Key takeaways
- ChatGPT's source list mixes three outcomes: pages it fetched, pages it cited, and brands it only mentioned. Treat them as separate signals.
- Being fetched means your page was considered. Being cited means it was the best support for a specific sentence. The gap between them is where most optimization work lives.
- Almost half of ChatGPT responses cite 10 or more unique domains, so you are always sharing the answer. Position and description matter as much as presence.
- Your server logs, citation data and a fan-out viewer together show where in the pipeline a page drops out.
- Treat small-sample studies as direction, not measurement, and check dates. ChatGPT's behavior changed a lot in 2026.
Why the source list is more than a link list
Most people look at the sources under a ChatGPT answer, see whether their domain is there, and move on. Present means good, absent means bad. That reading is too thin to act on.
Behind that list sits a pipeline. ChatGPT first decides whether a question needs the web at all. If it does, it fans the question out into several searches, pulls back a pile of results, reads some of them, and then attaches citations to specific sentences in the answer. Your page can drop out at any step, and each exit point has a different fix.
This guide walks through how to read the list, what each state says about your site, and how to find where your pages are leaking out of the funnel.
The three states: fetched, cited, mentioned
Suganthan, an SEO who read ChatGPT's raw network traffic across roughly 1,240 source records, found that every web result carries a hidden label, and that fetched, cited and mentioned are three different outcomes. His sample is small (one logged-in account, a few dozen queries, mostly SaaS and tech), so his structural findings are firm but his percentages are directional. Still, the framing is useful.
| State | What happened | What it tells you |
|---|---|---|
| Fetched | ChatGPT retrieved or opened your page while building the answer | Your page was findable and considered |
| Cited | A specific sentence in the answer links to your page | Your page was the best support for that claim |
| Mentioned | Your brand is named in the answer, with or without a link to you | Others are describing you, or the model knows you |
You can be mentioned without being cited, and cited without being mentioned. A vendor page, for example, tends to get cited for its own pricing and specs, while the verdict on which product is best gets attributed to a third-party review. That split is the real lesson: ChatGPT reads your page for your facts, and other people's pages for opinions about you.
The fetched-but-not-cited gap
The starkest example in Suganthan's batch: Reddit was fetched 278 times and cited 11. YouTube was fetched 201 times and cited zero. His explanation is mechanical. A citation has to bind to text the model actually pulled. A Reddit thread is all there in the page. A YouTube page in search returns metadata, not the transcript.
A word of caution on the Reddit numbers. That sample predates a big shift. According to Promptwatch Data, reddit.com's share of ChatGPT citations fell from about 4% to about 0.5% on August 14, 2026 (Reddit citations are dropping in ChatGPT). In the Sep 28 to Oct 4, 2026 window, ChatGPT drew just 0.46% of its citations from social platforms, against 12.2% for Google AI Overviews (social media citations by AI model, October 2026). The mechanism (fetched is not cited) still holds. The Reddit-specific ratios are stale.
What you are competing against in every answer
Before you interpret your own list, know the baseline. In Promptwatch's data for Sep 28 to Oct 4, 2026, 47.6% of ChatGPT responses with citations cited 10 or more unique domains, and 66.1% cited at least 8. Only 6% cited three domains or fewer (unique domains per response). Several URLs from one domain count once.
The breakdown by prompt type matters:
| Prompt type | Cite 10+ domains | Cite 5 or fewer |
|---|---|---|
| Organic (no brand named) | 52.5% | n/a |
| Competitor comparison | 49.1% | n/a |
| Brand-specific | 21.3% | 40.7% |
Profound's analysis of about 700,000 conversations from October to December 2025 reached a similar conclusion from a different angle: sources travel in packs, and ChatGPT cites competitors side by side. You are fighting for share of an answer, not ownership of it.
The practical upshot: a single citation in a list of twelve is a weak signal by itself. Look at where you sit, how the answer describes you, and which neighbors share the list.
Which of your pages gets cited, by type
The content type of a page predicts its odds. Promptwatch's report on Sep 28 to Oct 4, 2026 gives the share of ChatGPT citations by type (citation share by content type):
| Content type | Share of citations |
|---|---|
| Product pages | 22.4% |
| Landing pages | 16.8% |
| Listicles | 13.6% |
| How-tos | 8.7% |
| Documentation | 7.5% |
| General articles | 6.1% |
| Comparisons | 5.9% |
| News | 5.7% |
| FAQ | 3.3% |
| Reviews | 1.7% |
Product and landing pages together make up 39%. For commercial prompts, product pages, landing pages, listicles and comparisons cover almost 70% of citations. For branded prompts, product pages (23.3%) and landing pages (22.8%) lead, with reviews at 5.6% and FAQ at 4.8%.
If your source list shows only blog posts while competitors appear with product and landing pages, that is a clue. For brand prompts especially, make sure your own site covers pricing, features, reviews and alternatives, because that is what ChatGPT reaches for.
Also consider shelf life. Promptwatch's half-life study of pages ChatGPT cited on July 6, 2026 found landing pages and docs stayed cited longest, while news lost half its pages within 36 days (content half-life). A citation to a news post is rented. A citation to a doc page is closer to owned.
How fan-outs decide what gets fetched
ChatGPT does not search your prompt verbatim. It generates fan-out queries, short searches that approximate sub-questions. Promptwatch's fan-out data shows average query length falling from about 117 characters in December 2025 to about 53 in April 2026 (ChatGPT query fanouts). On August 8, 2026, ChatGPT also began using the site: operator at scale, going from about 0.4% to about 17% of fan-out queries overnight (site: operator report).
Two things follow. First, headings that read like short search queries match the way ChatGPT searches. Second, if your domain is the target of site: queries, ChatGPT is going looking for you specifically, which is worth seeing in your logs.
Suganthan also notes results are deduped by domain, so 20 thin pages from your site collapse into one. Don't build a page for every fan-out. One strong page per claim beats a pile of weak ones.
Reading the evidence: where to look
No single view shows the whole pipeline. Combine four.
1. The source list itself
Open a ChatGPT answer where you expect to appear and expand the sources. Note which of your pages show, which competitors sit beside you, and whether the cited text is your claim or someone else's claim about you. Free browser extensions that display fan-outs and fetched versus cited sources can speed this up. Promptwatch's free GEO Inspector Chrome extension does this, and Suganthan describes a similar tool called FanoutFox.
2. Your server logs
Windsor.ai's team joined 30 days of server logs with 90 days of Search Console data and offers a clear way to read bot traffic:
- A ChatGPT-User or Claude-User/1.0 request is the assistant opening your page to answer a specific question. That is a potential citation, not a guaranteed one.
- Index crawlers (OAI-SearchBot, Claude-SearchBot) build the index the assistant queries.
- Training crawlers (GPTBot, ClaudeBot) tell you almost nothing about visibility.
One wrinkle for ChatGPT: Windsor.ai reports, citing Peec AI, that ChatGPT now answers largely from its own search index rather than fetching pages live, and their logs show the effect from late August 2026. ChatGPT-User counts are falling for everyone as a result. Count OAI-SearchBot coverage, which pages it crawls and how often, as the main ChatGPT-side measure, and read ChatGPT-User as partial.
DarkVisitors is one way to watch AI agents and bots hitting your site without parsing raw logs yourself.
3. Search Console and Bing data
Windsor.ai's main finding was that pages that get cited are pages that rank. Claude fetched every page with over 1,000 Google clicks and fewer than one in five pages with under 10. The top 20 pages took 80% of Google clicks, 62% of Claude fetches and 59% of ChatGPT fetches. That is one site, so treat it as a pattern to test on yours, but it matches Promptwatch's observation about domain authority: sites in the DR 46-75 range earned almost half of ChatGPT citations in August 2026 (citation share by domain rank). You do not need to be a giant. You do need to rank.
4. Prompt-level citation tracking
Manual checks do not scale past a handful of prompts. A tracker that records which prompts cite which pages, and how you are described, turns the source list into a time series.
Diagnosing the pattern you see
Use the combination of signals to pick the fix.
| What you see | Likely cause | First move |
|---|---|---|
| Competitors cited, you never fetched | Not in the index or not ranking | Check crawl access, rank the page, confirm OAI-SearchBot coverage |
| Fetched often, rarely cited | Page lacks a quotable, specific claim | Open with the sentence you want quoted, add original data |
| Cited for facts, absent from verdicts | Third parties describe you, not you | Earn mentions on review, listicle and forum pages |
| Fetch requests return 301 or 404 | Wrong URL served to the bot | Fix redirects and serve the requested URL directly |
| Cited, then dropped within weeks | News-style content with short half-life | Move key claims onto landing pages or docs, refresh regularly |
| Mentioned, never linked | Model knows you but has no page to bind to | Publish a page that directly supports the claim |
Fetched but not cited: the most fixable problem
A citation binds to a sentence, not a whole answer. So being topically relevant is not enough. You need to be the best support for a precise claim. Windsor.ai's practical tips fit this well: open the page with the sentence you want quoted, and give each page at least one fact, figure or measured result only you have. Their pages that kept citations through the summer were built on their own numbers, while generic definition pages lost out to the original source or the vendor.
Cited but not mentioned
If ChatGPT links your page but the answer text never names you, you are a supporting source, not a recommendation. Track position in the answer and sentiment, not just presence. Promptwatch's advice here is to track your position and how ChatGPT describes you, and to earn mentions on third-party review, listicle and forum pages.
Tools that help you read the list at scale
The manual method gets you a first diagnosis. After that, you want repeatable tracking. A few options from the catalog, each with a different emphasis.
Promptwatch tracks real ChatGPT and other AI interfaces, shows which of your pages are cited, and pairs that with IP-verified AI crawler logs so you can see fetches next to citations. It also covers fan-outs and position in the answer, and has an Action Board and Content Agent for fixing gaps once found. That fetch-to-citation join is the part most trackers skip.
Profound tracks brand visibility across AI engines and is a common enterprise choice for citation monitoring.
Peec AI is a lighter prompt tracker with multi-language coverage, and the company has published useful work on how ChatGPT's index changed.
Otterly.AI is a budget monitoring tool if you mainly need to know whether you appear.
Ahrefs Brand Radar adds AI mention tracking to an SEO suite you may already use.
| Tool | Fetched vs cited visibility | Best for |
|---|---|---|
| Promptwatch | Crawler logs plus citation and page tracking | Teams that want to diagnose and fix |
| Profound | Citation and brand monitoring | Enterprise monitoring |
| Peec AI | Prompt tracking | Lean multi-language tracking |
| Otterly.AI | Basic monitoring | Low-budget presence checks |
| Ahrefs Brand Radar | Mentions inside an SEO suite | Existing Ahrefs users |
For wider comparisons, the GEO software directory at bestgeosoftware.com lists many more.
A weekly routine that works
- Pick 20 to 30 prompts that matter commercially. Mix organic, competitor comparison and branded ones, since each behaves differently.
- Record, for each prompt, whether you are fetched, cited or mentioned, and who else is in the list.
- Match against logs. If OAI-SearchBot never reaches a page, nothing downstream matters. If it does and no citation follows, work on the page itself.
- Fix one thing per page and note the date. Windsor.ai's suggestion is to check whether citation fetches rise within the month after a change.
- Re-check after model updates. Citations per response dropped around the GPT-5.3 rollout on March 4, 2026 (citation drop report), and behavior shifted again in August. Your baseline is a moving target.
What not to conclude
A few traps worth naming.
- Do not treat a ChatGPT-User fetch as a citation. It means the page was considered.
- Do not copy numbers from small studies into a strategy deck. Suganthan says as much about his own: one account, a few dozen queries.
- Do not respond to a missing citation with more pages. Dedupe by domain means thin pages dilute rather than add.
- Do not build a format strategy around markdown files. They make up 0.05% of AI search citations (markdown in AI search).
- Do not assume what worked three months ago still works. Reddit's collapse in ChatGPT citations happened in a single day.
The source list is a diagnostic, not a scoreboard. Read it as a funnel: indexed, fetched, quoted, named. Find the step where your pages disappear, fix that step, and check the next week's list.