A recommendation, measured honestly.
Mentionloom's headline number is how often an AI answer engine shortlists or recommends you, against the buyer questions that carry revenue. This page defines that number, the rules that keep it honest, and the three signals that sit around it.
The metric
AI Recommendation Share is the proportion of observed answers in which a brand is shortlisted or recommended, across its question set and engine set, over a fixed window.
Every answer resolves to exactly one state per brand, and only the last two count toward the metric:
| State | Definition | Counts |
|---|---|---|
| Absent | You do not appear, and no competitor is named. | No |
| Mentioned | Your name appears in any context, including a dismissal. | No |
| Shortlisted | You are named as one of several viable options. | Yes |
| Recommended | You are endorsed directly, named first, or given as the primary answer. | Yes |
Mention rate and recommendation share are different numbers. Monitoring tools stop at the first. Recommendation share is the one that changes a decision.
The sampling rules
These exist so the number survives a skeptical question. Omit one and the metric collapses under it.
Five runs per question per engine per cycle. One run is an anecdote. Five gives a variance estimate and makes the noise band honest.
Publish the band, not the point. "45.4%, ±2.5" is a different claim from "45.4%".
Hold back a control set. Roughly 20% of questions receive no intervention, so movement in the treated set is compared against the control.
Segment by intent stage. Discovery, comparison and decision move at different speeds, and decision questions are worth more.
Declare the surface. Every report states whether it measured an API surface or a consumer surface, and which model, region and date.
Check the consumer surface monthly. Re-run a subset of prompts directly in the consumer products and record the divergence.
Keep the raw answer. Engine, model, timestamp, question, raw text and cited URLs are retained so any number can be traced to the answer behind it.
Never imply causation from one cycle. Movement is reported with the control comparison, the window, and the other changes in that window.
Three signals, three jobs
The metric is one signal. Two others are reported beside it, never folded into it.
| Signal | What it is | Its job | In the metric |
|---|---|---|---|
| A — Answers | Observed recommendations across engines and questions | The metric | Yes |
| B — Referrals | Attributed sessions from AI sources | The outcome | No |
| C — Crawlers | Bot requests in server or CDN logs | Diagnostics: has the content been fetched at all | No |
Bot traffic is not a recommendation, and referral traffic is an outcome rather than a visibility measure. Mixing them into one score produces a number nobody can explain when it moves.
What this is today
Measure your own questions.
Join the waitlist to get a real workspace built from your site, competitors and analytics as the first cohort opens.