Methodology
How does TAMGIO measure visibility?
By Shamil Guliyev · Updated 2026-11-17
For every project TAMGIO asks 30-40 real buyer questions (prompts) to the selected AI engines repeatedly (N-sample), stores the answers, and counts only grounding-verified ones — answers where the engine actually searched. Every sample's status is shown openly; a memory answer or a missing answer surface is never counted as «the brand is absent».
Measurement steps
- Prompt set: 30-40 questions across five intents (best-of, comparison, how-to, local, pricing), approved by the user.
- N-sample: the same question goes to the same engine several times — one answer can be a coin flip.
- Grounding check: does the answer metadata confirm a live search.
- Analysis: does the answer actually recognize the brand — or only repeat the name, say it cannot find it, or confuse it with another company; in which position, in what tone, next to which competitors, was its own domain cited. Only recognition counts as a mention. Since 17 November 2026 a second, rule-based layer with an independent classifier double-checks this: in comparison and trust questions that name your brand, an answer that only repeats the name or confuses it with another product is not counted (one that describes your brand correctly and clearly tells it apart from a similarly named product still is), and a cautious verdict on a safety question is marked as a warning; an answer to a question without your name can count as recognition when it cites your own pages for what it says about you and the classifier agrees. From the same day, when the analysis misses tracked competitors that are clearly named in the answer text, the competitor list is completed from that text — this can only change share of voice. On the last scan before the switch (16 November 2026, 144 answers) recognition, mention rate, citation rate and coverage did not change, and share of voice moved from 16.84% to 16.61%.
- Aggregation: daily metrics per engine and segment, then the report.
Sample statuses
Every sample gets one of four statuses; only the first one counts.
grounded_ok— live search verified — the ONLY status that enters the metricsmemory_answer— answered from memory — shown, not countedno_surface— no answer surface (e.g. no AIO block) — not counted as «brand absent»provider_error— provider/network error — our side, not a measurement
Why the API, not UI scraping?
Every answer TAMGIO counts comes from the engine's official API — not from a bot that types into the ChatGPT or Perplexity app and copies the screen. That keeps the request repeatable, every answer timestamped and the access within the providers' terms.
- Repeatable
- Every scan sends the same question to the same model with the same settings. No logged-in session, cookie banner or page layout sits in between that could change the result or break the run.
- Timestamped
- Each answer is stored in full with the moment it was received and its citations. The evidence layer shows exactly that record, so any number can be traced back to the answer and the time.
- Within the providers' terms
- OpenAI, Perplexity and Gemini are queried through their official APIs with our own API key — no automated logins, no consumer accounts, no getting around bot protection. Google AI Overviews has no model API: in Audit Mode we read it through a commercial search-results (SERP) API, not with a browser of our own.
What the API does not show
An API answer is not a copy of what a signed-in person sees in the ChatGPT, Perplexity or Gemini app. The apps can run a different model or settings and can tailor the answer to the person — ChatGPT, for example, can use saved memory, past chats and an approximate location. Our calls carry no account and no history: only the question, web search switched on and, where the API supports it, the project's country (for AI Overviews, also its language). Read the numbers as a clean, comparable baseline of how the engine answers — not as a guarantee of what one particular customer will see.
Metrics and formulas
The formulas are public and the same code produces the dashboard, the shared report and the PDF.
| Metric | Formula | Note |
|---|---|---|
| Discovery visibility | mentions in unbranded / competitor-named prompts | headline metric — branded questions are a separate segment |
| Mention Rate | Σ(per-prompt mention share) / number of prompts | averaged per prompt, not per sample |
| Share of Voice | brand mentions / (brand + competitor mentions) | each competitor counts once per sample; your brand counts only when recognized, a competitor when named |
| Citation Rate | samples citing your domain / all valid samples | the citation link comes from the answer metadata; for Gemini the domain is read from the source name Google returns — the redirect link is never opened |
| Prompt Coverage | prompts with at least one mention / all prompts | shows the «we never show up» gap |
How are tone and warnings measured?
Tone answers one question: how does this answer talk about your brand? It is labelled only for measured answers that actually recognize the brand — an answer that does not name it, cannot find it or confuses it with another company gets no tone.
| Label | What it means |
|---|---|
| recommended | explicitly advised, named as the top pick or clearly favoured |
| neutral | listed or described without preference |
| negative | criticised or described negatively |
| warning (separate flag) | the answer explicitly warns against your brand, raises a safety or scam concern about it, or advises caution about it specifically. Generic caution for the whole category («always check reviews») does not count. It is independent of tone: a negative answer is not automatically a warning, and both can be true at once. |
How it is computed
A separate language model (currently Google Gemini) reads the question together with the full answer, your brand name, aliases and the competitor list, and returns fixed labels in a strict JSON format. Code then checks the result: if your brand is not actually in the answer text, recognition, tone and warning are dropped. Tone does not affect any of the visibility metrics above. In the dashboard and the report, «Sentiment breakdown» is the share of recommended, neutral and negative among the recognized answers, summed across all prompts; the warning flag appears on that answer's evidence card in the dashboard.
Limits
- It is a model's reading, not a person's. Borderline wording («a solid option») can land on either side of recommended and neutral.
- Tone describes how one answer is worded — not your real reputation, reviews or customer satisfaction.
- The warning flag is deliberately strict. In our human-labelled checks (83 answers, September 2026) it raised no false warning but caught only 2 of 7 real ones. A missing flag means «not detected», not «no warning» — read the answer itself.
- Answers analysed before 28 September 2026 have no warning flag — it did not exist yet.
Evidence layer
Open the dashed underline beneath any metric and you see the raw API answer that produced it, the citations and the measurement timestamp. No screenshots — the answer itself is stored, so you can verify the claim yourself.
Honesty rules
- No percentage-improvement promises — AI answers are non-deterministic.
- Failed and empty results are shown, not hidden.
- A number about a competitor is published only with an official source and a date stamp.
Measure your own brand
The one-off audit runs on the same methodology — we do the setup.
Measure your own brand