This page states how Salience Index measures AI visibility. It is the standing method: each published study also carries a pre-registration stating the exact sample, prompts, engines, pass counts and dates it used, so the study declares its instance of what this page defines. The page is versioned, and changes are logged at the bottom.
What we measure
When a person asks an AI engine a buying question, the answer names a small number of brands. For a defined brand and a defined prompt set, we measure five things. Presence: is the brand named in the answer at all. Salience rate: the share of answers that name it, across repeated runs. Position: where it appears among the brands named. Share of voice: who else is named, and how often. Sources: which domains the engine cites or draws on, where the engine exposes them. Brand matching includes obvious name variants and excludes ambiguous mentions; each study's pre-registration lists its brand set and matching rules.
The engines
We measure ChatGPT, Google AI Overviews, Perplexity, Claude and Gemini. Not every study covers every engine, and each pre-registration names its engine set. Engines retrieve and answer differently, so results are reported per engine before any aggregate is shown.
Prompts
Prompt sets are written from how buyers actually ask: discovery questions, constraint questions, comparisons. Prompts are tagged by intent family and, where relevant, by geographic frame, because whether a question mentions a market changes who gets named. Every study publishes its full prompt list. If you cannot see the questions, you cannot trust the score.
Repeated runs, because answers move
The same prompt returns different answers on different days, and sometimes within the same day. A single screenshot is an anecdote. Every prompt in a study runs multiple times, and we report the range across runs rather than the best one. Stability is itself a finding: a brand named in nine answers out of ten is in a different position from a brand named in five.
Collection discipline
Runs use clean accounts with no personalization and no history. Every answer is archived in full with its timestamp, engine and prompt version. Quotes in our reports come from archived answers, never from memory.
Pre-registration
Before a study's full data collection begins, its design is published: the sample, the prompt set, the engines, the pass counts, the dates, and the analysis we commit to running. Pilot runs conducted before registration are disclosed as pilots and used only to validate the instrument. Pilots are never reported as findings.
Product scores
Our tools report composite scores built from the same measurements. Wherever a single number appears, the components behind it appear beside it. Score thresholds are provisional until they are re-derived against our first full research scan; until then, treat the components as the signal and the single number as a summary.
Whether AI can read a site at all
Being named and being readable are different measurements. The Report Card layer fetches a site the way AI crawlers do: it evaluates robots.txt rules per crawler, compares what a non-rendering crawler receives against what a browser renders, and checks structured data and discovery files. It separates training crawlers, the bots that collect pages for model training, from retrieval crawlers, the bots that fetch pages to answer live questions, because blocking the wrong family silently removes a brand from AI answers while classic Google rankings look untouched. The two families are always reported separately.
What this method cannot see
We measure logged-out, unpersonalized answers, and real users increasingly get personalized ones. We measure at points in time, and the engines change continuously, so every number carries a date. Geographic frames approximate location effects rather than reproduce them. Measurement in this category is young everywhere, including here; where our confidence is low, the study says so.
Provenance labels
Every material claim in our research carries one of three labels. Verified: supported by a primary source or multiple independent sources. Claimed: asserted by a party with a commercial interest in the claim, which includes everything any company says about itself, ours included. Inferred: our reading of the evidence, marked as ours.
Data and corrections
Study data ships under CC BY 4.0, free to reuse with attribution, so anyone can re-run our numbers without trusting us. Corrections go to [email protected] and are applied with a note on the page they correct.
Version log
v1, published 9 August 2026. First public version.