The number of sources an AI answer cites is set mostly by the engine's route, its model version, and whether it searched at all. In the Machine Relations Index answer ledger, 15,381 successful runs of 1,008 buying questions between 2026-10-08 and 2026-11-07 averaged 9.98 cited sources on Perplexity, 4.53 on ChatGPT, 4.36 on Claude, 2.04 on Google AI Mode and 1.21 on Gemini. Gemini's low figure is not a small answer. When Gemini cited anything it cited 11.4 sources; it cited nothing in 89.4 percent of runs.
The same engine's number also moves when its model does. Over the four months of the ledger, Gemini ran on five recorded model versions, and its mean fell from 13.8 sources per run to 1.3 while the sources per citing run stayed between 10 and 15. What changed was the share of answers that cited anything.
So a sentence such as "ChatGPT cites five sources" is a statement about a route, a model and a date. This page gives those for five engines. It does not say what the consumer apps do.
What was counted #
- Population. Successful runs in the Index's buying-question panel from 2026-10-08 to 2026-11-07, on its 1,008 buying questions (168 in each of six question shapes). Questions flagged as naming a brand in the prompt are excluded by rule; none of the 1,008 fell out. The ledger export is dated 2026-11-08 08:10 UTC. The window starts after Perplexity's last change of source count (2026-10-07), and no engine changed its recorded model inside it.
- Engines. ChatGPT (gpt-4.1-mini with web search required), Claude (Claude Haiku 4.5 with web search available), Gemini (gemini-3.8-flash), Google AI Mode and Perplexity (sonar), called through their APIs. These are routes the Index uses, not the consumer apps. Google AI Overviews are left out. Google AI Mode answered 634 runs on 144 questions, so its intervals are wide.
- Unit. A run is one engine's answer to one question on one day. A cited source is one distinct URL in the run's recorded citations. A run that cited nothing counts as zero, so the first table is the average over all runs, and the "when it cited" column drops the empty ones.
- Runs. Perplexity 3,995, ChatGPT 3,979, Claude 3,922, Gemini 2,851, Google AI Mode 634. These are the same runs as the engine-overlap study, and the cited-run counts match it exactly.
- Uncertainty. 95 percent intervals resample the questions with replacement, keeping each question's runs together, 2,000 draws, fixed seed.
Five engines, five numbers #
Sources per run, 2026-10-08 to 2026-11-07 (95% interval over questions):
| Engine | Runs | Mean sources per run | Runs with no source | Mean sources when it cited | Distinct domains per run |
|---|---|---|---|---|---|
| Perplexity | 3,995 | 9.98 (9.97 to 9.99) | 0.0% | 9.98 (range 7 to 10) | 8.63 (8.56 to 8.70) |
| ChatGPT | 3,979 | 4.53 (4.36 to 4.69) | 8.0% (6.7 to 9.3) | 4.92 (middle 80% from 2 to 9) | 4.29 (4.13 to 4.44) |
| Claude | 3,922 | 4.36 (4.21 to 4.50) | 22.6% (20.2 to 24.8) | 5.62 (middle 80% from 4 to 8) | 4.04 (3.90 to 4.18) |
| Google AI Mode | 634 | 2.04 (1.81 to 2.29) | 29.3% (23.5 to 35.6) | 2.89 (middle 80% from 1 to 5) | 1.90 (1.68 to 2.12) |
| Gemini | 2,851 | 1.21 (0.99 to 1.43) | 89.4% (87.5 to 91.1) | 11.36 (middle 80% from 5 to 20) | 1.01 (0.83 to 1.19) |
Three readings follow.
Perplexity's number is a setting. It returned 10 sources in nearly every run (never fewer than 7 when it cited), and its most common count in August and September was 20. The route's source count changed on 2026-10-07; this page cannot see any change in Perplexity itself.
ChatGPT and Claude land in the same place for different reasons. ChatGPT's route requires a search and sends the buyer's sentence as its one query (how the engines search), so 92 percent of its runs cite. Claude chooses whether to search. Every Claude run with no search (885 of 885) cited nothing, and every run with at least one search cited something. Claude with one search averaged 4.8 sources, with two 6.0 and with three 6.5.
Gemini mostly does not cite on this route. In 2,548 of its 2,851 runs it recorded no search and no sources. In the 303 runs where it did, it cited 11.4 sources, and the count rose with the searches it ran: 5.1 sources with one search, 9.0 with two and 15.6 with four. A mean of 1.2 describes a mixture of many empty answers and a few long ones.
The question changes the count, and the engines differ in how #
Mean sources per run by question shape (95% interval):
| Question shape | ChatGPT | Claude | Claude: no source | Gemini | Gemini: no source |
|---|---|---|---|---|---|
| Best X | 7.04 (6.65 to 7.42) | 4.94 (4.70 to 5.19) | 6.1% | 2.27 (1.63 to 2.97) | 78.3% |
| Top lists | 6.62 (6.18 to 7.05) | 5.58 (5.40 to 5.76) | 1.4% | 3.69 (2.80 to 4.68) | 72.0% |
| Problem first | 3.98 (3.75 to 4.23) | 1.25 (0.96 to 1.58) | 77.2% | 0.05 (0.00 to 0.14) | 99.6% |
| How to choose | 3.53 (3.22 to 3.84) | 4.78 (4.39 to 5.15) | 23.9% | 0.37 (0.11 to 0.69) | 96.6% |
| Is it worth it | 3.13 (2.92 to 3.33) | 4.26 (3.99 to 4.52) | 17.2% | 0.41 (0.17 to 0.73) | 95.2% |
| X vs Y | 2.89 (2.68 to 3.09) | 5.37 (5.12 to 5.62) | 8.7% | 0.43 (0.17 to 0.73) | 94.6% |
Perplexity returns 9.96 to 9.99 for every shape.
On ChatGPT, a "best" or "top" question carries about 2.4 times as many sources as an X-versus-Y question, and the intervals do not overlap. Claude's highest counts are on top lists and X versus Y (5.6 and 5.4), and on problem descriptions it mostly does not search: 77 percent of those runs, against 1 to 6 percent for best-X and top-list questions. Gemini cites mainly on list-shaped questions. For Claude and Gemini the question decides whether anything is cited at all. For ChatGPT, which is required to search, it decides how many.
In practice, a buyer asking ChatGPT for the best tools in a category gets about seven sources to appear among. The same buyer asking it to compare two named products gets about three. That is a count of places, not a statement about who fills them.
Inside a run, sources repeat domains. Among runs that cited, the share of cited URLs that were an extra page from a domain already cited in the same answer was 14.8 percent for Gemini, 13.5 for Perplexity, 6.9 for Claude, 4.8 for Google AI Mode and 4.3 for ChatGPT. Ten Perplexity sources are on average 8.6 different sites.
One engine name, several models, very different counts #
The ledger records the model string each route reported. Over 2026-07-06 to 2026-11-07, Gemini ran on five of them:
| Gemini model (as recorded) | First to last run | Runs | Mean sources per run | Runs with no source | Mean sources when it cited |
|---|---|---|---|---|---|
| gemini-2.5-flash | Jul 6 to Jul 9 | 254 | 13.77 | 8.3% | 15.0 |
| gemini-3.5-flash | Jul 10 to Jul 21 | 436 | 11.59 | 18.1% | 14.2 |
| gemini-3.6-flash | Jul 22 to Aug 13 | 592 | 4.06 | 61.0% | 10.4 |
| gemini-3.7-flash | Aug 14 to Sep 2 | 588 | 0.54 | 94.7% | 10.2 |
| gemini-3.8-flash | Sep 3 to Nov 7 | 4,787 | 1.28 | 88.6% | 11.3 |
Each model ran the question set in roughly even shares across the six shapes, so the fall is not a change of question mix. But each model also ran on different dates, so this table cannot separate a model from anything else that changed on the web or in the route over the same weeks. It shows the size of the swing and where most of it sits: the sources per citing run moved from about 15 to about 11, and the share of runs that cited anything moved from 92 percent to 11.
The other routes moved less. ChatGPT's route changed from gpt-4o-mini-search-preview (2026-07-06 to 08-19; 1,486 runs, mean 4.84, 4.2 percent with no source) to gpt-4.1-mini (from 08-22; 6,201 runs, mean 4.54, 7.3 percent). Claude Haiku 4.5 stayed the same model throughout: monthly means 4.06, 4.15, 4.49, 4.30 and 4.44 for July to November. Google AI Mode records no model, and its mean fell from 17.4 sources per run in July to 4.3 in August, 1.1 in September and 2.2 in October, as the share of runs with no source went from 8 percent to 79, 59 and 32 percent. Perplexity's monthly mean was 12.3, 19.1, 19.3, 11.6 and 10.0.
What this changes #
- Quote a count with its engine, model, route and date. "AI engines cite four sources" and "AI engines cite twenty" were both true on some route in this ledger within the same quarter. The published range of 2 to 22 is mostly this.
- Track whether the engine searched before reading a drop in citations. A fall in cited sources, or in a brand's presence, can come from an engine searching on fewer questions. On this ledger that was most of Gemini's change between 3.5 and 3.7 and most of Google AI Mode's change from July to August.
- Count slots by question shape. On ChatGPT, list questions give the most places to appear. On Claude and Gemini, a problem described in plain words often returns no sources at all, so there is nothing to appear in.
- A pooled "sources per answer" moves when an engine's setting moves. The 6.99 pooled figure in the September 25 release covered months in which Perplexity's most common count was 20. The per-engine rows above are the part to carry forward.
What this page does not show #
- Consumer products. These are API routes with the settings the Index uses. ChatGPT, Claude, Gemini and Google AI Mode in their own apps, and Google AI Overviews, may cite more or fewer.
- The routes' ceilings. Perplexity's count is capped by the request, and Claude's and Gemini's search tools may have limits the ledger does not record. A larger mean for an engine here is not a larger appetite for sources.
- Why Gemini and Google AI Mode stopped citing. Gemini recorded no search in most runs. Whether that is the model's choice, a route setting or a change in the search tool is not in the data.
- Sources consulted. An engine may read pages it does not list. The sources it cites without naming are a separate matter.
- Buyer demand. The 1,008 questions are the Index's own panel for coverage of buying decisions, not a sample of what buyers type.
- Cause. Nothing here tests whether more sources in an answer changes which brand a buyer picks.
Method note #
Source: the Index answer ledger's run table, daily export of 2026-11-08 08:10 UTC, read with DuckDB, joined to the citation-events table for distinct URLs and domains per run. Filters: the Index's buying-question panel, successful runs, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity, run dates 2026-10-08 to 2026-11-07 for the engine and shape tables and 2026-07-06 to 2026-11-07 for the model and monthly tables. Sources per run: the run's recorded citation count, which equals the number of distinct cited URLs in the events table (a URL repeated within a run is counted once). Domains per run: distinct root domains among the run's cited URLs. Search counts are those the route recorded; they were not recorded for Claude and ChatGPT before September and are not reliably recorded for Gemini before its 3.7 model, so search-count figures use the window only. Shares and means are pooled over runs, not averages of per-question values. Intervals resample questions with replacement within each engine or shape, 2,000 draws, fixed seed. The repeated-domain share is the mean over citing runs of one minus distinct domains divided by distinct URLs.
Frequently asked questions #
How many sources does ChatGPT cite? #
On the Index's API route, 4.5 per answer on average (4.4 to 4.7), 4.9 when it cited at all, with 8 percent of runs citing nothing. Best-X and top-list questions carried about 7 sources and X-versus-Y questions about 3.
How many sources does Perplexity cite? #
10 per answer on the Index's route since 2026-10-07 (8.6 different domains on average), and most often 20 in August and September. The count is set by the request, so it says little about Perplexity's own preference.
How many sources does Claude cite? #
4.4 per answer on average, 5.6 when it cited. It cited nothing in 22.6 percent of runs, because it ran no search: every run with no search had no sources, and every run with a search had some. On problem descriptions it did not search in 77 percent of runs.
How many sources does Gemini cite? #
1.2 per answer on average on the current route, because it recorded no sources in 89 percent of runs. When it cited, it cited 11.4 on average. The same engine name averaged 13.8 sources per run on gemini-2.5-flash in July.
Why do studies report different numbers of sources per answer? #
Because the number depends on the engine's route, the model version, the date, the question and whether answers with no sources are counted as zero. In this ledger, one engine's mean moved by a factor of 25 across its model versions while its sources per citing run moved from 15 to 10.