How Bersyn measures

A buyer market is one category, in one country, in one language — for example accounting software, in Iceland, in Icelandic. Bersyn measures the market once, and every company in it reads exactly the same recorded answers.

Country and language are matched exactly. A measurement taken in en-US never stands in for one in is-IS, because the answers differ and a buyer in one market is not a buyer in the other.

Each market has a fixed, versioned pack of buyer questions — the questions someone actually types when deciding what to use. The pack is frozen for a measurement, and the wording is recorded verbatim, because with these systems the wording is part of the instrument.

Questions are written neutrally. A question that names a company invites an answer about that company, so anchored phrasing is kept out of the shared pack.

Every question is asked of the same four assistants — ChatGPT, Claude, Gemini and Perplexity — through their APIs, with the model, the request configuration and the web-search mechanism recorded on every answer.

Results are reported one assistant at a time and are never blended. The assistants frequently disagree, and that disagreement is a finding rather than something to average away.

Each question is asked more than once per assistant. The number of repeats is recorded with the measurement.

The same assistant, asked the same question, does not always answer the same way. That is why a result reads 2 of 3 rather than yes or no: the assistant named the company in two of the three recorded answers. Bersyn shows that split rather than rounding it into a verdict.

One recorded answer is one buyer question, asked of one assistant, on one repeat. The full text of every answer is stored, along with the model that produced it, when it was produced, and whether it was usable.

A measurement of twelve questions across four assistants with three repeats is 144 recorded answers. Every count you see is a count of those.

Named in is how many recorded answers named a company, out of how many were recorded — always with its denominator, always per assistant. Companies are matched against a curated list of known entities for that market, including the name variants assistants actually use.

Named instead is which other companies appeared in those same recorded answers, with the same denominator. It is not a rank and carries no suggestion that any of them displaced anybody.

A company named in zero answers is a result, not a missing measurement. 0 of 144 is one of the most useful things this product can tell you.

Bersyn records whether an assistant retrieved from the web before answering, and which domains it referenced.

🔴 Being named, searching the web, and citing a source are three separate observations. They are recorded separately, reported separately, and never combined into one number. Bersyn does not claim that searching produced a recommendation, or that being cited caused a company to be named.

Retrieval also means different things for different assistants. For some it is a choice the model makes each time; for others it is simply how the product works, and there was never an occasion on which it might have declined.

Every count traces back to the answers it came from, and those answers are readable. If Bersyn says a company was named in two of twelve answers to a question, the twelve answers are there to read.

Each measurement also records the question pack version, the contract it was run under, the assistants and models used, and the number of repeats — so any figure can be placed and checked later.

Two measurements can be compared when they are methodologically comparable: the same market, the same question wording, the same assistants and configuration, the same number of repeats.

🔴 A difference between two measurements is an observed difference. It is not, by itself, evidence that anything caused it. We have measured one market twice, days apart, with every recorded setting identical and nothing done to any company in between — and counts moved on their own.

So Bersyn will tell you what it recorded, and what changed since a comparable measurement. It will not tell you that something you published caused the change. That is a different claim, and it needs different evidence.

There is no score, no index, no grade and no rank. Bersyn previously published composite figures — a blended visibility score and percentage summaries — and withdrew them in August 2026 after testing the method against itself: a single number hid four assistants that disagreed, and the percentages implied a precision the measurement did not have.

It is not a rank tracker, not an SEO tool, and not a way to manipulate what AI systems say. It is a record of what they answered, with the answers attached.