Home / Methodology Disclosed The complete procedure, including the rules for borderline cases and the limits of the method. We disclose it because a measurement nobody can verify is not a measurement. meit is the GEO and SEO agency for ambitious companies in Zurich. We measure before we make an offer. Last updated 5 August 2026 · Procedure unchanged in use since 2025How we measure AI visibility.
Prompt set
Where the questions come from.
A prompt set is only as good as its origin. Invented questions produce invented results.
Source 1
Real customer enquiries
Wording taken from the client’s enquiries, pitch conversations and tenders. These questions are demonstrably real, because they were actually asked.
Source 2
Search terms with purchase intent
Taken from classic search and converted into natural language. They connect the measurement with what people actually search for in Google and Bing.
Source 3
The uncomfortable questions
Trustworthiness checks, direct provider comparisons, shortlist building. These are exactly the questions that produce the most revealing — and for the client most uncomfortable — answers.
Scope depends on the mandate: 40 to 60 prompts in stage 01, 80 to 120 in stage 02, 150 to 250 multilingual in stage 03. The set is not changed between measurements.
Systems
Four answer systems, two search engines.
Every system sources its content differently. Measuring one and inferring the others would not be serious work.
AI answer systems
- ChatGPTThe most used assistant, historically connected through the Bing index.
- Google AI OverviewThe widest reach, because it appears inside normal search results. Live in Switzerland since 25 March 2025, in four languages.
- PerplexityThe highest citation density, its own crawler, the strongest overlap with Google’s top 10.
- ClaudeUsed above average in B2B settings, with its own retrieval behaviour.
Measurement conditions
- No logged-in historyEarlier conversations and saved preferences distort the answer.
- From a Swiss IPLocation influences provider searches considerably. For German and Austrian mandates, from the country in question.
- Comparable weekdaysTo keep weekend and update effects constant.
- Classic search in parallelGoogle and Bing, desktop and mobile separately, 30 to 300 core terms.
Sources: Google, announcement of the AI Overview rollout in Switzerland, 25 March 2025 · StatCounter Switzerland, Bing desktop share, June 2026
- Taken from real customer enquiries
- Search terms with buying intent
- The question of credibility
Unchanged for the whole engagement.
- ChatGPT
- Claude
- Perplexity
- Google AI Overview
- Bing
30 core terms with buying intent.
- The wording of every answer
- Your sources, counted
- Positions in Google and Bing
- All of it as a table with timestamps
The raw data belongs to the client.
- From a Swiss IP
- No logged-in history
- Every answer checked by hand
- The same procedure each month
An unchanging procedure is what makes change measurable in the first place.
Rules
What counts and what does not.
Six rules fixed before the first measurement. Anyone who sets them afterwards sets them in their own favour.
What counts as a mention
The company name appears in the answer, spelled correctly or unambiguously attributable. A mention inside a list counts the same as an explicit recommendation. Mentions in paid elements do not count.
What counts as a citation
The company is linked as a source or explicitly given as evidence. A bare mention with no source reference does not count as a citation.
What counts as an independent source
Sources the company does not control itself. The company’s own website, its own social media profiles, paid directory listings and press releases on its own channels are not counted.
How competitors are selected
Confirmed by the client, not guessed by us. We propose a list, the client removes and adds. Only confirmed competitors enter the benchmark, so the comparison does not flatter our result.
How prompts are created
From three sources: the client’s actual customer enquiries, search terms with purchase intent from classic search, and the uncomfortable questions people really ask — about trustworthiness, say, or in a direct provider comparison. We do not phrase questions that favour a result.
How often we measure
Monthly, on comparable weekdays, with an unchanged set. Changes to the set are documented and the affected values are reported separately from the month of the change onwards.
Metrics
Six values, reported separately.
We deliberately do not combine them into an index. A visibility score hides which part actually moved.
| Metric | What it measures | How it is collected |
|---|---|---|
| Mention rate | Share of answers in which the company name appears | Per prompt, per system, monthly |
| Citation rate | Share of answers in which the company is given as a source | Per prompt, per system, monthly |
| Position in the answer | Where in the list the company is named | Only for lists, otherwise not reported |
| Source density | Number of independent sources about the company | Quarterly, checked manually |
| Classic position | Ranking in Google and Bing per core term | Monthly, desktop and mobile separately |
| Local visibility | Presence in the map pack and the state of the Google Business Profile | Monthly, per location |
Limits
What this method cannot do.
The section most providers leave out. Without it, a methodology is an advertising claim.
- We measure outputs, not causes. Why a system picks a source cannot be observed from outside. Any statement about causes is a reasoned assumption and is marked as one.
- Individual values fluctuate. There is a random component in answer generation. We assess quarters, not weeks, and never report monthly values in isolation.
- Models change without notice. An update can shift values overnight without anything changing at the company. We flag such breaks in the report.
- Our sample is your market, not the market. The figures apply to your prompt set and your competitors. They do not transfer to other industries.
- No measurement is free of judgement. Whether an ambiguous mention counts is decided by a person, following the rules above. We document borderline cases.
Verifiability
Everything you need to check it yourself.
Every client receives these five components. They belong to them, including after the contract ends.
- The complete prompt setAs a spreadsheet, with the origin noted for each question.
- The confirmed competitor listWith the date the client confirmed it.
- All raw data per measurementAnswer text, system, date, assessment, borderline cases flagged.
- The counting rules in the version that appliesChanges documented with dates.
- The analysis logicHow the six metrics are derived from the raw data, as a formula you can follow.
Read on
The definition, with the distinctions and the three levers.
Where the two overlap, evidenced with figures.
The same methodology, applied once and without obligation.
Frequently asked
Common questions about this.
Founder of meit, in digital marketing since 2006. Forbes Agency Council, LinkedIn Top Voice. He carries out every first analysis personally. Reachable directly at [email protected] – no account team in between, Monday to Friday, 09:00 to 18:00.
