Research: we measure it ourselves. And we also publish what speaks against us.
This is our research department, and it holds two kinds of pieces that must not be mixed up. The first kind is research: the AI Index, a measurement that records afresh every 25 days which companies a language model names, with the full file included as CSV and JSON under CC BY 4.0. No paywall and no registration. You can download it, recount it and arrive at a different outcome.
The second kind is analysis
Two arguments about how AI divides the market and what happens to authenticity. There is no dataset behind them and no measurement of our own was ever carried out for them. The figures that once accompanied them were withdrawn on 18 August 2026, and on 19 August 2026 the research trappings were removed as well: no methodology block, no registration number, no study. What remains is the argument, and you can pick it apart; that is the point.
- 01We measure with a setup you can repeat yourself: the same questions, the same model, the same settings.
- 02Where there is a file, we publish the whole file and not just the line that suits us. Where there is no file, the figure comes out.
- 03We leave our own mistakes standing, alongside what was wrong and why.
ongoing measurement
and remain visible
across nine sectors
model has named so far
Why we measure it ourselves
Almost everything you read about AI and marketing has been passed along. A figure from a presentation that traces back to a blog that in turn points to another blog, until nobody can point to the source any more. We wanted to know what actually happens, so we started measuring it ourselves: with a machine that puts the same customer questions to a language model as soon as the previous answer is more than 25 days old, and keeps every answer.
Measuring only makes sense if someone else can check it
That is why the full file comes with the ongoing measurement. You don't have to take our word for it, you can download it, count it yourself and arrive at a different outcome. That isn't a risk for us, that's the point.
On 18 August 2026 that rule fell hard on the two analyses
We then checked every figure on this site against its source. For those two, no file and no record exists, either with us or elsewhere. The figures have therefore been withdrawn, here and on the two pages themselves, with a note of what used to be there.
On 19 August 2026 a second step was taken, and it goes further than the numbers
Those two pieces still carried the form of research: a number, a registration mark, the word sector study and a block describing a method with a survey and eye tracking in it. Describing the method of research that was never carried out is a claim we cannot substantiate, even without figures. That form has therefore been removed. The two are now called analyses, because that is what they are: an argument with reasoning, not a measurement with a file. Anyone who teaches others to distrust figures nobody can substantiate should be the first to drop their own research trappings.
That rule also works against ourselves
Our own AI Index once carried figures that didn't hold up: three answer engines, 1,250 brands, thirty sectors, a national score of 41.2 with a trend line attached. None of it was actually running. We brought it back to what is really running and left the old story visible next to it, word for word, instead of quietly removing it. Read what was corrected on that page NL. Anyone who teaches others to distrust figures nobody can substantiate should be the first to withdraw their own.
Analysis 01
2025 · no datasetWhy an AI answer always mentions the same names
That's how many names per sector keep recurring in AI answers. A classic search result shows dozens of providers, an AI answer names a handful. There is no room in that answer for the middle of the pack, even if it isn't any worse. This used to state an exact number of 10 to 15; that was withdrawn on 18 August 2026.
This used to show two percentages from the same piece, about how often people use an AI tool to find a supplier and about the share of purchases that start in an AI conversation. Both were withdrawn on 18 August 2026. What you can measure yourself is in your own statistics: put the referral traffic from the known AI domains side by side with your organic traffic, month on month.
Analysis 02
2026 · no datasetThe post-text economy and the end of authenticity
That's how long a visitor aged 18 to 29 sticks around on a page without a dynamic, visual format. After that, they're gone. Text is no longer read as a source of information, but treated as the threshold in front of it. This used to state an exact time of 3.4 seconds and an exit percentage; both were withdrawn on 18 August 2026.
The measurement that doesn't stop
reg. O.001 · AI IndexThis is where the research begins. The two pieces above are analyses: reasoning without a dataset of its own. What follows below is the other kind, and the difference is precisely that there is a file behind it that you can download. What runs here is something that is never finished: the AI Index. Every sector has a fixed set of questions, every question goes back to the same model as soon as the previous answer is more than 25 days old, and the old answer is kept rather than overwritten. That is how history builds up. You see not only who is named now, but also who was there six months ago and isn't any more.
slide across for the older measurement points →
Nine sectors visible, seven measurement points per sector. Choose a row to highlight one.
The figures alongside it, all traceable in the file itself: 360 fixed customer questions, 40 per sector, nine sectors, seven measurement points per question and hence 2,520 recorded answers. Of those, 2,501 are valid. In 19 cases the model refused or asked a counter-question, and those rows remain with a marker instead of disappearing. Across all valid answers, the model named 3,807 different company names. The model queried is gpt-5-mini via the API, with fixed settings and no web access.
And the awkward side, because that belongs here too
For 184 of the 360 questions, Month -5 and Month -6 show exactly the same answer, simply because those questions don't yet have six real measurements behind them. So there are fewer real measurement points in the history than there are columns. We could have papered over that by leaving those columns empty. It is deliberately left in, with the explanation alongside it, because a dataset whose weak spot you don't know is useless.
Where the methodology and the full file are
Exactly how a measurement comes about, which normalisation rules apply to the company names and how much randomness sits in an answer, is fully written out in the methodology NL. The full file, including the literal answer text alongside the normalised list, is on the open data page NL as CSV and JSON. What the measurement means for a company and what you can practically do with it, we explain on the AI Index itself NL.
How we measure, and what you're welcome to check
reg. O.002 · accountabilityThe measurement setup is deliberately boring
A measurement is only worth something if the setup doesn't move along with the outcome. That's why everything is fixed for the ongoing measurement: one model, one prompt, one setting, a question list that doesn't change and a repetition on a fixed schedule. The questions are full customer sentences and not keywords, so the answer resembles what a real user gets back. The prompt is on the open data page word for word, without shortening, so you can redo the query yourself.
This section is only about the AI Index, and that is deliberate. For the two analyses there is no measurement setup to account for, because there isn't one. So those two don't state how many participants there were or which group they came from, because that setup never existed. Anyone who wants to check a claim from those two pieces can only do so by testing it themselves: in their own statistics, in their own advertising account, or by putting the question to a language model themselves. We say there is no file with them because we don't have one.
We publish the raw data, not just the conclusion
A chart without a file behind it is an opinion with formatting. That's why the file contains the raw answer text alongside every normalised company list, together with the sector, the question number, the measurement point, the date, the model and a column stating whether the measurement was valid. So you can not only check our count, you can redo it with different choices. If we merge two spellings that you'd rather keep separate, you can see exactly where that happened.
You may do with it what you like
The file marked as open data is under Creative Commons Attribution 4.0 International, CC BY 4.0 for short. Copying, adapting and publishing is allowed, including commercially, as long as you credit the source, link to the licence, name the dataset's last measurement point and state what you changed. Not covered by that licence are the trademarks and logos of the companies named, personal data, third-party material and the full output or terms of the model provider. Crediting us, moreover, does not mean we endorse your reuse or that a named company endorses us.
Prove us wrong
If you find a mistake, we want to hear about it. Is a company name spelled wrong, is there a name in there that doesn't exist, or is a count wrong? Email sales@theseo.nl with the word correctie in the subject line, and state which sector and which question number it concerns, which edition of the dataset you used together with the date of your download, and what your judgement is based on. We record receipt, assessment and decision. If a mistake turns out to be in our processing, we correct it and the dataset gets a new version with a line in the version history. Old editions are not silently overwritten.
One thing we do not correct: what the model answered. If the model doesn't name a company, that stays as it is in the data, even if that company feels it should be included. The measurement records what the model says, not what is true. A name ranking high in a list is a count of model output, not a recommendation from us; a name ranking low or being absent is likewise not a judgement.
What we're currently measuring
reg. O.003 · state of playBelow is the honest state of play: what's running today, what has been built but doesn't measure anything yet, and what we deliberately don't measure. There are no dates attached. Promising a date for research that depends on keys, budget and permission is exactly the kind of figure we warn others about.
- The AI Index keeps running. The process keeps picking up whichever questions are due, so the file grows on its own. Every edition gets its own latest measurement point.
- The history is filling in. For the 184 questions where Month -5 and Month -6 are still identical, a real measurement gets added each round. Those columns become true on their own.
- Corrections are processed. Every report is logged with receipt, assessment and decision, and a genuine mistake leads to a new dataset version with a line in the version history.
- The same question to multiple models. The measurement point has been built and tested, but there is no key yet for the other models. As long as nothing is being measured, no model comparison appears on the site and we don't name an outcome.
- A version number in the metadata. Since edition 2026.1, the published files carry a dataset ID, a version number and a publication date in the meta block. Cite that version, along with the date of your download.
- Adding a sector. Possible, but drawing up a question list that matches how customers in that sector really search is manual work. One only gets added to the list once it's actually being measured.
- No score from 0 to 100. No national average and no trend figure, because there is no calculation that produces one. What can be extracted is a mention share, and that is a count.
- No traffic loss. There is no link between this measurement and what you see in Search Console. Anyone who makes that promise is making it up.
- No tracking people outside our own site. That clashes with privacy law and with the reason this research exists. The growth is in open, aggregated data, not in watching over people's shoulders.
What you can do with this
This page isn't here to sell you anything. The two analyses are free to read without registration, the AI Index dataset is free to download and reuse, and the measurement setup is open. Use it in your own presentation, in an article, in a lecture slide, or to show your management that something is shifting. Take along the distinction that keeps recurring throughout this page: you can cite the dataset as a measurement, you cite the two analyses as our viewpoint.
If you want to know what the figures mean for your sector, a conversation is the shortest route. The answer differs per sector, and the dataset only says what the model said, not what you should do about it.
That same measurement for your own company is something we do provide
This page stays free and without registration. That measurement is the service having your AI visibility measured and improved: we put the questions to a language model that are genuinely asked in your sector, record which companies it names and repeat that, so you see whether you're included and for which questions you're not.
What that costs is on that page itself and on pricing NL, and it is deliberately separate from everything else. If you'd rather start with something that costs nothing, take the free scan, or grab one of the skills you can use to run the measurement work yourself.
The distinction on this page applies there too. The AI Index is a measurement and we publish it openly. What we then do for a client with such an outcome is work, not measurement, and that comes with a price rather than a dataset.