Why don't most of ChatGPT's links count as citations?

Most links ChatGPT shows in a chat weren't the sources it actually used. Research puts that gap at 63% of linked URLs going uncited.

Sited blog title card: Why don't most of ChatGPT's links count as citations?

Most links ChatGPT drops into an answer are not the sources it used to write that answer. Peec AI's research found that 63% of the URLs ChatGPT links to in a chat do not appear among that same chat's cited sources. ChatGPT is increasingly happy to name or link a brand without ever having retrieved that brand's page to ground its answer.

What "citation" actually means inside a ChatGPT answer

A citation should mean a source ChatGPT actually pulled during its web search and used to support what it wrote. A plain in-answer link is different: it's a clickable URL inside the text, not necessarily tied to retrieval at all. Peec AI researcher David Konitzny, quoted by PPC Land, defines external links as "clickable links to external websites directly within ChatGPT's answers, not citations or links that open a sidebar." Tools that only count links shown are measuring something different from sources used.

The core finding

Peec AI's tracking found ChatGPT now links to websites in 48% of chats, up from 10% previously, and conversations with links now carry about four of them on average instead of two. Yet 63% of those linked URLs do not show up among the chat's cited sources. That's not a rounding error. A separate fan-out study based on 60 conversations, reported by PPC Land, found the gap runs deeper: of 3,554 pages ChatGPT retrieved, only 110 were cited, a rate of 3.1%. Most of what ChatGPT pulls in during search never surfaces as a visible source at all.

Why links and citations diverge

Retrieval and generation are not the same operation. PPC Land reported that ChatGPT's cited sources are tied to the pages its web search retrieves to ground an answer, while the in-answer links it writes often point somewhere else entirely. Konitzny put it plainly, as quoted by PPC Land: "This means that websites can appear as direct links in ChatGPT without being included in its cited sources."

The fan-out study adds a layer. In 21 of 27 conversations, the first query ChatGPT ran already contained brand names the user never typed, for example naming Roborock Saros, Dreame X50, Narwal Freo Z10 and Eufy S1 Pro for vacuum questions, or Car and Driver, Edmunds and Top Gear for car questions, according to PPC Land's reporting. The system decides which brands matter before it's retrieved anything, which helps explain why some brands get named consistently while others, retrieved but never queried by name, barely surface. For more on how often ChatGPT skips retrieval altogether, see Sited's piece on ChatGPT skipping web search.

Brands get named without being cited, and vice versa

The same fan-out research found that of 119 brands named in ChatGPT's own search query, 68.9% were then mentioned in the final answer, compared with just 2.1% of the 515 brands that were fetched but never named in the query, as PPC Land reported. Getting into ChatGPT's own query phrasing appears to carry far more weight than simply being present in the retrieved results.

This mismatch isn't unique to ChatGPT. A Semrush study by Kevin Indig analysing 3,981 domain appearances across ChatGPT, Gemini, Google AI Overviews and Google AI Mode found that 62% of AI citations are "ghost citations": the brand was named without a linked source, or linked without being named. The same study found Google itself was named in AI answers nearly three times more often than it appeared as a linked source, and found almost no overlap between which brands ChatGPT cited and which brands Gemini named for identical prompts. If you track visibility across engines, that's worth following up: see Sited's piece on why AI engines barely agree and the related Semrush data on platform disagreement in the UK.

How citation sourcing has shifted in 2026

Citation patterns aren't static either. An Ahrefs study of 1.4 million prompts found ChatGPT retrieves dozens of URLs per query but cites only about half of them. That same study found 67.8% of all non-cited URLs came from Reddit, despite Reddit having its own dedicated retrieval category and over 16 million data points in the dataset, with Reddit cited at just 1.93%.

That Reddit pattern then shifted. Peec AI's Q3 2026 tracking found Reddit citations fell to roughly a tenth of their second-quarter level between July and September, while brands' own product pages were cited about two and a half times as often and "best of" listicles about half as often. Sessions running multiple searches rose from 2% in July to 61% in September, with the average session now drawing on 19 sources instead of 11. This coincided with ChatGPT making GPT-5.6 its default on 8 August 2026, a change Peec AI notes was not publicly announced to marketers, per promptthemarket.com's reporting. For more on that period, see Sited's coverage of Reddit's citation drop and the GPT-6 Astra citation changes piece.

Matching the exact pages ChatGPT cites is hard even with the tools built for it. A preprint from NUS and Xi'an Jiaotong University, reported by Relevant Audience, found Google's organic results on their own contained fewer than one in ten of the exact pages ChatGPT cited, and across 20 repeat runs, 40.8% to 44.5% of retrieved URLs appeared only once. Rotation is fast too: SISTRIX research cited by PPC Land measures ChatGPT citation rotation at 74% per week.

What this means for tracking AI visibility

If you measure "AI visibility" by counting links in ChatGPT transcripts, you're measuring link behaviour, not sourcing behaviour, and the two now diverge on a majority of chats. A vendor study by MentionBird found gpt-5-mini returned no visible web citations on 29.7% of API answers, compared with 0.6% for gemini-2.5-flash, and across four months 8.4% of qualifying questions never produced citations at all, as reported by TheGeoLab. Some share of your "zero citation" results may simply reflect how the model was queried, not whether your content was found. See Sited's piece on ChatGPT's zero-source answers for more on that pattern.

A limitation worth being honest about: none of the studies cited here count every page retrieved behind every answer across the full ChatGPT user base. They are samples, sometimes small (27 or 60 conversations), and a vendor's own product usage. The direction is consistent across several independent sources, but exact percentages should be read as indicative rather than universal.

What founders and marketers should do about it

Our view: chasing citation count as a single metric is a losing game when the same engine cites differently week to week and disagrees with other engines on the same prompt. What seems to matter more consistently is being the kind of source that gets named in ChatGPT's own generated query, and being the kind of source engines are structurally inclined to pull from. Omniscient Digital's analysis of 23,387 citations, using Peec AI data for branded queries, found earned media accounts for 48% of all citations, commercial brand content for 30%, and owned brand content for just 23%. That lines up with Sited's piece on why AI engines trust earned media. A practitioner at Passionfruit advises treating earned coverage, not owned landing pages, as the priority channel for AI citation work.

Track brand mentions and citations separately. Expect week-to-week volatility rather than a stable ranking. Treat Reddit, listicles and your own product pages as separate channels with separately shifting odds of being cited. If you want a structured starting point, Sited's AI search readiness checklist walks through what to check first.

Frequently asked questions

Does a link in a ChatGPT answer mean ChatGPT used that page as a source?

Not necessarily. Peec AI's research found 63% of in-answer links do not appear among that chat's cited sources, meaning the link and the actual retrieval are often separate events.

Why does Reddit get retrieved so often but cited so rarely?

Ahrefs' study of 1.4 million prompts found 67.8% of ChatGPT's non-cited URLs came from Reddit, which was cited at just 1.93% despite having its own retrieval category. That pattern then shifted again, with Peec AI's Q3 2026 data showing Reddit citations falling to roughly a tenth of their earlier level.

Does being retrieved by ChatGPT mean my brand will get mentioned?

Not reliably. The fan-out study found only 2.1% of brands that were retrieved but never named in ChatGPT's own search query were mentioned in the final answer, compared with 68.9% of brands that were named in the query itself, per PPC Land. This counted only a small sample of 27 conversations, not every ChatGPT answer.

Is this citation gap the same across ChatGPT, Gemini and Google AI Overviews?

No, and the engines do not even agree with each other. Semrush's analysis of 3,981 domain appearances found 62% of citations across these engines were "ghost citations", and found almost no overlap between which brands ChatGPT and Gemini cited for the same prompts.

Sources