DOC.02 How we measure it
We do not guess and we do not ask a panel. We put the question to the model itself and keep the answer. Those same systems may use our own public content in return, on the two conditions set out in our AI policy. This is the whole method, and it has run since March 2026 as the AI Index.
There is a question list per sector. Those are not keywords but whole sentences, the way somebody would put them to an assistant, and that list is handwork: a cleaning company gets different questions from a car garage. At the moment nine sectors are running with 360 questions between them. Per question the model gets an instruction that does not change: name ten Dutch companies, names only, no explanation. That word Dutch is not a detail: the index that runs today is a measurement of the Dutch market, and what it says about the British or Irish market is nothing.
A question becomes due again when its previous answer is more than 14 days old. We publish only after a complete round has been checked. Each question retains its current dated answer and six earlier answers without individual dates. These are seven stored answers, not a verified monthly time series.
We are questioning a model here, not the market. Today that is gpt-5-mini from OpenAI, through the API. That means we say nothing about Gemini or Perplexity, even though you might expect that of an index like this. We write that down rather than leaving it out, and the full account is in the methodology.
The whole dataset is open. Every question, every measuring moment and every answer can be downloaded as CSV and JSON under CC BY 4.0 through the open data, with the rankings per sector beside them, so you can check every count yourself. A measurement you are not allowed to check is a claim.
FIG.03: Where the number 2,520 comes fromsheet 3/7
Three multiplications, and a window that explains why the number does not keep growing
The paragraphs above name four numbers in four places: nine sectors, 360 questions, seven measuring moments and 2,520 recorded answers. Separately they look like loose facts. As a chain you can see it is a sum, and which link you would have to change if you wanted to know what happens with a tenth sector.
ai index · how the count is builtthe numbers appear in the text above
9sectors with a question list of their own
×
40questions per sector, because 360 divided by nine
=
360questions running at the moment
×
7measuring moments kept per question
=
2,520recorded answers, the number from the text
The window of seven, per question
new answer
now
-1
-2
-3
-4
-5
-6 drops out
The window shifts; the file does not grow. Once the current answer is more than 14 days old, the question can be measured again and its previous answer moves into history. Nine sectors, forty questions and seven stored answers give 2,520 answer records. Only the current answer round is dated.
FIG.04: How an answer comes aboutsheet 4/7
Ten places filling up, and then a question that waits
The paragraph above describes what happens per question: the model gets an instruction that does not change, gives ten names back, and that answer is stored. Below that happens too, again and again, because that is what separates an index from a snapshot. There are deliberately no names in the boxes: which companies the model names is a measurement, and that belongs in the open data and not in a drawing.
ai-index/answermechanism, not a measurement
the instruction that does not changeName ten Dutch companies, names only, no explanation.
- 0 of 10
- 1 of 10
- 2 of 10
- 3 of 10
- 4 of 10
- 5 of 10
- 6 of 10
- 7 of 10
- 8 of 10
- 9 of 10
- 10 of 10
measure again once the answer is more than 14 days oldpublish after checking the complete round
The model is asked for ten names. This illustration explains the method; it is not a live result. Each question becomes due when its previous answer is more than 14 days old. Only a complete, checked round becomes the latest public edition.
