How to Get Cited by ChatGPT (Citation Mechanics)

ChatGPT cites your page when its live-search pipeline retrieves your content and then picks it out of the roughly 15% of retrieved pages it actually names, rather than the much larger set it reads silently and never attributes. To get cited by ChatGPT, a page has to clear four steps in order: OAI-SearchBot has to index it, a retrieval pass has to pull it for the query, a relevance step has to score it near the top, and a final selection step has to name it. One fork sits in front of all of that, and most guides skip it. If ChatGPT answers from static training data, nothing was retrieved and there is no real source to cite. If it runs a live search, a Sources button with clickable links appears, and only then are the mechanics below in play.

We write this blog with an answer-first, gap-analysis editorial process, and that shaped this post. Most pages ranking for this topic either summarize a bigger report without its methodology or repeat a reverse-engineered ranking framework whose own author admits it is unconfirmed. This one holds a hard line between what OpenAI has documented and what third parties have inferred. If you landed here without the base definitions, generative engine optimization is the entry point. Everything after it is the ChatGPT citation mechanics underneath.

Training-Data Answers vs. Live-Search Answers: the Fork Every Citation Depends On

ChatGPT answers in one of two modes, and only one of them can cite anything.

The default mode answers from parametric memory: the patterns baked into the model's weights during training. GPTBot collected that training data, and OpenAI's documentation says GPTBot "is used to crawl content that may be used in training our generative AI foundation models." Once training ends, that knowledge is frozen to a cutoff date. When ChatGPT answers from it, nothing is fetched in real time, so there is no live source behind the sentence. What looks like a fact is recall, and recall carries no citation.

Live-search mode runs when ChatGPT decides a query needs current information. Only then does the retrieval pipeline fire against OAI-SearchBot's index, and only then does a Sources button with real, clickable links show up. OpenAI describes OAI-SearchBot as the crawler that "is used to surface websites in search results in ChatGPT's search features."

The practical consequence is simple. A citation only exists when live search ran. Every mechanic in the rest of this guide describes the live-search path, because it is the only path that produces a real citation. The strategic question of which content type to prioritize given this fork belongs to the program in ChatGPT SEO: How Brands Get Recommended. Here we stay on the internals.

The Crawl-to-Citation Pipeline, Stage by Stage

Five stages sit between your published page and a citation in a ChatGPT answer: crawl, index, retrieval, relevance scoring, and synthesis with citation. OpenAI has documented the first two directly. The scoring internals in the middle come from third-party teardowns, not from OpenAI, and this section flags which is which at every step.

Crawl: OpenAI's Three Named Crawlers

Three OpenAI crawlers touch this pipeline, and they do different jobs. Confusing them is the most common mistake on this topic, because two of them sound interchangeable and are not.

Source: OpenAI crawler documentation, developers.openai.com/api/docs/bots, accessed July 7, 2026.

For citation, OAI-SearchBot is the crawler that counts. It builds the index ChatGPT Search reads from, and OpenAI states plainly that sites opting out of it "will not be shown in ChatGPT search answers." GPTBot governs training use only; disallowing it, in OpenAI's words, "indicates a site's content should not be used in training generative AI foundation models." ChatGPT-User is different again. It fetches a page live when a user's prompt sends it there, it "is not used for crawling the web in an automatic fashion," and it "is not used to determine whether content may appear in Search."

This stage is also where a page can fail before any ranking factor is relevant. AI crawlers do not execute JavaScript, so a page that renders its content client-side can hand the crawler an empty shell. We ran into this ourselves: we migrated our own React single-page app to prerendered static HTML for 20 marketing pages because AI crawlers do not execute JavaScript. If OAI-SearchBot cannot read the content, the page never becomes a candidate, and none of the scoring that follows can rescue it.

Index: What OAI-SearchBot Builds

OAI-SearchBot's crawl feeds ChatGPT's own search index. ChatGPT Search does not rely on that index alone; OpenAI has said it partners with outside search providers for web coverage, and Microsoft's Bing is widely reported as the primary one. The takeaway for citation is direct: your page has to sit in an index ChatGPT can query. Being crawlable by OAI-SearchBot is the front-door path in, and conventional search visibility feeds the same retrieval step through the partner index. The same crawlability precondition holds across every AI engine, not just ChatGPT; the cross-engine version is in our LLM optimization guide.

Retrieval: Where One Prompt Becomes Many Queries

Retrieval is where a single prompt turns into a stack of searches. AirOps' March 12, 2026 report analyzed 15,000 original prompts and found ChatGPT rarely searches for just what you typed. It generated two or more fan-out queries on 89.6% of searches, expanding 15,000 prompts into 43,233 total queries. Derived from AirOps' own before-and-after query counts, that is roughly 2.88x (43,233 ÷ 15,000), so one prompt sets off close to three searches on average before any citation decision is made.

Fan-out changes which pages get pulled. AirOps found 32.9% of cited pages were discovered only through a fan-out query, never the original prompt. A third of citations came from searches the user never typed. Optimize only for the literal prompt and you miss the query set ChatGPT actually runs.

Relevance Scoring: the Least Documented Stage

What happens to a retrieved page next is the part OpenAI has not published, and it deserves an explicit hedge. The clearest public account comes from one third-party technical teardown by Rankly (March 11, 2026), and it is inference, not confirmation. Rankly describes ChatGPT search filtering roughly 40 to 50 retrieved URLs down to 10 to 20 candidates, fetching each with a short timeout, splitting the content into 128-token chunks, scoring those chunks by embedding similarity against the query, and selecting 3 to 5 sources for the answer based on the single best-scoring "audition chunk" per page.

Treat every number in that paragraph as reverse-engineered. Chunk size, timeout length, and the scoring method are not OpenAI-confirmed. The reason it is worth stating at all is the design implication, not the exact figures. If a page is scored on its strongest short passage rather than its whole body, then self-contained, quotable passages matter more than total word count. That is a hypothesis you can act on without treating "128 tokens" as fact.

Synthesis and Citation

In the last stage, the model drafts an answer from the top-scoring passages and surfaces a subset of them as inline, clickable citations. Two things happen here that explain most citation frustration. First, the model reads more than it shows, so passages that shaped the wording of an answer can go uncited. Second, the citation set is small, usually a handful of sources, so the bar to become one of them is high. This is where the real question, how does ChatGPT cite sources, turns from a mechanics question into a selection one, which the next two sections take apart.

Why Retrieval Isn't Citation: the 15% Gap

ChatGPT citations are far rarer than retrieval, and AirOps measured exactly how much rarer. This is the ChatGPT retrieval vs citation distinction, and the report put real numbers on it. Across 15,000 prompts, ChatGPT retrieved 548,534 pages and cited only 15% of them: 82,108 citations. The other 466,426 retrieved pages shaped answers without ever being named.

That ratio is worth stating on its own. Derived from AirOps' own retrieval and citation counts, for every page ChatGPT cites, roughly 5.7 more get pulled into retrieval and never named (466,426 ÷ 82,108 ≈ 5.68). Retrieval is common. Citation is rare.

A second study, run independently and measuring a different unit, points the same way. Profound analyzed roughly 730,000 real ChatGPT.com conversations from October to December 2025 (U.S., English, each containing at least one web citation). Instead of page-level retrieval odds, it measured citation likelihood by conversation turn. The first turn produced a citation 12.6% of the time. By turn 10 that fell to 4.5%, and by turn 20 to 3.0%. A citation is roughly 2.8 times more likely on the opening turn than on turn 10, and about four times more likely than on turn 20.

The two studies measure different things and must not be blended. One counts pages; the other counts conversation turns. Set side by side, they agree on the shape: citation is scarce and front-loaded.

Profound's data also shows how unevenly AI citations spread. Concentration runs high (a Gini coefficient of 0.8), yet the top 10 cited domains together capture only 12% of all citations. Both facts hold at once because there is an enormous long tail: most domains are cited rarely, a few are cited often, and no small club dominates. Wikipedia leads at 5%, Reddit follows at 3%, then Reuters and NIH at 1% each. There is room in the tail, and that is where most sites compete.

One Number That Circulates Wrong: Wikipedia's Citation Share

This is also the right place for a concrete "verify before repeating" example, because the same Profound study is the source of a number that circulates incorrectly. A figure claiming "Wikipedia accounts for 7.8% of citations," attributed secondhand to Profound, shows up in other write-ups on this topic. Profound's own dated primary page (Brandon Punturo, Research Lead, published February 3, 2026) says something different and more specific: Wikipedia is 5% of all citations and appears in 18% of conversations that carry any citation at all. The 7.8% figure does not match Profound's published number, so we name the discrepancy rather than repeat either version as settled.

It gets slipperier, which is the point. A separate, larger analysis (Averi/Profound, 680 million AI citations, August 2024 to June 2025, published January 7, 2026) puts Wikipedia at 47.9% of ChatGPT's top-10 citation sources. That is a third Wikipedia number, and it is not a contradiction of the 5% figure. It measures a different thing over a different window: share of the top-10 source mix across a huge citation set, versus per-conversation share in a three-month sample. The 47.9%, the 5%/18%, and the secondhand 7.8% are not interchangeable, and averaging them produces nonsense. "Wikipedia's citation share" means nothing until you say which of them you are counting.

How ChatGPT Selects and Ranks Sources for a Given Prompt

The practical question behind all of this is blunt: how does ChatGPT choose sources for a given prompt? Part of the answer is backed by disclosed methodology and real sample sizes. Much of what circulates online is reverse-engineered guesswork dressed up as a formula. We keep those in separate columns and use only the verified one for factual claims.

The verified findings reward a close read, because they are the only ones with a sample behind them. AirOps' report shows Google ranking correlates strongly with citation: pages ranking #1 in Google were cited 3.5 times more often than pages beyond position 20 (a 43.2% citation rate versus roughly 12%). Title-query overlap matters too. Pages whose title overlapped the query by 50% or more were cited 20.1% of the time, against 9.3% for pages under 10% overlap, a 2.2x lift (20.1 ÷ 9.3 ≈ 2.2). Domain Authority clusters in a middle band rather than at the ceiling: 63.6% of citations went to sites with an estimated DA between 20 and 80, not overwhelmingly to the biggest domains. And higher-readability pages, at a Flesch Reading Ease of 50 or more, were cited more often than harder-to-read ones.

The unverified column exists to name what is circulating, not to endorse it. ZipTie's widely-referenced "40% domain authority, 35% content quality, 25% platform trust" weighting reads like a settled ranking formula wherever it gets quoted, but ZipTie's own article calls it reverse-engineered and "not officially confirmed by OpenAI." A secondary summary of AirOps' report floats per-DA-tier average citation counts (8.4 versus 1.6) and a "content updated within three months is 3x more likely to be cited" freshness claim; neither appears in AirOps' own disclosed findings, so we leave both out. The standard we hold is one line long: verified numbers drive the guidance, flagged numbers get named and nothing more.

Page-Level Engineering to Get Cited by ChatGPT

None of the mechanics above are levers you pull directly. What you control is the page, and a few page-level moves line up with the verified findings instead of the guesswork.

Match your headings to how people actually phrase the query, then answer directly underneath. This ties straight to AirOps' title-query-overlap finding: the pages cited most were the ones whose titles overlapped the query. A heading that mirrors a real question, with a self-contained answer in the first sentence or two below it, gives the retrieval step an obvious match and the synthesis step a clean passage to quote.

Write short, self-contained passages. This one is an inference from the chunk-level scoring described earlier, not a confirmed rule, so hold it as a design choice. If pages are scored on their best short passage, a page built from tight, standalone chunks has more chances to produce a high-scoring one than a wall of context-dependent prose.

Make the page crawlable, because crawlability is a precondition, not a ranking factor. If OAI-SearchBot cannot index the page, ranking never enters the picture. One direct move is to hand AI crawlers a clean, plain-text version of your key information. We publish our own llms.txt and llms-full.txt for exactly this reason: a curated, crawlable knowledge base an AI crawler can read without fighting through markup. You can check whether your own file is present and valid with our free llms.txt checker.

Keep ranking in conventional search healthy. AirOps' data makes Google position the single strongest verified correlate of citation in the set, so the SEO work you already do feeds the same retrieval index ChatGPT draws on. That correlation is measured, not assumed.

What This Means for Your Content

The mechanics come down to a short list. A citation only exists in live-search mode. Your page has to be crawlable by OAI-SearchBot before anything else. Retrieval pulls far more pages than it cites, so being retrieved is table stakes, not the finish line. Google rank and title-query match are the strongest verified pulls toward citation, and short, self-contained passages give the synthesis step something clean to name.

Turning that into a repeatable process is a strategy job, not a mechanics one, and it lives in ChatGPT SEO: How Brands Get Recommended. To find out whether any of it is actually working for you, How to Track ChatGPT Mentions of Your Brand covers the measurement side. If you want more sourced numbers on how this behaves across AI engines, our AI SEO statistics roundup collects them.