# Independent check of "What Britain starts now" (10 October 2026)

*This is the independent check of the figures and method behind [What Britain starts now](https://www.salenovo.com/research/what-britain-starts-now/), made on 10 October 2026 against the draft findings, before anything was published. "The note" below means that draft. Every must-fix and should-fix item was applied before publication, so some of the draft wording quoted here isn't in the published piece. The names of the files and scripts used during the check have been taken out; nothing else has been changed.*

Checked against the raw counts by trade code and period, the downloaded company rows for the twelve high street trades, the postcode lookups, the live Companies House advanced search, and the official Companies House "Incorporated companies in the UK" releases on gov.uk.

Severity key: **must fix** = wrong or unsupported as written; **should fix** = materially misleading or incomplete; **note** = worth knowing, no change required.

## Summary

The counts are sound. Every SIC-level number in the note reproduces from the raw files, and 14 of 14 live spot-checks match (two differ by 1–2 companies, which is the live index moving). The company-level totals are exactly right and agree with the official statistics to within 1%.

The problems are in interpretation and in the geography:

1. **Geography has a bug, not a coverage limit.** The geocoding step read the downloaded rows with a standard CSV reader, which mis-splits every row whose SIC list has more than one code (`[86101, 87200, 96020]` is unquoted). Companies with one code were geocoded at 95–98%; companies with two or more at 33–57%. Re-geocoding the missing 28,023 postcodes lifts coverage to 97–99% and changes several published regional growth rates by 30–80 points (sports facilities North West +251% becomes +193%; driving schools North East +6% becomes +44%; other education South West +37% becomes +58%). The "London grew slowest in nearly every consumer trade" claim fails: with full geocoding London is slowest in 2 of the 12 trades.
2. **Codes per company rose from 1.31 (2019) to 1.47 (2025).** Code assignments grew 36% while companies grew 21%. Every "+X%" in the note is measured against a baseline of +36%, not +21%; hairdressing (+49%) and vehicle repair (+46%) are barely above it. The note never states this.
3. **"Almost all of that growth went to three kinds of company" divides code growth by company growth.** The four named codes grew by 128,986 assignments against 142,714 extra companies, but total code assignments grew by 324,290. The honest share is a third of the extra code assignments, not 90% of the extra companies.
4. **"The only thing on the register that's growing this year" is contradicted by the note's own table** (social work +43%, sports facilities +31%, games publishing +30%, holding companies +21%, taxis +13%, plus reservation services +50%, exhibition organisers +44%, facilities support +44% that the note omits).
5. **The 2026 fall is not "mostly in the shell codes".** The two largest absolute falls are property codes (68100 −7,179; 68209 −6,519). Online retail, the note's archetypal shell code, is down only 7%, less than the register (−7.4%). The note also does not mention mandatory director identity verification (ECCTA, in force for new incorporations from November 2025), which is the obvious untested candidate cause for the collapse in agent-formed codes.
6. **Takeaways geography row is wrong** even on the original geocoding (North East is +52%, not "around +100%"; slowest is Scotland +46%, not South West/East of England).
7. **The 2022 cohort exposure is described backwards**: January 2022 companies had four years to the end of 2025 and December 2022 companies three, not "three years for the first and 3.5 for the last".

The headline as proposed is not supportable. A version that is appears in section 5.

---

## 1. Re-derivation of every quoted figure

Method: combined the two passes of raw counts (7,975 rows, no duplicate (sic, period) pairs), set every count of exactly 5,000 to missing (59 rows), pivoted, and recomputed each table with my own code.

### Risers and fallers table (2019, 2025, change, 2022 cohort gone)

All 27 code rows reproduce exactly: 2019 count, 2025 count, rounded change, rounded "gone". Exact values behind the rounding, for the record:

| SIC | 2019 | 2025 | change | 2022 formed | dead by end 2025 | gone |
|---|---|---|---|---|---|---|
| 47910 | 22,613 | 78,340 | +246.4% | 53,742 | 41,070 | 76.4% |
| 68209 | 36,752 | 69,151 | +88.2% | 51,984 | 14,675 | 28.2% |
| 68100 | 25,562 | 49,466 | +93.5% | 34,829 | 11,204 | 32.2% |
| 62012 | 10,859 | 27,815 | +156.1% | 13,508 | 7,084 | 52.4% |
| 64209 | 9,512 | 21,133 | +122.2% | 13,522 | 2,974 | 22.0% |
| 56103 | 14,628 | 23,581 | +61.2% | 18,315 | 11,868 | 64.8% |
| 73110 | 6,590 | 14,698 | +123.0% | 9,593 | 5,918 | 61.7% |
| 62090 | 10,401 | 18,284 | +75.8% | 12,542 | 7,080 | **56.5%** (note says n/a) |
| 47990 | 8,830 | 16,206 | +83.5% | 21,069 | 17,510 | 83.1% |
| 85590 | 6,936 | 14,119 | +103.6% | 8,534 | 3,955 | 46.3% |
| 46900 | 5,455 | 11,984 | +119.7% | 9,546 | 6,706 | 70.2% |
| 96020 | 10,634 | 15,836 | +48.9% | 14,704 | 8,272 | 56.3% |
| 56102 | 7,280 | 12,221 | +67.9% | 8,819 | 4,993 | 56.6% |
| 81210 | 3,725 | 7,733 | +107.6% | 7,431 | 5,664 | 76.2% |
| 47110 | 4,086 | 7,793 | +90.7% | 5,033 | 2,460 | 48.9% |
| 45200 | 6,701 | 9,774 | +45.9% | 7,546 | 3,945 | 52.3% |
| 43910 | 1,828 | 3,997 | +118.7% | 3,073 | 1,778 | 57.9% |
| 93110 | 739 | 1,540 | +108.4% | 749 | 301 | 40.2% |
| 85530 | 446 | 1,005 | +125.3% | 549 | 241 | 43.9% |
| 47730 | 367 | 940 | +156.1% | 986 | 233 | 23.6% |
| 86230 | 1,364 | 2,270 | +66.4% | 1,903 | 315 | 16.6% |
| 49320 | 2,646 | 4,115 | +55.5% | 2,885 | 1,353 | 46.9% |
| 96090 | 31,369 | 19,694 | −37.2% | 23,416 | 13,095 | 55.9% |
| 99999 | 12,417 | 7,967 | −35.8% | 15,367 | 8,921 | 58.1% |
| 82990 | 30,322 | 21,552 | −28.9% | 21,628 | 10,495 | 48.5% |
| 49410 | 16,938 | 13,826 | −18.4% | 17,539 | 12,690 | 72.4% |
| 69201 | 4,850 | 4,106 | −15.3% | 3,508 | 1,260 | 35.9% |

- **should fix**: 62090 "gone" is shown as n/a but the data gives 56% (7,080 / 12,542). There is no missing value for this code.

Division rows: 42 (6,764 → 4,580, −32.3%, gone 63.3%), 61 (3,310 → 2,382, −28.0%, 57.5%), 78 (12,517 → 12,417, −0.8%, 56.9%), 41 (34,897 → 36,653, +5.0%, 58.6%) all reproduce. Division sums treat the 5000-artefact cells as zero, which is correct if those cells are true zeros (see section 4, artefact).

### "Trades that last" table

All 22 figures reproduce (dental 16.6, holding 22.0, pharmacies 23.6, specialist medical 26.9, letting 28.2, electricity 28.1, solicitors 31.1, bookkeeping 32.6, vets 33.1, architects 32.7, accountants 35.9; packaging 86.8, non-store retail 83.1, electrical appliance retail 79.7, wholesale clothing 77.6, market stalls clothing 76.8, road construction 76.9, online retail 76.4, general cleaning 76.2, window cleaning 75.6, sales agents clothing 75.9, freight 72.4).

- **should fix**: the table is presented as the ranked extremes of codes with 500+ formations in 2022 but is a selection. Omitted from the "last" side, in rank order: television programme production (32.1%, 1,495 formed), general medical practice (32.7%), motion picture production (33.8%), head offices (34.1%), mixed farming (34.8%). Omitted from the "don't" side: administration of financial markets (84.0%), leather goods retail (78.8%), household goods wholesale (78.4%), computer retail (77.3%), computer wholesale (76.6%), watches and jewellery wholesale (76.6%). TV and film production surviving as well as architects weakens the "licence, premises or professional body" reading, and the table should either include them or be labelled as selected.

### 2026 table

All 18 rows reproduce (58290 3,667 → 7,348 +100.4%; 62012 20,364 → 33,171 +62.9%; 63120 +53.1%; 72190 +49.2%; 63110 +47.5%; 88990 +42.7%; 93110 +30.9%; 58210 +29.9%; 64209 15,278 → 18,526 +21.3%; 49320 3,061 → 3,468 +13.3%; 46160 3,359 → 990 −70.5%; 46190 5,809 → 3,283 −43.5%; 47750 3,161 → 2,033 −35.7%; 47990 12,760 → 8,674 −32.0%; 47730 716 → 499 −30.3%; 96090 15,471 → 10,792 −30.2%; 47710 6,512 → 4,647 −28.6%; 99999 6,193 → 4,488 −27.5%). Division 46: 48,794 → 36,110 (−26.0%) ✓. Division 87: 7,452 → 5,781 (−22.4%) ✓.

- **should fix**: nine cells in the 2026 table are blank; the figures above fill them. Pharmacies 716 → 499 is a small base and the note already says so.
- **should fix**: the "growing" column skips 79909 other reservation services (667 → 1,001, +50.1%), 82301 exhibition and fair organisers (+44.5%), 81100 combined facilities support (+43.7%), 66190 auxiliary financial (+41.1%), 70100 head offices (+38.6%), which all rank above sports facilities and games publishing. Omitting them makes the software story look cleaner than the data is.
- **note**: by-division percentages all reproduce; division 62 is +25.49%, so "+25%" is fine.

### Short version

| Claim | Check |
|---|---|
| 21% more companies in 2025 than 2019 (829,933 vs 687,219) | 829,933 / 687,219 = +20.8% ✓ |
| Online retail 22,613 → 78,340; three in four of 2022 cohort gone | ✓ (76.4%) |
| Hairdressers +49, takeaways +61, cafés +68, cleaners +108, roofers +119, driving schools +125, gyms +108, pharmacies +156, a quarter gone | all ✓ |
| Dentists 17, pharmacies 24, specialist medical 27, solicitors 31, bookkeepers 33, vets 33, architects 33, accountants 36 | all ✓ |
| Packaging 87, non-store 83, online 76, wholesale clothing 78, cleaning 76 | all ✓ |
| Jan–Sep 2026 down 7.4% (595,287 vs 643,185) | −7.45% ✓ |
| Software publishing +100, dev +63, portals +53, hosting +47, R&D +49 | all ✓ |
| Register-wide 2022 cohort 55% gone | live: 439,246 / 802,467 = 54.7% ✓ |
| "shell codes are down 15 to 70%" | **must fix** — see section 4(e) |
| "the only thing on the register that's growing" | **must fix** — see section 4(h) |
| "almost all of that growth went to three kinds of company" | **must fix** — see section 4(a) |
| "Construction ... stood still or fell" | **should fix** — see section 4(h) |

## 2. Live spot-checks (Companies House advanced search, 10 October 2026)

| Query | Note / data | Live | |
|---|---|---|---|
| 47910 inc 2019 | 22,613 | 22,613 | ✓ |
| 47910 inc 2025 | 78,340 | 78,339 | ✓ (−1, live index) |
| 47910 inc 2022 | 53,742 | 53,742 | ✓ |
| 47910 inc 2022, dissolved 2022–2025 | 41,070 | 41,070 | ✓ |
| 47730 inc 2019 | 367 | 367 | ✓ |
| 47730 inc 2025 | 940 | 940 | ✓ |
| 86230 inc 2022 | 1,903 | 1,903 | ✓ |
| 86230 inc 2022, dissolved 2022–2025 | 315 | 315 | ✓ |
| 62012 inc Jan–Sep 2025 | 20,364 | 20,366 | ✓ (+2, live index) |
| 62012 inc Jan–Sep 2026 | 33,171 | 33,171 | ✓ |
| 96020 inc 2019 | 10,634 | 10,634 | ✓ |
| 96020 inc 2025 | 15,836 | 15,836 | ✓ |
| 58290 inc Jan–Sep 2026 | 7,348 | 7,348 | ✓ |
| 46160 inc Jan–Sep 2026 | 990 | 990 | ✓ |
| All companies inc 2019 | 687,219 | 687,219 | ✓ |
| All companies inc 2025 | 829,933 | 829,933 | ✓ |
| All companies inc Jan–Sep 2025 | 643,185 | 643,185 | ✓ |
| All companies inc Jan–Sep 2026 | 595,287 | 595,287 | ✓ |
| All companies inc 2022 | — | 802,467 | |
| All companies inc 2022, dissolved by end 2025 | 55% | 439,246 (54.7%) | ✓ |

Note on how the page reports counts: a normal page says "22,613 results"; a zero page says "No results found"; a single result says "1 result" (singular). See the artefact finding in section 4.

- **note**: the counts drift by a company or two between collection and today because the index is live (a 2025 company's incorporation date can be corrected; a dissolved company can be restored). Nothing in the note depends on that precision.

## 3. Arithmetic of percentages

Every change % and gone % in the three tables and the short version reproduces to the rounded integer from the raw counts. No arithmetic errors found. Two presentation points:

- **should fix**: the rounding is half to even, so a value of exactly x.5 can round down. None of the published figures is affected, but 56102 gone (56.6) and 62 division 2026 (+25.49) sit close to the edge. Use one decimal in the CSVs even if the piece quotes integers.
- **note**: driving schools regional percentages are computed on bases of 9–71 companies (see section 4(g)).

## 4. Method

### (a) Per-code counts, multi-coding, and what "growth" means — must fix

A company carries up to four codes, and the number it carries has risen. From the counts files, total code assignments divided by the live company totals:

| Period | Code assignments | Companies | Codes per company |
|---|---|---|---|
| 2019 | 896,682 | 687,219 | 1.305 |
| 2022 | 1,075,138 | 802,467 | 1.340 |
| 2024 | 1,197,973 | 844,294 | 1.419 |
| 2025 | 1,220,892 | 829,933 | 1.471 |
| Jan–Sep 2025 | 943,740 | 643,185 | 1.467 |
| Jan–Sep 2026 | 889,129 | 595,287 | 1.494 |

So code assignments grew **+36%** 2019 → 2025 while companies grew +21%. A trade whose code count rose 36% has simply kept pace. Against that baseline, hairdressing (+49%) and vehicle repair (+46%) are modest, cafés (+68%) and takeaways (+61%) moderate, and the note's "steady growth" on the high street is partly founders ticking more boxes. The rows files confirm this inside the trades: the share of companies carrying more than one code rose from 19% to 30% for pharmacies, 31% to 40% for other human health, 49% to 60% for general cleaning, 11.5% to 14.4% for hairdressing.

The sentence "almost all of that growth went to three kinds of company" is arithmetically the four codes' growth (47910 +55,727; 68209 +32,399; 68100 +23,904; 62012 +16,956 = 128,986) set against the 142,714 extra companies. That divides apples by oranges. The sum of all positive code changes is 383,190; the four codes are 33.7% of that. The register does not support a multi-code OR query (I tested `sicCodes=47910,68209` and the two-parameter form; both return no results), so the share of *companies* carrying any of those codes cannot be measured with this method.

Required changes: (1) state the 1.31 → 1.47 codes-per-company figure and the +36% code-assignment baseline next to every growth table; (2) replace "almost all of that growth" with "the four codes account for a third of the rise in code assignments"; (3) add a column or footnote giving each trade's growth relative to the +36% baseline, or at least say which trades fall below it.

The italic preamble ("a company carries up to four codes, so a formation in a trade means a company registering that code, and the trade totals add to more than the number of companies") is honest about the definition but not about the drift, and nothing in the body carries the warning forward.

### (b) Reclassification from the catch-all codes — should fix

The three catch-alls fell by 11,675 + 8,770 + 4,450 = 24,895 assignments between 2019 and 2025, against +383,190 of positive code growth. Even if every one of them was a founder who would once have picked a catch-all and now picks a real code, reclassification can explain at most 6.5% of the rise in risers. Within the twelve downloaded trades, the share of companies also carrying 96090/82990/99999 is small and roughly stable (0.4–8.5%). So the note's worry is overstated for the 2019 → 2025 risers and the sentence "some of every riser is reclassification" should be quantified: "at most a fifteenth of the growth".

The bigger reclassification risk is the one the note itself raises under "what to check": formation agents' code-picker lists. The 2026 falls of 71% (46160) and 43% (46190) in two "agents involved in sale of..." codes, with no matching rise elsewhere in division 46 (the division is down 26%), look like a channel changing, not 2,500 fewer clothing agents. That belongs in the body, not the to-do list.

### (c) Is "gone by end 2025" for the 2022 cohort a fair survival measure — must fix the description, should fix the method

The note says the window is "the three-year mark for the first of them and 3.5 years for the last". It is the other way round and the numbers are wrong: a company formed on 1 January 2022 had 48 months to 31 December 2025; one formed on 31 December 2022 had 36. The script docstring ("36+ months") is right; the prose is not.

Does it matter? From the 2019 cohort in the twelve downloaded trades (68,249 companies), 35.8% had dissolved within 36 months and 41.9% within 48 months, so a company's exposure window moves "gone" by about six points. Within a single cohort year the mix of January-to-December formations is similar across trades, so the *ranking* of trades is fair; the *level* (55% register-wide, 76% online retail) is a blend of 36–48 months and should be described as "three to four years".

Caveat 3 says "most dissolutions are shells the registrar removed at seventeen months". The rows data do not support seventeen: the median time to dissolution is 27 months, the mode is 18–24 months (first confirmation statement at 12 months, first accounts due at 21 months, then strike-off), and a quarter of dissolutions come after 48 months. Say "most come between 18 and 30 months, when the first filings fall due".

### (d) Companies dissolved before 2010 are not in the search index — note

Irrelevant here: every cohort is 2019 or later. The exclusion that could matter is the reverse one, companies dissolved and later *restored* show as active and fall out of the "gone" numerator; this is small (official statistics show about 500–600 restorations a month across the whole register) and biases "gone" down by well under a point.

### (e) Is Jan–Sep 2026 vs Jan–Sep 2025 fair — must fix two claims

Windows: both are 1 January to 30 September, 273 days each. Fair.

Indexing lag: I pulled the all-company count month by month from the live index. 2026 against 2025: Jan +3.5%, Feb −10.3%, Mar −3.1%, Apr −6.5%, May −10.1%, Jun −9.0%, Jul −7.5%, Aug −14.1%, Sep −10.3%. The fall is spread through the year, not concentrated in the most recent month, so index lag is not the explanation. Incorporations appear on the index within hours; the Sep 2025 index figure (76,211) agrees with the official quarter to 1%.

Fee rise: both windows are after 1 May 2024, so the comparison is clean. But caveat 4's claim that "2024 against 2025 is the cleaner comparison for recent change" is wrong: 2024 has 98,328 incorporations in April alone (the rush before the £50 fee) against 60–65k a month afterwards, so 2024 is the *dirtiest* year in the series. The clean comparisons are 2025 vs 2026, or any post-May-2024 window against the same months a year later.

Shell codes: the note says "the shell codes are down 15 to 70%" and the 2026 fall is "mostly in the shell codes". From the counts: 47910 online retail 60,524 → 56,218 (**−7%**, less than the register's −7.4%); 82990 **−8%**; 46900 **−6%**; 99999 −28%; 96090 −30%; 47990 −32%; 46190 −43%; 46160 −71%. "15 to 70%" is true only of the five codes it is true of. And the two largest absolute falls are property: 68100 buying and selling own real estate 38,563 → 31,384 (−7,179) and 68209 letting 53,481 → 46,962 (−6,519), followed by 96090 (−4,679), 47910 (−4,306) and 47990 (−4,086). Residents' property management 98000 is down 24%, housing associations 68201 down 34%. The 2026 story has a property leg the note does not mention.

Untested cause the note omits entirely: the Economic Crime and Corporate Transparency Act identity-verification regime. Identity verification became compulsory for directors and PSCs of newly incorporated companies from November 2025, and formation agents must now be registered Authorised Corporate Service Providers (gov.uk published an ACSP list on 2 October 2026 and a "directors warned to verify identities" release on 17 September 2026). A near-total collapse in the agent codes (46160 −71%) in the first full year of IDV is exactly what that would produce. The note cannot test it either, but under its own rule ("anything about why without a comparator") it should not float the AI boom as a cause and leave IDV unmentioned. Verify the exact commencement date before quoting it.

### (f) Official Companies House monthly incorporation statistics — note (they agree)

Source: "Incorporated companies in the UK" quarterly releases, gov.uk. The latest is April to June 2026 (published 30 July 2026, first in a new format with monthly figures). The July to September 2026 release is scheduled but not yet published as of 10 October, so the Q3 2026 figure cannot be checked officially.

Official UK incorporations, all companies (from the January–March 2026 release's back series, and the April–June 2026 CSV):

| Period | Official | Search index | Index / official |
|---|---|---|---|
| 2019 (4 quarters: 182,742 + 169,976 + 169,751 + 155,950) | 678,419 | 687,219 | +1.3% |
| 2025 (211,112 + 209,798 + 215,982 + 184,888) | 821,780 | 829,933 | +1.0% |
| Jan–Sep 2025 | 636,892 | 643,185 | +1.0% |
| Jan–Mar 2026 | 204,612 | 206,349 | +0.8% |
| Apr 2026 / May 2026 / Jun 2026 | 67,059 / 62,091 / 63,137 | 67,637 / 62,569 / 63,625 | +0.8% |
| Jan–Jun 2026 | 396,899 | 400,180 | +0.8% |

Official 2019 → 2025: +21.1% (note: +20.8%). Official H1 2026 vs H1 2025: 396,899 vs 420,910 = **−5.7%** (index: −5.9%). The index's Q3 2026 is 195,107 against 218,097 for Q3 2025, −10.5%, which is what pushes the nine-month figure to −7.4%; it is consistent with the published monthly pattern but is unverified until the Q3 release (expected late October). The index runs about 1% above the official series throughout, most likely because it counts bodies the statistical series excludes or counts by a different date; the ratio is stable so growth rates are unaffected.

Recommendation: quote the official −5.7% for the first half as the anchor and the index −7.4% for nine months as provisional, and say the official Q3 figure is due.

### (g) Geography — must fix

**The geocoding bug.** The geocoding step built the set of postcodes to look up with a standard CSV reader over the downloaded rows. Those files are the advanced-search CSV export, in which the nature of business field is an unquoted bracketed list, so a row with two or more codes (`[86101, 87200, 87300, 96020]`) has extra commas and the reader takes the wrong column as the address. The result, from the rows files and the cache:

| Trade | Geocoded, one code | Geocoded, 2+ codes | Share with 2+ codes 2019 → 2025 |
|---|---|---|---|
| 43910 roofing | 95.4% | 39.9% | 29.7% → 29.3% |
| 45200 vehicle repair | 96.2% | 37.2% | 27.3% → 33.4% |
| 47730 pharmacies | 98.3% | 53.3% | 19.3% → 29.8% |
| 81210 cleaning | 97.0% | 36.3% | 49.4% → 59.7% |
| 85590 other education | 95.6% | 49.4% | 52.5% → 64.5% |
| 86900 other human health | 96.6% | 42.3% | 30.9% → 40.2% |
| 93110 sports facilities | 96.7% | 38.5% | 51.7% → 44.8% |
| 96020 hairdressing | 97.3% | 44.4% | 11.5% → 14.4% |

Only 1,227 of 131,733 cached postcodes (0.9%) are genuine postcodes.io misses; 99.8–100% of rows have a postcode. The "63 to 90% coverage" in the note is this bug, not a data limit. Because multi-code companies are both more likely to be agent-formed and more common in 2025 than 2019, the geocoded sample is not a random sample of either year.

**What changes when it is fixed.** I looked up the 28,023 missing postcodes (leaving the original lookups untouched) and re-ran the region and council logic. Coverage rises to 97–99% for every trade. Published vs corrected regional growth, own-address formations 2019 → 2025:

| Trade, region | Note | Corrected | 2019 → 2025 corrected counts |
|---|---|---|---|
| Hairdressing NI | +109% | +105% | 113 → 232 |
| Hairdressing North East | +101% | +111% (now fastest) | 283 → 598 |
| Hairdressing Wales | +76% | +73% | 320 → 553 |
| Hairdressing Yorkshire | +72% | +75% | 669 → 1,170 |
| Hairdressing London | +25% | +26% | 2,600 → 3,288 |
| Takeaways North East | "around +100%" | +58% (+52% on original) | 599 → 946 |
| Takeaways West Midlands | "around +100%" | +97% | 1,083 → 2,138 |
| Takeaways East Midlands | "around +100%" | +79% | 744 → 1,333 |
| Takeaways slowest | "South West, East of England" | Scotland +51%, South West +53%, Wales +54% (East of England is +68%, mid-table) | |
| Driving schools East of England | +217% | +194% | 34 → 100 |
| Driving schools Scotland | +211% | +204% | 23 → 70 |
| Driving schools East Midlands | +169% | +124% | 37 → 83 |
| Driving schools London | +152% | +158% | 80 → 206 |
| Driving schools North East (slowest) | +6% | +44% | 18 → 26 |
| Sports facilities North West | +251% | +193% | 68 → 199 |
| Sports facilities East Midlands | +200% | +165% | 40 → 106 |
| Sports facilities Yorkshire | +200% | +120% | 60 → 132 |
| Sports facilities slowest | South West +31% | North East +25% (South West +41%) | 28 → 35 |
| Other human health NI | +88% | +93% | 103 → 199 |
| Other human health Scotland | +43% | +61% | 418 → 674 |
| Other human health East of England | +3% | +12% | 1,200 → 1,346 |
| Other human health London | +6% | +11% | 2,756 → 3,059 |
| Other education Wales | +85% | +94% | 181 → 352 |
| Other education Scotland | +74% | +93% | 264 → 509 |
| Other education West Midlands | +70% | +93% | 500 → 966 |
| Other education slowest | South West +37% | South West +58% (London +59%) | 460 → 728 |

Council counts are undercounts throughout: hairdressing Birmingham 321 → 369, Manchester 229 → 257, Glasgow 201 → 220, Leeds 187 → 199; other human health Birmingham 309 → 403; other education Birmingham 195 → 302. Rankings mostly hold, but "biggest rises by council" changes in places (sports facilities: Glasgow and Cheshire West enter; other human health: Glasgow and Edinburgh enter, Belfast drops out of the top seven).

**"London grew slowest in nearly every consumer trade"** — must fix. With full geocoding London is the slowest region in 2 of the 12 trades (hairdressing +26%, other human health +11%). In roofing London is second fastest (+138%), in driving schools fourth (+158%), in sports facilities fifth (+121%), in pharmacies third (+157%). On the original geocoding the claim also fails (London slowest in the same two). "The growth is in the North, the Midlands and the devolved nations" holds for hairdressing, takeaways, cafés, convenience stores, cleaning, vehicle repair and other education; it does not hold for roofing, driving schools, pharmacies or sports facilities, where London and the South East are among the fastest.

**Small bases** — should fix. Driving schools and sports facilities regional rates rest on 2019 bases of 8–80 companies per region (Northern Ireland driving schools: 1 → 5). Quote counts, not percentages, for any region under about 50 companies, or merge regions.

**Own-address definition** — should fix. "Crowded" is computed over the combined 2019 and 2025 rows of that one trade, so an address with six companies in 2019 and five in 2025 is excluded from both years, and an address with nine of this trade and five hundred of other trades is kept. The piece says "addresses with fewer than ten companies of that trade", which is right as far as it goes, but should add "among the companies formed in 2019 or 2025". The Companies House default address (CF14 3UZ) appears on 2–368 rows per trade and is caught by the crowded filter, so it does not inflate Wales.

**Registered office is not trading address** — note. The caveat is present. For hairdressing and takeaways the registered office is often the owner's home or accountant; the own-address filter removes the big agent addresses but not the small accountancy practices with eight clients.

### (h) Claims the tables don't support, and untested causes — must fix

1. **"Whatever the AI boom is doing to the economy, it is the only thing on the register that's growing this year."** The note's own 2026 table lists social work +43%, sports facilities +31%, games publishing +30%, holding companies +21%, taxis +13%, and the data add reservation services +50%, exhibition organisers +44%, facilities support +44%, auxiliary finance +41%, head offices +39%, technical testing +31%. Software is the largest grower and the only large division growing strongly (62 +25%, 58 +38%, 63 +32%); it is not the only thing growing. "AI boom" is a cause with no comparator and should go, under the note's own rule.
2. **"most of it was never a business"** (online retail). The data say 76% of the 2022 cohort left the register within three to four years. Dissolved is not "never a business" (caveat 3 says so). Say "three in four were gone within four years".
3. **"Construction, haulage, recruitment and accountancy stood still or fell."** Division 41 +5%, 42 −32%, but division 43 specialised construction (the largest, 47,569 → 54,597) +15%, including the roofers the same paragraph celebrates, plumbing +20%, electrical +14%, domestic building +17%. Section F as a whole: 89,230 → 95,830, +7.4%. Against the +36% code baseline construction fell behind; it did not stand still in absolute terms. Accountancy: 69201 −15% but bookkeeping +9% and tax consultancy +22%; division 69 −2%.
4. **"the shell codes are down 15 to 70%"** and **"2026 is lower again, mostly in the shell codes"** — see (e).
5. **"A licence, a premises or a professional body is worth thirty points of survival."** Holding companies (22% gone) and electricity production (28%) have none of the three; TV production (32%) and film production (34%) survive like architects. The pattern is real for the licensed professions but the one-line law overreaches; say "the licensed professions lose a quarter to a third; the no-premises retail and wholesale codes lose three quarters".
6. **The 5000 artefact** — should fix the caveat. The page reports zero results as "No results found" and a single result as "1 result" (singular); neither matches a "N results" regex. The 59 affected rows therefore stand for codes with **zero or one** formation in the period, not zero. The counting script's pattern still fails on both: it matches ",000 results" inside "first 5,000 results" and returns 0, so the caveat's "the fix is a one-line regex" is not yet true. None of the headline codes is affected; division 05 (coal) Jan–Sep 2026 shows as 0 because all three codes hit the artefact.
7. **"Formed is not trading. ... the piece never quotes a rise without it [the survival column]."** The 2026 table and the geography section quote rises with no survival figure. Either drop the promise or keep it.

## 5. The headline

Proposed: **"Britain's start-up boom was mostly online shops that never opened. The real growth is barbers, takeaways, cleaners and, this year, software."**

Not supportable as written:

- *"mostly online shops"*: online retail's code grew by 55,727; the two property codes grew by 56,303 between them (68209 +32,399, 68100 +23,904), and property's survival (28–32% gone) is the opposite story. Online retail is the biggest single code but not "most" of anything measurable, and the division behind "most" is invalid (section 4(a)).
- *"never opened"*: untested. A quarter of the 2022 online-retail cohort is still on the register, and dissolved is not "never traded".
- *"the real growth is barbers, takeaways, cleaners"*: barbers +49% against a +36% code baseline; cleaners +108% but with 76% of the 2022 cohort gone, identical to online retail, so by the note's own rule cleaners cannot be a "real growth" poster child.
- *"this year, software"*: supportable, with the qualification that it is the largest grower, not the only one.

A version the tables support:

**"Britain registers a fifth more companies than in 2019. The growth is online shops and property vehicles; three in four of the online shops formed in 2022 had gone within four years, while the trades that keep their companies are dentists, pharmacists and solicitors. In 2026 formations are down 6 to 7%, and software is the one large sector still growing fast."**

Or shorter: **"Britain's start-up boom was online shops and property vehicles. Three in four of the online shops are already gone. The trades that last are the licensed ones, and the one thing still growing in 2026 is software."**

Each clause maps to a figure the note has and this check reproduces: +21% (829,933 vs 687,219, official +21.1%); 47910 and 68209/68100 the three largest risers; 76.4% of 53,742; 16.6%, 23.6%, 31.1%; −5.7% official H1 / −7.4% index nine months; division 62 +25%, 58 +38%, 63 +32% as the only large divisions above +20%.

## Checklist before publication

Must fix:
1. Re-run the geocoding with a reader that handles the bracketed code list, so all postcodes are looked up; regenerate the region, council and trade summary tables; redo the geography table. The 28,023 missing lookups are already done.
2. Rewrite the takeaways geography row; drop "London grew slowest in nearly every consumer trade".
3. State codes per company 1.31 → 1.47 and the +36% code-assignment baseline; replace "almost all of that growth went to three kinds of company".
4. Remove "the only thing on the register that's growing" and "AI boom"; list the other growers or say "the largest".
5. Correct "shell codes down 15 to 70%" (47910 −7%, 82990 −8%, 46900 −6%) and "mostly in the shell codes" (property codes are the two biggest falls); mention identity verification as an untested candidate cause alongside the fee rise.
6. Fix the cohort exposure description: four years for January 2022 formations, three for December.
7. Rewrite the headline (section 5).

Should fix:
8. Fill 62090 gone (56%) and the nine blank 2026 cells.
9. Label the "last / don't" and 2026 tables as selections, or include the omitted codes (TV production 32%, financial-market administration 84%, reservation services +50%, exhibition organisers +44%).
10. Caveat 3: "most dissolutions come at 18 to 30 months", not seventeen.
11. Caveat 4: 2024 is the dirty year (April rush); the clean comparison is 2025 vs 2026.
12. Caveat 6: the artefact covers zero *or one* result, and the current regex still returns 0 for both; fix the parser before re-use.
13. Quote the official H1 2026 figure (−5.7%) and flag Q3 as pending the late-October release.
14. Replace regional percentages on bases under about 50 companies with counts (driving schools, sports facilities, pharmacies).
15. "Construction stood still": section F +7%, division 43 +15%; say "fell behind the register" instead.
16. Own-address definition: add "among the companies formed in 2019 or 2025".

Notes:
17. Live counts drift by 1–2 from the stored ones; immaterial.
18. The search index runs ~1% above the official statistical series throughout; growth rates are unaffected.
19. Pre-2010 dissolution exclusion is irrelevant to these cohorts; restorations bias "gone" down by well under a point.
20. Reclassification from the catch-alls explains at most 6.5% of 2019 → 2025 riser growth; the real reclassification risk is agents' code pickers, visible in the 2026 agent codes.