A June Audit Finds Almost No Shared URLs Across Four AI Engines
On 6 June 2026, 15 prompts produced a mean pairwise URL Jaccard of 0.0079 across ChatGPT, Copilot, Google, and Perplexity.
Direct answer
The September 2026 paper says a 6 June audit of 15 commercial prompts produced 589 citations, and the mean pairwise exact-URL Jaccard across the four engines was 0.0079.
A September 2026 arXiv paper by Benjamin Tannenbaum, numbered 2609.22655, measures whether four live AI search products cite the same pages when the prompt is held fixed. The abstract reports a 6 June 2026 audit of 15 commercial prompts. That day produced 589 citation observations across ChatGPT, Microsoft Copilot, Google, and Perplexity, which the paper maps to 528 unique exact URLs and 356 domains.
Same-prompt overlap is the result. Mean pairwise Jaccard similarity on exact URLs is 0.0079, with a prompt-clustered bootstrap 95 percent confidence interval of 0.0037 to 0.0132. The median is zero. The abstract says 84.9 percent of engine pairs shared no cited URL. On the ten prompts observed on all four engines, mean exact-URL Jaccard is 0.0072, reciprocal-rank-weighted overlap is 0.0027, and top-five exact-URL overlap is zero in all 60 pairwise comparisons.
One engine does not stand in for the set
A single engine captured 11.4 to 42.6 percent of the four-engine URL union. The union held 2.43 times as many URLs as the broadest single engine, on average. Among URLs seen on 6 June, 96.4 percent appeared in only one engine. The author says this does not throw out engine-free page scores. It limits what those scores estimate: page fit, not a stable chance of being cited by a named engine.
Source: arXiv 2609.22655.
FAQ
- How many URLs were unique to one engine?
- The abstract says 96.4 percent of the URLs observed on 6 June appeared in only one engine.
- Did the top five results overlap?
- On the ten prompts seen on all four engines, top-five exact-URL overlap was zero in all 60 pairwise comparisons.