The typical publication's citation rate on Gemini fell by a factor of eight between this summer and this autumn, and nothing in the ledger suggests the publications changed. In the Machine Relations Index answer ledger, 378 buying questions were asked of Gemini on two model generations: gemini-2.5-flash and gemini-3.5-flash from June 2 to July 21, then gemini-3.7-flash and gemini-3.8-flash from August 14 to November 9. In the first period 85.5 percent of the answers cited at least one source (590 of 690). In the second, 14.1 percent did (281 of 1,999). The typical domain's citation rate fell 8.1-fold (5.6 to 10.3). Its share of the answers that did cite moved 1.33-fold (0.93 to 1.54), which is not distinguishable from no change.
So a Gemini citation rate on this route is mostly a measure of whether Gemini searched. It is a poor measure of whether a publisher lost standing. The ledger cannot say why Gemini stopped searching, and it cannot separate the model from other things that changed over the same weeks.
What was counted #
- Population. Successful Gemini runs in the Index's buying-question panel, restricted to questions the Index marks eligible for public scoring. Gemini is called through its API with the Google Search tool attached to every call that recorded a model. This is not the consumer Gemini app and not Google AI Mode or Google AI Overviews. The ledger export is dated 2026-11-09 08:10 UTC.
- Two periods. The Index records the model string each run reported. Period A is gemini-2.5-flash (June 2 to July 9) and gemini-3.5-flash (July 10 to July 21). Period B is gemini-3.7-flash (August 14 to September 2) and gemini-3.8-flash (September 3 to November 9). gemini-3.6-flash, which ran July 22 to August 13, is the half-way model and is left out of the contrast, as are 215 older runs with no model recorded.
- Same questions. 378 questions were asked in both periods: 690 runs in A, 1,999 in B. Only these are compared, so a change in which questions were asked cannot produce the result. Period A ran news-topic questions that period B did not, and they are not in the pairing.
- A citing answer. A run whose recorded citation count is above zero. A cited domain is counted once per answer however many of its pages the answer listed.
- Intervals. 95 percent intervals resample the 378 questions with their runs and cited domains kept together, 2,000 draws, fixed seed.
The same questions, asked of two Gemini generations #
Share of Gemini answers that cited anything, on the 378 questions asked in both periods:
| Period | Models | Runs | Answers with a cited source | 95% interval |
|---|---|---|---|---|
| A: June 2 to July 21 | gemini-2.5-flash, gemini-3.5-flash | 690 | 85.5% (590) | 82.4 to 88.5 |
| B: August 14 to November 9 | gemini-3.7-flash, gemini-3.8-flash | 1,999 | 14.1% (281) | 11.0 to 17.5 |
Across every eligible question rather than the paired set, the model-by-model figures are in the sources-per-answer study. The pattern repeats here: when a Gemini answer cited at all, it cited about the same number of sources in every generation (a median of 15 on gemini-2.5-flash, 14 on 3.5, 9 on 3.6, 9 on 3.7 and 11 on 3.8). What changed was how many answers cited anything. Gemini 3.8's monthly figure has been steady since it arrived: 87.4 percent of answers cited nothing in September, 89.0 in October and 88.7 in November to date.
Question shape decides which answers still cite. Share of answers with a cited source, across all eligible questions (period A from 110 to 120 runs per shape, period B from 884 to 926):
| Question shape | Period A | Period B |
|---|---|---|
| Top list | 99.1% | 28.3% |
| Best X | 100.0% | 21.9% |
| X versus Y | 84.5% | 6.0% |
| Is X worth it | 90.8% | 4.9% |
| How to choose | 80.8% | 2.9% |
| Problem-first (a situation, no product named) | 56.8% | 0.3% |
Industry decides too. On gemini-3.8-flash, 41.0 percent of the AI visibility and GEO questions (210 runs) still cited a source, against 0.5 percent of cybersecurity questions and 0.6 percent of martech and advertising questions. The industry cuts are small (about 200 runs each) and shape and industry are not independent, so treat this as a pointer.
What happened to each publication #
The table follows the domains that Gemini cited most in period A. "Rate" is the share of all Gemini answers that cited the domain. "Share of citing answers" is the same count divided only by the answers that cited something. On the paired questions:
| Domain | Answers citing it, A to B | Rate in A | Rate in B | Share of citing answers, A | Share of citing answers, B |
|---|---|---|---|---|---|
| reddit.com | 199 to 117 | 28.8% | 5.9% | 33.7% | 41.6% |
| youtube.com | 158 to 66 | 22.9% | 3.3% | 26.8% | 23.5% |
| medium.com | 125 to 20 | 18.1% | 1.0% | 21.2% | 7.1% |
| forbes.com | 51 to 37 | 7.4% | 1.9% | 8.6% | 13.2% |
| semrush.com | 20 to 21 | 2.9% | 1.1% | 3.4% | 7.5% |
| tryprofound.com | 12 to 41 | 1.7% | 2.1% | 2.0% | 14.6% |
Forbes was cited in 7.4 percent of Gemini answers in period A and 1.9 percent in period B. Read on its own, that is a fall to a quarter. Read against the answers that cited anything, Forbes went from 8.6 percent to 13.2 percent. Reddit's rate fell from 28.8 to 5.9 percent while its share of citing answers rose. Nothing in a rate alone tells those two stories apart.
Across the 76 domains cited in at least 8 answers in period A and at least once in period B:
- The median rate fell 8.1-fold (5.6 to 10.3).
- The median share of citing answers moved 1.33-fold (0.93 to 1.54).
The share is not a stable ranking either. Of the 109 domains cited in at least 8 answers in period A, 33 were cited in no answer in period B. The rank order of the 109 correlated at 0.44 between periods (Spearman, 0.22 to 0.49), and the top ten domains by answers shared 5 members (3 to 6). The answers that still cite are a different subset of questions, mostly lists and best-X questions, and the domains they cite are a different mix. tryprofound.com is the clearest case: cited in 12 answers in period A and 41 in period B, with about a sixth as many citing answers per run. That fits the questions Gemini still searches being disproportionately about AI visibility, though this ledger does not test it.
This site's own domains are in the data but too thin to read. authoritytech.io was cited in 6 paired answers in period A and 3 in period B, and machinerelations.ai in none and then 1.
What this changes #
- Do not read a Gemini rate as a publisher's standing. A falling Gemini citation rate for a publication, a brand or a whole category across these months is explained first by the share of answers that cited anything. Check that share before reading the rate.
- Report two numbers. The share of answers that searched, and the brand's share among the answers that did. A single rate mixes them.
- Pin the model. A per-engine comparison across a date range that spans a model change is two instruments joined end to end. Report each model period beside the others, or compare inside one.
- Where Gemini still cites, the question is a list. On the current route, Gemini cited sources on about 3 in 10 top-list questions and on almost none of the problem descriptions. A brand's route into Gemini's cited answers runs through list and best-X questions, and for AI visibility topics.
What this page does not show #
- Why Gemini stopped. The Google Search tool is attached to every call, and in the runs with no source the response records no grounding, so the model or the tool declined to search. A model choice, a route setting or a change in Google's tool would look the same in the ledger.
- Model against date. Each model ran in its own weeks. The pairing holds the questions fixed and cannot hold the calendar fixed. Google AI Mode, which the sources-per-answer study covers, and Google AI Overviews, which the Index also collects, went quiet over the same summer on their own schedules. Claude's and Perplexity's mean citations per answer did not move on August 14.
- The consumer product. The Gemini app may search more or less than this API route.
- Domain ranks at small counts. Domains with fewer than 8 citing answers in period A are not in the rank comparison, and a count of 3 or 6 is a pointer, not an estimate.
- Buyer behaviour. Nothing here says what a buyer sees, or that a citation in Gemini changes a purchase.
Method note #
Source: the Index answer ledger's run table, daily export of 2026-11-09 08:10 UTC, read with DuckDB, joined to the citation-events table for cited domains and to the question table for eligibility, shape and industry. Filters: engine gemini, successful runs, questions marked eligible for public scoring, a recorded model string, run dates 2026-06-02 to 2026-11-09. Periods and models as defined above. The 378 paired questions are those with at least one run in each period. A domain counts once per answer. Intervals and rank statistics were computed by resampling the 378 questions (2,000 draws, seed 7). Dropping the 19 days from August 14 onward that the Index's collection-health rule flags as silent (ten on 3.7, nine on 3.8) moves the share of answers with no source from 94.7 to 92.4 percent on 3.7 and from 88.6 to 88.3 percent on 3.8.
Frequently asked questions #
Why did Gemini's citations drop? #
In the Machine Relations Index, Gemini's API route stopped citing any source in most answers. On the same 378 buying questions, 85.5 percent of answers cited something with gemini-2.5-flash and gemini-3.5-flash and 14.1 percent did with gemini-3.7-flash and gemini-3.8-flash. When it did cite, the count per answer barely moved. The ledger cannot say whether the model, the route or Google's search tool is the cause.
Did Forbes or Reddit lose authority with Gemini? #
There is no sign of it in this ledger. Forbes was cited in 7.4 percent of Gemini answers in June and July and 1.9 percent from August, but its share of the answers that cited anything rose from 8.6 to 13.2 percent. The rate fell because far fewer answers cited anything.
Can I compare a brand's Gemini citation rate across months? #
Only inside one model period. Across a model change the rate moves for every domain at once. Report the share of answers that cited anything alongside the brand's share among them.
Which Gemini questions still cite sources? #
On gemini-3.8-flash, 30.2 percent of top-list questions and 23.0 percent of best-X questions cited at least one source in this ledger, against 3.2 percent of how-to-choose questions and 0.4 percent of problem descriptions. Among industries, AI visibility and GEO questions cited in 41.0 percent of runs.