ObsurfableSurfacedBy's earlier finding — only 2.7% of cited domains appearing across all five major engines — already killed the "one GEO score" idea. Foglift's Q3 2026 AI Search Citation Benchmark makes the same point with a frozen buyer-prompt panel and a quarter-over-quarter comparison.
Between 1–10 August 2026, Foglift sent the same 75 brand-neutral buyer questions (25 verticals, three intents each) to ChatGPT, Claude, Gemini, Google AI Overviews, and Perplexity with live web retrieval. That is 375 answers. Cited URLs were normalized to domains and counted once per response.
The headline: of 92 domains in the five engines' combined top-25 lists, 69 (75%) appeared in only one engine's top 25. Zero domains appeared in all five. Mean pairwise Jaccard overlap on cited domains was 0.094 — down from 0.180 in Q2 on the same prompt set.
A citation win on Gemini is not a citation win on Claude. The Q3 panel is the cleanest recent proof.
What the panel measured
| Item | Q3 2026 |
|---|---|
| Prompts | 75 (frozen from Q2) |
| Engines | ChatGPT, Claude, Gemini, Google AI Overview, Perplexity |
| Responses | 375 |
| Distinct cited domains | 1,510 (vs 1,119 in Q2) |
| Mean pairwise domain Jaccard | 0.094 (vs 0.180 in Q2) |
| Top-25 union | 92 domains |
| Exclusive to one engine's top 25 | 69 (75%) |
| In all five top 25s | 0 |
Foglift is explicit that models and indexes changed between quarters, so Q2→Q3 movement is observational, not a controlled causal study. That is the right caveat. It does not weaken the operational conclusion: an authority map from May did not describe August.
Overlap collapsed while the source pool expanded
Distinct cited domains rose 34.9% (1,119 → 1,510). Mean overlap fell 47.8%. Engines are citing a wider web and agreeing less about which domains belong in the lead pack.
How many engines share a leading source:
| Engines containing the domain | Domains | Share of the 92-domain union |
|---|---|---|
| 1 | 69 | 75% |
| 2 | 15 | 16.3% |
| 3 | 6 | 6.5% |
| 4 | 2 | 2.2% |
| 5 | 0 | 0% |
Same prompts, five different leaderboards
Every engine answered the same 75 questions. Their top cited domains did not.
| Engine | #1 in this panel | Share of that engine's 75 answers |
|---|---|---|
| ChatGPT | HubSpot / Consumer Reports / Tom's Guide (tie at 5.3%) | Vendor docs and consumer press, not Reddit |
| Claude | Forbes | 14.7% |
| Gemini | YouTube | 49.3% |
| Google AI Overviews | YouTube | 44% |
| Perplexity | YouTube | 69.3% |
Perplexity cited Reddit on 58.7% of its 75 answers in this panel. ChatGPT's top 10 in the same window is first-party help centers and vendor sites. That is CiteLens's "independent vs Google-aligned" split showing up as domain lists, not just organic-rank overlap.
Overall (all 375 answers, domain counted once per response):
| Rank | Domain | Share |
|---|---|---|
| 1 | YouTube | 32.8% (13.9% in Q2, +18.9 pp) |
| 2 | 21.3% (3.2% in Q2, +18.1 pp) | |
| 3 | Forbes | 11.7% |
| 4 | Healthline | 5.3% |
| 5 | G2 | 4.3% (24th in Q2) |
YouTube stayed #1. Reddit jumped from 11th to 2nd on this frozen panel. G2 moved from 24th to 5th. Vendor first-party sites occupied five overall top-25 slots, up from two in Q2.
Read this carefully against the ChatGPT-specific Reddit cliff later in August: panel-level Reddit share can rise on Gemini/Perplexity/AIO while ChatGPT Search simultaneously dumps Reddit. Blended "Reddit is winning AI search" charts hide that.
Engines cite pages, not homepages
Only 1.5% of 3,412 cited URLs were homepages. 74.7% were at least two path segments deep. Domain-level mention dashboards will tell you "YouTube won" and hide whether the cited object was a 12-minute product walkthrough, a Short, or a channel home.
If you only track brand.com, you will miss the comparison page, help article, or G2 profile that actually entered the answer.
How this sits next to earlier overlap research
Only 2.7% of domains are cited by all five engines used a much larger citation crawl. Foglift uses a small, frozen buyer panel. Different methods, same direction: cross-engine source agreement is the exception.
CiteLens explained why Google AI Mode and ChatGPT cannot share a playbook (93% vs 30% of citations from Google's top 10). Foglift shows the inventory consequence: even the leaderboard of "who gets cited" is engine-private.
A separate August study from Kojable, on four provider stacks and ten B2B questions, found pairwise URL Jaccard around 0.009–0.020 and citation density from 5.28 events per 1,000 words (OpenAI) to 39.79 (Gemini). Raw citation counts are not interchangeable across products. More footnotes is not the same as more agreement.
What to do with this
- Stop using one engine as a proxy. ChatGPT-only retainers will systematically miss YouTube-heavy Google/Perplexity paths and Claude's Forbes-shaped stack.
- Track cited URLs, not just domains. Homepages are almost never the object.
- Re-baseline quarterly on a frozen prompt set. Foglift's Q2 map was already stale by early August.
- Interpret Reddit and YouTube per engine. A blended rank will lie during weeks when one engine's policy changes.
How Obsurfable fits
Obsurfable runs the same prompts across engines and keeps mention, citation, and competitor columns separate. That is the only reporting shape this benchmark supports.
Bottom line
In Q3's frozen buyer panel, no domain made every engine's top 25, 75% of leading domains were exclusive to one engine, and overlap halved versus Q2 while the cited-domain pool grew. Multi-engine AEO is not a slogan. It is what the citation graph requires.