Every number on this site, where it came from, and when we last checked it.

Claimeda one-year surge in demand for multi-agent systems, credited on our own homepage to Gartner

WhyRemoved, and then explained wrongly. On 7 September we called it a false attribution: a software company’s press release carrying Gartner’s name, with no Gartner publication behind it. There is one. Gartner’s own article, published 18 December 2025, reports “a 1,445% surge in MAS inquiries from Q1 2024 to Q2 2025”: inquiries to its analysts, over five quarters. The name was right. We were wrong twice: our sentence turned inquiries into demand and five quarters into a year, and our check then failed to find Gartner’s page and blamed Belitsoft, whose release had quoted Gartner correctly. Corrected on 24 September 2026, by re-checking this register.

Claimedstandard AI personalisation performing worse than sending nothing at all, across 500,000+ cold emails

WhyNo source found. The largest public dataset we could find (Hunter’s 31 million emails) reports the opposite: two custom attributes lift reply rate from 3.6% to 5.6%. A figure we could not source, contradicted by one we could.

Claimedthe same personalisation performing better than baseline after a deliverability rewrite pass

WhyNo source found. It appeared in our own plan document with no citation, which is how an internal guess becomes a public claim.

Claimedone human running two AI seats booking more meetings per dollar than pure AI

WhyNo source found. The argument it was making survives (Hunter’s data shows manually edited emails beating fully automated ones by 18%) but that is a smaller, truer claim with 31 million emails behind it.

Claimedset in the typography of a statistic beside three real figures

WhyNot a statistic at all: a description of the offer, dressed as evidence. It belongs in the copy, and that is where it went.

Claimedskill bars for Development, Design, Marketing and Consulting

WhySelf-scored, unfalsifiable, and inherited from the WordPress template. At least they were honestly ours.

Claimedthe best computer-use score on OSWorld

WhyNot removed: corrected. The real figure was in the mid-80s against a 72.36% human baseline, so the sentence had quietly stopped making its own point: agents now beat people at bounded desktop tasks. Being out of date made the claim too modest, which is the failure mode nobody checks for. (We recorded it as 86.1%; on 24 September 2026 the leaderboard showed no such entry and a best model at 86.0%, and the figure was corrected again.)

Claimedthe success rate of cross-application agent workflows, as written in our own plan

WhyNot removed: corrected. The source it pointed at, the TheAgentCompany benchmark from Carnegie Mellon, reports the most competitive agent completing 30% of long-horizon tasks autonomously. The direction was right and the number was not, and a number that is only pointing the right way is still a wrong number. (The paper’s 30% has since been overtaken on the benchmark’s own leaderboard, at 42.9%; since 24 September 2026 the site quotes the leaderboards and gives the date it read them.)

Claimeda week lost to admin by letting agents, reserved for our back-office pages

WhyNot found. 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: a sentiment, not an hours figure.

Claimeddental appointments that are a no-show, reserved for the dental page

WhyAn anecdote promoted to a statistic. It traces to one NHS dentist describing his own practice over one year in a trade interview: no sample, no method, not a British Dental Association study. Usable as his story, labelled as his; not as a rate.

Claimeda month lost to no-shows by UK salons

WhyNot found, anywhere we could reach. A number in wide circulation that traces to nothing is one claim with many echoes.

Claimedthe reduction in no-shows from taking a deposit

WhyNot found, and replaced with something stronger. The first of two randomised controlled trials in PLOS ONE, among 10,111 hospital outpatients, found that stating an appointment’s cost in the reminder cut did-not-attend rates from 11.1% to 8.4%. A cheaper intervention, properly measured, arguing against buying software rather than for it.

Claimedcalls missed, bookings made after hours, cost per missed appointment: the specified defaults for our missed-call calculator

WhyEvery published figure we found came from a company selling call answering, and they disagree with each other by a factor of two. So the calculator carries none of them: it opens on a worked example of our own, labelled as one, and asks you to replace every number with yours.

Claimedpages combining text, images and structure being cited that much more often than text alone: the figure our thumbnails and diagrams were first justified with

WhyNo checkable method. We had it second-hand from an aggregator that itself marked it unverified; the origin, a December 2025 vendor blog, asserts it from 15,847 AI Overview results and never says how the multiplier was computed. The thumbnails stay, because they never needed a statistic: they are machine-readable and they are not stock photography. The number goes.

Claimedof calls to all 3,047 UK property firms rung between 6pm and 9pm going unanswered: on our evening-calls note, the letting agents page and the missed-call calculator

WhyNot removed: corrected. 94.1% is the study’s figure for sales and letting agencies. Across the whole sample of 3,047 numbers, deduplicated from 3,169 branch records, it was 93.8%. Close enough to change no argument, and wrong all the same: the right document, the wrong row.

Claimedof decision makers bothered by AI-written email, quoted beside the 31 million emails as if it came from them

WhyNot removed: corrected. Hunter’s page gives it as US-based decision makers, from a survey separate from the email dataset, and states no sample size. It is still worth quoting; it supports a smaller claim than the one we were making with it.

ClaimedNivo’s research described as covering UK lenders and brokers, with its message rounds and hours given as the cost of every case

WhyNot removed: corrected. As Mortgage Solutions reported it, the research covers more than 70 specialist lenders and brokers and names no country; the 15 to 20 message rounds, five-plus people and five-plus hours describe a typical secured loan case. We had added a country and widened the scope, wherever we quoted it.

Claimedmanaged properties and employees at a typical UK letting agency

WhyNot removed: corrected. Goodlord’s 2025 report gives both as the average. An average and a typical agency are different firms when a few large ones sit in the sample, and by the same report more than half of agencies employ five people or fewer.

Claimeddescribed on our plumbing and heating page as a vendor figure with no method stated

WhyNot removed: corrected, and this error ran against the source. The study states its sample, its calling window, its dates and its commercial interest. What it measured was evening calls to sales and letting agencies, which is the real reason it cannot stand in for a plumber’s phone; the chart now says that instead.

Claimedas the best score on the OSWorld desktop benchmark, credited to a named model: on our homepage, three notes, the financial services page and the triage tool

WhyNot removed: corrected. Read in a browser on 24 September, the OSWorld-Verified leaderboard shows no such entry: its best model scores 86.0% and its best agentic framework 90.2%. We cannot reproduce our own figure, which is the worst kind of error to find on a page about checking. Everywhere it stood now says 86.0%, read that day.

Claimedof long-horizon office work finished by the best AI agent, given as current: on our homepage, three notes and the financial services page

WhyNot removed: replaced. It is TheAgentCompany paper’s result from December 2024, and the benchmark’s own leaderboard now shows 42.9%. The long-horizon figure we quote is now OSWorld 2.0’s best, 31.4% of 108 tasks, from the team behind the first OSWorld: a newer benchmark read on the day, rather than an old one presented as live.

ClaimedGOV.UK designated standards lists, which our standards checker, the rules matcher and two notes said we had read as if thirteen were all there are, with five of the thirteen list links going to pages GOV.UK had renamed

WhyNot removed: scoped and repaired. GOV.UK publishes designated standards for 39 product areas; we carry the thirteen our ingest reads, which hold 3,517 standards, and the pages now say so and name the other 26. The five broken links sent every EMC, low voltage, PPE, pyrotechnics and conformity assessment result to a missing page. Every list link now comes from GOV.UK’s own index, and a test fails any that does not.

Claimedcalled excluded from CE recognition in Great Britain, and “where the deadlines still live”

WhyNot removed: corrected. The GOV.UK guide we cited lists them as having specific rules, in its section on where a UKCA mark may be placed. Its sector guidance shows construction products and cableways still accept CE, and medical devices is the only one with a CE deadline. We had also cited the guide by a title GOV.UK replaced in March.

Claimedas URL accessibility’s score among AI citation factors: a score out of ten, on our AI-search note and the visibility checker

WhyNot removed: corrected. Zyppy’s synthesis scores twenty-three factors from 9.5 down to 2 and states no scale beyond that. The “out of ten” was ours. The ranking stands; the denominator does not.

Claimedpages cited by AI assistants, including AI Overviews, measured against Google’s top ten

WhyNot removed: corrected. Ahrefs’ 1,064-day average covers ChatGPT, Perplexity, Copilot and Gemini, not AI Overviews, and it is compared with organic search results rather than a top ten.

Claimedof AI citations going to syndicated press releases, attributed to BuzzStream’s 4-million-citation dataset

WhyNot removed: corrected. The 0.04% comes from a separate BuzzStream study of news citations, whose size the page does not give: syndicated press releases were 0.32% of its news citations and 0.04% of that whole dataset.

Claimedfor the landlord database and the ombudsman under the Renters’ Rights Act, on our clocks note

WhyCorrected. It was the government’s phrasing in April. On 9 September, two days before we published, the department fixed the dates (registration opens on 15 December 2026 in the West Midlands, and every letting landlord must be registered by 14 November 2027) and its November 2025 roadmap already expected ombudsman membership in 2028. We missed both.

Claimedfor FCA guidance on AI under its existing rules: the Treasury Committee’s request, quoted on our Mills Review note without the answer

WhyCorrected. The FCA’s reply, published on 16 April 2026, five months before our note, did not promise the guidance: it offered more examples of good and poor practice and an evaluation of its AI Live Testing towards the end of 2026 or early 2027. Quoting the request alone made the guidance read as coming.

Claimedbusy, said of the worked example on the waiting-time calculator

WhyCorrected. Forty ninety-minute jobs are sixty hours of work against thirty available, which is 200%. The calculator always showed 200%; the sentence above it said 100%. Found while moving the site in October 2026.

Claimedcredited to Gartner as “a figure Gartner never published”, on the financial services page

WhyCorrected. The 24 September correction above reached the field note and this register, but not one sentence on the financial services page, which still repeated our first, wrong finding. Gartner did publish the figure, as a count of inquiries to its analysts; the sentence now says so. Found while moving the site in October 2026.

Claimedradio equipment standards said to stop counting on 15 November 2026, on our designated standards note

WhyCorrected. Six end on 15 November. Sixteen is every radio equipment deadline still ahead on that list, running to May 2028. Found on 5 October 2026, when we read the thirteen lists again for the note’s scheduled re-check.