How to get cited by ChatGPT: build sources worth retrieving

There is no citation switch

ChatGPT may use training data, live search, retrieval systems, or other sources depending on the product and prompt. Citation work therefore starts with the public sources those systems can discover and evaluate.

A business improves its citation readiness by making its information accessible, specific, consistent, and corroborated across its public presence.

How often live search even enters the picture is worth knowing before investing heavily in citation work. Semrush's analysis of more than a billion lines of U.S. clickstream data found ChatGPT enabled its web-search feature on about 34.5% of queries as of February 2026 — down from roughly 46% in late 2024 — meaning most responses still draw on the model's training data alone, with no live retrieval and therefore no live citation opportunity at all. Search tends to trigger for four reasons: the user turns it on explicitly, the user asks for sources, the model is uncertain, or the question involves information after the training cutoff. A buying question about current pricing, availability, or a recent product change is exactly the kind of prompt likely to trigger a live search; a general definitional question is more likely to be answered from memory, where the more relevant target is being present in the training corpus rather than being retrieved live.

This means citation strategy actually has two separate audiences to win: the live retrieval system, which behaves like a search engine and rewards freshness and accessibility, and the pretraining process, which behaves like publishing a book and rewards being widely and consistently stated across the web long enough before the next model's training cutoff. The first is something you can influence within weeks. The second is not observable or controllable on any useful timeline — nobody knows which pages a future model will be trained on, and no vendor can promise inclusion in a training run.

Publish information only your business can provide

Generic summaries compete with thousands of similar pages. First-party material gives a system a reason to select the source: product specifications, operating methods, original data, expert explanations, policies, tested comparisons, and documented outcomes.

This is also the fastest-failing first check most companies run. Ask the assistant your own category's core buying question in a fresh session. The common result on a first try is not a wrong answer or a competitor's answer; it is no usable answer at all, because no page anywhere states the specific fact plainly enough to retrieve. That gap, not a ranking deficit, is usually the actual starting point.

The three crawlers, and why blocking the wrong one costs you the citation

OpenAI operates three separate, independently configured crawlers, and confusing them is the most common way a business accidentally makes itself uncitable. Each is controlled by its own robots.txt user-agent line, and allowing or blocking one has no effect on the others.

• GPTBot — collects content to train OpenAI's generative AI foundation models. Blocking it opts a site out of training data; it has no effect on whether ChatGPT can cite the site in a live search answer
• OAI-SearchBot — crawls and indexes content specifically to power ChatGPT's search feature. This is the one that matters for citations: OpenAI states directly that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as plain navigational links
• ChatGPT-User — fetches a page in real time when a user's own prompt triggers browsing or a GPT Action. It is not used for indexing, and OpenAI states it is not the setting that determines whether content can appear in search results

Resolve entity and source ambiguity

Use a consistent organization name, describe what the company does, identify the people responsible for expert content, link official profiles, and keep key facts aligned across the site. Contradictory dates, names, and claims make retrieval less reliable.

A robots.txt file written to block “AI scraping” in a blanket way, or one that only names GPTBot because that was the only bot that existed when it was written, can silently block OAI-SearchBot too if it disallows all user agents by default. The practical fix is to name each bot explicitly rather than relying on a wildcard rule, and to re-check after any robots.txt change, since OpenAI notes it can take roughly 24 hours after an update for its systems to adjust.

OpenAI's own help documentation states that ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information, and that placement is not guaranteed — the same posture Google takes toward organic ranking, and the same reason no vendor on either side can sell a guaranteed outcome. Reliability, per OpenAI, is explicitly tied to a user's ability to open a cited source and confirm it actually supports the claim; a business whose public facts contradict each other across its own site and its listings elsewhere makes that confirmation step fail, which is a direct reason to be dropped rather than cited.

How ranking in ChatGPT is different from ranking in Google

There is no ranked list to climb. ChatGPT composes an answer and names a handful of sources, so the outcome is binary per question: you were included or you were not. Position one and position eleven do not exist. This is why the familiar tactics of moving up a SERP do not transfer, and why a page that ranks well in Google can be absent from the answer to the same question.

The selection that matters happens at the passage level. When the assistant retrieves, it is looking for a specific span of text that answers the question directly enough to quote or paraphrase. A page can be authoritative overall and still lose because the answer is spread across four paragraphs, buried under a narrative introduction, or phrased in language nobody searches with.

What ChatGPT search is doing mechanically

When a prompt triggers live search, ChatGPT reformulates the question into one or more targeted queries and sends them to third-party search providers, alongside content provided directly through OpenAI's publisher partnerships. It reviews the returned results and may run further queries against the same or other providers before writing a response. OpenAI's own help documentation adds a caveat worth repeating to a client who wants guarantees: search results and citations can be incomplete, outdated, or incorrect, and a reader should open a cited source to check that it supports the answer.

The practical implication is that a page's crawlability and indexability for OAI-SearchBot functions like a prerequisite, not a strategy. If the search layer OpenAI queries has never indexed the page, it cannot be retrieved regardless of how well-written the passage is. This is the same technical floor as classic SEO: canonical URLs, working status codes, and content that renders in HTML rather than only after client-side JavaScript.

What moves the result, in rough order of leverage

The highest-leverage change is usually not technical. It is publishing an answer to a buying question that your site currently does not answer anywhere. Most companies discover on a first check that assistants cannot recommend them for their main category question because no page on the site states the answer plainly.

This is measurable, not just observed anecdotally. An analysis of 30 million cited sources across ChatGPT, Google AI Mode, Gemini, Perplexity, and Google AI Overviews found Reddit, YouTube, and LinkedIn were the most-cited domains overall, with Wikipedia and Forbes also in the top five and review platforms like Yelp and G2 appearing heavily in recommendation-style queries; the same analysis found ChatGPT specifically favored Wikipedia, Reddit, and editorial sites like Forbes. None of those are vendor sites talking about themselves. The practical reading is not “go post on Reddit” as a tactic — the analysis credits real user discussion and independent editorial judgment, not brand participation, for why those sources get trusted. It is that being accurately represented in the third-party sources your category already leans on is frequently a faster route to a citation than any change to your own domain.

Use a reproducible citation baseline

Define a panel of buyer questions and record the model, market, date, response, cited URLs, mentions, and competitors. Repeat the panel at a controlled cadence. The useful output is a trend with source-level evidence—not a single percentage detached from the underlying answers.

Expect variance. The same prompt returns different sources on different days, so a single check cannot tell you whether a change worked. Sample each question repeatedly and compare distributions across weeks rather than reading one run as a verdict.

Store enough detail per observation to answer why as well as whether: the exact prompt, whether search was enabled or the model answered from memory, the full response text, every cited URL, and which competitors appeared instead. Over enough weeks this also tells you whether your category is a live-search category or a training-data category for the assistant, which changes what kind of investment is worth making next.

Common questions