The method
Named, or absent.
One answer, a few brands named, the rest absent. Bermee measures who makes it in.
On this page · 18 sections
Method v0.9
The Bermee measurement standard: what is counted, how, under which rules, and what is disclosed with every issue. Metric definitions, formulas, sampling rules, the source taxonomy and the disclosure items are public under CC BY 4.0.
What the number is
Mention Rate is the share of AI answers to a category’s buying questions that name the brand.
What n means
n is the number of AI answers behind the figure. Bigger n, steadier number.
What the 95% confidence interval means
Example
Say a brand is named in 380 of 1,000 answers. Its Mention Rate is 38%. Asking again with the same questions would not give exactly 38% — answers vary. The 95% confidence interval, here 35–41%, is the range where the true rate sits 19 times out of 20. When two brands’ ranges overlap, the index does not call one ahead of the other.
Why we show it: a single number without its range invites false precision.
How the number is made
Questions
Each category has at least 40 questions people ask when they are deciding what to buy. Bermee writes them, labels each by intent, and refreshes the set every quarter.
≥ 40 prompts; intent mix ≥ 40% recommendation, ≥ 20% comparison, ≥ 20% informational, ≤ 20% navigational
AI systems
Every question goes to 5 AI systems: OpenAI, Perplexity, Gemini and Anthropic through their APIs, and Google AI Overviews through Google search results. Each issue names the model, version and surface of each one.
Five engines, declared per issue with model and version
Runs
Each question is asked five times on each assistant, every week. Repeating the question shows how much the answers vary.
≥ 5 executions per prompt per engine per collection week
Answers
Every answer is logged as it came back, then coded: which brands it names, in what order, and which sources it cites. Failed calls are left out and counted in the disclosure.
Provider errors excluded from denominators and counted in the disclosure block
Mention Rate
Mention Rate is the answers that name the brand divided by all answers. It is published with its 95% range and n, so you can see how steady it is.
Mention Rate = named ÷ answers; 95% confidence interval (Wilson); ranked at ≥ 2% of answers
How this index is built
Answers are collected live every week and logged, not modelled. Every published rate carries a 95% confidence interval and n, the number of answers behind it. Prompts are sampled and published: a labelled sample at launch and every retired prompt after each rotation. Ownership is disclosed on every page and in every issue.
0. Design rules
| Rule | Why |
|---|---|
| Measurement only. Bermee never sells optimisation, content or tooling to improve a score. | The referee cannot coach a team. |
| IAB names for the numbers; AMEC logic for the report. | Borrowed neutrality. |
| What other measurement tools do not disclose, Bermee publishes. | Disclosure is the standard. |
| Brands are ranked; agencies are never ranked or scored. | Partners must be able to trust the index. |
| One headline number, monthly movement, confidence stated. | A number that moves for unexplained reasons is not a measurement. |
Bermee states what moved and by how much. It never states why. Interpretation belongs to partners.
1. Public and proprietary
| Public (CC BY 4.0) | Proprietary (trade secret) |
|---|---|
| Metric definitions and formulas (§2) | The live prompt sets, per category (sample and retired prompts are public — §3) |
| Prompt-set construction rules: intent mix, counts, refresh cadence (§3) | Brand alias dictionaries |
| Sampling rules: executions per prompt per engine, period, rolling window (§3) | Source-classification model (the taxonomy is public; the classifier is not) |
| Source taxonomy (§2.7) | Any weighting applied in composite scores (v0 uses none) |
| Disclosure items per issue (§5) | Raw answers (available under license) |
| Headline index table per category, CSV | Full dataset: prompt-level, answer-level, URL-level |
2. Metrics
All metrics are computed per measurement cell — a category (a defined market, e.g. “running shoes — US”) × region (national by default) × persona (neutral by default; a disclosed panel persona otherwise) × surface (chat answer by default; AI Overview, AI Mode or shopping surface where collected) — per engine, per period, and aggregated across engines with equal weight unless an issue states otherwise. v0 issues publish the default cell only; every cell is named in the disclosure block. Each metric is reported with its 95% confidence interval (Wilson) and the number of answers behind it.
2.1 Presence
Mention Rate — the headline number
Share of responses in the category prompt set in which the brand is named.
responses naming brand ÷ all responsesStatus: Headline. Published in issue 1 with 95% CI and n.
Mention Rate (competitive)
Same numerator; denominator = responses naming at least one brand.
responses naming brand ÷ responses naming ≥ 1 brandStatus: Published alongside P1.
Share of Voice
Brand’s share of all brand mentions. Brands below the 2% threshold are aggregated as “other” so Share of Voice sums to 100%.
brand mentions ÷ all brand mentionsStatus: Published.
Citation Rate (owned)
Share of responses citing a URL on the brand’s owned domains.
responses citing owned domain ÷ all responsesStatus: Published.
Attribution Gap
The brand’s works, products or people appear without the brand being named.
mention rate of brand’s works/products − P1Status: Published. Bermee’s own metric.
Visibility Momentum
Change in P1 versus the prior official issue, with a significance flag.
P1(t) − P1(t−1); significant if the 95% CIs do not overlapStatus: Published from issue 2. Issue 1 shows “new” for every brand.
2.2 Prominence
Average Position
Mean rank of the brand among brands named in a response.
mean over responses naming brandStatus: Published.
First-mention rate
Share of responses where the brand is the first brand named.
first-named responses ÷ responses naming ≥ 1 brandStatus: Published.
Recommendation rate
On recommendation-intent prompts only: share of responses that name the brand as a pick.
recommended responses ÷ recommendation-prompt responsesStatus: Defined. Not published in issue 1 (publish gate — see below).
2.3 Portrayal
Sentiment polarity
Share of brand mentions positive / neutral / negative, with themes. Measured from issue 1 as an optional module under evaluation; never in the headline. Publication is decided after issues 1–2 on stability.
Status: Under evaluation. Not published in issue 1.
Factual inaccuracy rate
Share of brand claims that contradict a declared fact sheet.
Status: Future module; named, not built.
Hallucination rate
Share of responses attributing to the brand products or facts that do not exist.
Status: Future module.
2.4 Persuasion
Recommendation strength
Covered by R3 in v0.
Status: See R3.
Post-citation click-through
Out of scope for a public standard (needs brand analytics); declared.
Status: Out of scope.
2.5 Sources
Citation Share (domain)
Share of all citations in the category pointing to a domain. The Source Map.
citations to domain ÷ all citationsStatus: Published.
Citation Coverage
Share of responses citing the domain at least once.
responses citing domain ÷ all responsesStatus: Published.
Citation Share by source class
C1 aggregated by taxonomy class (§2.7).
Status: Published.
Citation Share (URL)
C1 at page level.
Status: Audit deliverable, licensed.
2.6 Composite
No composite score in v0. The headline is P1 Mention Rate, reported with P5 and P6. A weighted score is deferred until two issues of data exist; any weights would be published.
2.7 Source taxonomy
| Class | Examples |
|---|---|
| Owned | The brand’s own domains |
| Earned editorial | Newspapers, magazines, specialist press, review sites with editorial staff |
| Social & community | Forums, Reddit, YouTube, LinkedIn |
| Reference | Wikipedia, directories, review aggregators |
| Institutional | Government, academic, standards bodies |
| Commerce | Retailers, marketplaces |
| Other | Everything not classified above |
The classifier is proprietary; the taxonomy and examples are public. The taxonomy version is named in every disclosure block.
3. Measurement rules
| Rule | v0 value |
|---|---|
| Engines | Declared per issue with model and version string. v0 set, 5 AI systems: OpenAI, Perplexity, Gemini, Anthropic, Google AI Overviews. AI Mode when a collection path exists. |
| Prompt set per category | ≥ 40 prompts; intent mix ≥ 40% recommendation, ≥ 20% comparison, ≥ 20% informational, ≤ 20% navigational. |
| Prompt authorship | Written per category by Bermee; refreshed quarterly; 20% rotation at most per refresh, with the overlap reported. |
| Prompt disclosure | Live set private. Published per category: a sample of 5–10 prompts, intent-labelled, at launch; every retired prompt after each quarterly rotation, dated. Full live set available under the dataset license. |
| Executions | ≥ 5 per prompt per engine per collection week. The CI width shows the sample. |
| Region, persona, surface | v0: US national, neutral persona, chat answer. State-level cells and disclosed persona panels follow for later categories, published as experimental until stable. |
| Collection cadence | Weekly. |
| Publication cadence | Official issue: monthly (calendar month; frozen snapshot; P6 significance versus the previous issue; published by the 3rd business day). Provisional: a weekly 4-week rolling figure, labelled provisional, for paid accounts only; never a second official number. |
| Brand universe and the 2% threshold | Discovered, not hand-picked: a seed list builds the alias dictionary; every brand the engines name in ≥ 2% of a category’s valid responses enters the ranked table. Below the threshold, brands are counted as “other” so Share of Voice sums to 100%. |
| “Also named” | Every brand below 2% is still published under the leaderboard as “Named in fewer than 2% of answers”: name and the number of answers it appeared in, unranked, sorted by count then name; no CI, no movement. Aliases merged in review count as one name; extraction errors do not appear. |
| Brand aliases | Alias list per brand, versioned and disclosed on the brand page. |
| Movement and significance | Reported only with a significance flag. Official P6 compares calendar-month issues (independent samples); a change is significant when the two 95% intervals do not overlap. The weekly provisional view shows movement versus the last official issue, flagged provisional. |
| The measurement cell | Category × region × persona × surface, per engine and aggregated with equal weight. v0 publishes the default cell; every cell is named in the disclosure. |
| Errors | Provider errors are excluded from denominators and counted in the disclosure. A failed extraction excludes its answer. |
3.1 Sample prompts — Running shoes — US
8 of the 40 live prompts, intent-labelled. The live set stays private; retired prompts are published, dated, after each rotation.
What are the best running shoes right now?
recommendation · added 2026-10-11What are the best cushioned running shoes for long distances?
recommendation · added 2026-10-11What running shoe brands do podiatrists recommend?
recommendation · added 2026-10-11Hoka vs Brooks: which is better for daily running?
comparison · added 2026-10-11Which running shoe brands are most popular with marathon finishers?
comparison · added 2026-10-11How often should I replace my running shoes?
informational · added 2026-10-11Are carbon-plated running shoes worth it for amateurs?
informational · added 2026-10-11Where can I buy Hoka running shoes?
navigational · added 2026-10-11
Retired prompts: none yet.
4. Publish gates
A metric is published only when every field it depends on scores ≥ 0.90 against a human-verified gold set. Repeatability is checked per field at the same bar; the all-fields composite is disclosed, not used as a gate.
| Field | Feeds | Issue 1 |
|---|---|---|
| brands named + position | P1, P2, P3, R1, R2 | published |
| products | P5 | published |
| citations + owned domains | P4, C1, C2, C3 | published |
| is recommended | R3 | withheld — below the bar at the time of the pre-issue draft |
| polarity | T1 | withheld — under evaluation, never in the headline |
A withheld metric is named as withheld in the issue, never silently missing. A published field that later drops below 0.90 on a gold check is withheld from the next issue until it clears.
5. Disclosure block
Published with every issue:
Engines and model versions · prompt count and intent mix · executions per prompt per engine · collection dates · region, persona (with preamble version) and surface per cell · brand alias versions · source taxonomy version · error count excluded · method version · ownership statement (common ownership) · what changed since last issue.
License and versions
The method (this page), the headline index table per category and the headline CSV are published under Creative Commons Attribution 4.0. Attribute as “Bermee Index, bermee.com” with the issue number. Everything else is under the terms.
| Version | Date | Status |
|---|---|---|
| v0.9 | 2026-10-22 | Pre-issue draft, published with site v0 |
| v1.0 | 2026-10-27 | With issue 1 — Running shoes — US |