How we measure the AI shelf.

§1The shelf model

We track 185 supplement shelves: product categories like magnesium glycinate, prenatal multivitamin, lion's mane or collagen. Each shelf is sampled with a battery of shopping prompts, the questions a real buyer types, repeated hundreds of times on every AI assistant. A few shelves are a shopping goal rather than a single ingredient, such as immune support; we keep them because shoppers ask for them by that name.

Every question is hand-curated per shelf: a human decides what a real buyer would ask, in buyer language. That is only feasible because AI Shelf Space covers one industry; a tool serving every industry cannot hand-curate every question.

§2Sampling: why we ask the same question hundreds of times

AI answers are probabilistic. In an independent study of roughly 3,000 prompt runs, fewer than 1% of identical prompts returned identical brand lists, and it took on the order of 60 to 100 runs for the brand set to stabilize. A tool that asks once is reporting a coin flip.

AI Shelf Space samples every shelf hundreds of times on every AI assistant, every edition, and the number of answers behind every figure is shown next to it. An answer that names no brand counts as recommending no brand. Visibility is computed on the answers that named at least one brand, and that share is published per assistant below, so the denominator is never hidden.

Why your own test will look different. One chat is one draw. Two independent batches of the same question, run on the same day, moved a brand's share by 16 points on average (8 points on ChatGPT, 22 on Gemini). A dashboard built on hundreds of answers shows the weight of the coin, not one toss, which is also why a rank is the more robust figure: it cannot be contradicted by a single chat.

§3The seven AI assistants

We currently sample ChatGPT, Google Gemini, Google AI Overviews, Google AI Mode, Microsoft Copilot, Perplexity and Claude. Consumer apps are sampled as shoppers see them, not through developer APIs. Claude is the exception: no vendor exposes its consumer UI at scale, so it is sampled via Anthropic's API with web search enabled, and flagged accordingly. Model versions are re-verified every edition and pinned in the change log (§8).

Share of answers naming at least one brand, per assistant, this edition:

§4Reading the answers

Each raw answer is parsed by an extraction model into brands, products and list positions. Brand names are then canonicalized: "GOL," "Garden of Life" and "gardenoflife.com" count as one brand, matched against an alias map covering 9,857 brand names this edition. Every data point carries the brand-map version that produced it, so results are reproducible.

§5The metrics

• Visibility. Of the brand-carrying answers on a shelf, the percentage that named your brand. Always shown with the assistant's brand-carrying rate.
• Rank. Your position on the shelf when you appear, and your standing ordered by visibility against every other brand on that shelf.
• Movement. Change between measurements. An alert fires when your brand drops out of a shelf's top ten, when a rank change holds across two consecutive weekly measurements, or when a new competitor enters your shelf. A percentage alone never triggers an alert, because a single week's percentage moves with sampling.

§6Cadence

The public Index publishes a new edition every quarter, each built on a full measurement of every shelf on every assistant. Between editions, customers' shelves are re-sampled weekly for the alerts. Every edition states the week its data was collected, and a published edition is never silently rewritten.

§7What we never do

• No pay-for-placement, ever. The Index is not influenced by who subscribes.
• No vendor-supplied numbers. Everything we publish comes from our own sampling.

§8Change log

• 2026-09-09 First public edition, Autumn 2026, published 2026-09-09. 185 shelves × 7 AI assistants, 178,786 answers, 1,366,316 brand recommendations.

See the methodology in action: the Index