Blog · how it works
How ChatGPT decides which sources to cite
ChatGPT answers from two different places: a live retrieval index when search is on, and trained memory when it is off. Only the first can cite a URL. When it does cite, it names about 15 sources per answer and repeats only about one in five of them the next time the same question is asked.
The basics
Two modes, two different questions
Almost every confused conversation about ChatGPT visibility comes from mixing these up.
| Search on (retrieval) | Search off (memory) | |
|---|---|---|
| What happens | Runs searches, reads pages, then answers | Answers from training data |
| Can it cite a URL? | Yes | No. It has no document to point at |
| What moves it | What is published and crawlable now | What the wider web said before the training cut-off |
| Response time | Days to weeks | Only when a new model ships |
| Who can influence it | You, fairly directly | Mostly other people writing about you |
If a report tells you your ChatGPT visibility is 0% and does not say which mode it measured, it has not told you anything. Being absent from retrieval is a publishing and crawlability problem you can fix this quarter. Being absent from memory is a reputation problem measured in model generations.
How it works
The retrieval pipeline, step by step
With search on, ChatGPT is a retrieval-augmented system, and every stage is a place you can be excluded.
-
01
A crawler builds the index
OpenAI runs separate bots for separate jobs. OAI-SearchBot builds the search index that ChatGPT answers from. GPTBot collects training data. ChatGPT-User fetches a page live when a user's request requires it. Blocking the wrong one costs you the wrong thing.
-
02
The question becomes searches
The model rewrites the user's question into one or more queries. Small changes in the user's phrasing produce different queries, and therefore different sources.
-
03
Documents are retrieved and reranked
Candidates come back and are scored for relevance to the rewritten queries. This is the stage where the passage-level writing decisions pay off or do not.
-
04
The answer is composed with citations attached
The model writes from the documents in its context window and links the ones it used. Roughly 15 links per answer, far more than Gemini's 3.
ChatGPT cites an average of about 15 sources per response; Gemini cites about 3.
From an index built on 126 million US AI search prompts recorded January to April 2026. A five-fold difference in citation slots means the same visibility work has very different odds on each engine.
That difference in citation slots is worth planning around. Fifteen slots means a mid-sized brand can realistically appear alongside the incumbents. Three slots means winning is close to winner-take-all. The same content programme has very different odds on the two engines.
The uncomfortable part
Four in five sources change overnight
This is the single most important and least-reported fact about ChatGPT visibility, and two independent studies agree on it closely.
Asked the same question twice, ChatGPT repeated only 21.2% of its cited domains and Google AI Overviews 31.5%. Within a one-week window overlap rose only to 26.7% and 36.8%.
693,509 answers across 16,143 ChatGPT prompts and 15,805 Google prompts, 26 March to 25 April 2026. The practical consequence: a single screenshot of an AI answer is a noisy sample, not a measurement.
Day-over-day source churn ran 88.3% on Gemini, 79.2% on ChatGPT, 75.9% on Google AI Mode and 44.4% on Perplexity. Sources cited on all seven consecutive days: 11.1% on Perplexity, 2.6% on Google AI Mode, 1.1% on ChatGPT, 0.4% on Gemini. 84% of the sources for a question were cited by only one of the four engines.
530,875 citations across 181,225 distinct URLs, 2,398 real queries, seven consecutive days in June 2026, 46,259 day-over-day comparisons. Independent of the Parse study and agreeing with it closely on ChatGPT.
Ask ChatGPT the same question twice and roughly 79% of the cited sources are different. Only 1.1% of sources survive seven consecutive days. Across four engines, 84% of the sources for a given question are cited by exactly one of them.
One screenshot is not a measurement. It is one sample of a noisy distribution.
Three consequences follow, and all three are commercial rather than technical.
- A single check proves nothing, in either direction. Not appearing once is not absence. Appearing once is not visibility.
- Anyone reporting a week-on-week change from a handful of prompts is reporting noise. With ~79% daily churn you need repeated runs across a fixed question set before a movement means anything.
- Engines must be measured separately. With 84% of sources unique to one engine, an averaged 'AI visibility score' hides the only thing you could act on.
What helps
What actually raises the odds
Given that volatility, the goal is not to hold a slot. It is to be in the candidate pool often enough that the dice land your way at a useful rate.
- Be fetchable by the right bot. OAI-SearchBot is the one that decides whether ChatGPT can cite you at all. Many sites block it without knowing, because a platform default did it for them.
- Be quotable at passage level. Self-contained paragraphs, question-shaped headings, one claim per paragraph, the answer in the first sentence.
- Carry evidence, not adjectives. The strongest controlled study on generative-engine visibility found that adding statistics, citations and quotations outperformed rewriting, simplification and keyword work, with keyword stuffing performing worse than no change at all.
- Be present off your own domain. Roughly 85% of AI references point at third-party sites. Reddit alone accounts for around 40% of aggregate multi-engine citation frequency, with Wikipedia second.
- Stay current. Retrieval favours recently updated pages heavily on commercial questions.
Targeted content changes raised visibility in generative engine responses by up to 40%.
GEO-bench: roughly 10,000 queries across nine datasets. Authority-style edits, adding citations, statistics and quotations, outperformed cosmetic edits such as rewriting, simplification and keyword work. Effect size varied by domain.
Reddit is the single most-cited domain across generative engines, at roughly 40% of multi-engine aggregate citation frequency. Wikipedia is second, appearing in 26–48% of ChatGPT top-10 answers; YouTube third at about 19% of Google AI Overviews top-source share.
Synthesis of six independent studies covering more than 680 million citations recorded between August 2024 and April 2026.
Be careful with second-hand versions of the GEO study. Per-tactic percentages circulate that are not in the paper. The safe claim is the direction: evidence-heavy content beats polished content.
Limits
What this cannot do
- No submission endpoint exists. You cannot request a citation. You can only be crawlable, quotable and present in the sources retrieval already favours.
- Memory mode is not addressable this quarter. If ChatGPT does not know who you are with search off, publishing more of your own pages will not change it before the next model.
- Volatility caps how good the news can get. With ~1% seven-day persistence, 'we hold position for query X' is not a claim anyone can honestly make about ChatGPT.
- Blocking decisions are not reversible retroactively. A page excluded from the index while a bot was blocked is not cited during that period, and re-crawling is on OpenAI's schedule, not yours.
Sources
Where these numbers come from
Every figure on this page is external and linked. The studies measure different populations over different windows, so read them as converging evidence rather than one authoritative number.
- Semrush AI Visibility Index 2026ChatGPT cites an average of about 15 sources per response; Gemini cites about 3. www.semrush.com ↗
- ParseAsked the same question twice, ChatGPT repeated only 21.2% of its cited domains and Google AI Overviews 31.5%. Within a one-week window overlap rose only to 26.7% and 36.8%. parse.gl ↗
- GetMentionsDay-over-day source churn ran 88.3% on Gemini, 79.2% on ChatGPT, 75.9% on Google AI Mode and 44.4% on Perplexity. Sources cited on all seven consecutive days: 11.1% on Perplexity, 2.6% on Google AI Mode, 1.1% on ChatGPT, 0.4% on Gemini. 84% of the sources for a question were cited by only one of the four engines. www.getmentions.ai ↗
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, ACM SIGKDD 2024Targeted content changes raised visibility in generative engine responses by up to 40%. arxiv.org ↗
- AI Platform Citation Source Index 2026Reddit is the single most-cited domain across generative engines, at roughly 40% of multi-engine aggregate citation frequency. Wikipedia is second, appearing in 26–48% of ChatGPT top-10 answers; YouTube third at about 19% of Google AI Overviews top-source share. everything-pr.com ↗
- SimilarwebChatGPT's share of generative AI web traffic fell from about 76% a year earlier to around 53%, as Gemini passed a quarter of traffic and Claude grew fastest in the category. www.similarweb.com ↗
- Omnibound, AEO statistics compilation 2026Roughly 85% of AI references point to third-party platforms rather than brand-owned sites. www.omnibound.ai ↗
Keep reading
FAQ
Common questions
How does ChatGPT choose which sources to cite?
With search enabled it rewrites the question into search queries, retrieves candidate documents from its own search index, reranks them, and composes an answer from the passages it selects, linking the ones it used. Selection happens at passage level, so a page is chosen for the specific paragraph that answers the rewritten query.
How many sources does ChatGPT cite per answer?
About 15 on average, measured across 126 million US prompts between January and April 2026. Gemini cites about 3. That difference means ChatGPT has far more room for non-incumbent brands to appear in a given answer.
Why does ChatGPT cite different sources every time I ask?
Because retrieval is probabilistic and the index moves. Two independent 2026 studies found ChatGPT repeats only about 21% of its cited domains when asked the same question again, and that only about 1% of sources are cited on seven consecutive days. Any single check is one noisy sample.
Which OpenAI crawler do I need to allow?
OAI-SearchBot builds the search index ChatGPT cites from, so blocking it makes citation impossible. GPTBot collects training data and ChatGPT-User fetches pages during a live user request. They are separate directives in robots.txt and should be decided separately.
Can I get my brand into ChatGPT when web search is turned off?
Not directly and not quickly. With search off the model answers from training data, so what moves it is what third parties published about you before the training cut-off. That is a reputation and earned-coverage problem, and it only resolves when a new model is trained.
Does ChatGPT use Google's index?
No. ChatGPT search runs on OpenAI's own retrieval layer, historically built with Bing infrastructure and crawled by OAI-SearchBot. Ranking well in Google does not by itself make you retrievable in ChatGPT, which is why the two need measuring separately.
How many prompts do I need to track for a reliable reading?
Enough that daily churn of roughly 79% averages out. In practice that means a fixed set of buyer questions run repeatedly over weeks, reporting a rate rather than a position, not a handful of prompts checked once a month.
Does ChatGPT name you when your buyers ask? <span class='tone'>Lifewood Data Technology will find out.</span>
A fixed question set, run repeatedly in both modes, so the answer is a measurement rather than a screenshot.
- 100+ languages
- 50+ delivery centres
- Measured monthly
Lifewood Data Technology publishes this reference.