How we measure
Five questions, five AI providers, repeated five times. No more — and nothing less.
AI assistants give a slightly different answer every time. One screenshot therefore proves nothing. So NorthAnswer doesn't measure once, but twenty-five times per provider — 125 planned answers in total — with questions that are locked in before the measurement starts. You get an outcome you can check for yourself.
Step 1: your five customer questions.
Together we settle on five questions your customers really ask an AI assistant. For example: “Which accountant in Leiden is good for sole traders?” No questions about your brand name — the AI usually knows that. Questions a customer would use to find you. The five questions are frozen and given a fingerprint (hash), so nobody can change them afterwards.
Step 2: five official API measurements.
We ask each question five times through the official APIs of five AI providers — the API is the connection software uses to talk to an AI, without your personal history or settings — OpenAI, Anthropic, Google Gemini, Perplexity and xAI. Each time in a fresh, empty conversation, without memory. Where the API allows it, we measure two lanes separately: the answer from the model's own knowledge, and the answer with web search. Note: these are API measurements. What you see yourself in the ChatGPT app or another consumer app can differ — your history, location and settings play a part there. That is exactly why we measure through the API: it is the same for everyone and repeatable.
Step 3: every answer recorded with an evidence ID.
Every answer is stored in full, with model, date, time and a unique evidence ID. In the report, every figure and every quote points to such an ID. On request you receive the complete answer file. And whatever was not measured is stated just as literally: “not measured”.
Step 4: counting, not interpreting.
We count in how many of the answers your business is mentioned by name, and which other names come up. No overall score, no average, no “AI score”. Certainty comes in two layers: first per provider over five repetitions, then one main label per question out of six fixed labels. Every main label carries its count, and a provider that was not measured never counts as a zero. The six labels and what they mean are below.
Step 5: three pieces of advice — and what we don't know.
You get at most three pieces of improvement advice, ranked by substantiated priority: with the reason, how feasible it is and what has to come first. Each one points to the finding it rests on. A measurement of AI answers usually proves no cause, and what a piece of advice will yield we don't know in advance; only a re-measurement shows that — the same five questions, the same method, run again later — and even that shows change, not proven cause. A re-measurement is not part of the pilot; it belongs to the Build package and Maintenance. We don't promise the numbers will go up. We do promise you'll see them honestly.
The six certainty labels.
Five providers can disagree. That is why certainty comes in two layers, and why a difference between providers always outranks a majority. Every statement holds for this measurement round and no further.
Layer 1 — per provider, per question. We count over that provider's five repetitions; we always count out of five. Named four or five times is consistently visible, never named is consistently absent, anything in between is a mixed picture. Fewer than five usable answers is called insufficient coverage.
Layer 2 — the main label per question. The question then gets one main label, over fully measured providers only (five usable answers). The rules run from top to bottom and the first one that fits wins. If no provider was measured at all, it is not measured. Fewer than four fully measured and it is insufficient coverage. If at least one AI mentions you consistently and at least one other consistently does not, it is a split picture — that rule comes before any majority, because otherwise 3–2 would be summarised as consistently visible while two AIs never mention you at all. After that, three or more providers consistently visible makes it consistently visible, and three or more consistently absent makes it consistently absent. Everything else is a mixed picture.
Every main label carries its count. For example: “split picture — mostly visible: 4 of 5 consistently visible, 1 consistently absent.” That way you can see straight away where the label comes from. A provider that was not measured never counts as a zero: the report says so in as many words.
- consistently visible
- This AI mentioned you in four or five of its five answers in this measurement round.
- consistently absent
- This AI mentioned you in none of its five answers in this measurement round.
- split picture
- At least one AI mentioned you consistently, at least one other consistently did not.
- mixed picture
- You turned up sometimes and sometimes not; no fixed pattern in this measurement round.
- insufficient coverage
- Too little was measured to say anything about this question.
- not measured
- Nothing was measured here. This is never a zero.
The same six names, the same order and the same explanation appear in the report and on the report page. One source, no second truth.
What this is — and what it is not.
It is
- a snapshot on one date
- checkable down to the individual answer
- the same method at every re-measurement
- your own customer questions
It is not
- a prediction
- a ranking
- a judgement on the quality of other businesses
- identical to what one person sees in an app
How it works
Measure. Explain. Improve.
A baseline: first know where you stand, then change something — and whoever re-measures later compares against this fixed point. One measurement, three steps, the same 125 answers. Watch it run and click any point.
Example — not a real measurement.
The same 125 points, rearranged under the six certainty labels. Five repetitions per question show how often you are mentioned — not why, and not what tomorrow brings.
Each point is one answer. Click it.
Example measurement with invented answers and a fictional client, meant to show how a measurement works. The figures are not a result.
The same three steps, in text:
Measure
We ask the five buying questions your customers literally ask. Each one five times: AI answers can differ. That is why we repeat each question five times per provider, so you can see how often your business is mentioned. 5 questions × 5 runs × 5 providers = 125 planned answers — everything recorded, everything traceable.
Explain
You get the lost-customer story: for every question, who the AI recommended, the exact arguments it used, and where your business appeared — or quietly vanished.
Improve
At most three pieces of advice for Monday morning, ranked by substantiated priority: with the reason, how feasible it is and what has to come first. Each one says what it rests on and what we don't know — including what it will yield; only a re-measurement shows that, and even that shows change, not proven cause.
What you can check
- Every figure clicks through125 answers, each with an ID
- Reproduciblemeasurement-plan hash on page 1
- Public questions onlyzero uploads of company data
- Independent5 systems, 0 partner ties
- Fixed pilot price€197 for the pilot; follow-up separate
- No ranking promisemeasurement, not placement
- Dutch, with an addressKvK 62675567 · GDPR
Timeline
From request to report in four steps.
No weeks-long intake. The ten working days start at receipt of your payment and your website address: by working day 10 the report is with you — with evidence behind every point.
BY WORKING DAY 10.
- RequestYou requestThe form starts no measurement. Within one working day you get a personal reply: if your market fits, the measurement plan follows — and only then an invoice.
- After receipt of payment125 answers5 questions × 5 runs × 5 providers. Same questions, same providers, fixed in the measurement plan with the hash on page 1.
- We traceEvery point to evidenceEvery statement in the report has the literal question and the literal answer as evidence, with date and time. What was not measured is marked as such.
- By working day 10Report + 30-min walkthroughPDF with source map and three improvement actions with a certainty label, plus a 30-minute personal walkthrough.
Four questions we hear often.
Why not just look in ChatGPT myself?
You can, and the knowledge base explains how. But one look in one app says little: the answer changes per run, per person and per setting. Five repetitions per question, per provider, through the API show how often your business is mentioned — not why, and not what tomorrow brings.
Do you also measure Google's AI overviews?
No. We measure through the official APIs of five providers. What a provider shows in its own search engine or app falls outside the measurement and appears in the report as “not measured”.
Can I verify the measurement?
Yes. Every figure points to an evidence ID; on request you receive the complete answer file with question, answer, model, date and time.
What if my business isn't mentioned anywhere?
Then that is what it says: 0 out of 125. That is a finding, not an error — and usually the clearest starting point for the three actions.