Six AI Search Engines, One Question, Almost No Shared Sources
Key findings
When we sent the same query to several AI search engines, they named similar brands but cited almost entirely different websites. Across 4,206 queries, the average domain overlap between any two engine families was 8.5%, while the average brand overlap was 24.2%.
Exact pages barely overlap. Mean URL-level Jaccard overlap was 3.6% (median 0.0%) across 38,320 engine-pair comparisons.
Domains overlap slightly more. Mean domain overlap was 8.5% (median 6.7%).
Brands overlap about 2.8x more than domains. Mean brand-set overlap was 24.2% (median 23.1%) across 39,810 comparisons.
Most sources are single-engine. On average, 83.7% of the domains cited for a query appeared in only one engine family. Only 0.7% appeared in every family that answered that query.
88.1% of queries had no domain cited by all engines.
The practical read: engines broadly agree on who the relevant players are, but they disagree on which pages to trust as proof. Earning a citation in one engine tells you little about the others.
How we measured it
We pulled every successful experiment execution created between 2026-10-19 and 2026-10-25 from the XLR8 AI platform, across all organizations, and kept only the latest execution per experiment. That left 120 executions from 32 organizations and 22,270 query-by-model results.
Models were grouped into seven engine families: Claude, GPT (including ChatGPT), Gemini, Google AI Mode, Grok, Perplexity (Sonar) and Copilot. For each query within an execution, we kept one result per family and kept only queries answered by at least two families. That gave 4,206 qualifying queries, averaging 4.9 families and 30.7 distinct cited domains each.
For each pair of families on the same query we computed Jaccard overlap, meaning shared items divided by the union, on three sets: cited URLs (normalized with the platform's URL cleaner), cited domains (lowercased, www. stripped), and brands extracted from the response. Pairs where either side had an empty set were skipped. All figures are our own computations from the citations and brands columns of visibility_experiment_results.
Results
The three layers of agreement fall away sharply as you move from brands to pages.
Layer compared | Mean overlap | Median | Pair comparisons |
|---|---|---|---|
Brands mentioned | 24.2% | 23.1% | 39,810 |
Cited domains | 8.5% | 6.7% | 38,320 |
Cited URLs | 3.6% | 0.0% | 38,320 |
Agreement also varies by pairing. Domain overlap was highest between Claude and Perplexity (16.6%, n=2,432 queries) and between Claude and Gemini (12.0%, n=2,873). It was lowest between Copilot and GPT (3.5%, n=187), GPT and Grok (4.4%, n=1,629) and Google AI Mode and Grok (4.5%, n=1,491). Brand overlap followed a similar order: Claude and Perplexity led at 29.0%, GPT and Perplexity followed at 28.1%, and Copilot with Google AI Mode was lowest at 18.2% (n=166).
In plain terms, even the most similar pair shared roughly one cited domain in six. Within a single query, the typical cited domain was unique to one family (83.7% on average). Domains cited by a majority of the answering families made up just 8.2% of a query's domains on average, and domains cited by all of them just 0.7%.
Citation volume differs a lot by engine, which matters for reading these numbers. In the same window, Grok averaged 39.4 citations per response, Perplexity 10.0, Claude Sonnet 5 6.7 and Google AI Mode about 3.3 (all from the 7-day set of 27,673 results, before our latest-execution filter). Engines that cite few sources have fewer chances to overlap, so part of the low overlap reflects list size rather than disagreement.
What this means for GEO
Brand visibility and source visibility are different problems. Being named in answers is the part engines agree on most (24.2% brand overlap), so strong category positioning travels across engines. Being cited is not portable: with 83.7% of cited domains appearing in a single engine family, a page that wins in one engine is unlikely to be the page another engine reaches for.
Two implications follow from the data. First, measure citation share per engine rather than blending engines into one score, because a blended number hides which engines never cite you. Second, spread proof across several kinds of pages and domains (your own site, third-party reviews, comparisons, community threads) instead of optimizing a single flagship page.
Limitations
One week of data (2026-10-19 to 2026-10-25), 120 executions, 32 organizations. Results may shift with other time windows.
Query mix is not random. Experiments reflect what customers chose to track, which skews toward B2B software and commercial-intent queries.
Jaccard overlap is sensitive to list size. Engines that return few citations (Google AI Mode, Copilot) have lower ceilings for overlap than heavy citers (Grok, Perplexity). We did not normalize for this.
Engine families merge several models (for example GPT includes ChatGPT and GPT-5.x variants). Where a family had more than one result for a query, we kept one, so the choice of model within a family is arbitrary.
Domain matching is exact after lowercasing and stripping
www.; subdomains count as different domains, which slightly understates overlap.Brand overlap depends on the platform's brand extraction from response text, and naming variants may be counted as different brands.
Several queries recur across organizations in the same category, so observations are not fully independent.
