Who does the AI recommend?

Status: draft for review, 19 September 2026. Every figure in this document is taken from stored evidence; none is pending.

This page is the published method. It is written so that a third party with the same keys could reproduce the result, and so that any firm named in the table can see exactly how its figure was produced.

---

1. The question

When a prospective client asks an AI assistant which solicitor to instruct, which firms get named? We asked four assistants the same set of questions repeatedly, over several days, and counted.

We did not ask the assistants to rank firms, and we do not present the result as a measure of legal quality. It measures one thing: how often a firm is named in answer to a client-phrased question.

2. The prompt set

268 prompts in three families. Each was put to each engine eight times: 268 x 4 x 8 = 8,576 calls. Every raw response is retained for twelve months.

Cities: Birmingham, Bristol, Cardiff, Edinburgh, Glasgow, Leeds, Leicester, London, Manchester, Newcastle.

How the set was built. 72 prompts were written first, piloted, and frozen on 16 September 2026 together with the assessment thresholds, after an 80-answer pilot and before the main collection began. After the first results, 196 more were written: practice areas the first 72 could not reach at all - immigration, insurance, shipping, investment funds, tax, criminal defence, the Court of Protection among them - and six phrasings of every corporate and consumer area, each asked as a different kind of client.

Both sets are pinned by a hash, and one was pinned late. The registered 72 were hashed when they were frozen, and still match. The 196 were meant to be hashed before collection and were not; the hash in data/prompt_set_wave2.lock.json was recorded on 18 September, after collection. The evidence that they did not change while being asked is that their source file was last modified at 17:10 UTC on 17 September and the first wave-2 call was made at 17:16 UTC. The assessment now refuses to pass if either set no longer matches its hash - until 18 September nothing checked the lock at all.

The published index uses all 268. That decision was made on 18 September, after data existed, and is recorded in the amendment log with its reasons and its consequences. The registered 72 remain reproducible and are reported as a check on the published figures.

The mix of questions shapes the results. All 268 are 54% corporate, 31% consumer and 15% local; the registered 72 alone were 28%, 17% and 56%. Firms that win corporate work rank higher on the published set than they would on local questions alone, which is why the headline findings are stated by question type.

Scotland is included although it is a separate jurisdiction, because clients there ask the same question and the Scottish firms deserve a fair hearing.

No system prompt beyond the provider default. No memory, no location, no prior context. Provider default sampling. UK English.

3. The engines

Web search or grounding was enabled on every engine, because that is what a real user gets.

The exact model ID is recorded on every individual call, so a later model change cannot silently contaminate the dataset.

Search was not invoked equally

Search or grounding was enabled on every call, but three of the four engines decide for themselves whether to use it. The rate differs by engine and by who is asking. Computed on all 8,576 published calls; earlier versions of this table were computed on the first five runs, then on the registered 72 prompts.

Google Search grounding cannot be forced - it is not a function-calling tool, and setting the tool-choice mode to ANY produces no grounding at all - so this is a property of the engine, not a defect in the method. It is reported rather than corrected, and search_used is recorded on every call so any figure can be recomputed on searched calls alone.

A second stratum: search-embedded AI

Version 1 recorded that Google AI Overviews had no API. That was true of the direct route and turned out to be false of every route: a SERP data provider reaches them. Three further surfaces were therefore measured and are reported separately from the four assistants above, because they answer a different question shape - the assistants are asked, these are searched.

Computed on all 268 prompts, the published basis. On the registered 72 alone the AI Overview rate is 77.7%: lower mainly because more than half of those prompts are local, and local searches are where Google holds back.

Calls are counted the way the scoring counts them: one per prompt, surface and run, and where a call failed and was retried the successful record is used. 103 of 2,144 AI Overview calls and 1 of 2,144 AI Mode calls never succeeded and are excluded from the denominator. An earlier version of this section kept the first record for each call instead, counted retried calls as failures, and published 78.6% of 412 on the registered 72; the rate barely moved, but the denominator was wrong.

The AI Overview rate is itself a finding, because it is not even across the prompt set:

Google is markedly less likely to put an AI answer above a local solicitor search than above a corporate or consumer one. On roughly a third of local questions Google showed no AI summary, only its ordinary results.

All calls pinned to the United Kingdom and English, which is a controlled variable the assistant APIs do not offer at all.

On Bing, and why it is withdrawn rather than scored zero. Across 1,300 distinct calls the provider returned no Copilot answer element at all - not once, on any prompt family. Every one of those calls succeeded, returning status 20000 Ok. with ten organic results and nothing else.

Two checks establish that this is a limit of the data source and not a finding about Bing:

1. A control test with three short search-shaped queries ("best divorce solicitors leicester", "who are the best employment law solicitors in the UK", "top law firms for commercial litigation uk") returned only organic results, with no AI element, so it is not an artefact of our long conversational prompts.
2. The provider's three candidate Bing AI endpoints - serp/bing/ai_mode, serp/bing/copilot and serp/bing/ai_overview - each return 40402 Invalid Path. No Bing AI surface exists in this provider's catalogue.

Collection was stopped at that point rather than continued.

This surface is therefore withdrawn from the index and carries no score. It would be straightforwardly wrong to publish "0%" for Bing: that number would read as "no firm is visible in Copilot", when what actually happened is that Copilot was never observed. A firm's absence from a measurement that was never taken is not evidence about the firm.

Whether Bing shows a Copilot answer to a human in the UK remains not determined. Settling it needs a person opening bing.com, which is recorded as outstanding rather than quietly assumed either way. This matters because the question that prompted the work - law firms use Copilot internally, so does Copilot see them? - is a fair question that this index cannot answer. What it can say is that the two Google search-embedded surfaces, which are measurable, show the same firms dominating as the assistants do.

Do the two strata agree?

Both strata were scored on the same 268 prompts, so the comparison is like for like. They are reported separately and never averaged.

They largely agree. Across every frame firm the two orderings correlate at +0.93, and 15 of the two top 20s are the same firms. This says the two Google surfaces we could measure name much the same firms as the assistants. It does not answer the question about Copilot, which this index could not measure (see Bing, above).

Where the two top 20s differ: in the assistants' but not Google's, ALLEN OVERY SHEARMAN STERLING LLP, BURGES SALMON LLP, LEIGH DAY, MACFARLANES LLP and SHOOSMITHS LLP; in Google's but not the assistants', BRISTOWS LLP, CLYDE & CO LLP, Lewis Silkin LLP, Mishcon de Reya LLP and Slater and Gordon UK Limited. It is reported as an observation, not a finding: a handful of names at the edge of a top 20.

Model pinning

Model IDs were pinned on 2026-09-16 after querying each provider's live model list. Note that gemini-2.5-pro, named in the original specification, was found to be retired on that date ("no longer available to new users"), which is why pins are verified against the live API rather than assumed.

4. The frame

Definition: every SRA-authorised law firm whose most recent accounts filed at Companies House show turnover of £15m or more, as on file in the week of 16 September 2026, plus a small number of firms retained by hand and firms whose business is split or moved between registrations (both below). £15m is the index's threshold, chosen because firms of that size must publish a profit and loss account. Small companies need not publish one; the small-company turnover limit is £15m for financial years beginning on or after 6 April 2025 and was £10.2m before that, so some firms below £15m do publish one and are left out by the threshold rather than by the law.

Retained by hand, because they are large firms that publish no usable turnover: HUGH JAMES; LATHAM & WATKINS (LONDON) LLP; LEIGH DAY; SKADDEN ARPS SLATE MEAGHER & FLOM (UK) LLP; SLAUGHTER AND MAY.

The population comes from the SRA's own register of regulated firms, retrieved in a single call on 16 September 2026: 25,076 organisations, of which 8,909 are currently-authorised SRA-regulated firms. Each firm's CompanyRegNo from the register links it to Companies House.

Where the register holds no company number - true of 1,139 authorised companies and LLPs - the firm was matched to Companies House on an exact name and an active company, and read like every other firm (corrections, 19 September). That field matters more than any other. Matching firms to Companies House by name alone is actively wrong: "Mills & Reeve" resolves to a micro-entity shell that files no turnover, "Irwin Mitchell" to a dormant company, while the trading entities are Mills & Reeve LLP and Irwin Mitchell LLP. Published on name matching, both would have shown no revenue.

The frame is 269 firms. 255 carry a turnover figure filed in sterling. Of the 14 that do not, 3 file a profit and loss account in euros or US dollars, 5 file no profit and loss account at all, 2 have their filed figure withheld because it is not the firm's size (see corrections, 19 September), and 4 split or moved their business between registrations and are assessed on the last full year of the whole business (see the amendment of 19 September). The distinction matters: a firm reporting in euros has filed its turnover, it is simply not on a sterling axis, and reporting it as a non-filer is untrue.

Firms whose business is split or moved between registrations

The frame reads one registration at a time. Most firms trade through one, so that works. Two kinds of firm fall through it, however large they are:

• a firm that runs parts of its business through two registrations under common control, so that neither files the whole firm's turnover;
• a firm that has moved its business to a new registration, which has only a part-year on file or files no UK accounts at all, while the old registration is no longer authorised.

Such a firm is assessed on the last full twelve-month year for which the whole business is on file. Where two registrations held it in that year, their turnover is added. Where one registration held all of it, its figure is used. Part-year and longer periods are never used and never annualised. The split or move must be stated in the firms' own accounts; a shared name is not enough.

4 firms are in the frame on this basis:

• Proskauer Rose (London) LLP, GBP 101.1m for the year to 31 October 2023, filed by Proskauer Rose (UK) LLP before the business moved: the UK LLP transferred its business on 5 April 2025 to a Delaware LLP that files no UK accounts.
• Roythornes LLP, GBP 23.2m for the year to 30 June 2023, filed by Roythornes Limited before the business moved: the limited company transferred its whole business to the LLP on 2 January 2025.
• Hodge Jones and Allen LLP, GBP 19.8m for the year to 30 April 2024, its LLP and limited company combined: legal aid work ran through the limited company, and everything else through the LLP, from 15 October 2021 until the LLP acquired the legal aid business on 28 February 2025.
• FBC Manby Bowdler (Midlands) Limited, GBP 16.5m for the year to 31 March 2024, filed by FBC Manby Bowdler LLP before the business moved: the LLP sold its whole business to the limited company on 25 October 2024 and ceased to trade.

These years are older than most firms' latest accounts, so each figure is shown with its year end and with whose accounts it comes from. The firms are in the table and on the largest-firms list. They are not in the turnover correlation, and no change figure is given for them. The evidence for each is in data/split_business.csv; how they were found is in the amendment of 19 September (section 7).

Firms excluded by judgement

KPMG, Deloitte, PricewaterhouseCoopers and Ernst & Young are all SRA-authorised and carry the same constitution code as any law firm, so nothing in the register distinguishes them. Their filed accounts cover the whole accountancy business, which is predominantly not legal work, and their turnover runs to billions. They are excluded by an explicit, published list rather than a silent filter.

Thurcom Limited is excluded for a related reason. It is SRA-authorised as a corporate partner in Skadden's UK partnership, and its own accounts say it has no employees or trade; its GBP 115.2m of revenue is its share of that partnership, not a law firm's fee income. Skadden's UK practice is in the frame in its own right, as a partnership that files no accounts, like Latham & Watkins.

How turnover was obtained

Most large firms file flat scanned PDFs with no text layer at all - 35 to 70 pages yielding under 40 extractable characters - so tagged extraction returns nothing for exactly the firms the index is about. Those figures were read by a vision model constrained to report only a value printed in the document, and to answer NOT_FOUND rather than estimate. Every reading stores the page number, the value as printed and the multiplier applied, so any figure can be checked against the filed document in seconds.

Firms that report in a currency other than sterling

3 firms in the frame present their accounts in euros or US dollars, because that is their functional currency. Bird & Bird LLP is the largest, at EUR 665.6m.

Both figures are published: the turnover exactly as filed, and beside it a sterling figure converted at the European Central Bank euro foreign exchange reference rate for that firm's own accounting reference date. Rates come from the ECB Data Portal, dataflow EXR, daily series EXR.D.<CCY>.EUR.SP00.A, which quotes units of each currency per euro; sterling conversions from a third currency are therefore cross rates through the euro, with both legs taken from the same observation day. Every conversion carries its rate, its series key and the date it was retrieved, so a reader can reproduce the arithmetic.

The converted figure sits in its own column and is never substituted for the filed one, and no converted figure enters the correlation. The revenue/visibility correlation is computed only on turnover filed in sterling, so that adding a conversion cannot move a pre-registered result.

One honest caveat. IAS 21 and FRS 102 translate profit and loss items at the average rate for the period, not the closing rate. Converting at the accounting reference date is simpler, reproducible and tied to one citable observation, but it is not what the firm's own accountants did. The difference is small at these rates and the figures are indicative; the filed figure is the reported one. The ECB also publishes these rates for information only and discourages their use for transactions.

Figures for firms that also appear in the 2020 UK 200 were cross-checked against their 2020 published revenue - 146 of the 255 sterling figures; the other 109 are firms with no 2020 figure to check against. The check is crosscheck_2020.py: each 2020 name reaches a frame firm through the same alias table the scoring uses. An earlier version of this section gave 109 of 254 and one figure outside the bounds; that count was made by hand, could not be reproduced, and is replaced by the scripted one. The check caught Stewarts, which came back as GBP 111,614 where the statement was printed in thousands; transcribing the full income statement settled it - staff costs of "31,096" are impossible in literal pounds - and the figure was corrected to GBP 111.6m.

On the final data 4 figures sit outside 0.4 to 3 times the 2020 baseline:

• Bryan Cave Leighton Paisner LLP, at GBP 211.2m against GBP 685.9m: the accounts give consolidated revenue of GBP 211,197,000 for the 12 months to 31 December 2024 beside GBP 195,320,000 the year before, so the filed figure stands; the 2020 figure may measure a wider scope of business.
• Womble Bond Dickinson (UK) LLP, at GBP 129.0m against GBP 375.9m: two independent reads agree that GBP 128,959,000 is the turnover printed in the consolidated profit and loss account for the 12 months to 31 March 2025, so the figure is published as filed; the 2020 figure may measure a different scope of business.
• SETFORDS LAW LTD, at GBP 62.9m against GBP 17.6m: the accounts print GBP 62,932,797 in whole pounds beside GBP 45,028,020 the year before, so the filed figure is consistent with its own prior year.
• STOWE FAMILY LAW LLP, at GBP 49.3m against GBP 13.0m: the accounts print GBP 49,312,837 in whole pounds beside GBP 37,954,364 the year before, so the filed figure is consistent with its own prior year.

Every figure, with or without a 2020 baseline, also passes the scale-consistency check described under corrections: the printed value, the units the statement is headed in, and the turnover reported must agree.

Firms that file no profit and loss account

Under UK company law, micro-entity, small, dormant and total-exemption filings contain no profit and loss account. Only medium, full and group filings carry turnover. Separately, roughly 2,200 SRA-regulated firms are traditional partnerships or sole practices not registered at Companies House at all — Slaughter and May among them.

Those firms are in the frame — so that an assistant naming them counts as naming a frame firm rather than an outsider — but they carry no turnover figure and are excluded from the revenue/visibility correlation. The number excluded is reported alongside the correlation. They are never silently dropped.

Filing lag

The accounts used end between 31 December 2024 and 30 April 2026 (the split or moved businesses are assessed on older years, from 30 June 2023), so a table published in October 2026 mixes financial years. Every firm's turnover is published with its own made_up_to date, so the reader can see the vintage of each figure rather than assuming a common one.

Periods other than twelve months

12 firms' latest accounts cover a period other than twelve months, usually because the firm moved its year end: 8 shorter, from 9 to 11 months, and 4 longer, from 15 to 18 months. Their turnover is published as filed, with the period length beside it, and is never scaled to a year. The GBP 15m threshold is applied to the figure as filed. A shorter period understates a firm's annual turnover and a longer one overstates it; 1 firm admitted on a longer period would fall below GBP 15m if its figure were scaled to twelve months. A split or moved business, above, is treated differently because there the latest filings do not cover the whole business at all, whatever their length.

Year-on-year change

The change column compares the turnover in a firm's most recent filed accounts with the comparative printed beside it in the same accounts, in the firm's own reporting currency - a euro reporter is compared in euros, so an exchange-rate movement is never mistaken for growth. 212 firms carry a change figure. A figure is withheld, never estimated, when:

• the two periods do not end about a year apart (350 to 380 days; 358 to 372 where the accounts give no period start, because then the length test below cannot run);
• the periods differ in length by more than 5%, where both starts are known;
• the accounts record a period other than twelve months, or a change of year end. Many firms moved their year end to 31 March in 2023-24, leaving an 11-month (or 7- to 18-month) period beside a full year, and comparing the two overstates or understates growth by a fraction of a year;
• the firm merged inside the window, when the change is an accounting event rather than trading. Three firms, named in the table.

The change is descriptive. It is not evidence that visibility caused it: visibility was measured in September 2026 and every change figure ends by 30 April 2026, before it, so the effect would precede the cause. The 2027 edition registers a test that gets the order right.

The registered 72 and the published 268

The published index is all 268 prompts. The 72 registered before any data existed are kept as a check on it.

The top of the table does not depend on the extra questions. 15 of the top 20 firms are the same on the registered 72 and on the published 268, and the two index scores correlate at rho = +0.84 across every frame firm.

The pre-stated check. Before wave 2 ran, the plan named four large firms the new prompts had to reach - firms the registered 72 could not name because no question touched their work - and said that if they were still not named, the new prompts were not doing their job.

What the mix of questions does to the headline. The largest-firms test - how many of the 25 largest firms by turnover are named in fewer than one answer in ten - gives 16 of 25 on the registered 72 and 8 of 25 on the published 268. The assistants did not change; the 72 are 56% local questions and the 268 are 15%. Split by question type within the published set, the answer is not in doubt:

The published finding is stated that way - by who is asking - rather than as a single count that the question mix can move from 16 to 8. For the same reason turnover tracks visibility more closely on the published set (rho = +0.65, n = 255) than on the registered 72 (+0.50): corporate questions are where the largest firms dominate.

Collection. 6,272 wave-2 calls, all eventually successful. 187 Perplexity calls first failed with HTTP 401 when the account ran out of credit mid-run - the same status an invalid key returns - and were re-run once it was topped up. 19 OpenAI calls returned no text because the model spent its whole output allowance on 12 to 22 web searches; all succeeded on retry, at the same settings as every other call. Three calls failed on a local DNS error and were retried. No call was re-run to change an answer: resume retries only calls with no successful record.

5. How a firm is counted as "named"

Firm names are extracted from each response, normalised, and matched against an alias table built from each firm's registered name, trading names and previous names.

• An exact alias match is accepted.
• A fuzzy match requires agreement between two scorers — token_set_ratio >= 90 and token_sort_ratio >= 85. One scorer alone is not sufficient, because token_set_ratio scores any token subset as 100 by construction: the single word "Mills" would otherwise score 100 against "Mills & Reeve".
• Anything that does not match goes to a review queue, and each name in it is decided against the SRA register by a published rule: outsider (a currently authorised firm that is not in the frame), not a firm (prose, directory and accreditation labels, words taken from our own questions), or held for review (no confident match - mostly Scottish and overseas firms, which the register does not cover). A name held for review is never scored and never named anywhere in the index. These decisions are made by rule, not by a person reading each name; the index cannot be published while any name is undecided.

Firms named that are not in the frame are recorded separately as outsiders.

The unit of truth

Every firm-level statement is expressed as "named in X of Y runs", where Y is the number of calls actually executed, not the number that happened to produce a recognisable firm name. Runs in which an assistant refused, hedged, or named only unrecognised firms remain in the denominator.

How the score is built

Visibility. Each answer gives a firm a weight by where the firm first appears in it: 1 for first, 0.8 for second, 0.6 for third, 0.4 for fourth, 0.4 for fifth, and 0.2 for sixth or later. A firm named twice in one answer is weighted once, at its first position. A firm's weighted presence is the sum of its weights divided by the number of answers actually collected, and the index rescales weighted presence so that the most visible firm scores 100.

Named in. The number of answers, out of all answers collected, that name the firm at least once, with a 95% Wilson confidence interval.

Numbered or banded. A firm carries a numbered position only when it is named in 30 or more answers. Below that, the interval is too wide for a position to mean anything, and the firm is shown in a band by how often it was named.

Stability. The eight runs are split into runs 1-3 and runs 4-8, each half is scored on the answers collected in it, and the stability figure is the number of firms in both halves' top 20.

Agreement between assistants. Each assistant's top 20 is taken by how often it names each firm; agreement is the mean Jaccard index across the six pairs of assistants (the number of firms two lists share, divided by the number in either). The same Jaccard index compares the corporate and local top 20s.

6. What we do not claim

• This is not a quality ranking. It does not measure legal skill, outcomes or client service.
• Assistants are non-deterministic. Eight runs measure a distribution, not a fact. Run-to-run stability is reported.
• Search results change. The dataset is a snapshot: the assistants' answers were collected between 16 September 2026 and 18 September 2026, and Google's between 17 September 2026 and 17 September 2026 (UTC).
• Self-citation (whether an assistant cited a firm's own website) is computed only where the firm's domain comes from its SRA record, and is not determinable where a provider returns redirect URLs rather than publisher domains. It is not part of the published table.

7. Pre-registered assessment criteria

Thresholds for deciding whether the data supports a particular finding were fixed on 16 September 2026, before any data existed, and are published in full. Amendments made after that date are listed below with their date and reasoning. No threshold has been changed after seeing results. The set of questions the thresholds are applied to was changed after results were seen - from the 72 registered prompts to all 268 - and that amendment is recorded below.

The thresholds, as registered:

Amendment log

2026-09-16 — revenue bands replaced by continuous turnover. Made before any data existed. The original criteria bucketed firms into four bands (Top 25 / 26–50 / 51–100 / 101–200) derived from 2020 revenue. Bands were a proxy for current revenue we did not have. With filed turnover available per firm, four buckets over roughly 250 firms create large numbers of ties, and ties attenuate the Spearman correlation and discard statistical power. The correlation is therefore computed on continuous turnover. Two tests written in band language were restated in turnover terms with their numeric thresholds unchanged: "101–200 firms beating the Top 25 median" became "firms in the bottom half by turnover beating the median index of the 25 largest by turnover", and "Top 25 named in under 10% of runs" became "the 25 largest by turnover, named in under 10% of runs".

Consequences, stated openly. Widening the frame is not neutral to the outcome. It adds low-revenue, low-visibility firms, which reinforces monotonicity and tends to raise the correlation — that is, it makes the "revenue does not predict visibility" finding less likely, not more. The frame-coverage gate is a rate (30%), so the absolute bar rises as the frame widens, and outsider share shrinks mechanically for the same reason.

2026-09-18 - the published index moves from the registered 72 prompts to all 268. Made after data existed, and recorded here for that reason. Results on both sets had been seen, including that the largest-firms test then gave 19 of 25 on the registered 72 and 8 of 25 on the 268 (on the corrected data of 19 September: 16 of 25 and 8 of 25).

Reasons. The 72 could not reach whole practice areas, so large specialist firms were published as never named because no question touched their work, not because assistants ignore them - Fragomen went from 0 of 2,304 answers to 196 once asked about immigration. The 268 gave 187 firms enough mentions to carry a numbered position, against 109 on the 72 (199 and 119 on the corrected data), and one index is easier to read than two.

What did not change: every numeric threshold keeps its 16 September value. They are applied to a larger set of prompts. On 18 September the assessment on the published 268 returned CENTREPIECE, spine D (strong); on the registered 72, spines D and E (strong). Re-run on the corrected data of 19 September, it returns CENTREPIECE, blockers: none; spines supported: B. The directories are the new front door (strong); D. It depends entirely who is asking (strong). On the registered 72: D. It depends entirely who is asking (strong).

Consequences, stated openly. The question mix moves from 28% corporate / 17% consumer / 56% local to 54% / 31% / 15%, and tests that depend on who is asking move with it. The published findings are therefore stated by question type; the section on the registered 72 gives both sets of results so a reader can judge the change. Choosing a basis after seeing which one supports a finding is the risk pre-registration exists to prevent. This change weakens the registered headline rather than strengthening it.

A prediction in the 16 September amendment was wrong. It said widening the frame would tend to raise the correlation between turnover and visibility. It fell: from +0.52 on the first 171-firm frame (166 firms with sterling turnover) to +0.45 on the 268-firm frame, both on the registered 72 prompts. (An earlier version of this entry gave +0.55 on a 179-firm frame; no stored assessment holds that figure, and the archived one is quoted instead. On the corrected frame of 19 September the registered 72 give +0.50.) The firms added - mostly GBP 15-40m, read from filings that had been downloaded and not yet read - were low-revenue but more visible than the prediction assumed.

2026-09-19 - firms whose business is split or moved between registrations. Made after data existed, and recorded here for that reason. It began with a question about one firm: why Hodge Jones & Allen was missing. Assistants named it in 571 of 8,576 answers, and neither of its two registrations filed GBP 15m, though together they turned over GBP 19.8m.

The rule is set out in section 4. So that it was not a rule for one firm, it was applied to the whole register in two passes, neither of which looks at how often a firm is named:

• every SRA registration, of any status, was grouped by its registered name with legal and generic words removed (split_business_sweep.py); 88 of 8,809 groups hold a currently authorised registration outside the frame and another registration with a company number;
• every authorised registration outside the frame was linked to other registrations through its own trading and previous names (split_business_links.py): 143 linked companies, 3 with full, medium or group accounts since 2023;
• every filing since 2023 that could carry a profit and loss account was listed (out/_audit/split_business_candidates.csv). A group, or a company linked in the second pass, was read where a registration other than the authorised one filed full, group or medium accounts, or where the authorised registration's own accounts had not been read: 22 filings in 8 groups (out/_audit/split_business_readings.jsonl). The rest filed abridged accounts, which only small companies may file and which cannot show GBP 15m, with no two registrations filing for the same year; or had accounts only on the authorised registration, already read and under GBP 15m;
• each candidate was then checked against the filed pages: that the registrations were one business, the date of the split or move, and the figure itself.

4 firms qualify and join the frame (section 4). In the other 4 groups the whole business was under GBP 15m in its last full year, or its accounts carry no profit and loss account. They are not named. The check also corrected one reading: a three-month period in the accounts of Roythornes' former company had been read as GBP 41.2m, which is the comparative column for the 18 months before it; the three-month figure is nil. That reading was never used in the frame.

What did not change. No threshold moved. The turnover correlation keeps its registered basis - each frame firm's own latest accounts, filed in sterling - so these firms are not in it: +0.655, n = 255. With their whole-business figures added it would be +0.647, n = 259.

Limits. The sweep finds registrations linked by name or by the register's own trading and previous names. A business that moved to a registration with an unrelated name, and no trading name linking the two, would not be found.

Effect. The frame moved from 265 to 269 firms. Proskauer Rose (London) LLP is named in 166 of 8,576 answers (position 116); before the rule its mentions were counted as naming a firm outside the index; Roythornes LLP is named in 45 of 8,576 answers (position 162); before the rule its mentions were counted as naming a firm outside the index; Hodge Jones and Allen LLP is named in 571 of 8,576 answers (position 39); before the rule its mentions were counted as naming a firm outside the index; FBC Manby Bowdler (Midlands) Limited is named in 12 of 8,576 answers (shown in a band). Adding a firm to the frame does not change any other firm's score, so this rule moves no other firm and not the correlation; the extraction correction made the same day does, and its effect is recorded with it (corrections, 19 September).

Corrections made during the build, and what each one changed

These are recorded because the figures moved. Each was found by a check that disagreed with another source, not by inspection, and each is reproducible from the stored evidence.

2026-09-17 - the SRA and Companies House spell company numbers differently. The SRA register strips leading zeros from numeric company registration numbers: 1,938 of its records hold a number shorter than eight characters, so the register's 5743784 and Companies House's 05743784 are the same company and did not join. Every firm whose registration number takes that form was therefore screened, downloaded, read and then dropped, because the turnover could not be matched back to the firm.

Nine firms were restored by normalising both sides to the Companies House eight-character form, among them Keystone Law (GBP 115.2m), Fletchers Solicitors (GBP 76.9m) and Boodle Hatfield (GBP 43.6m). With the other frame changes made the same day, the frame moved from 177 to 183 firms, and further reading of filed accounts took it to 268.

This also settles a question the project had already met once. Fletchers Solicitors matches the register by name at a perfect score to FLETCHER DERVISH SOLICITORS, because that firm holds "FLETCHERS" as a previous name. The company number resolves it correctly to SRA 469026, its own registration. Name matching would have attributed a GBP 77m firm to the wrong practice; this is why company numbers are used and names are not trusted.

2026-09-17 - a model reported a multiplier and did not apply it. The vision reader asks for the figure as printed, the units the statement is headed in, the multiplier those units imply, and the turnover with that multiplier already applied. On 12 of 239 documents the first three came back right and the fourth came back unmultiplied - Linklaters as 2,278.9 rather than GBP 2.279bn, Irwin Mitchell as 299,865 rather than GBP 299.9m.

The error was silent in both directions: the figures were internally consistent, and both firms simply fell below the GBP 15m threshold and left the frame. It surfaced because Irwin Mitchell had been the second most visible firm in the index and disappeared from the table.

The reader now checks the model's arithmetic against the model's own observations. printed_value and units are read off the page; turnover is a calculation. Where they disagree and the stored figure is the printed one, the figure is recomputed from the observations and the correction is written into the record's note. Where they disagree in any other way the figure is left alone and flagged, because a silent third answer would be worse than either. All 12 were corrected from the cached fields without re-reading any document.

Affected firms, all verified against the corrected figure: Linklaters, Freshfields, Clyde & Co, DAC Beachcroft, Gowling WLG, Irwin Mitchell, Stephenson Harwood, Knights, Stevens & Bolton, Co-operative Legal Services, and two accountancy practices already excluded from the frame.

2026-09-17 - one business was counted twice. Hage Aaronson Limited files a group turnover of GBP 35,273,121 and Joseph Hage Aaronson & Bremen LLP files the same figure on an entity basis. Both are separately SRA-authorised with different company numbers, so both entered the frame and the same GBP 35.3m appeared twice. The curated brand map now drives the collapse as well as admission, and the two are one row.

2026-09-18 - generic words were being matched as firms, and genuine mentions were being lost to brackets and footnotes. Two defects, found while folding in the wave-2 answers, and fixed together because each exposed the other.

Generic register names. The SRA register records trading names, and the name normaliser strips words such as "the", "law", "solicitors" and "firm". A trading name such as "The National Law Firm" therefore reduced to a single ordinary word, and that word matched every answer containing it: "FCA investigations", "SFO investigations" and the bare word "investigations" were all filed as mentions of one small firm, "conveyancing" and "immigration" as another, "Chambers UK and Legal 500" as a third. 930 outsider verdicts rested on a one-word key.

The line is now drawn from the register itself. A single word counts as a firm's name only if it appears in six or fewer authorised registrations and is not a word from our own prompts; real one-word names sit at one to six (Fieldfisher 3, Shoosmiths 5, Kennedys 5), ordinary words start at twelve. 1,269 register keys were dropped as generic, the outsider matching threshold was aligned with the frame matcher's (token_sort_ratio 85, up from 80), and every machine-made outsider verdict was re-decided under the new rule. Outsider verdicts fell from 2,948 to 1,851. None of this can change a frame firm's own score - an outsider verdict cannot take a name away from a frame firm - but it changes the outsider counts behind the coverage figures.

Brackets and footnotes. Assistants attach things to a name that are not part of it: citation markers ("Addleshaw Goddard[2][3][7][9]") and a trailing bracketed qualifier, closed or not ("Anthony Gold (South London)", "Baker McKenzie (Band 1"). Left on, they pulled the match score under the threshold and a plain mention of a frame firm was parked for review instead of counted. The extractor now strips citation markers and a trailing bracket, offers the bracket's contents as a candidate in their own right (so "VWV (Veale Wasbrough Vizards)" still reaches the firm), and splits "X or Y" into two candidates, since no firm's name contains the word "or". "And" and "&" are not split: Slaughter and May, Mills & Reeve and Bird & Bird all contain them.

One weight per firm per answer. The position-weighted score summed a weight for every mention row, so a firm named twice in one answer under two spellings was rewarded twice. It now takes the firm's earliest position in each answer.

Effect, measured like for like on the registered 72 prompts. Against the table as first published, 131 frame firms gained mentions and 4 lost them. The gains are the recovered bracket-and-footnote mentions, concentrated on consumer and regional firms where a city in brackets is common: Irwin Mitchell (357 to 393), Stowe Family Law (175 to 207), Stewarts Law (229 to 259), Burges Salmon (111 to 140), Thompsons Solicitors (147 to 171). 19 of the top 20 are the same firms.

2026-09-18 - about thirty growth figures compared a full year with a part year. Scanned accounts give period ends but not starts, so the period-length test could not run on them, and a wave of firms moved their year end to 31 March in 2023-24. The result was growth computed across a 12-month year and a 7- to 18-month comparative - one firm shown at +67.6% against a 7-month prior period, another at -26.5% against an 18-month one. The model reading the accounts had recorded each period length in its note; those notes now withhold the figure, and the length tolerance where both starts are known was tightened from 15% to 5%, because 15% admitted an 11-month period. Growth figures fell from 244 to 214. A withheld figure shows its reason in the table, and its prior-year cell is left blank rather than showing a part-year figure.

2026-09-19 - names credited to the wrong firm, one brand listed as several firms, and large firms the frame could not see. Found by an audit of every name credited to a frame firm and every row of the table, each flag checked by a second reviewer against the stored answers and the SRA register. The verdicts, with their evidence, are kept in out/_audit/.

Ordinary words credited to a firm. The register records descriptive trading names, and the normaliser strips "the", "law" and "firm", so a name such as "The Yorkshire Law Firm" (Ison Harrison) reduced to "yorkshire" and every answer that said "Yorkshire" was counted as naming the firm. The same mechanism credited "DBA" (a damages-based agreement) to Excello Law, "property" to Property Legal, "serious injury" to Fletchers, "housing disrepair" to Antony Hodari and "Resolution" (the family lawyers' association) to DWF.

Different firms with similar names. The fuzzy rule accepts a one-letter difference, so "Wilson Solicitors LLP" (London, SRA 520695) was credited to Wilsons Solicitors LLP (Salisbury), "Neilsons" (Edinburgh) to Nelsons and "Fosters" (Norwich) to Forsters.

463 name-and-firm pairs are now barred, recorded in data/alias_exclusions.csv. 271 of them apply only to the questions where the wrong reading occurred, because the same words also name the frame firm elsewhere: "Wilsons Solicitors" is the Salisbury firm in one answer and a Leeds conveyancer in eleven others.

Two firms in one string. "Kingsley Napley and BCL Solicitors" was one candidate, and only one firm was credited. A candidate is now split when it does not match a firm as a whole and every part matches a different firm exactly. Names that contain "and" or "&" - Slaughter and May, Mills & Reeve, Bird & Bird - match whole and are never split.

One brand, several rows. Sister registrations with a geographic suffix, or a curated brand name that did not match the full registered name, escaped the collapse onto one row. Nine brands were affected - BCLP, DAC Beachcroft, Fieldfisher, Goodwin Procter, HSF Kramer, HFW, Hogan Lovells, Watson Farley & Williams and Weil - and three of them had a large registration published as never named while the brand's mentions went to a smaller one: Herbert Smith Freehills Kramer Global LLP, Goodwin Procter (UK) LLP and WFW Global LLP. Each brand is now one row, headed by the registration with the largest filed turnover, as for every other firm.

Figures that were not a law firm's size. Thurcom Limited is SRA-authorised as a corporate partner in Skadden's UK partnership; its accounts say it has no employees or trade, so it is excluded, like the Big Four, by a published list. Kirkland & Ellis LLP files accounts for a small euro-reporting entity with branches in Paris and London; Kirkland's London practice is a partnership that files no UK accounts. The firm stays in the index, but its filed figure is withheld (data/turnover_withheld.csv).

Firms the frame could not see. The frame joins the register to Companies House on the register's company number, and 1,139 authorised companies and LLPs have none. Each was matched to Companies House on an exact name and an active company (screen_missing_company_numbers.py); those whose latest accounts carry a profit and loss account were read like every other firm. 10 cleared GBP 15m and joined the frame: GATELEY PLC (GBP 124.7m), HCR Legal LLP (GBP 93.2m), RWK Goodman LLP (GBP 71.5m), EXPRESS SOLICITORS LIMITED (GBP 64.9m), SETFORDS LAW LTD (GBP 62.9m), BOND TURNER LIMITED (GBP 58.8m), DUNCAN LEWIS SOLICITORS LTD (GBP 31.8m), RRADAR LIMITED (GBP 22.3m), CLINICAL NEGLIGENCE SERVICES LIMITED (GBP 18.6m), SPECTARE LIMITED (GBP 17.5m). FRAGOMEN (GB) LIMITED, whose consolidated accounts include Fragomen LLP, now heads the Fragomen row.

Register numbers that belong to another company. Two SRA records carry the company number of a different, dissolved company. Carter & Carter Solicitors Ltd's number is an adult-education company dissolved in 2014, whose accounts to July 2006 had been published as the firm's latest turnover; the firm's own company files accounts with no profit and loss account, so it leaves the frame. Bates Wells & Braithwaite London LLP's number is a Sudbury LLP dissolved in 2013; the firm's own LLP (OC325522) was found by exact name and read like every other firm. Accounts made up before 2023 are now never used as a current figure.

A figure not shown to be legal fee income. CD&R Galaxy UK Opco Limited, which trades as Vialto Partners, stays in the index but its GBP 133.5m is withheld: the SRA register records immigration as its only area of legal work, and Companies House also classifies the company as a holding company and under other professional activities. It is the same reasoning that excludes the Big Four.

Merger labels that were not merger years. Five firms had their change figure suppressed as merger-affected, from a list typed by hand. Four of those years were not merger years: Hogan Lovells Cadwalader, HSF Kramer and Ashurst Perkins Coie file accounts for years that ended before their combinations, and Eversheds Sutherland's two years are twelve-month years of the same business. Their change is now published like any other firm's. The same review found, in the firms' own accounts, two combinations the list had missed: Morton Fraser MacRoberts' members' report calls the year to April 2025 its first full year as the merged firm and gives like-for-like growth of 7% (the unadjusted figure was +39.3%), and Rothera Bray's accounts record the acquisition of Massers on 30 September 2024. Their change is now suppressed, with the reason shown, as is A&O Shearman's.

Effect. The frame moved from 268 to 265 firms. 156 frame firms gained mentions and 17 lost them; 18 of the top 20 are the same firms. The turnover correlation moved from +0.58 to +0.65.

2026-09-19 - firm names written as headings, or listed after a colon, were not read. Found by a check of the answers in which no frame firm was credited. Names were taken from bold text and list items, and the whole answer was swept for capitalised names only when it had neither. Gemini writes most recommendations as a heading per firm - "### 1. Family Law in Partnership (FLiP)" - with bold labels beneath it ("Location:", "Best for:"), so the labels were read and the firms were not. The same gap lost names listed in a line after a colon ("Also highly rated: Kingsley Napley, Russell-Cooke, Osbornes ...") in every engine.

Three changes, applied to every answer:

• headings are read as names, each side of a colon or spaced dash offered separately;
• a line that enumerates - two or more commas - is swept for names in every answer. A name found only that way counts when it is a known firm name exactly, never on an approximate match, and is otherwise discarded rather than queued;
• a candidate made only of generic words ("Legal", "Solicitors") can no longer match a firm approximately. "Legal", mostly from "Legal 500", had been credited 603 times to one firm outside the index; it is now credited only where its own name is written (2 answers).

Reading more of each answer also meant more strings reached old matching decisions made for single words. Two reviews read the stored answers before anything was published: every name-and-firm pairing the change created (234), and every firm outside the index credited through a single word (726, 11,850 mention rows), each verdict challenged by a second reviewer. They found ordinary words, headings and places credited to firms that share the word - "Manchester" to a firm called Manchesters, "Midlands", "Cambridge", "Liverpool", "STEP" (the trust practitioners' society), "True", "The Premium" - and fragments of other firms' names: "Slate" from Skadden, Arps, Slate, Meagher & Flom, "Watkins" from Latham & Watkins, "Shearman" from A&O Shearman. In each such case the index firm was already credited in the same answer. 270 of the single-word pairings were wrong in every answer (5,179 rows) and 32 in some. 3,229 of those 5,179 rows were already counted before the change, so the published share of names going to index firms had been understated since the first build. 399 exclusions were added to data/alias_exclusions.csv, each with its evidence (261 of them limited to the questions where the words were wrong, because the same words also name the firm elsewhere).

Barring a word from one firm was not enough: the automated register step then matched it to the next firm that shared it ("Conveyancing", barred from one conveyancing firm, reached another). So a word decided not to be a firm is now refused by the matcher outright, approximate matches included, unless it is also an index firm's own name. Checked against every stored answer before it was applied, that refusal removes 1 credit to an index firm (Fox & Partners and Herbert Smith Freehills Kramer - Herbert Smith Freehills Kramer Global LLP).

Effect. Answers crediting no frame firm: ChatGPT 66 to 60, Claude 74 to 64, Gemini 322 to 49, Perplexity 77 to 54. Frame-firm mention rows rose from 83,896 to 92,909. Share of names going to frame firms: corporate 89.1% to 91.8%; consumer 64.3% to 73.2%; local 44.2% to 49.0% (this also includes the four firms admitted the same day). The turnover correlation moved from +0.648 to +0.655 (n = 255); engine agreement from 0.50 to 0.58; top-20 stability stayed at 19; 20 of the top 20 are the same firms.

8. Reproducibility

The evidence log retains, for every call - 8,576 to the assistants and 5,588 to the search surfaces - the prompt ID and full prompt text, the engine, the exact model ID, the run number, a UTC timestamp, whether search was invoked, the complete response text, every cited URL, and any error.

For every firm, the Companies House response and the accounts document itself are stored alongside the parsed turnover, with the source URL and the UTC time it was fetched. Each turnover figure records how it was obtained: ixbrl, vision_llm, ixbrl_text_llm, does_not_file or unavailable.

Everything downstream is rebuildable from these two stores.

9. Attribution

The UK Law Firm AI Visibility Index 2026 includes data supplied by the Solicitors Regulation Authority.

Company information is supplied by Companies House under the Open Government Licence.

---

Crescat Digital. Questions about any figure in this index, including a correction to your own firm's entry, go to [contact address to be confirmed]. We will show you the filed document and the date we read it.