The study covers two exact prompts, each run once in Gist’s native Deep mode across ChatGPT, Perplexity and Google AI Overviews. All six runs completed and converged. They produced 128 valid stored answers between 2:45 and 2:55 p.m. Pacific on October 6, 2026. Convergence describes the native stopping rule; it does not establish independent observations or semantic scoring accuracy.
best accounting software for small businesswhat's the best facial cleanser for sensitive skin
| Question · engine | Answers | Distinct texts | Named-family distribution | Cited hosts |
|---|---|---|---|---|
| accounting · Google AI Overviews | 12 | 2 | 4 families: 10 answers, 5 families: 2 answers | 9 |
| accounting · ChatGPT | 12 | 12 | 4 families: 5 answers, 5 families: 7 answers | 11 |
| accounting · Perplexity | 12 | 12 | 5 families: 12 answers | 9 |
| skincare · Google AI Overviews | 32 | 4 | 3 families: 29 answers, 4 families: 3 answers | 10 |
| skincare · ChatGPT | 32 | 32 | 2 families: 28 answers, 3 families: 3 answers, 4 families: 1 answers | 5 |
| skincare · Perplexity | 28 | 28 | 2 families: 1 answers, 3 families: 4 answers, 4 families: 23 answers | 8 |
Presence, preference and sources are separate
A named family is a literal mention of an eligible product brand or family in the stored answer. We exclude publishers, authorities and generic corporate parents. QuickBooks and QuickBooks Online count once as the QuickBooks family; Wave and Wave Accounting count once as the Wave family. Multiple cleanser products under one brand count once. The first named family is its earliest literal appearance, not an inferred ranking.
All 128 answer-level family sets matched between the native presence fields and a separate text/alias check. This checks the mapped families, not whether discovery found every possible entity. The native recommendation scorer has unresolved discrepancies, so these two articles use validated name counts and citation-host lists, not a recommendation league table.
What a citation list can tell us
Hosts are lowercased and an initial www. is removed; other subdomains remain distinct. A host is not necessarily a separate publisher. Raw citation lists can include shopping or platform links. Each Perplexity run attached the same ten-URL citation set to all of its answers. Repetition does not establish independent source discoveries or diagnose a collection fault.
Stored answer text is normalized by the collection adapters. Original provider bodies and inline citation anchors were not part of this export, so the articles do not assign particular links to particular sentences or infer source influence. The public observation table preserves the resulting names, first-name label, host sets and response-text hashes without internal account or run identifiers.
Limits of the pilot
Google’s 12 accounting observations contained two distinct texts; its 32 cleanser observations contained four. Other runs had a distinct text hash per answer, which still does not establish independence. These are one-batch results for two prompts, not category-wide rankings, consumer demand, product-quality assessments or trends. No Claude measurements are included. The prompts did not specify a country; the results are not presented as a controlled US-market sample.
Counting each distinct Google text once changes the mean named-family count from 4.17 to 4.50 for accounting and from 3.09 to 3.50 for skincare. This sensitivity check is not a replacement population estimate.
Read and reuse the evidence
Download the 128-row observation table (CSV) · Download the aggregate counts (JSON)
The two stories share this dataset. They are not independent studies. Gist sells AI-visibility services. The articles report measured output patterns and disclose their limits.