GEOforge Research · Citation Data
Ask Google’s AI Overviews and AI Mode the same question and they back their answers with largely different websites. About 21% of cited sites are shared.
Google runs two AI answer surfaces, AI Overviews on top of the results page and AI Mode in its own tab, and most tracking treats them as one “Google AI”. For the same prompt, we found they cite the same websites only about 21% of the time, well below the 56% and 34% that repeat runs of each surface share with themselves. Being cited in one is not a reliable sign of being cited in the other.
GEOforge runs each brand’s tracked prompts against Google AI Mode and AI Overviews on a schedule and stores every source each answer cites. For each prompt we compared the websites (domains, not individual pages) cited in answers collected on different days. Google’s own pages are excluded. We scored each pair with the Jaccard overlap: websites cited by both answers, divided by websites cited by either.
Three comparisons per prompt and week: an AI Mode answer against another AI Mode answer; an AI Overviews answer against another AI Overviews answer; and an AI Mode answer against an AI Overviews answer collected in the same run. The third is the one that matters, and the first two tell us how much overlap to expect from a surface that agrees with itself.
Quite. In no week did the surfaces share more than 27% of their cited websites, and in the four most recent weeks the figure was about 18%. By comparison, AI Mode’s repeat runs shared between 48% and 67% of theirs.
Two surfaces, one prompt, different sources
Average share of cited websites two answers have in common, by week. Hover a point for the number of brands and prompts.
| Week of | Brands | Prompts | AI Mode vs AI Overviews | AI Mode vs itself | AI Overviews vs itself |
|---|---|---|---|---|---|
| 14 Sep | 10 | 239 | 24% | 56% | 46% |
| 21 Sep | 13 | 314 | 27% | 53% | 47% |
| 28 Sep | 13 | 304 | 19% | 48% | 32% |
| 5 Oct | 14 | 423 | 18% | 60% | 27% |
| 12 Oct | 9 | 383 | 17% | 52% | 24% |
| 19 Oct | 9 | 215 | 19% | 67% | 31% |
Average of per-brand figures, so no single brand dominates. Brands with at least 15 prompts measured in the week.
No, and the self-overlap lines are the control for that. If answers were simply random draws from the same pool, two runs of one surface would overlap no more than an AI Mode answer and an AI Overviews answer. Instead, AI Mode repeats itself much more (56% on average). AI Overviews did too until late September (46%), then fell to about 28%, closer to the cross-surface figure of 18%, but still above it. We did not find a change in our own collection code in that window, and we cannot say whether the shift came from Google or our data provider.
No. Each week we repeated the comparison brand by brand for brands with at least 15 prompts measured that week, between 9 and 14 brands per week. In 64 of 68 brand-weeks the cross-surface overlap was lower than both self-overlaps, and the weekly figure barely moves if we drop the most concentrated week (21% against 21%) or the most recent one (21%). These brands track different topics, so this is not one client’s prompt set.
Data. Six full weeks, 14 September to 25 October 2026. For each prompt that had at least two runs on both engines in the week, we compared the websites cited in the second and third runs of each engine, and the second runs of both engines against each other. That is 1,878 prompt-weeks from brands with at least 15 prompts that week, between 9 and 14 brands a week. Prompts where either answer had no cited websites were skipped.
Normalization. The overlap is a ratio, so it does not depend on how many answers we collected. Headline figures are averages of per-brand figures, and the pooled prompt-weighted figure is within two points of them.
Checks. Week-by-week split; brand-by-brand agreement; unweighted versus pooled averages; dropping the most concentrated week and the latest week; confirming that sources per answer stayed between about 6 and 9 for both engines. We excluded the week of 26 October because Google AI Overviews collection changed that week. We did not use July and August: the earlier data holds too few brands with repeat runs, and Google also halved AI Mode’s source lists in late August.
Limitations. Prompts are the ones our clients track, so they lean B2B and are not a random sample of searches. We compared websites, not pages, so two different pages on one site count as a match. AI Overviews’ own consistency changed during the window for reasons we cannot see. Every figure is aggregated; no brand or prompt is identified.
GEOforge tracks every source your brand and your competitors earn across ChatGPT, Google AI Overviews and Google AI Mode, week by week. 14-day free trial, no sales call.
Start a free trial →Sources & method. GEOforge production citation database, 14 September to 25 October 2026, 1,878 prompt-weeks across brands with at least 15 prompts a week. See also our analysis of how Google and ChatGPT cite “best” and “vs” pages.
