Research & field notes
How ShelfSignal measures visibility and shopping choices
Definitions, sample designs and limits for AI mentions, simulated seller choices and storefront-readiness inspections.
Vladimir Fedorov · Published 26 October 2026
Three measurements, three different questions
| Study | Published design | What it can say |
|---|---|---|
| AI Shelf Index | 20 questions ×5 assistants ×5 repeats =500 answers | Which brands are named, in what order and with which sources. |
| Order Shelf | 6 purchase briefs ×5 assistants ×2 runs =60 decisions | Which brand and seller an assistant selects in a simulation. |
| Agent-Ready pilot | Six observed storefront layers,0–3 points each; maximum 18 | A dated inspection score, not an observed purchase or a predictive sales score. |
Sample design and repeatability
The visibility index uses non-branded buyer questions. The published magnesium study lists the 20 questions and the five platforms: ChatGPT, Claude, Gemini, Perplexity and Grok. Runs were collected through official model APIs with web grounding. The published method describes retaining model version, prompt, run index and timestamp.
For a new comparison, freeze the market, language, prompts, assistants, available model versions, web-search mode, repetitions and brand-matching rules. A later model release or a different search source may change the result even when the storefront has not changed. Preserve that context instead of silently treating runs as interchangeable.
Definitions
Visibility share is sampled answers naming a brand divided by all answers. Several brands can occur in one answer, so the shares overlap. Named-first rate counts answers where the brand is named first. A per-platform metric uses only that assistant’s answers.
Choice share describes selections within the simulated purchase sample. Brand, seller and channel are different dimensions. The reason an assistant states for choosing an offer is a reported rationale, not proof that changing one attribute will cause a sale.
Citation share uses the study’s stated link-count denominator. A read page and a credited source are not necessarily the same thing. Links, brands and answers must not be substituted for one another when comparing percentages.
What the research does not claim
No real orders, payment, customer retention or revenue are measured by these studies. The Agent-Ready pilot did not perform checkout; labels such as “bought” should be read as historical shorthand, not a transaction record. New research will use selection or readiness wording.
Published tables are aggregates. Original prompts and source/model metadata are retained in the measurement system as described by each study, but not every raw run is public. Where a public page lacks a version or timestamp, do not assume independent historical reproduction is possible.
Sources and authorship
This method note is by Vladimir Fedorov, co-founder and CTO of ShelfSignal. It describes the formats already published; it is not a fresh run or a new result.
Request your report
Tell us about your website and goals. We will reply with the appropriate scope and next steps.
From findings to approved improvements
Use the measurement on its own, or ask us to turn it into website recommendations and implementation. Optional managed hosting adds chat-based editing through an agreed messenger, with you in control of publication.