1. The questions
Every industry has 30 fixed questions, worded the way American shoppers type them: “best X for Y”, “which brand makes the best Z”, “what should I switch to if…”. None of them contains a brand name, and none was written from a brand’s point of view. The set stays the same from one window to the next, so when a ranking moves, the answers moved and the questions did not.
2. The platforms
ChatGPT — OpenAI gpt-5 through the Responses API with web search on, the way a shopper gets it in the app. Perplexity — Perplexity Sonar through the Agent API; search is built into the model. Gemini — Google Gemini 3.8 Flash with Google Search available; the model searches when it judges the question needs it, and each answer records whether it did. Each assistant gets each question once per window: no system prompt, no persona, no follow-up turn. The consumer apps can answer slightly differently from the APIs. We use the APIs because they can be put the identical question in the identical way week after week.
3. The population
Brands join an industry either because we seeded them or because an assistant mentioned them in an earlier window. No brand is ranked until its store passes a Shopify check: first the public /products.json endpoint, then the homepage headers and markup. Stores that fail are still logged when mentioned and shown under “named, but not on Shopify”, but they get no score. Stores we cannot reach sit out until the check can run.
The brand list is locked when a scan begins. Nothing is added, dropped or renamed mid-scan; a brand first spotted during a window enters the list for the following one.
4. The score
Visibility Score: answers that mention the brand, divided by all answers in the window, times 100. Mentioned in 40 of 90 answers gives 44.4. Average position: across the answers that mention the brand, the mean of its place in the list of brands given. Questions hit: the number of different questions that produced at least one mention. Movement: rank this window against rank last window, within the same industry. The per-assistant figures apply the same arithmetic to one assistant’s answers.
No weights, no normalisation, no model sits on top of the tally. Sentiment is not scored, there is no “authority” factor, and a brand we did not measure gets no estimate.
5. Mentions
A mention is a brand’s name or a listed alias appearing in an answer. Names that are also ordinary words (Seed, Ritual, Bubble, Hero) are matched case-sensitively to avoid counting the word rather than the brand; multi-word names are matched loosely. When an answer names a brand more than once, it counts once, at the position of its first appearance.
6. Cadence
Every index page states the date of the window it shows. The index is rescanned on a rolling basis and republished when a window completes; the previous window is kept for movement. Errors and time-outs are excluded from the answer count and reported per window.
7. What the Shortlist is not
- It is not a ranking of who has the best products. It measures what the assistants say, not whether they are right.
- It is not paid. No brand can buy inclusion, removal or position. ForgeCRO sells audits and implementation, and publishes this because the audit needs a benchmark.
- It is not a consumer-app screenshot. API answers and app answers can differ; we choose repeatability over theatre.
8. Corrections
If a brand is misnamed, wrongly verified or missing an alias that costs it mentions, write to info@forgecro.com. Alias changes apply from the next window; scores in a published window are not edited.
