How every number is made, and where it stops

Definitions, sampling limits, measurement uncertainty and what is observed versus inferred. If a number on this site or in the product cannot be traced back through this page, tell us.

Definitions

The words used on every chart and in every report, in the order they occur in the product.

Tracked question
One buyer question in one market, collected on the engines your plan includes. The same wording in two markets is two tracked questions, because each engine answers it from inside a different country.
Market
A country and language pair from which answers are collected, for example United States / English or Switzerland / German. Answers are collected from inside the market, so the same question in two markets is two measurements.
Answer
One response from one engine to one tracked question at one collection time, stored in full with the sources it cited. Every rate on this site and in the product is counted over answers.
Mention
A brand named in an answer, detected in the stored answer text. A mention says the brand appeared; the role says how.
Role
How an answer treats a brand it names. Recommended: put forward as the pick for the question. Listed: one option among several, without preference. Mentioned: named in passing, not as an option. Caveated: named with a warning or a condition. Absent: the engine was asked and did not name the brand.
Citation
A source URL or domain the engine attached to an answer. Citations are counted per answer, so a domain cited three times in one answer counts once for that answer.
Visibility score
The share of answers that name the brand, over the tracked questions in the selected engines, markets and period. 30 answers naming the brand out of 120 collected is a visibility score of 25, shown as 25 with 30 of 120 next to it.
Recommendation rate
The share of answers in which the brand’s role is Recommended, counted over the same answers as the visibility score. It is always lower than or equal to visibility.
Intent weight
An estimate of how close a question sits to a purchase, used to order the questions. It is inferred, labelled as an estimate, and never mixed into an observed rate. The product lists the inputs behind each weight next to the weight itself.

Sampling

How often answers are collected, why the same question gives different answers, and when a rate is shown at all.

  • One collection a day

    Every tracked question is collected once a day on each engine and market in your plan. On-demand scans are collected the same way.

  • Answers vary between runs

    The same question asked twice can return different brands, roles and sources. Bluemoon does not average this away; it keeps every answer and counts across them, which is why a count sits next to every rate.

  • Minimum sample

    A rate is shown once enough answers exist for the selection. Below that you see the count and a dash where the rate would be. A dash means not measured; a zero means the engines were asked and never named the brand.

  • Trends

    Trend points are computed on a rolling window. Each point is the rate over the answers collected in the window ending that day, with its count. Two periods are compared as two windows of equal length, never one day against another.

Measurement uncertainty

Four reasons the same brand can score differently from one day to the next without anything having changed on your side.

  • Day-to-day variation

    Engines sample their own models, retrieve fresh pages and change wording between runs. A brand named in 5 of 6 engines today and 4 of 6 tomorrow has not necessarily lost anything. A run of days tells you more than any two.

  • Engine changes

    Engines update models, retrieval and answer formats without notice. When we detect a change we date it in the changelog below. A step in a chart that lines up with a dated change should be read with that in mind.

  • Logged-out collection

    Answers are collected without an account and without history, so nothing personal shapes them. This is the closest thing to a neutral baseline; it is not what a signed-in user with a long history sees.

  • Geo-targeting

    Answers are collected from inside each market. Where an engine also uses signals we cannot set, such as an inferred language preference, the answer may still differ from what a resident sees.

A count next to every rate

Every rate in the product carries the number of answers behind it, in the same place, at the same size. The period and the answer count are printed on every report. When you filter to one engine or one market, the count shrinks with the selection and the rate is hidden below the minimum sample.

Finding

Ridgewell is named in 3 of 6 engines for alpine hardshell questions and recommended in 0.

Observed over 12 questions · 144 answers · Aug 10–Sep 7, 2026
Question
best hardshell jacket for alpine climbing
Engine
ChatGPT, web, not personalised
Market
United States, collected from inside the market
Collected
Sep 8, 2026, 06:12 UTC
Answer
#a7f3-2026-09-08-us-chatgpt · stored in full
Sources in answer
gearbench.example · trailnotes.example · alpstore.example
ObservedCounted from stored answers.EstimateNever shown as a rate. Where something was not measured you see a dash.

Observed versus inferred

Two lists that never mix. The left column is counted from stored answers; the right column is derived from other inputs and labelled wherever it appears.

Observed, counted from stored answersInferred or estimated, labelled as such
Mentions: which brands an answer namesIntent weight: an estimate of commercial value, labelled with its inputs
Roles: recommended, listed, mentioned, caveated, absentPrompt volume: how often a question is asked, estimated from search demand
Citations: the sources an answer relied onCompetitor classification: whether a newly named brand is a competitor or a retailer
Answer text: stored in fullInfluence on purchases: how much an answer moves a buyer
Collection time, engine and market

How each inferred item is labelled

Intent weight
Shown with an “Estimate” pill, together with the market and the inputs it was derived from. It orders questions; it is never added to an observed rate.
Prompt volume
Shown with an “Estimate” pill and the source it came from. It is never presented as a count of answers.
Competitor suggestions
Shown as “Suggested” until someone on your team accepts or rejects the brand. A suggested brand is not counted as a competitor in any rate.
Influence on purchases
Shown as “Inferred”, and only where traffic or conversion data is connected. It is never presented as a measured rate.

Known limits

What the method does not cover. We would rather list these than have you find them.

Changelog of the method

Every change to a definition, threshold or collection rule, dated. Charts that cross one of these dates should be read with it in mind.

DateChange
2026-08-31Microsoft Copilot added to the collected answer surfaces.

Read the answers behind your own numbers

Seven days free. Every rate you see will carry its count, and every finding its question, engine, market and collection time.