Methodology, in the open.
Exactly how Mentioned turns AI answers into a visibility score — every rule, threshold and limitation in one place. The same rules run in the product.
- 01Buyer prompts25 proposed from your site · you edit
- 02Four enginesOfficial APIs · web search on
- 03Structured readingClaude · strict JSON · validated
- 04Rolling metrics7-day window · 95% margin
- 05Alerts & planOnly when a change is real
4
AI engines
ChatGPT, Perplexity, Claude and Gemini through their official APIs, with web search on.
7
Day rolling window
The latest answer for every prompt and engine in the last 7 days makes up the score.
95%
Margin of error
Every score ships with a ± range, and alerts ignore moves smaller than it.
1
Public formula
A mention rate weighted by position. No black box, no secret sauce.
Ask what buyers actually ask.
A score is only as good as the questions behind it. We track buyer prompts — what a potential customer would ask an AI assistant while choosing a product.
- During onboarding we read your homepage and propose 25 prompts in your market’s language, phrased the way a buyer in your country would ask.
- Four intents: “best X for Y”, “X vs Y”, “alternatives to X” and “how to …”.
- Most prompts leave your brand name out, so we measure unprompted recommendations.
- You edit everything: add, tag, pin, pause, or import and export by CSV.
- Each prompt is sent verbatim as a single user message — no system prompt, no hint about your brand.
- Best forCategory shortlists
“best crm for a 10-person sales team”
- VersusHead-to-head choices
“northwind vs fieldline for startups”
- AlternativesSwitching moments
“alternatives to kestrel crm”
- How toJobs your product does
“how to automate sales follow-up emails”
The engines buyers use, through the front door.
Every check goes through the engine’s official API with live web search enabled — the closest repeatable equivalent of a buyer asking the assistant today.
| Engine | API | Default model | Web search | Location |
|---|---|---|---|---|
| ChatGPT | OpenAI Responses API | gpt-5-mini | Web search tool, low search context | Country passed |
| Perplexity | Sonar chat completions API | sonar | Built-in web search, low search context | Country passed |
| Claude | Anthropic Messages API | claude-haiku-5-5 | Web search tool, up to 2 searches | Country passed |
| Gemini | Gemini API | gemini-3-flash-preview | Grounding with Google Search | Not supported by the API |
Models in use on this deployment. They’re configurable and change as providers ship new ones.
Localized where possible
Your brand’s country is passed to ChatGPT, Perplexity and Claude as an approximate location. Gemini’s API has no location setting for grounded search.
Engines by plan
Starter: ChatGPT & Perplexity · Growth: ChatGPT, Perplexity, Claude & Gemini · Agency: ChatGPT, Perplexity, Claude & Gemini.
Built for bad days
Each call has a 25s timeout and 2 retries. After 5 consecutive failures an engine is paused briefly, and checks resume later.
Every answer, read the same way.
A small Claude model reads each answer at temperature 0 and returns strict JSON that is validated against a schema before anything is stored.
best crm for a 10-person sales team?
For a small team, these come up most:
- 1. Fieldline — strong pipeline automation.
- 2. Northwind — fastest to roll out, fair per-seat pricing.
- 3. Kestrel — deep customization.
g2.comreddit.comcapterra.com
{
"brandMentioned": true,
"brandPosition": 2,
"sentiment": 0.6,
"competitorMentions": [
{ "name": "Fieldline", "position": 1 },
{ "name": "Kestrel", "position": 3 }
],
"citedDomains": ["g2.com", "reddit.com", "capterra.com"],
"shortReason": "Listed second as the fastest to roll out."
}Invalid JSON
The validation error is sent back once so the model can correct itself.
Still failing
The check is retried as a job — up to 3 attempts, with exponential backoff.
Never guessed
A check that still fails is left out of scores, not counted as “not mentioned”.
What counts as a mention
- Your brand, one of its aliases or your domain is named as an option, recommendation or subject — a passing word doesn’t count.
- Rank is the 1-based position in any ranked or bulleted list of options; unlisted mentions have no rank.
- Sentiment is the tone toward your brand, from −1 (very negative) to +1 (very positive).
What else is recorded
- Tracked competitors in their tracked spelling, plus other vendors the answer recommends — which powers competitor discovery.
- Cited sources as bare domains, merged with the URLs the engine returned.
- A one-sentence reason, so you can see why the verdict was made.
- The reader sees up to 12,000 characters of the answer and 40 cited URLs.
One number, one public formula.
The visibility score is a mention rate weighted by position: being named first is worth more than being named fifth, and being left out is worth nothing.
- Ranked #11.00
- Ranked #20.85
- Ranked #30.70
- Ranked #40.55
- Ranked #50.40
- Ranked #6 or lower0.30
- Mentioned, not in a list0.60
- Not mentioned0.00
| Prompt | Engine | Outcome | Weight |
|---|---|---|---|
| best crm for small sales teams | ChatGPT | Ranked #1 | 1.00 |
| best crm for small sales teams | Perplexity | Ranked #2 | 0.85 |
| best crm for small sales teams | Claude | Mentioned, no list | 0.60 |
| best crm for small sales teams | Gemini | Not mentioned | 0.00 |
| alternatives to kestrel crm | ChatGPT | Ranked #3 | 0.70 |
| alternatives to kestrel crm | Perplexity | Not mentioned | 0.00 |
| alternatives to kestrel crm | Claude | Ranked #1 | 1.00 |
| alternatives to kestrel crm | Gemini | Ranked #5 | 0.40 |
100 × mean(1.00, 0.85, 0.60, 0.00, 0.70, 0.00, 1.00, 0.40)
56.9± 28.0
Which answers count
The latest run for every active prompt × engine within the trailing 7 days — every sample from that run, nothing older.
Overall and per engine
The same formula produces an overall score across engines and a score for each engine, every day.
Paused prompts drop out
Deactivated prompts stop counting immediately, so the score always reflects the set you track today.
Honest error bars, not false precision.
AI answers vary from one ask to the next. So next to every score we show a ± range — the 95% margin of error of the average answer weight.
- Computed with a normal approximation: 1.96 × the standard error of the answer weights, in score points.
- It captures variation across prompts, engines and repeat samples, and shrinks as you track more answers.
- Shown once there are at least two answers; ranges under half a point are hidden.
- In the worked example above, 8 answers give 56.9 ± 28.0 — far too wide to act on. Breadth fixes that.
answers in the window
Same spread as the worked example. 200 answers is, for example, one site with 50 prompts checked once on 4 engines.
How often prompts are checked
Self-serve plans re-check every prompt on every engine once a week. Because the score uses the latest answer per prompt within a 7-day window, the trend stays smooth between checks.
Every prompt on every engine, once a week. Plus quick checks that re-ask your 10 least-recently-checked prompts on demand.
Pinned priority prompts every day; the rest rotate in slices so each one is still covered every 7 days.
- StarterEvery prompt checked weekly + 2 on-demand checks per brand/month
- Growth25 priority prompts weekly · every prompt at least every 14 days + 2 on-demand checks per brand/month
- Agency10 priority prompts weekly · every prompt at least every 14 days + 2 on-demand checks per brand/month
- EnterpriseDaily tracking on pinned priority prompts, optional repeat samples per check
A quick check re-asks your 10 least-recently-checked prompts. Self-serve plans take one sample per check.
Every metric, defined.
The numbers around the score use the same window and the same answers, so they always add up.
Only when it’s real.
Scheduled runs compare today’s rolling metrics with those from 7 days earlier. The default drop threshold is 10 points — and it’s raised to the margin of error whenever that’s larger.
Visibility drop
Score falls by at least the larger of 10 points and the margin of error (now or 7 days ago). Critical at twice that.
Visibility gain
The same rule, upward — so good news is also real news.
Competitor overtook you
A tracked competitor now appears in more answers than you, after being at or below you 7 days earlier.
Lost or gained prompts
A prompt on an engine flips between mentioning you and not (majority of samples) between its last two runs.
New gap domains
Sites that started citing competitors but not you, compared with a week earlier.
Run problems
A run failed or finished incomplete — so missing data is never mistaken for a drop.
What the numbers can’t tell you.
We’d rather you trust a smaller claim than doubt a bigger one. Here is what we don’t claim.
01
API answers aren’t a screenshot of an app
Consumer apps can use different models, system prompts, memory, personalization and location. We measure how the models answer consistently — not what one person sees.
02
Answers are a sample
A score is only as representative as the prompts you track. Review them as your market and positioning change.
03
Engines change without notice
Model and retrieval updates can shift many answers at once. Watch for step changes on a single engine.
04
Automated reading can be wrong
Similarly named products occasionally confuse extraction. Aliases help, and every verdict links to the full answer and the reason.
05
Outages aren’t zeros
Checks that fail after retries are left out of scores rather than counted as “not mentioned”. See the status page.
06
Recommendations are suggestions
The weekly plan is written by an AI model from your data. It can’t guarantee how any engine will answer.
Every change, on the record.
When the way we measure changes, it’s logged here and in the changelog.
v2.1 ·
- Every self-serve plan re-checks every prompt weekly, with on-demand quick checks of the 10 least-recently-checked prompts.
- Cost-efficient web-search models by default on every engine (configurable per deployment).
v2.0 ·
- Scores moved to a rolling 7-day window using the latest run per prompt and engine.
- 95% margin of error on every score; drop alerts require a change larger than the margin.
- Brand aliases count as mentions; your country is passed to engines that support location.
- Competitor discovery from vendors recommended in answers.
v1.0 ·
- Position-weighted visibility score.
- Structured answer reading with schema validation.
- Cited sources and gap domains.
Why use the APIs instead of the ChatGPT or Gemini apps?
APIs give a consistent, repeatable measurement: the same question, asked the same way, every week, with no personal history, memory or logged-in state shaping the answer. Scraping consumer apps is fragile, often against their terms, and still only shows what one account sees. The trade-off is that API answers can differ from what a particular person sees in an app — so read the score as a trend across many prompts, not a screenshot.
Does Mentioned tell the engines which brand it is tracking?
No. Each prompt is sent exactly as written, as a single user message, with no system prompt and no hint about your brand. Your brand, aliases and competitors are only used afterwards, when the answer is read.
How big a change should I take seriously?
Compare the move with the ± range next to the score. A change smaller than the range is usually noise. Drop and gain alerts already apply this rule: they only fire when the change is larger than both your alert threshold and the margin of error.
Why does one engine sometimes move all at once?
Providers update models and retrieval without notice. A step change across most prompts on a single engine, on the same day, is more often a model update than a change in how the market sees you. Open a few answers before and after to check.
Can I check a verdict myself?
Yes. Every data point links to the full answer with your brand and competitors highlighted, the cited sources and a one-sentence reason for the verdict. If a similarly named product is being counted, add or remove an alias.
Free AI audit
Put the method to work.
Run a free audit: five buyer questions on two engines, scored exactly as described on this page.
Your website
- 01Where you appear across AI answers
- 02Who AI recommends instead of you
- 03Every answer, with the sources cited
Free · results in about a minute · or start a 3-day trial