Nuggets and fool’s gold: the figures we took off our own site.

This site publishes a register of every figure it uses, and of every figure it removed. That second list is the useful one, and it is the one nobody else keeps.

28 figures have been struck or corrected so far, and the register lists every one. The ones that failed on their sourcing did not fail randomly. They failed in four recognisable ways, and once you can name the four you can test almost any number you are shown.

One: the real source, stretched

The one we were proudest of catching was a 1,445% one-year surge in demand for multi-agent systems, credited to Gartner, on our own homepage. We traced it to a press release by Belitsoft, a software development company, could find no Gartner publication carrying it, and wrote it up here as the worst failure on the list: a vendor’s number wearing an analyst’s name.

It was not. Re-checking our own register on 24 September 2026, we found the Gartner article, published on 18 December 2025: “Gartner reports a 1,445% surge in MAS inquiries from Q1 2024 to Q2 2025.” The name was right. What was wrong was the sentence: inquiries to Gartner’s analysts had become demand, and five quarters had become a year. Belitsoft had quoted it correctly; we had blamed the messenger because our search came back empty.

That makes it the commonest failure on the register rather than the rarest. The source is real, and the sentence quoting it is bigger than the source: a base widened, a population swapped, a period stretched. Almost every correction we have made since is this kind.

Two: one person, promoted

A widely repeated figure held that one in seven dental appointments is a no-show. It traces to a single NHS dentist describing his own practice over one year, in a trade interview. No sample. No method. Not a British Dental Association study.

That is a perfectly good anecdote and we would happily quote it as one. It is not a rate. The same interview is also the origin of a £56,000 annual cost for that one practice — again true of him, again not a national figure.

The tell is grammatical. An anecdote that has been promoted loses its owner: "one dentist found" becomes "one in seven appointments". If a statistic has no subject, go and find who it originally belonged to.

Three: nothing at the end of the trail

Three of them simply have no source. £39 per missed UK appointment: none found. 40% of bookings made after hours: none found. £200,000 a month in salon no-show cost: none found. A fourth, 11 hours a week lost to admin by letting agents, was struck the same way — the nearest real source, Goodlord's industry survey, says only that one in five letting agents felt they were spending too much time on admin, which is a sentiment rather than an hours figure.

These circulate because each repetition looks like corroboration. Five blog posts citing each other is one unsourced claim with four echoes, and the only way to tell is to follow every link to the end. Most of the time the end is a slide.

Four: a range that is not a number

The subtlest failure. We wanted a figure for missed calls in the trades — one in five was the claim. What we found was that the published range for unanswered UK business calls runs from roughly half to 94%, and that every source was a company selling call answering. Only one stated a method, and it measured evening calls to sales and letting agencies, not the trades.

A spread that wide is not a measurement with uncertainty. It is a marketing range, and picking a value from it — even a conservative one — launders a sales claim into a fact. So we did not pick one. We built a calculator that carries none of those figures instead, where you enter your own number and it can tell you to buy nothing.

What we did with the two that could be saved

Striking a figure is not the only move. Two were replaced with something stronger.

The claim that deposits cut no-shows by 55–65% had no source. But two randomised controlled trials, published in PLOS ONE, tested a cheaper intervention on hospital outpatients, and in the first, 10,111 patients, stating the appointment's specific cost in the reminder message cut did-not-attend rates from 11.1% to 8.4%. That is a better finding than the one we lost, it is properly sourced, and it argues against buying software rather than for it. It became a note of its own.

What this costs, and why we do it anyway

Every struck figure is an argument that got weaker. The dental page we intended to write lost all three of its numbers. The trades page lost both of its. That is a real cost and we are not going to pretend it was painless.

But a firm that publishes the figures it removed is making a checkable claim about the ones it kept, and a firm that publishes only its best numbers is making no claim at all. Every remaining figure on this site names its source, its date and how much weight it carries, and the whole register is public at what we checked. If one has gone stale before we got to it, the page says so and asks you to tell us.

Apply the four questions to the next AI vendor deck you are shown. In our experience roughly half the numbers in one will not survive the first question, and the ones that fail are almost always the ones doing the persuading. We ran that exercise on five published AI claims, and on forty-five citations, with much the same result.

What this does not tell you

Whether a well-sourced figure is still right. The four tests catch stretched sources, promoted anecdotes, missing sources and marketing ranges; they do not catch a source that has been overtaken. The 12–20% on the register described a genuine benchmark and simply had its number wrong — the published figure is 30% — and every one of the four tests would have passed it, because the source was real.

On 24 September 2026 we opened every source this site had cited without a link and read each against the sentence citing it. Five of our sentences had the document right and its scope wrong.

A figure for sales and letting agencies was given as the figure for every firm the study rang. A US survey was quoted as if it covered every buyer. A vendor’s numbers for a typical secured loan case were stretched to every case, and given a country the report never names. An average agency was called a typical one. The fifth ran the other way: we had told readers that a well-documented study stated no method. Each is corrected where it appeared and listed on the register.

The same day we re-read every linked source against its sentence, and then looked for anything newer than each. That found the other kind of error: sources that were right when we read them and had since been overtaken — a consultation answered eight months before we wrote about it, a start date that had already passed, a leaderboard entry we could not find again. Those are on the register too, and they are why every source here carries the date it is next re-read.

Nor is the list complete for anyone else. The taxonomy comes from the figures that failed on our own site, so it describes our mistakes; another site’s would add modes we never made.