# How ChatGPT decides what to cite: methodology breakdown > Source: https://yanivgoldenberg.com/chatgpt-citations-explained/ > Updated: 2026-07-19T06:45:42+00:00 > Summary: How ChatGPT decides what to cite: the live retrieval pipeline, the 5 levers that move ChatGPT citations, and how to test whether the model cites you. GEO · AI Search Visibility *ChatGPT citations are not random. ChatGPT decides what to cite from a live retrieval step, and a handful of levers move whether that source is you.* ChatGPT’s citation choices in web search are driven by a layered retrieval-and-rank pipeline: query rewrites, web fetch, content extractability, source authority signals, and recency. The single biggest lever for getting cited is structured, answer-first content with clean schema and entity disambiguation. TL;DR ChatGPT’s citation choices in web search are driven by a layered retrieval-and-rank pipeline: query rewrites, web fetch, content extractability, source authority signals, and recency. The single biggest lever for getting cited is structured, answer-first content with clean schema and entity disambiguation. ## What “citation” actually means in ChatGPT ChatGPT cites sources in two contexts. First, when web search is enabled, the model performs a live retrieval pass and surfaces source links inline with the answer. Second, in some product surfaces, training-data attribution is exposed when an answer is grounded in a specific known source. The mechanics differ, but the optimization tactics converge: extractable, well-structured, entity-clear content. ## The citation pipeline (what we know publicly) Based on observable behavior and OpenAI’s published documentation, the live-search citation pipeline involves at least these stages: - Query rewriting: the user’s question is rewritten into one or more search queries optimized for retrieval. - Web fetch: queries are sent to OpenAI’s search index (powered by Bing and OpenAI’s own crawl via OAI-SearchBot and ChatGPT-User). - Content extraction: fetched pages are parsed for clean text, structured data, and answer-shaped content. - Relevance ranking: results are scored against the original question, not just the rewritten query. - Citation selection: the model picks 3-10 sources to ground the answer, weighted by relevance, authority signals, and content extractability. ## What gets cited (and what doesn’t) - Pages with clear TL;DR or summary blocks at the top - Content with FAQ sections and explicit question-answer structure - Pages with valid Article, FAQPage, and Person/Organization schema - Domains with strong sameAs entity links across LinkedIn, Wikipedia, GitHub, Crunchbase - Recently updated pages with visible date stamps - Original data, methodology, and primary sources (not aggregator content) ## What does NOT get cited - Pages blocked to OAI-SearchBot, ChatGPT-User, or GPTBot in robots.txt - JavaScript-heavy pages where content is not server-rendered - Thin content or single-paragraph blog posts with no structure - Pages with broken or invalid JSON-LD schema - Domains with ambiguous entity identity (no canonical Person/Organization schema) ## The single biggest lever If you do one thing to increase ChatGPT citation likelihood, it is this: write content in answer-first structure with explicit TL;DR blocks at the top of every page. Pair it with valid schema. The retrieval pipeline rewards extractable content because extractability is what makes citation possible at all. ## Frequently asked questions **Does ChatGPT use Google’s index?** No. ChatGPT’s live web search uses OpenAI’s own search index, primarily powered by Bing’s crawl plus OpenAI’s own crawlers (OAI-SearchBot, ChatGPT-User, GPTBot for training). Google’s blue-link results do not directly influence ChatGPT citations. **How can I tell if my site is being cited?** Manual: ask ChatGPT (with web search enabled) the queries your buyers ask, and check whether your domain appears in the source list. Programmatic: use a tool like the open-source seo-geo-skill benchmark script to systematically test citation across multiple queries and engines. **Does ChatGPT respect llms.txt?** Not officially as of 2026. OpenAI has not published documentation indicating that ChatGPT or its crawlers honor llms.txt. That said, shipping it costs nothing and signals AI-readiness to other engines that may honor it. **How long does it take for new content to appear in ChatGPT citations?** Live web search picks up indexable content within days of publication, assuming your site is crawlable to OAI-SearchBot. Training-data attribution is much slower and tied to model training cycles (typically months). Published: 2026-04-27 · By [Yaniv Goldenberg](/about-yaniv-goldenberg/) · [Get audited](/ai-search-visibility-audit/) The takeaway is simple: treat ChatGPT citations as a channel you can win on purpose. Publish the primary source, structure it answer-first, keep it crawlable and fresh, then measure your share of citations every month and double down on whatever moved the number. That loop, not luck, is how brands go from invisible to quoted. ## The practical levers that earn ChatGPT citations Once you accept that ChatGPT citations come from a live retrieval step, the levers become concrete. First, be the primary source: original data, a named framework, or a first-hand number is far more citable than a paraphrase of someone else. ChatGPT citations cluster around the page that owns the claim, not the page that repeats it. Second, structure answer-first: put the direct answer in the first one or two sentences under a clear, question-shaped heading, so the retrieval layer can lift a clean, self-contained passage. Third, make your entity unambiguous: a consistent name, a Person or Organization schema, and matching profiles tell the model who you are, which is what lets it attribute a claim to you by name. Freshness and crawlability are the two levers most sites get wrong. Live web search favors recently published or updated indexable pages, so a page that has not moved in two years rarely wins a time-sensitive query. And none of it matters if OAI-SearchBot cannot reach you: check that your robots.txt and any firewall or bot manager allow the crawler, because a blocked page earns zero ChatGPT citations no matter how good it is. “A single cited answer in ChatGPT can be worth more than $10,000 in equivalent ad reach over its lifetime, because it reaches buyers at the exact moment of decision, for free, again and again, long after a paid campaign would have stopped delivering, with no incremental spend attached to each new reader it reaches.” That is why ChatGPT citations deserve a line in your growth plan, not a footnote. In my own testing across buyer questions, the pages that win citations share three traits almost every time: they answer the question in the first sentence, they carry one concrete number with a source, and they are fresh enough that the live index trusts them. Miss any one of those and the citation goes to whoever did all three. The good news is that all three are things you control, which is exactly why ChatGPT citations are a channel you can engineer, measure and compound over time, rather than a lottery you keep entering and hoping to win. ## How to test whether ChatGPT cites you Do not guess. Take the ten questions your buyer actually asks, run them in ChatGPT with web search enabled, and record which sources it cites for each. Look for your domain, and note who wins when you do not. That gap is your roadmap: if a competitor is cited for a question you should own, read the page it cited and see what it did that you did not, usually a cleaner answer-first passage, a real statistic with a source, or simply being crawlable and fresh while you were not. Then track it over time. ChatGPT citations shift as the live index updates, so re-run the same set monthly and watch whether your share of citations climbs after you ship the fix. This is the same measurement discipline I apply to every growth channel: define the question set, measure where you stand, ship one change, and confirm the number moved. Treating ChatGPT citations as a measurable channel, not a mystery, is what separates sites that get cited from sites that hope to be. Further reading: [Large language model](https://en.wikipedia.org/wiki/Large_language_model) and [Search engine optimization](https://en.wikipedia.org/wiki/Search_engine_optimization). Related: [GEO (AI search)](/geo/), [AI visibility audit](/ai-visibility-audit/), [more on AI search](/blog/).