ChatGPT Cited 40% Fewer Sources per Answer in October: Data of 6,700 AI Answers

In the second week of October, ChatGPT changed how it shows its sources.

We track the same buyer questions every day on ChatGPT, Google AI Overviews and Gemini, logged out. That lets us compare ChatGPT before and after the switch with two other engines answering the same questions over the same weeks.

Trackers reported the change in opposite directions. Some saw fewer citations, others saw more links. In our data both are true. ChatGPT's source list got shorter, links written into the answer itself rose sharply, it searched less, and its source list leaned more on reference and news pages.

Here is what changed, what did not, and how to read the flood of numbers that followed.

The day ChatGPT's answers changed

Our tracker runs every question once a day. In the first nine days of October, about 15% of ChatGPT answers wrote a link into the text itself. On 10 October that became 3 in 4, overnight, for both brands we track.

Other trackers saw the same switch. Sistrix reported the new answer format in all six countries it tracks from 9 October, and found the end of answer source list missing from most answers until 22 October (Sistrix, October 2026). Data from Peec AI across a million prompts showed links written into the text in 42% of responses after the switch, against 8% to 10% before (Search Engine Journal, October 2026). Peec AI is also the tool we track with, so that figure is not fully independent of ours. Our tracker first shows the new format in its 10 October runs.

GPT 6 and a redesigned answer layout reached logged out users within the same few days, so no outside study can say which of the two caused what. We treat it as one product change.

Your analytics may have dipped too, for a different reason. For about a week from 9 October, links from ChatGPT lost their tracking tags. In the first days, analytics tools showed ChatGPT visits down by about 46% while direct traffic rose. Sistrix found tagging back to normal by 17 October (LLM Pulse, October 2026; Sistrix). If your October report shows a ChatGPT traffic drop, check those dates before acting on it.

Did ChatGPT start citing less?

To answer that fairly, we took the 12 questions (6 per brand) that ran every day on all three engines, and compared a clean week before the switch with the latest full week after. The weeks in between are left out. On many days between 8 and 20 October, a large share of ChatGPT answers showed no sources at all, which matches what Sistrix saw in the same window, and that temporary gap would distort a before and after comparison.

The source list shrank, and the links moved. Peec AI, the tool we track with, counts a link written into the answer text separately from a cited source, never as both (Peec AI docs). Counted together, ChatGPT linked 9.2 pages per answer before the switch and 8.2 after. Some pages may appear in both lists in other answers, so treat the combined figure as an upper bound.

A typical answer after the switch shows the pattern. Asked how to send money abroad, ChatGPT listed seven providers, each followed by a link labelled Official website. Its four cited sources were all independent comparison guides. None of the seven homepages appeared in the source list.

ChatGPT now retrieves fewer pages, and cites a smaller share of them. Across each week it also drew on a wider spread of websites: 204 against 144 for one brand, 98 against 82 for the other.

Being named did not change. A separate test by Proofsource of 188 buyer questions found brands named in 12.4% of answers before GPT 6 and 13.0% after, a difference inside the margin of error (Proofsource, October 2026).

Other trackers count differently and report different directions. Sistrix counted 3.3 linked pages per logged out answer before the switch and 4.1 after. Proofsource counted 5.1 citations before and 8.4 after. A tracker that counts only the source list will show a fall. One that counts every link may show a rise. Both are measuring something real.

What ChatGPT searches for now

Before it answers a buyer question, ChatGPT usually runs its own web searches. We read every search our tracker recorded for both brands. The tracker does not record searches for every answer, so the last row shows how complete each period is.

The clearest test is the middle two columns. On 9 October and on 10 to 11 October our tracker recorded the searches for every answer that searched, and the number of searches per answer fell by more than half overnight. The clean weeks either side show the same direction. Counted per answer, including answers whose searches our tracker did not record, the drop between those weeks is smaller, about 28%, partly because the tracker recorded more answers after the switch. ChatGPT still ran more than one search in 80% to 90% of answers that searched. What fell most was searching inside a specific website, a habit that appeared in our data after ChatGPT's August update.

An analysis by David Konitzny of Peec AI across 5.7 million searches found the same direction: about 3 searches per chat down to 2, and searches inside one website down from 22.6% to 8.3% (via Masters of Search, October 2026). Our level of website searches is higher, likely because our questions are commercial, and buyer questions tend to push ChatGPT to check specific company and review sites.

Where ChatGPT's citations go now

The mix of ChatGPT's source list moved at the switch. On the same questions, Google AI Overviews barely moved.

Read this table together with the links in the text. On our 12 questions, almost all of those links went to company and product sites. Peec AI's wider data found 54% of them going to homepages (Search Engine Journal). Count both kinds of link and the company share falls only from about 80% to about 75%. So much of the shift in the source list is company links moving into the text, not ChatGPT turning away from company sites. What changed is the job each list does. The links in the text point at the companies named, and the source list points at the pages that back up what the answer says.

If this pattern holds, it splits the job for brands in two:

1. Get named. A link in the text now comes with being named in the answer, so a mention is worth more than it was. Being named itself did not become more common.
2. Be on the pages ChatGPT checks. Reference, comparison and news pages now carry a bigger share of the source list, so coverage there counts for more in that list than another page on your own site.

Why AI search numbers disagree

October produced a wave of confident numbers that seem to contradict each other. Most of the disagreement comes from four sources.

Different questions. How often does ChatGPT go to the web for an answer? Similarweb puts the share of US ChatGPT responses with citations at about 7% across all chats, on data to May 2026 (Similarweb, 2026). Ahrefs found search queries in 14.4% of 66.6 million responses to broad prompts (Ahrefs, October 2026). On commercial comparison questions, Graphite measured 88.5% in late September (Graphite, October 2026). On our buyer questions in the last week of October, it was 99.5%. All four can be right.

Different denominators. A share of answers, a share of citations and a share of answers that searched are three different numbers. A figure without its denominator can't be compared with anything.

Different surfaces. Many studies query the engines through developer APIs instead of the apps people use. Researchers who compared the two found that Google results plus OpenAI's API together recovered only 26% of the pages ChatGPT cited in its app (Jiang et al., October 2026).

Randomness. Ask the same question twice and you get different sources. Two answers to the same prompt on the same day repeat 41% of their cited websites on ChatGPT and 45% on Gemini, against 94% on Perplexity (Muck Rack, October 2026). In our own data, a consumer brand's daily visibility moved by 4 to 8 points from one day to the next on average across 12 to 25 October, depending on the engine.

The surfaces and the trackers change too. Between 8 and 20 October, as few as a third of ChatGPT answers on some days showed any sources, then it recovered fully. A study that happened to sample those days would report a very different ChatGPT. That is why our before and after comparisons use the clean weeks either side, and why the search figures above report how complete each period is.

The research itself is young. A review of 38 published brand visibility studies found 35 had no peer review, only 2 checked their results against an outside benchmark, and only 7 repeated their measurement over time (Ibrahim, Research Square, October 2026). An audit of 13 AI visibility trackers found that none publishes a validation of its own data (Jacob Whitmore, October 2026).

How to read an AI search study

Before you act on a number, ask eight questions.

How much data a claim needs

Most AI visibility reports are built on far less data than their claims need. Standard sampling math gives a rough guide.

By the same math, our own 12 question panel can show large shifts, like links in the text jumping from under 1 in 5 answers to 3 in 4. It is too small to measure fine differences between brands or categories, which is why the full study widens it.

The useful lesson is in the third row. Running one question a hundred times tells you a lot about that question and very little about your market. Answers to the same question are alike, so extra runs of it add less than new questions do. Spread your effort across many real buyer questions, run each a few times, and report a range, not a single number.

What we are testing next

October left three open questions, and we are working on each:

1. The full October comparison. The same design as this article on daily data through the end of November, with ranges on every estimate and checks against random dates, published in December.
2. App versus API. The same 100 questions asked through the ChatGPT, Gemini and Perplexity apps and their APIs for 30 days, to measure how far the API drifts from what users see.
3. Does the page say that? A check of about 3,000 claims in AI answers against the pages cited for them, with people reading a sample, to see how often the cited page supports the claim.

We will publish what we find, whichever way it goes.

How we gathered this data

• Setup: daily tracking with Peec AI, logged out, on ChatGPT, Google AI Overviews and Gemini. The B2B brand's questions run from the US; each of the consumer brand's panel questions runs from a single country. Perplexity was tracked until mid September and is not in the October comparison.
• Brands: two client brands tracked throughout, one B2B software company and one consumer finance brand. Names are withheld, and the questions are not published because they name the brands and their competitors.
• Before and after: 12 questions (6 per brand) taken from those that ran every day on all three engines in both weeks. One of the 12 names a brand in the question. Before is 28 September to 4 October, after is 26 October to 1 November, 84 answers per engine per week (80 to 83 for AI Overviews, which does not appear for every search). The weeks in between are left out because of the source gap described above.
• Daily and weekly series: all questions for both brands, 1 October to 2 November for the daily share of answers with links in the text, and 28 September to 1 November for searches.
• Definitions: a citation is a source in the answer's source list. A link in the text is a URL written into the answer itself, counted separately. Website types follow Peec AI's classification, with brand, competitor and vendor sites grouped as company sites.
• Limits: two brands, two categories, mostly commercial questions, and one tracker. These are counts from one week each side of the switch, not a significance test. The December study adds ranges. Google also changed how links show in AI Overviews on 15 October (Google, October 2026). Google described its change as moving links higher in the answer. In our data the share of AI Overview answers with links in the text fell from about 25% to about 12% from that week, while its cited sources per answer rose from 8.6 to 10.1. Gemini also cited a smaller share of the websites it retrieved, 79% against 97%, and less community content. Neither comparison engine was perfectly still, which is why we read them as a check against web wide shocks, not as an exact baseline.

Frequently asked questions

Every AI search number is a measurement of a moving system, so ask what it counted and when.

Sources

1. Sistrix, ChatGPT in October, GPT 6, UTM parameters and sources, October 2026.
2. Search Engine Journal, ChatGPT traffic drop appears to be a tracking problem, October 2026.
3. LLM Pulse, Why ChatGPT referral traffic fell in October, October 2026.
4. Peec AI documentation, Get links report, 2026.
5. Proofsource, GPT 6 made ChatGPT answers longer and named the same brands, October 2026.
6. Masters of Search, GPT 6 reverses GPT 5.6 search patterns, Konitzny finds in 2.2M responses, October 2026.
7. Similarweb, AI search stats 2026, 2026.
8. Ahrefs, How ChatGPT does query fan out, 15 million queries studied, October 2026.
9. Graphite, Parametric Knowledge Optimization, October 2026.
10. Jiang et al., Understanding ChatGPT citations, evidence from production web search across model generations, arXiv, October 2026.
11. Muck Rack, What is AI reading, October 2026.
12. Ibrahim, Brand visibility in generative AI answers, a scoping review, Research Square, October 2026.
13. Jacob Whitmore, AI visibility tracker methods audit, October 2026.
14. Google, Updates to AI Mode and AI Overviews to help you explore the web, October 2026.
15. THREADSLAB tracking in Peec AI, two brands, 28 September to 2 November 2026, logged out.