We respect your privacy.

We use strictly necessary cookies to keep you signed in and to protect against CSRF. With your permission we also use a small amount of first-party analytics to improve the product. We do not sell your data and we do not use third-party advertising trackers. See our cookie policy and privacy policy .

← All posts

ChatGPT's links are not its citations

Crawlmind Engineering··5 min read

A ChatGPT in-answer link is a clickable URL placed inside the text of a reply, and it is a different object from the cited sources ChatGPT attaches to the same answer.

That distinction used to be academic, because inline links were rare. New tracking data from Peec AI suggests they are now common, and that they mostly point somewhere other than the citation list. If your AI visibility reporting only reads citations, it is missing a surface that ChatGPT now uses in close to half of the conversations Peec tracks.

#What Peec AI measured

David Konitzny, a GEO researcher at Peec AI, posted daily figures on LinkedIn that PPC Land summarized on November 16, 2026. The metric is the share of ChatGPT conversations with at least one external link, where an external link means a clickable link inside the answer itself, not a citation and not a link that opens the sources sidebar.

The reported numbers, all via PPC Land's write-up:

  • Around 10% of chats had an in-answer link until early October (PPC Land).
  • The share reached 41% on October 9, peaked at 49.5% on October 14, and stood at 47.8% on November 14 (PPC Land).
  • Among chats with links, the share containing exactly one link fell from about 43% to 18%, and chats with five or more links became the largest group at almost 39% (PPC Land).
  • The average number of external links per linked chat went from roughly 2 to 4 (PPC Land).

The finding that matters most for measurement is the overlap test. According to the same article, 63% of external URLs did not match any cited source URL in the same conversation, and 47% of linked domains did not match any cited source domain.

#Why the timing points to a model change

Peec's chart marks October 7 as the day GPT-6 became the default, and the rise starts two days later. OpenAI's ChatGPT release notes record the GPT-6 rollout in ChatGPT starting October 7, 2026, alongside a new interface layer that adds interactive elements to replies. A step change that lines up with a model launch is the same pattern we described in model swaps are the new core updates: broad, sudden, and not announced as a change to sourcing.

Correlation with a launch date is not proof of cause. OpenAI has not said anything about inline linking behavior, and the Peec post does not claim to know why the share moved. The dip in late October and the recovery in early November are also unexplained.

#What the data does not tell you

Read the limits before you plan around these numbers. PPC Land notes that the post gives no sample size, prompt set, market, language, ChatGPT plan, or login state, and that the figures have not been independently replicated. Peec also sells AI visibility monitoring, which gives it a commercial interest in the story.

Three further gaps matter:

  • No click data. Konitzny himself said more external links do not automatically mean more clicks or referral traffic. A link in an answer is an opportunity, not a visit.
  • Different denominators. In Peec's figures, the 63% figure is about URLs and the 47% figure is about domains. The gap between them suggests some inline links go to a different page on a site that is already cited, but that is an inference, not a measurement.
  • No link-type breakdown. In replies, Konitzny said the direct links usually point to a homepage or a product page. No share was published, so treat that as an observation from one researcher.

#Why this changes how you track ChatGPT

Most AI visibility tools, including the citation checks most teams run by hand, read the source list ChatGPT returns with an answer. That list answers one question: which documents did the model ground on. Inline links answer a different one: where did the model send the reader.

If the two lists overlapped heavily, reading only citations would be a fair proxy. The Peec numbers say they do not. A brand can be linked in the answer text without appearing in the citations, and a page can be cited as evidence while the click opportunity goes to a competitor's homepage in the paragraph above it.

The practical consequences:

  1. Citation share understates presence. A brand that is named and linked in the body but never cited shows up as zero in a citation-only report.
  2. Citation share can overstate traffic potential. Being one of several cited sources, often tucked behind a sources panel, is a weaker position than an inline link in the sentence that recommends you.
  3. The landing page differs. If inline links lean toward homepages and product pages, the page ChatGPT sends people to may not be the page it read. That has consequences for which pages need clear pricing, product facts, and next steps.

#What to do this week

Capture both lists separately. For each tracked prompt, store the cited source URLs and the in-answer link URLs as two fields, with the model name and timestamp. Do not merge them into one "mentions" count. This is the same discipline we argued for in count verified mentions, not mentions.

Report three numbers, not one. Share of answers where you are cited, share where you are linked inline, and share where you are both. The "linked but not cited" bucket is the one that a citation-only tool hides.

Check where the inline links land. Group your in-answer links by landing page. If most go to the homepage, make sure the homepage states what you sell, for whom, and at what price in plain text a model can read. Product pages need the same facts near the top.

Confirm the click side in analytics. OpenAI's publishers and developers FAQ says ChatGPT adds utm_source=chatgpt.com to referral URLs from search results. Segment that traffic by landing page and compare it with the pages your tracking shows as inline links. If a page is linked often but gets no tagged visits, the link may not be prominent, or the referral may not be tagged on that surface.

Annotate October 7, 2026. Any ChatGPT trend line that crosses that date should carry a note. Comparing September citation share to November citation share without one mixes two different answer formats.

Keep engines apart. The Peec data covers ChatGPT only. Do not apply the inline-link pattern to Gemini, Perplexity, or AI Mode without measuring each one.

#The short version

A citation tells you what ChatGPT read. An inline link tells you where it pointed the reader. Since early October those two lists have diverged enough that tracking one tells you little about the other. Store both, report both, and check whether the pages that receive the links are ready for the visitors they may send.

Related field notes

Share or discuss

Field notes in your inbox

New posts, no spam. Roughly monthly. Unsubscribe with one click.