How to Get Cited by ChatGPT and AI Search

When I tested 48 AI-built startups against ChatGPT, only 4 were recognized by name, yet 28 surfaced when a buyer asked for the best tools in their category. The nine that appeared in neither answer had something in common, and it was rarely the content. It was eligibility: pages the answer engines could not fetch, could not parse, or could not file. Citation advice usually starts at "write better content." The data says start earlier, at whether the machine ever got to read yours.

What follows is the whole path, in the order the machines actually walk it. If you want the study behind those numbers first, it lives in the 48-startup recognition test; this page is the playbook it points to.

Where ChatGPT's citations actually come from

ChatGPT cites pages it retrieves at answer time, not pages it memorized during training. That distinction decides everything you do next. An AI citation is the linked source an assistant attaches to a claim in its answer, and it comes from a live fetch: the model runs a web search mid-conversation, reads a handful of pages, and quotes the ones it can extract an answer from. Training data gives a model vague memory of big brands. Retrieval gives it sources, today, including sources it has never heard of.

That is good news for anyone whose product is younger than the last training run. You do not need the model to know you. You need to win the fetch.

Two OpenAI crawlers matter here, and they are not interchangeable. GPTBot collects pages for training future models. OAI-SearchBot fetches live pages when a user's question triggers a search, and it is the one that produces citations in ChatGPT today. Block GPTBot and the next model generation learns less about you. Block OAI-SearchBot and today's answers cannot cite you at all, no matter what you publish. Perplexity (PerplexityBot) and Google's AI features (Google-Extended, and the ordinary Googlebot feeding AI Overviews) follow the same split between training and retrieval. Plenty of "protect your content from AI" templates block all of them in one line, which is how a site ends up invisible to the exact channel its roadmap promises to win.

The Four Gates to an AI Citation

Here is how to get cited by ChatGPT and AI search, compressed to four steps:

1. Open the gate: allow OAI-SearchBot in robots.txt and serve your content in real HTML.
2. Front-load answers: give every key page a self-contained passage a model can lift whole.
3. Attach evidence: original data, a named author, a visible date.
4. Align the echo: use one plain category description on your site and every external profile.

The Four Gates model says a page earns an AI citation only when it is fetchable, extractable, evidenced, and corroborated, in that order. Every citation failure I have seen in the exam corpus lands at one of those four gates, and the order is strict. A beautifully structured answer behind a blocked crawler is a tree falling in an empty forest. Original data on a page no model can parse is the same. So work the gates in sequence, and do not spend a minute on gate three while gate one is shut.

Gates one and two live entirely on your site and close in an afternoon. Gates three and four take longer, which is exactly why the afternoon should come first.

Gate 1: the crawler has to get in

Retrieval eligibility is binary. Either OAI-SearchBot receives your content when it asks, or your page does not exist for citation purposes. Three checks decide it.

The served HTML has to contain the content. Most AI crawlers, OAI-SearchBot and PerplexityBot included, execute no JavaScript. A page that assembles itself in the browser sends them a title tag and an empty div. In the exam corpus of 37 AI-built sites, nearly three quarters carried at least one failure on this front, and the founders were always surprised, because the page looked complete in every browser they owned. If you check nothing else today, request your homepage with JavaScript off and read what comes back.

robots.txt has to let answer engines through. One exam finding from an AI-built analytics product makes the case better than any argument. The founder wanted AI-driven signups. The template they shipped on had other ideas, and the evidence line in the finding was their own robots.txt:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Disallow: /

The first block is a defensible choice about training. The second one turned off citations entirely: every ChatGPT search that would have quoted their pricing or docs came back empty for them and full for their competitors. The exam flags it as a contradiction between what the site wants and what it tells machines, and ships the correction as a prompt for the agent that built the site:

My robots.txt blocks OAI-SearchBot and PerplexityBot site-wide. Remove the blanket disallow rules for answer-engine crawlers so they can fetch the marketing pages. Keep genuinely private paths (app, api, account) disallowed by path for all agents, not by bot name. Leave the GPTBot decision as a separate, deliberate line. After deploying, fetch /robots.txt and confirm the only disallows left are path-scoped.

llms.txt tells AI readers where to look. llms.txt is a plain text file at your site root that lists what your product is and which pages matter, written for language models rather than search engines. It is young as a convention and adoption among answer engines is still uneven, which I say out loud because most guides oversell it. It costs ten minutes, more than half the AI-built sites we examined lacked one, and when a crawler does read it, it is the cheapest self-description you will ever publish. Write it, keep it honest, move on.

You can test this whole gate in one pass: the free AI readability check reads your site the way an answer engine does and reports what a model can actually determine from the served HTML, before you touch anything.

Gate 2: write passages a model can lift

A model building an answer does not read your page the way a person does. It hunts for a self-contained passage that resolves the question, and it weights the top of the page hard: in Search Engine Land's analysis of the content traits LLMs quote most, 44.2% of ChatGPT citations pointed at the first 30% of the cited page. The single strongest trait among cited pages was what that study calls an answer capsule: a short block, right under the heading, that answers the question completely without needing anything around it.

The mechanics are unglamorous:

• Put the answer first, then justify it. Every section should resolve its heading within the first two sentences. Suspense is for novels; retrieval reads the opening and moves on.
• Shape headings like the questions people ask. A model matching "how do I get cited by chatgpt" to a page full of clever section titles has to guess. Do not make it guess.
• Keep passages self-contained. Forty to eighty words that stand alone quote cleanly. A paragraph that opens with "as we saw above" cannot be lifted without dragging the whole page along.
• Define your terms in one sentence. "X is Y" sentences are the easiest structures for a model to extract and reuse, and they tend to come back verbatim in answers.
• Use tables and lists for anything comparative. Structured blocks survive extraction; prose comparisons get paraphrased into mush.
• Mark up FAQs with FAQPage schema and articles with author and date. Schema does not create citations by itself, but it removes ambiguity about what each block is, and ambiguity is what gets you skipped.

The quick-answer box at the top of this page is the pattern doing its own job: if an assistant quotes this article, that block is the likeliest passage. Structure every important page so its equivalent exists.

Gate 3: give the model a reason to pick you

Passing gates one and two makes you quotable. It does not make you the quote. When five fetched pages all answer the question, the model picks the passages with the most specific support, and specificity is the one axis where a small site can outgun an incumbent.

Original data is the strongest card. Not "studies show": your number, from your corpus, that exists nowhere else. The 48-startup test earned more third-party references than anything else we have published, because anyone writing about AI visibility needs a concrete figure and "4 out of 48 recognized by name" is one. You have an equivalent sitting in your product analytics or your support inbox. One real number beats ten adjectives.

Names and dates do quiet work too. A page signed by a person, with a visible publish date and a real update history, reads as accountable; an anonymous undated page reads as generated. And write claims you can stand behind in declarative sentences. Models quote "X does Y" and skip "X may potentially help with Y", for the boring reason that a hedge is not an answer.

Honest limitation: a young domain will lose evidence contests against Wikipedia and the big review sites for broad questions, whatever it does. The winnable ground is the specific question in your niche where your first-hand data is the best available. Aim there.

Gate 4: the web has to agree on what you are

Answer engines cross-check. A product whose homepage says "AI customer support", whose Crunchbase entry says "conversational commerce", and whose one review listing says "chatbot builder" gives the model three shelves to file it on, and it often picks none. The fix costs nothing but discipline: one plain-language category sentence, used identically on your homepage, your profiles, your directories, and your llms.txt.

Third-party presence compounds slowly after that. A few review-site listings, a correct Crunchbase profile, real mentions in communities where your buyers ask questions. This is the slow gate, measured in months, and the reason startups in our test surfaced by category while remaining unknown by name: the category signal was corroborated, the brand was not. The name-versus-category mechanics, and why the category is the query that pays, are covered in the recognition test writeup.

How to tell if any of it is working

Citation work without measurement turns into superstition fast. Four signals, cheapest first:

1. Ask the questions yourself, on a schedule. Once a month, put your name question ("what is [product]?") and your category question ("best tools for [job]?") to ChatGPT with search enabled. Save the answers somewhere you can diff. Memory is a terrible instrument for noticing gradual change.
2. Watch referral traffic. Visits from chatgpt.com and perplexity.ai show up in your analytics as ordinary referrers. Small absolute numbers, unusually high intent.
3. Read your server logs for answer-engine fetches. OAI-SearchBot hits on a page are the receipt that gate one is open. Zero hits over weeks, on a site with traffic, points back to eligibility.
4. Automate the diff. This is what the AI mirror in Tabkeel exists for: every exam asks a model what your site is, stores the answer word for word, checks it against what your pages claim, and keeps the history, alongside share of voice for your category question. You ship a fix, run it again, and watch the answer move instead of guessing. The weekly scheduled mirror with the email digest is part of the Founder plan; running it by hand comes free with the exam.

The mirror is also where this connects to the rest of your funnel. Being cited brings the visit; what the visitor finds decides the rest, which is why the same exam that checks your citability also walks the other six fronts of an AI-built site, and why the click side of search has its own playbook once Google is sending you impressions. If you want the baseline before touching anything, point the exam at your URL now; it reads the whole surface in about two minutes and the findings arrive with the evidence attached, nothing to install and no signup standing in the way.

Four moves that waste a quarter

Blocking every AI bot "to protect content", then wondering where the citations went. The training question is a real debate. The retrieval question is not, if buyers ask assistants about your category. Separate the two lines in robots.txt and make each a decision instead of a default.

Optimizing for name recognition. Nobody types your brand into ChatGPT until they already know you. The query that converts is the category question, and category presence is winnable years before name recall is. Chasing the wrong one burns the quarter.

Paying for placement. There is no ad slot inside a ChatGPT citation, and services promising guaranteed AI mentions are selling weather. Citations are retrieved, passage by passage, at answer time. The gates are the only lever.

Publishing volume instead of evidence. Twenty thin posts add nothing a model wants to quote. One page with a number nobody else has outperforms them all at gate three, and it is the only kind of page other sites link to, which feeds gate four while you sleep.

Frequently asked questions