In short
- Checkable
- Every line points at something on this site. A changelog you cannot verify is a press release with dates on it.
- The mistakes too
- “Corrected” is one of the three types. A product log in which nothing was ever wrong is not credible — and it is not true either.
- No method
- What changed is here; how it was built is not. That is not secrecy but the same rule that applies everywhere on this site.
- No “coming soon”
- This is the only page that is purely about the past. What is coming sits on the product pages; what exists sits here.
Which changes are listed here, and in what order?
Three types, not eight. New, improved, corrected — and that third one is the most important. A product log in which nothing was ever wrong reads like a press release, and that is exactly what a changelog should not be.
| Date | Type | What changed |
|---|---|---|
| 2026-08-11 | New | The study pages now number 11. The newest sets side by side the three studies that underpin our blind jury, with their address and their limit: a model judging its own work approves its own mistakes more than 50% more often, a smarter model does not help, and even the name attached to a text shifts the verdict. Of the 10 sources in our register, 9 are now citable. |
| 2026-08-11 | Corrected | The argument underpinning our blind jury rested on three studies that appeared nowhere in our source register — while their figures had been on the site for months. All three have now been checked at the source and added; the counts above and below move with them. One sentence turned out to summarise the study more strongly than it allows and has been corrected: a name attached to a text shifts the verdict in both directions, not only upwards. |
| 2026-08-11 | New | Three study pages were added that lean on our own calculation rather than an external publication: how much of the market a measurement covers, a case where we got it wrong ourselves, and the Gartner prediction that came due this year. |
| 2026-08-11 | Corrected | On a number of pages, quotation marks and non-breaking spaces appeared on screen as code instead of as characters. Visible to every visitor, and found by no check at all — the delivered page was, after all, flawless. A check that does look for it is now in place. |
| 2026-08-11 | Corrected | One page quoted a percentage in Dutch that did not appear in the French and English versions and had no source anywhere. The figure is gone; the claim underneath is correct and remains, now with the study it rests on beside it. |
| 2026-08-10 | New | The MCP endpoint is in place: an address where an AI agent fetches your company data directly instead of crawling your website. 12 tools, 8 of which answer without anything being connected. |
| 2026-08-10 | Corrected | The four main pages carried no machine-readable data blocks at all, and the check meant to catch that only ran over the language folders. Both are fixed — the pages carry them now, and the check looks at the whole site. |
| 2026-08-10 | Improved | Every figure we take over from someone else has since sat in a register with the size of the measurement, the address of the publication and the date on which we found that figure back in the source text itself. Today that is 9 out of 10; whatever lacks that date does not appear on the site. |
| 2026-08-09 | New | German added as the fourth language. The site has run in 4 languages since then; the dashboard counts more, and a report is written in the client's language. |
| 2026-08-08 | Improved | The 13 agents sit on one page instead of one page each. For each, what it delivers and whether it may act without asking — and the agents without a working connection yet are listed separately and labelled as such. |
| 2026-08-07 | New | 8 free tools, no sign-up. They give a real answer to one question and do not try to push you to a form halfway through. |
| 2026-08-06 | New | A glossary of 92 terms from this field, each with a definition that stands on its own. Anyone meeting a term they do not know does not have to wade through a sales page first. |
Why is this public?
Because it is the cheapest proof there is. Everybody in this field says they move fast; a dated list of what actually shifted costs nothing and cannot be faked without being caught — every line points at a page you can open.
And because we recommend this kind of page to clients. A dated, factual page that changes regularly is one of the few things we can say with any confidence that an AI likes to fetch. It would be odd to advise that and not have one ourselves.
What does not go on here?
No announcements. What is coming belongs on the product pages, alongside the rest of the story; here you find only what already exists. That makes this the only page on the site that is purely about the past, and that is deliberate — a changelog with a “coming soon” in it is a wish list.
And no internal detail. What changed is here; how it was made is not. That is the same rule that applies across this site — we show what a measurement produces, not how it is put together — and it holds just as firmly here, even though the log would look busier for it.
Frequently asked questions
How often does this page change?
As often as something changes that you can check, and no more often than that. Padding a changelog to look lively is the same kind of untruth as a figure without a source: it is not accurate and it is meant to impress.
Why are mistakes listed?
Because they happened. And because we ask clients to believe us when we say a fix did not work — you only believe that from someone who says it about themselves too. One of those corrections got its own study page, because it was more instructive than a line in a log.
Are all changes listed here?
No, and it would add nothing. What is here is what a reader can see or check. Everything else — how something is built, which parts moved — does not belong on a public page, for the same reason that we describe our measurement method without writing it out.
Read on
Free, no account and no card.