Rank honestly on published criteria, say which tool is yours, and work on getting into other people’s lists.
Related topics
Go further
Where it sits in the plan
- 1.3.3Get ready to be quoted by answer enginesin Ship-ready
- 3.4Become citable by answer enginesin Authority baseline
- 4.1.7Read the AI Assistant channel as its own rowin Find the one channel that works
Sources
- Google, Optimizing for generative AI features on Google Search (Google Search Central)
- Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan & Ameet Deshpande, GEO: Generative Engine Optimization (KDD, 2024)
- Google, Search Central: SEO documentation & quality guidelines
Compiled by The Marketinque Team to our editorial standards.