THE AI RECOMMENDATION INDEX

This window the Index measured 311 software categories, 271 of them published with enough answers to rank. 438 sampled answers from 5 assistants (ChatGPT, Claude, Gemini, Perplexity and Grok), Sep 15, 2026 to Oct 13, 2026, each one a buyer-shaped question with no product names in it. Every product an answer recommended was extracted, resolved to one canonical entity and counted.

The assistants mostly disagree with each other. Across 27 categories where at least 3 of them answered, a single product led every engine in 0% of them (n=27, 28-day window). The pages those answers cite are largely the industry writing about itself. Of 5,812 citations, 54.3% pointed at a vendor competing in the category being asked about (28-day window).

A ranking reads like this. Share is the percentage of a category's sampled answers that recommended the product; the per-engine figures split that by assistant. A product named in passing, or named as a warning, does not add to its share. Every category page carries its own n and window, and each category is resampled roughly monthly on a rolling 4-week window, so a share is a claim about that window and nothing older. Prompt design, sample gates and entity resolution are in the methodology.