72% of small businesses said GOV.UK guidance did not help them. Then the page changed overnight

The short version

• The survey is real and unusually big: 10,195 responses, fieldwork 15–29 April 2026, 88% direct from small businesses. Published by HMRC on 15 September 2026.
• 72% said the guidance did not help them — 4,234 No against 1,646 Yes. That arithmetic is sound.
• ⚠️ The reported “increase of 15.4%” is percentage POINTS, not per cent. The relative rise is about 27%.
• ⚠️ The “No” bucket contains people who never read it — the report’s own list of reasons starts with people who use an accountant.
• ⚠️ The two years are not the same sample. 5,880 answers this year against 1,626 last year — 3.6 times more people.
• The same denominator switch happens twice more in the report, once in each direction.
• 🚩 And 623 characters of the report changed between our two fetches, on consecutive mornings, with no revision note on the page.

What the survey actually is

The Administrative Burdens Advisory Board is eleven people from business and the professions whose job is to tell HMRC what the tax system feels like from the outside. Once a year they run the Tell ABAB survey, and this year it landed a record number of replies. The chair’s foreword puts it plainly: “The response has been exceptional, with 10,195 businesses and agents sharing their views between 15 April 2026 and 29 April 2026.”

The executive summary adds the split: “The 2026 Tell ABAB Survey received a record 10,195 responses, with 88% directly from small businesses and 12% from agents representing businesses, providing ABAB with its largest and most comprehensive evidence base to date.”

That is a better sample of UK small-business opinion than almost anything you will be sold this year, and it is free. Which is exactly why it is worth reading properly rather than taking the headline.

The headline number, and the first problem

Section 7.3 of the report says this:

“Of those who responded, 72% said the guidance did not help them complete their task. This is an increase of 15.4% when compared with 2025.”

The first sentence is fine. The table beneath it gives 4,234 No and 1,646 Yes from 5,880 answers, which is 72.0% and 28.0%. Last year’s figures were 920 and 706 from 1,626, which is 56.6% and 43.4%. All four percentages are correct.

The second sentence is not. 56.6% to 72.0% is a rise of 15.4 percentage points. As a percentage change it is about 27%. The report presents the point figure as a percentage, and the column it sits in is headed “% change since 2025”.

This matters in one direction only, and it is the direction that flatters: 15.4% sounds like a smaller deterioration than 27% does. Nobody is being devious — percentage points and percentages get mixed up in boardrooms every day of the week — but if you are quoting this survey, quote the points.

⚠️ While you are in that table: the Yes row prints its change as “-15,4%”, with a comma where the decimal point belongs. We are quoting it exactly as published. It is still there.

The second problem, and this is the useful one

Read the paragraph immediately after the table. It lists why people answered No:

“Reasons given for why GOV.UK guidance did not help them complete their task included that they did not need to access it (as they relied on their agent); that guidance was too complex or too generic or that they experienced difficulties navigating GOV.UK.”

Look at the first reason on that list. “they did not need to access it”. Those people never opened the guidance. They pay an accountant to deal with it.

So the 72% contains at least two entirely different populations: people who read the guidance and could not finish the job, and people who were never going to read it in the first place. One of those tells you to rewrite the page. The other tells you the page was never the problem. The question as written cannot separate them, and therefore the number cannot tell you which.

We are not guessing at this, and we are not accusing anyone of hiding it — the report states it openly, in the very next sentence. It simply does not travel with the 72%, and the 72% is the part that gets quoted.

The third problem: the two years are not comparable

The guidance question drew 5,880 answers in 2026 and 1,626 in 2025. That is 3.6 times as many people. The record-keeping question in section 6.2 drew 10,140 against 2,193 — 4.6 times as many.

The report is upfront that this year is much bigger; it calls 2026 its “largest and most comprehensive evidence base to date”, which is true and is a good thing. But a sample that quadruples is very unlikely to be the same mix of people, and none of the year-on-year tables carry a note about it.

So the honest reading is: 72% is this year’s figure from a much larger and differently composed group. It is not a clean measurement of opinion moving. That is a genuinely useful thing to know before you put the number in a slide.

The same trick twice more, once in each direction

Once you are watching denominators, the report has two more.

Section 6.4, understating a negative finding. The text says: “Approximately 60% of respondents indicate that there are no benefits to MTD for Income Tax. In 2025, 65.2% of responses indicated no benefits.” Note that it says respondents in one sentence and responses in the next. The table beneath carries the note “Note: Respondents could select multiple responses to this question.” and its counts add to exactly 11,875 — we checked the sum. So 7,131 out of 11,875 selections is 60.1%; but 7,131 out of the 10,071 people who answered that section is 70.8%. The published figure is the kinder one.

Section 7.5, the other way round. The text says “The majority of respondents (51.5%) had not used the QR code”, while the table beside it reports 5,246 of 5,676 — 92.4%. Both are right: 5,246 is 51.5% of all 10,195 people who took the survey, and 92.4% of the 5,676 who answered that question. The word “respondents” changes meaning between the sentence and the table it sits next to.

None of this is fraud. It is the ordinary way numbers drift when a long document is assembled by several people. But it is why we re-run the arithmetic on anything we are going to quote — the same reason we published our own embarrassing click figures rather than describing them.

And then the report changed overnight

Here is the part we did not expect. We fetched the full report on 15 September, the day it was published, and again on the morning of 16 September, because we re-check every source before writing about it. 623 characters of the text had changed.

Before assuming the page had moved, we re-ran the identical stripping code over both saved copies — the difference could easily have been our own tooling. It was not: both versions came out 623 characters apart under the same pipeline. Then we looked at where.

The edits are housekeeping. Table after table had a “Total answered” row ending in bare zeros, and those zeros have been replaced:

That is a genuine improvement — a column of zeros where a percentage belongs is exactly the sort of thing that makes a reader distrust a table. Somebody spotted it and fixed it within a day, which is faster than most organisations manage.

Two things are worth saying plainly about it, though.

First, every figure this article relies on is identical in both versions. We checked eleven specific spans — the 72% sentence, the counts, the 2025 comparison row, the reasons paragraph, the 60% sentence, the multi-select note, the QR sentence, the sample size and the fieldwork dates. All present, all unchanged. Including the comma typo, which the tidy-up did not reach.

Second, the publication page does not say it happened. Its “Updates to this page” section shows one line — published 15 September 2026 — and the landing page came back byte-identical on every morning we fetched it. We checked again on the day this was published, three days after the edit: the report is now unchanged since the 16th, and the update history still records nothing. So this is not a log that had yet to catch up. We are not claiming anyone hid anything; an attachment edit may simply not trigger a revision entry. But the practical effect is that a reader who quoted a table on Tuesday has no way to discover it reads differently on Wednesday — which is the same problem, from the other side, as trying to establish when a web page last changed. Usually you cannot, unless you kept a copy.

Which is a rather pointed thing to happen to a report about whether published guidance helps people finish a task.

What to do with your own “did this help?” question

Most business websites now have a version of this question somewhere — a thumbs up and down at the foot of a help article, a one-question survey after a support chat, a star rating on a delivery note. Four things, learned at HMRC’s expense rather than yours.

• Ask whether they used it before you ask whether it helped, and report the two separately. Otherwise “no use” and “never opened it” land in the same box and you cannot tell whether to rewrite the page or stop linking to it.
• Write the denominator next to the number, every time. Of everyone who visited? Of everyone who answered? Of everyone who used it? Three very different figures, and the argument about whether 72% is bad is usually an argument about which one it was.
• Say points when you mean points. If satisfaction went from 56% to 72%, that is 16 points, or about 29% — and whichever you quote, name it.
• If your sample size changed, say so beside the comparison. A number that quadrupled its respondents did not just get more accurate; it got different.

What argues the other way

The direction is probably real even if the size is not. Pull every caveat above and you still have the largest sample this survey has ever had, and a clear majority of it saying the guidance did not get them through the task. Our complaint is with how the movement is described, not with the finding that a lot of people are stuck.

And ABAB is the body doing the criticising, not the one being protected. This report exists to tell HMRC uncomfortable things, it publishes its own raw counts so anyone can recompute the percentages — which is exactly how we found all of this — and it fixed a presentation fault within a day. A great many organisations publish a percentage with no counts at all, and you cannot check those even when they are wrong.

Our own side of it: the second fetch was routine, not clever. We only caught the overnight edit because we happened to keep the first copy, and we only trusted the difference because we re-ran the same code over both files first. Had we compared yesterday’s text against today’s using slightly different tooling, we would have reported a change that was entirely our own.

Sources, and what we checked

• The report: Tell ABAB report 2025 to 2026, HM Revenue & Customs, published 15 September 2026, Open Government Licence v3.0 — at gov.uk under administrative-burdens-advisory-board-tell-abab-report-2025-to-2026. Fetched raw and tag-stripped on 15, 16 and 18 September 2026; every quotation above asserted against the stripped text of the final fetch, and every load-bearing span checked in all three. The text changed once, between the 15th and the 16th, by 623 characters; the 16th and the 18th are identical to the character.
• Every percentage in this article was recomputed from the report’s own published counts rather than copied from its text. Where our figure differs from the report’s, we have said so and shown the working.
• The overnight change was established by re-running one identical tag-stripping routine over both saved copies of the page, so that the difference could not be an artefact of our own tooling, and by confirming the landing page was byte-identical on both mornings.
• This is a reading of a free public document, not tax advice.

The image at the top of this page was generated by AI.