Key takeaways
- ChatGPT cites roughly half of the URLs it retrieves. Ahrefs' study of 1.4 million prompts put it at 49.98%, so being fetched is not the same as being cited.
- There are three separate outcomes: fetched (pulled into context), cited (a numbered link tied to a claim), and mentioned (your brand named while another page gets the link).
- Reddit and YouTube are the extreme cases. One network-traffic analysis saw Reddit fetched 278 times and cited 11, and YouTube fetched 201 times and cited zero.
- Fanout queries decide which pages even enter the pool. Titles, headings and URLs that match short sub-queries do most of the work.
- Crawler logs alone will not tell you if you are winning. You need to compare crawls against citations per page.
What fetched and cited actually mean
Open a ChatGPT answer with search turned on and you see a handful of numbered links. Those are the citations. They are the visible tip of a much bigger pile of pages the retrieval system touched on the way to the answer.
It helps to separate three events that people tend to lump together:
- Fetched: ChatGPT's retrieval tool pulled your page (or its snippet) into the model's working context.
- Cited: a numbered footnote ties your page to a specific claim in the answer.
- Mentioned: your brand name appears in the text, but a different page carries the link.
A page can be fetched without being cited. It can be cited without the brand being named. And a brand can be named with no link to its own site at all. Each of these needs a different fix, which is why a single "AI visibility" number hides so much.
The numbers: about half of what ChatGPT retrieves gets credit
Ahrefs analyzed 1.4 million ChatGPT prompts and found that ChatGPT cited 49.98% of the URLs it retrieved. That is about 23.4 million cited pages out of roughly 47 million pulled. One caveat from the write-ups: some of the uncited half were probably never opened at all. The figure tracks the whole journey from retrieval to citation, not a deliberate read-then-discard decision.
The more interesting claim is where the gate sits. Ahrefs argues there is a gatekeeping layer before ChatGPT reads your actual page content, and that the title, snippet and URL do most of the heavy lifting in that first decision. That matches how people describe the behavior when they ask the model why it skipped a page it had plainly retrieved.
Two other findings cut against habit:
- Among search-index results, the median cited page was around 500 days old, about 1.3 years. Relevance seems to beat raw freshness here.
- Cited and non-cited URLs carry different metadata. In the Ahrefs table, cited search results had a snippet 2.52% of the time versus 0.09% for non-cited ones, while publication dates were more common on the non-cited side (49.00% vs 33.79%). Ahrefs itself flags that this kind of data is easy to over-read, so treat it as a clue rather than a recipe.
How many citations per answer?
Promptwatch's sources per response report (dated September 2025, so the figures drift) put ChatGPT at about 5 cited sources per web-search answer, against roughly 10 for Google AI Overviews and Perplexity. A later report on unique domains per response says almost half of ChatGPT responses cite 10 or more domains (Sep 28 to Oct 4, 2026). Those two numbers look like they disagree. They measure different periods and different things, and ChatGPT's behavior has moved a lot in a year, so quote the dated one for the period you care about.
The Reddit paradox
The clearest picture of fetched-versus-cited is Reddit. When Suganthan Mohanadasan read ChatGPT's own network traffic across a few dozen searches on one logged-in Pro account in June 2026, Reddit was fetched 278 times and cited 11. YouTube was fetched 201 times and cited not once. That is a small sample from one account, but the direction matches the bigger study.
Ahrefs reports that 67.8% of non-cited URLs in its dataset came from Reddit. A LinkedIn post by Jayson Demers describes a case where 84 Reddit threads were pulled for a single answer with no credit given. The model seems to use forums to gauge consensus and build context, then attach the citation to something that looks cleaner and more self-contained.
There is an update worth knowing. Promptwatch Data shows Reddit's share of ChatGPT citations falling from a steady 3.83% (July 18 to Aug 7, 2026) to about 0.52% over Aug 14 to 17, an 86.4% relative drop. Promptwatch flags that as provisional, since a data-collection issue can't be ruled out. Its October 2026 social media report puts all social platforms at 0.46% of ChatGPT citations for Sep 28 to Oct 4, versus 12.2% for Google AI Overviews. So the paradox was already real, and it has since gotten more extreme: Reddit may still shape what ChatGPT reads, but it is now almost never the link.
Why a page gets fetched and not cited
Nobody outside OpenAI knows the exact rules. What the evidence supports is a short list of plausible mechanisms.
Fanout creates the candidate pool
ChatGPT fans a prompt out into several web searches, and each fanout query is a separate retrieval your content can win. Promptwatch's query fanout data shows the average fell from 2.15 per response in early December to about 1.84 by early March, and fanout queries got shorter, from about 117 characters to the high 80s, then about 53 characters (six to eight words) when the data returned in April. On August 8, 2026, ChatGPT started using the site: operator at scale, jumping from about 0.4% to about 17% of fanout queries overnight, with searches per response nearly doubling.
Practical read: your page can be retrieved for a sub-query the user never typed. If your title and headings do not look like the short queries the model generates, you are fetched less, and cited less after that.
Selection looks like a similarity match
Ahrefs concludes that cited pages are the ones whose titles and content match the questions ChatGPT is asking behind the scenes, and that surface through the right retrieval channel, namely its general search index. Retrievable and rankable pages win the link even when news, Reddit and YouTube supplied context.
Single-source claims look unverified
One practitioner write-up on Medium reported that its pages were fetched 12,755 times in a week while holding about 4% of citations in its category. When the author asked the engine why it was skipping pages it had retrieved, the answer pointed to claims that appeared only on the site's own domain. They read as single-source and unverified, and repeating them across your own pages still counts as one source. That is one site's experience and one model's self-report, so hold it loosely. It does fit the idea that corroboration elsewhere matters.
The same author threw out two theories: structured data was not what separated cited from uncited pages, and the opening of a page could not be isolated as a cause because the compared pages sat in different demand pools.
Content type changes the odds
Promptwatch's content type report found product pages were the most cited type at 22.4% of citations (Sep 28 to Oct 4, 2026), followed by landing pages and listicles. The content half-life report tracked every page cited on July 6, 2026: landing pages and docs stayed cited longest, while news lost half its pages within 36 days.
Fetched, cited, mentioned: what to do about each
| Outcome | What it tells you | Likely cause | First thing to try |
|---|---|---|---|
| Not fetched | The pipeline never reached you | Blocked crawler, errors, weak title or snippet for fanout queries | Check crawler logs and robots rules, rewrite titles and H2s as short queries |
| Fetched, not cited | You were read but passed over | Another page looked more authoritative or self-contained, claim not corroborated | Make a focused page per sub-question, add original data, earn third-party coverage |
| Mentioned, not cited | The model knows you but links someone else | Brand appears in third-party pages that get the link | Look at which pages carry your name and improve your own page for that prompt |
| Cited | You won the slot | Match on query, channel and content type | Track it and protect it, refresh before it decays |
A practical workflow
- Confirm access. Check that the OpenAI crawlers you want are not blocked and that they get clean responses. Server-side crawler logs show who visited and what errors they hit. Note that blocking crawlers is a separate question from citation, so read your robots rules carefully before changing them.
- List the prompts that matter. Group them by topic, then look at the short sub-queries a prompt generates. A browser extension or a fanout generator can show these.
- Rewrite for the sub-query. Put the entity and category first in titles and H2s. Build focused pages for "X vs Y", pricing, alternatives and how-tos rather than one sprawling page.
- Compare crawl against citation, per page. Pages with plenty of AI crawler hits and few citations are your fetched-not-cited list. That list is the work queue.
- Add something a competitor cannot copy. Original data, a clear definition, or a quotable number gives the model a reason to attach a link to you instead of a rival's roundup.
- Get named elsewhere. If your claims only exist on your own domain, look for third-party pages ChatGPT already trusts and earn a place there.
- Re-audit after model updates. Fanout length, count and operators have changed several times this year, so last quarter's titles may be tuned to last quarter's behavior.
Tools that help close the gap
Most basic prompt trackers answer "was I mentioned?" Fetched versus cited needs two data sets side by side: what AI crawlers read, and what AI answers cite.
Promptwatch is built around that join. Its IP-verified AI crawler logs show which pages GPTBot, OAI-SearchBot and ChatGPT-User read and which errors they hit, and its citation analytics show which pages get linked, so the crawl-to-citation path per page is visible. It also surfaces query fan-outs, which is useful for the sub-query rewriting step above.
Other tools in the category approach it from the citation side:
| Tool | Angle | Useful for this problem |
|---|---|---|
| Promptwatch | Crawler logs, citations, fan-outs, content agent | Seeing crawled-but-uncited pages and fixing them |
| Ahrefs Brand Radar | Brand tracking inside an SEO suite | Context if you already live in Ahrefs |
| Profound | Enterprise AI visibility analytics | Large-brand monitoring |
| Peec AI | Prompt-level visibility tracking | Lightweight tracking of mentions |
| Otterly.AI | Affordable monitoring | Basic mention and citation checks |
| MentionBird | Which sources AI engines cite | Spotting the third-party pages that get the link |
MentionBird
If you want to compare more options, the directory at bestgeosoftware.com lists GEO platforms side by side.
What the data can and cannot tell you
A few honest limits. No Promptwatch report measures fetched-but-not-cited directly, because it tracks citations and fanout queries rather than retrieved URLs. The best public numbers for the gap are the Ahrefs study and the Mohanadasan network-traffic analysis, and the latter is one account across a few dozen searches. Several of the practitioner write-ups are single-site observations. And ChatGPT changes quickly: the Reddit collapse in August 2026 happened within weeks of the fanout change on August 8.
So use these findings to form hypotheses, then test them on your own pages. If you want proof that something works, change one title or one page, and watch whether crawls, citations and clicks move.
FAQ
Does being fetched mean ChatGPT will cite me eventually?
No. The Ahrefs data suggests roughly half of retrieved URLs never get a citation, and some may never be opened. Fetched is a precondition, not a result.
Is a mention better or worse than a citation?
A mention builds awareness but sends no traffic. A citation puts a clickable link next to a specific claim. For revenue, citations matter more, but track both.
Should I block OpenAI crawlers to stop being read for free?
That is a business decision with tradeoffs. Blocking removes you from the pool entirely, which is the opposite of what you want if citations are the goal. Make the choice deliberately, not as a reaction to fetch counts.
Why would Reddit be read so much and cited so little?
The evidence suggests ChatGPT uses forum threads for context and consensus, then links a cleaner source. Promptwatch Data now shows Reddit's ChatGPT citation share near 0.5% since mid-August 2026, with all social platforms at 0.46% in early October.
How do I cite ChatGPT itself in an APA or MLA paper?
That is a different question from this guide. Use your style guide's current rule for generative AI, and cite the primary sources the model points to rather than the chat itself where you can verify them.
Bottom line
Treat "fetched" as the start of the funnel, not the finish. Make your titles and headings match the short queries ChatGPT generates, publish pages that are easy to quote and hard to substitute, get your claims corroborated off your own site, and measure crawls against citations per page. The pages that are read constantly and cited almost never are your most useful to-do list.