The finding
Different engines produce different shortlists. The rebuilt archive contains 2,356 readable runs across 289 panel/question groups. Not every group has every engine, and not every group qualifies for a comparison.
For each question and engine, we combine the business names extracted from its archived runs. We then divide the names shared by two engine sets by the number of distinct names in their combined set. That is Jaccard overlap. If one set contains A and B, and the other B and C, the overlap is one third: one shared name among three distinct names. It is not the one-half chance that a name in the first set also occurs in the second.
The mean of the 24 eligible pair means is 12.4%. Weighting those means by the number of question groups gives 11.9%. Both describe this archive. Neither estimates your business’s odds.
| Engine pair | Mean Jaccard overlap | Question groups |
|---|---|---|
| Google AI Mode (app) / Google AI Overviews (app) | 21.4% | 36 |
| Google AI Mode (app) / ChatGPT | 11.1% | 42 |
| Google AI Mode (app) / Claude | 14.1% | 45 |
| Google AI Mode (app) / Gemini | 17.0% | 72 |
| Google AI Mode (app) / Perplexity | 15.0% | 58 |
| Google AI Mode (API) / Google AI Overviews (API) | 18.9% | 52 |
| Google AI Mode (API) / ChatGPT | 6.7% | 54 |
| Google AI Mode (API) / Claude | 10.1% | 68 |
| Google AI Mode (API) / Gemini | 12.4% | 79 |
| Google AI Mode (API) / Perplexity | 11.5% | 74 |
| Google AI Overviews (app) / ChatGPT | 10.9% | 21 |
| Google AI Overviews (app) / Claude | 13.5% | 26 |
| Google AI Overviews (app) / Gemini | 15.2% | 34 |
| Google AI Overviews (app) / Perplexity | 13.1% | 30 |
| Google AI Overviews (API) / ChatGPT | 10.1% | 30 |
| Google AI Overviews (API) / Claude | 11.4% | 47 |
| Google AI Overviews (API) / Gemini | 10.7% | 52 |
| Google AI Overviews (API) / Perplexity | 9.5% | 47 |
| ChatGPT / Claude | 17.0% | 81 |
| ChatGPT / Gemini | 8.5% | 100 |
| ChatGPT / Perplexity | 7.8% | 95 |
| Claude / Gemini | 9.7% | 113 |
| Claude / Perplexity | 9.5% | 106 |
| Gemini / Perplexity | 11.3% | 145 |
The full range is 6.7% to 21.4%. API rows are labeled as API rows; they do not stand in for the consumer app. The source archive contains six engine families and eight surface labels.
What a buyer can use
A report about one engine is evidence about that engine. Measure the other surfaces before deciding you are visible there. Keep the question wording and collection conditions consistent, and show each engine separately. An overall number can hide a gain on one surface and a loss on another.
Low set overlap does not prove that a blended metric is always useless, or that one engine provides no information about another. It does show why a single score needs a definition, its underlying engine results, and a clear account of missing observations.
Repeat answers change too
The archive also compares pairs of repeat runs within a question and engine. The measure below is again Jaccard set overlap. It is not a first-position hold rate, and its complement is not the fraction of businesses that disappeared.
| Surface | Mean repeat Jaccard | Run pairs |
|---|---|---|
| Google AI Mode (app) | 38.6% | 139 |
| Google AI Mode (API) | 44.9% | 232 |
| Google AI Overviews (app) | 30.0% | 51 |
| Google AI Overviews (API) | 37.7% | 51 |
| ChatGPT | 33.5% | 60 |
| Claude | 25.5% | 86 |
| Gemini | 23.2% | 400 |
| Perplexity | 48.6% | 307 |
These pairs share questions and runs, so they are not independent trials. We do not attach an independence-based confidence interval to them.
Recompute it
Download the anonymous counts, computed summary, and Python recomputation script. The data notes describe the filters and units. No business names, client identities, locations, prompts, answer text, or exact run dates are in the export.
The extractor and name normalization can merge distinct businesses or split one business into variants. A nine-firm recall check in one market does not establish extraction accuracy across the archive. We exclude empty name sets, require a combined set of at least three names, and show pair means only where at least 12 question groups qualify. Runs are pooled within each question/engine; they are not necessarily matched by date. These choices limit what the figures can support.
Correction — November 7, 2026
The earlier page called Jaccard overlap a probability, listed six selected pairs as though their 6.7–11.3% range covered the full table, and labeled two API results as consumer-app results. Those descriptions were wrong. The current rebuild yields 2,356 readable runs and 289 groups, replacing the unsupported 1,984-run/139-group headline. The pair means reproduce the archived aggregate; the run/group headline now comes directly from the export. This is a correction and reproducibility release, not evidence that engine behavior changed since August.