What Makes Chatgpt Cite A Website

• Citation starts with retrieval, and retrieval only happens when ChatGPT searches the web.
• No, ChatGPT does not always show sources.
• Access is the first technical gate, and it must be checked before any content work.
• Within-domain evidence points to alignment.
• Measurement works at the level of individual prompts. For instance, according to Discovered Labs' arXiv study (Moore and Dunne, submitted 28 September 2026), roughly 2.1 million citation records were collected from ChatGPT, Claude, Google AI Overviews and Gemini across nineteen B2B software companies, with 10,042 cited pages crawled and over sixty page features measured.

Citation starts with retrieval, and retrieval only happens when ChatGPT searches the web. According to Stridec, ChatGPT cites sources only when its browse tool has been invoked, and the candidate index for that tool has been Bing's web index. Browsing is triggered by explicit requests for sources, recency-dependent questions, and queries the model judges to fall outside its training confidence. When none of those conditions apply, the answer comes from model memory and there is no verifiable page to cite.

Once retrieved, a page must still survive selection. According to Usetellr, a page must be crawlable, relevant to the exact query, and easy to extract a passage from. Pages that are fast to parse and state their claims plainly give the model a clean passage to lift. Pages that bury the answer under navigation, interstitials or vague marketing copy rarely make it into the output, even when they were read.

The gap between reading and citing is large. According to Zyppy research from 2025, ChatGPT cites only around 15% of the pages it retrieves; the other 85% are read but not referenced. Teams should therefore treat retrieval eligibility as a gate and passage quality as the real competition.

Retrieval eligibility is the gate; passage quality decides which retrieved page is named. According to the Discovered Labs arXiv study, prompt-content alignment (overlap between page wording and the prompts buyers type) was the strongest page-level citation signal, with a standardised coefficient of +0.37, which the authors translate to roughly 30% more citations for a page one standard deviation better aligned.

No, ChatGPT does not always show sources. According to Stridec, ChatGPT cites sources only when its browse tool has been invoked. A study of 14,000 real LMArena conversations, reported by Stridec, found that 24% of GPT-4o responses were generated without fetching online content, while 76% relied on live retrieval.

That split means a brand's exposure depends heavily on whether a query triggers the browse tool. Queries about current pricing, recent releases or named vendors are far more likely to trigger browsing than conceptual questions answered from training data. For those conceptual questions, ChatGPT answers from its training data and typically shows no sources at all.

Access is the first technical gate, and it must be checked before any content work. According to SearchScore, OAI-SearchBot is the crawler OpenAI uses for ChatGPT search discovery, and it is distinct from GPTBot, which is used only for model training. Blocking GPTBot does not remove a site from ChatGPT search, but blocking OAI-SearchBot does.

A common failure is a legacy rule under User-agent: * in robots.txt, which can block search crawlers unintentionally. According to SearchScore's audit of 6,944 websites (July 2026), 6.9% block at least one major AI crawler, often in exactly this way.

A team should therefore check its robots.txt file for an explicit OAI-SearchBot entry and confirm that the crawler is not disallowed. Only after that access check is passed does optimizing page wording make practical sense, since a page the crawler cannot reach has no chance of being retrieved or cited.

Within-domain evidence points to alignment. According to the Discovered Labs arXiv study by Moore and Dunne (submitted 28 September 2026), roughly 2.1 million citation records were analysed across ChatGPT, Claude, Google AI Overviews and Gemini, with 10,042 cited pages crawled. Prompt-content alignment, the overlap between page wording and the prompts buyers type, was the strongest page-level signal, with a standardised coefficient of +0.37. The authors translate that to roughly 30% more citations for a page one standard deviation better aligned.

The same study found that FAQ blocks, schema and page speed looked helpful in pooled data but lost significance within the same domain, because large brands tend to have both more citations and these features. The practical implication is that wording a page to match real buyer questions is measurable, while adding schema alone is not shown to move citations. Perplexity was excluded from that analysis for having too few samples, so its findings should not be extended to it.

Position matters. According to KIME's 2026 guide, ChatGPT concentrates citations heavily in the first 30% of a page, and that section accounts for 44.2% of all citations. Authors should therefore place the direct answer, the defining facts and the key comparison near the top of the page, in passages that remain understandable when quoted on their own. Citations are also binary rather than ranked.

According to KIME, a brand is either cited or not, with no equivalent of a ranked position three as in Google's ten blue links. The goal for a page is to be the named source that ChatGPT cites, not to rank within a list of results. Because the citation is all-or-nothing, the opening section of each page carries most of the weight in earning that citation.

Concentration is high. According to analysis published in Search Engine Land, roughly 30 domains account for around 67% of ChatGPT citations within any given topic. For a brand-owned site, this means competing for citation is partly competing against established reference, review and editorial domains.

A brand gains ground by publishing narrow, specific passages that answer the exact questions those domains answer only in general terms, and by keeping them current. Measuring which domains appear for your target prompts is the fastest way to see who holds the citations.

Measurement works at the level of individual prompts. Teams should begin with a fixed set of buyer prompts. Next, they record which URLs ChatGPT names for each prompt across repeated runs. Then they note which of their own pages are retrieved but absent from answers. Three numbers matter most.

The first is the share of target prompts that name the domain. The second is the citation rate of pages rewritten for alignment. The third is the crawl access of OAI-SearchBot. Fastlook's Citation Analytics and Brand Memory report citations across engines, but the same method can be run manually.

OpenAI publishes no official ranking formula, so every model of citation behaviour should be treated as a hypothesis to test against the team's own prompt set. According to Zyppy research from 2025, ChatGPT cites only around 15% of the pages it retrieves; the other 85% are read but not referenced in the output.

Related guides

• Get ChatGPT to Quote Your Site: 8 Crawler and Content Checks
• Tools That Get Your Site Cited in ChatGPT Answers
• Make Perplexity Cite Your Website: What Actually Works
• Get Perplexity to Cite Your Website: What Works in 2026
• How ChatGPT Picks Sources: Retrieval, Bing, and Citations