Skip to content
traze

Methodology · version 1.1

How Traze measures AI visibility and organic traffic

Every number in the dashboard, in a case study and in the free check follows the definitions on this page. Read them, check them and hold us to them.

Last updated 17 September 2026

In short

  1. 01We measure AI visibility by asking a fixed set of questions per topic to 11 AI engines every week and recording, per answer, whether your brand is mentioned and whether your domain is linked as a source.
  2. 02Share of voice is your share of all brand mentions within the same answers, compared with a fixed competitor set.
  3. 03Organic traffic comes from Google Search Console, not from estimates. Case studies compare a start month with a frozen measurement month and are labelled as example data until verified.

1 · Prompt set

Which questions we ask

AI visibility starts with the questions. For every topic in your strategy, Traze builds a fixed set of questions: between 10 and 50, phrased the way buyers ask an AI, in the language and country of your market. Not 'shopify seo', but 'which agency helps a Shopify store with SEO?'.

Every question gets an intent: orient (what is, how does it work), compare (which is better, what does it cost) or buy (who supplies it, where do I order). So you see not only whether you are mentioned, but at which stage of the customer journey.

The set is built from your brand profile, search data, 'People also ask' questions and the questions your support and sales teams already get. You approve the set and can add or remove questions. We never change questions silently: every change is logged and shows as a visible break in the trend line.

Example questionsExample
  • OrientWhat is generative engine optimisation and do I need it as an online store?
  • CompareWhich SEO platforms publish automatically into WordPress and what do they cost?
  • BuyWhich company can measure and improve our AI visibility in the Netherlands?

2 · Engines

Which engines we track

Traze asks every question to 11 engines: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Claude, Copilot, Grok, Le Chat, DeepSeek and Bol Shophulp. Per engine we use the official API where one exists. For engines without a usable API, such as Google AI Overviews, we retrieve the answer through a search-results provider; those measurements are labelled as such in the dashboard.

Every query runs in a clean session: no login, no conversation history and no personalisation, with the language and country setting of your market. The full answer, the date, the sources and the model version (when the engine provides it) are stored, so every number can be traced back.

Tracked right now

  • ChatGPT
  • Google AI Overviews
  • Google AI Mode
  • Gemini
  • Perplexity
  • Claude
  • Copilot
  • Grok
  • Le Chat
  • DeepSeek
  • Bol Shophulp

The list grows with the market. When we add an engine it starts at zero and is shown separately; we do not add it retroactively to old scores.

3 · Definitions

Mention, citation, position, competitor

01Mention
Your brand name (or a registered variant, such as your product name) appears in the answer text. A mention does not need to contain a link.
02Citation
A page on your domain is listed by the engine as a source: as a link, footnote or source reference. A citation counts even if your brand name is not in the text.
03Position
Where your first mention sits in the answer: first paragraph, middle or bottom. Higher is worth more, because readers stop there.
04Competitor
A brand from the competitor set you define at the start (at most ten). Brands outside that set are visible in the answers but do not count towards share of voice.

4 · Calculation

How the numbers are calculated

01AI visibility (per engine)
answers with a mention of you ÷ all sampled answers × 100
The share of questions in which you are mentioned, per engine and per topic. Multiple samples of the same question in a week each count.
02Citation rate
answers with a citation of your domain ÷ all sampled answers × 100
The same, but for source references. Usually lower than visibility; the gap shows where you are mentioned but not linked.
03Share of voice
mentions of you ÷ (mentions of you + mentions of your competitor set) × 100
Your share of all brand mentions within the same answers. An answer that names you and two competitors counts as one for you and two for the rest.
04Weekly score
average of all samples in the calendar week, per engine
The dashboard shows weekly scores, not single samples. One sample is too volatile to build anything on.

5 · Frequency

How often we measure

Every question goes to every engine weekly, and several times per engine, because AI answers differ from run to run. Those repeats are combined into one weekly score per question, per engine. How many repeats each engine gets is shown next to every score in the dashboard.

Visibility, citation rate and share of voice are calculated and stored per calendar week. We show trends over at least four weeks; a one-week difference is a signal, not a result.

Want to measure a question right away, for instance after a publication? You can, manually. Such a one-off sample is kept apart from the weekly score and does not count towards the trend.

How performance monitoring combines rankings and AI visibility

6 · Caveats

What the numbers are not

01AI answers vary
The same question gives different answers on the same day. That is why we sample several times and average per week, but a weekly score remains a sample with noise. Small shifts of a few percentage points are usually not a change.
02We do not see personalisation
We measure without an account and without history. What a logged-in user with their own history sees may differ. We measure the neutral answer because it is verifiable and comparable.
03Engines change
Models get updated, source references change shape, an engine starts searching differently. A sudden jump across all customers points to such a change; we mark it in the chart and explain it in the changelog.
04Small sets are noise
Ten questions for one topic give a rough indication; fifty give a usable picture. The dashboard shows, next to every score, the number of samples it rests on.
05A mention is not revenue
Being mentioned is a necessary step, not a sale. That is why we connect AI visibility to Search Console and, if you connect them, to conversions from GA4 or your store.

7 · Free check versus platform

How the free AI visibility check relates to the platform

The free check uses the same definitions but is a snapshot with a small set. The platform measures every week with your full set.

How the free AI visibility check relates to the platform
AspectFree checkPlatform
QuestionsA small set of standard questions for your market, built automatically10 to 50 questions per topic, approved by you
RepetitionOne sample, at the moment you run the checkWeekly, several times per engine, with weekly scores and a trend
EnginesThe engines available in the check at that momentAll 11 engines, labelled per measurement method
CompetitorsThe brands that happen to show up in the answersA fixed competitor set with share of voice
CitationsLinked or not, per answerPer page and per engine, with the sources others get cited for alongside
OutcomeAn indication of where you standA diagnosis per missed question and a task in the content calendar

A good score in the check is not a guarantee and a low score is not a verdict: it is one sample. That is why we call it a check and not a report.

Run the free check

8 · Organic traffic

Where the organic numbers come from

Organic numbers come from Google Search Console: clicks, impressions, average position and click-through rate, per page and per query. That is the source Google itself reports, not a tool's estimate. Traze syncs daily and keeps the history, including beyond the sixteen months Search Console itself shows.

A 'ranking' in Traze is the average position from Search Console over a period, not a single manual check. For search engines outside Google we use that engine's equivalent source where available.

Conversions and revenue only appear if you connect GA4, your store or your CRM. Without a connection we show no revenue figures; we do not compute them from assumptions.

Numbers in case studies

A case study compares the start month with the last full month before a frozen measurement date. The period does not quietly keep growing. AI citations in a case are only shown if we tracked them from the start. Until we have verified customer numbers, the case studies on this site are labelled as example data: illustrative, not a real customer.

See the case studies and their labels

9 · Commitments

What we do not do

  • We do not pay any engine, source or platform to be mentioned, and we do not buy mentions.
  • We do not query engines through personal accounts or scraped user sessions.
  • We never show a single screenshot as proof of visibility; only weekly scores, with the number of samples next to them.
  • We do not change definitions without a version number. This page is version 1.1; changes are listed below with a date.

Changes

  1. Version 1.0: first publication of the measurement method, ahead of the platform launch.
  2. Version 1.1: sections on sampling modes (API versus browser), confidence intervals and the open-source runner added; engines extended with Google AI Mode, Copilot and Bol Shophulp through browser capture.

10 · Sampling modes

API sampling versus browser capture

Every measurement carries a label: api or browser. With API sampling we ask the question through a model's official API, with fixed settings and no personalisation. That is cheap, repeatable and good for trends, but it is not what a consumer sees in the app: no ads, no product cards, sometimes a different model than the consumer version.

With browser capture a clean browser session on our own servers opens the consumer product (ChatGPT without an account, Perplexity, Google AI Overviews and AI Mode, Copilot, Bol Shophulp) and reads the answer, the sources, the cards and any sponsored placements. That is what people really see, but it is slower, more expensive and more sensitive to interface changes.

In the dashboard both labels sit next to every data point and are never mixed into one figure. Sponsored placements and product cards come exclusively from browser captures. Without a capture cluster the platform falls back to API sampling and says so.

api
Official model API, fixed parameters, no session. Repeatable, cheap, no ads or cards.
browser
Logged-out browser session on our own infrastructure. Answer, sources, cards and ads as a consumer sees them.
simulated
Deterministic simulation without keys or cluster, always labelled. Only to show flows, never as a figure.

11 · Confidence

Confidence intervals next to every score

A weekly score is a sample: n answers of which k name the brand. Next to the percentage we therefore show a 95% confidence interval (Wilson interval), wide for small n and narrow for large n. Two weekly scores whose intervals overlap we do not call different.

Judgements by the decision layer (is this brand mentioned, is it recommended, is this source a citation) each carry their own certainty. We store that raw probability and show it as certain (≥ 80%), likely (50–80%) or uncertain (< 50%). Uncertain judgements do not count towards scores and never trigger an automatic action.

Wilson 95%: p̂ ± 1.96 · √(p̂(1−p̂)/n + 1.96²/(4n²)) / (1 + 1.96²/n), centred on (p̂ + 1.96²/(2n)) / (1 + 1.96²/n).

12 · Open source

Reproduce it yourself with the runner

The measurement logic behind API sampling is an open-source CLI in the repository under packages/traze-runner (MIT licence). You give it a file with prompts and a list of engines; the runner asks every question in a clean session, stores the full answer with date and model id as JSON Lines and computes mention rate with a Wilson interval.

So you can reproduce every score on this site and in the sector index with your own keys, or criticise the method where it matters: in the code.

npx traze-runner sample --prompts prompts.txt --engines openai/gpt-4o-mini,perplexity/sonar --brand "Your brand" --out results.jsonl

packages/traze-runner in the repository

Questions · methodology

Frequently asked questions

Because engines work differently. Perplexity cites almost every sentence, ChatGPT only searches live for current questions, AI Overviews leans on the Google index. One averaged number hides exactly the differences you can act on.

You can measure today

See where you stand in AI answers

The free check uses the same definitions as the platform. Want the full measurement on your own prompt set? See pricing or book a demo.