Obsurfable“If we win GEO, we win AI search” assumes one retrieval graph. Christian Strunk’s LLM SEO survey, October 2026 stress-tests that assumption with 1,032 answers, 4,141 cited sources, and 58 companies across the US and Germany.
Collected October 6, 2026 via real apps (DataForSEO), the survey shows engine-specific citation budgets, tiny domain overlap, and a persistent split between being named and having your site cited — especially outside ChatGPT.
Method snapshot
| Item | Detail |
|---|---|
| Companies | 58 (38 US English, 20 Germany German) |
| Industries | B2B software, agencies, manufacturers, online shops, local services, consulting |
| Questions | 6 per company (problem / solution / provider stages), generated from each site |
| Engines | ChatGPT (web search), Gemini, Google AI Mode |
| Repeat run | Same questions 3 hours later for 10 companies (173 comparable pairs) |
| License | CC BY 4.0 — CSV download |
Strunk warns: one-day snapshot, non-random company selection, tool-generated questions, small cells flagged below n=30. Company names are not published.
Finding 1: Citation density differs sharply by engine
| Engine | Answers | Answers with sources | Mean sources / answer | Median |
|---|---|---|---|---|
| All | 1,032 | 90% | 4.0 | 4.0 |
| ChatGPT | 348 | 99% | 6.0 | 5.0 |
| Gemini | 348 | 82% | 2.9 | 3.0 |
| Google AI Mode | 336 | 89% | 3.1 | 3.0 |
ChatGPT is citation-hungry: nearly every answer carries links, with 6 sources on average. Gemini omits sources on 18% of answers and cites half as many domains per response.
Strategies that assume “more citations = more chances to appear” favor ChatGPT-shaped surfaces — not Gemini-minimal answer cards. That rhymes with ChatGPT vs Perplexity separate AEO strategies and Foglift’s low cross-engine citation overlap — AI Engines Do Not Share Citation Lists.
Finding 2: Source-type mix is engine-specific
Share of 4,141 citations by rule-based type:
| Source type | All engines | ChatGPT | Gemini | Google AI Mode |
|---|---|---|---|---|
| Other companies’ websites | 44.7% | 57.6% | 31.3% | 32.0% |
| Lists & comparisons | 18.0% | 9.5% | 33.1% | 20.4% |
| Own website | 11.1% | 14.9% | 5.4% | 8.9% |
| Google properties | 6.2% | 1.2% | 10.9% | 11.5% |
ChatGPT disproportionately cites competitor and peer sites (57.6%) and leads on own-site share (14.9%).
Gemini leans on listicles and comparison articles (33.1%) and Google.com surfaces (10.9%).
Google AI Mode adds forums/community (3.9%) and video (3.9%) — a different third-party mix than ChatGPT’s near-zero video share in this sample.
Tactical read: winning Gemini may require comparison content placement (yours or partners’); winning ChatGPT may require being the cited vendor page in a dense source list — aligned with Onely’s October ChatGPT vendor vs listicle split.
Finding 3: Named in the answer ≠ your URL in the sources
| Engine | Answers naming company | Own site cited | Named without own site (of named) | Not named although cited (of cited) |
|---|---|---|---|---|
| All | 52% | 24% | 56% | 4% |
| ChatGPT | 54% | 43% | 22% | 2% |
| Gemini | 52% | 13% | 78% | 11% |
| Google AI Mode | 51% | 17% | 68% | 4% |
Overall, more than half of named answers (56%) do not cite the company’s own domain. Gemini is extreme: 78% of named answers lack an own-site source.
That is the survey-scale version of Sentic’s mention/citation races: brand awareness in prose decouples from owned grounding.
Cited but not named is smaller (4% overall, 11% on Gemini) but painful — your page supplies evidence while another brand gets the recommendation.
Finding 4: Cross-engine domain overlap is tiny
For 336 questions answered on both ChatGPT and Gemini:
| Comparison | Mean Jaccard overlap | Questions with no shared domain |
|---|---|---|
| ChatGPT ∩ Gemini | 6.5% | 73% |
| ChatGPT ∩ AI Mode | 8.1% | 67% |
| Gemini ∩ AI Mode | 9.8% | 64% |
| All three cite same domain (mean per question) | 0.14 domains | 87% share none across all three |
Only 12.7% of domains were cited by at least two engines in the three-way comparison.
You do not have one “AI SERP.” You have three largely disjoint evidence bundles for the same buyer questions.
Finding 5: Referring domains predict naming — but not uniformly
Share of answers naming the company by referring-domain tier (US + Germany combined):
| Referring-domain tier | Named rate (all engines) | ChatGPT | Gemini | Google AI Mode |
|---|---|---|---|---|
| Top third | 72% (n=340) | 78% | 68% | 71% |
| Middle third | 49% (n=355) | 45% | 50% | 53% |
| Bottom third | 35% (n=337) | 39% | 39% | 28% |
Top-tier link graphs correlate with being named — directionally consistent with domain-heavy citation models — but bottom-third sites still get named 35% of the time on average. Authority helps; it is not the only path.
Google AI Mode punishes the bottom third hardest (28% named vs 39% on ChatGPT).
Finding 6: Volatility inside a single afternoon
Repeat runs 3 hours apart (173 pairs):
| Metric | All engines |
|---|---|
| Naming same brand both runs | 82% |
| Naming switched | 18% |
| Mean shared-source Jaccard | 29.6% |
| Pairs with no shared domain | 30% |
Naming is stickier than citations. Source lists churn ~30% to 42% (Gemini highest) even within hours — reinforcing AI Citation Volatility.
Industry snapshots worth noting
| Industry | Named rate | Own-site citation share | Lists & comparisons |
|---|---|---|---|
| B2B software | 81% | 16.9% | 31.1% |
| Local services | 21% | 3.9% | 12.6% |
| Agencies | 36% | 5.7% | 22.9% |
| Manufacturers | 62% | 21.8% | 7.2% |
Local services show how generic buyer questions under-name providers even when citations exist — echoing Howseen’s SMB cold-start — 85.5% never named on first scan.
Playbook: three-engine reality
- Build three source strategies, not one checklist: listicle/comparison presence for Gemini, dense vendor citations for ChatGPT, Google property + forum/video tolerance for AI Mode.
- Track named rate and own-site citation rate separately — especially on Gemini where 78% of names lack your URL.
- Use fixed prompts and repeat runs; single screenshots lie within 3 hours.
- Invest in referring-domain diversity when naming is the goal; invest in answer pages when citations lag — see Page Optimization Only Helps High-Authority Domains for sequencing.
How Obsurfable fits
Strunk’s survey is reproducible science with transparent tables. Obsurfable applies the same discipline to your category: public observations in Explorer, matched metrics in the App — mention rate, citation presence, and competitors per engine.
FAQ
Can I cite the data in my own posts?
Yes — CC BY 4.0 with attribution: “Christian Strunk, LLM SEO survey, October 2026, christianstrunk.com.”
Why is google.com #1 in cited domains?
Gemini and AI Mode cite Google properties heavily (10.9% and 11.5% of their citations); ChatGPT still contributed 25 google.com citations in the top-domain table.
Does 52% named rate mean GEO is easy?
That is across selected companies with tool-generated questions — not random SMBs. Local services named only 21% of the time.
How does overlap compare to Pineprompt?
Pineprompt measures brand overlap across seven engines; Strunk measures domain overlap across three. Both show fragmentation.
Bottom line
Strunk’s October 2026 survey shows ChatGPT cites more, Gemini favors listicles and Google surfaces, and 73% of ChatGPT–Gemini pairs share no domains. More than half of brand names appear without an own-site citation — 78% on Gemini. Treat AI search as parallel retrieval systems, measure each, and sequence authority vs page work to the gap you actually have.