How accurate are AI business recommendations? The UK evidence, measured (2026 ledger)

Last Updated on September 5, 2026

The question this answers: how accurate are AI business recommendations in the UK, and what does the evidence say?

Ask ChatGPT, Gemini or Google AI Mode to recommend a plumber, an accountant or a dentist and you get an answer, not a list. That answer is now a buying decision for a lot of people. Almost nobody is measuring whether it is right.

Whito has been measuring it since June 2026. This page is the whole ledger in one place: every original measurement, with its base, its collection date and a link to the study it came from. Nothing here is borrowed from another publisher. Every figure was collected by Whito and is already published on its own page.

The headline numbers

1. Do the engines agree with each other?

No. Disagreement turns up in every sector tested. If three engines asked the same question name three different businesses, at most one of them can be the best answer.

Sources: Whito original studies, each linked. Bases differ and the metric differs between rows. See Known limits.

The metric changes between rows. The June test measured whether engines picked the same top business. The sector tests measured whether a business was named at all by more than one engine. Those are different questions with different bases, and the numbers must not be lined up as a trend. What survives both framings is the direction: agreement is the exception in every sector tested.

2. Are the recommended businesses real?

Two figures on this site look contradictory until you read what each one tested. Both are true. They answer different questions, and the gap between them is the point of this section.

Sources: Whito original studies, each linked. The last two rows are the same 728 recommendations under two different tests.

Why 51% and 5.2% are both correct

A name-match test asks a narrow question: does a company of this name exist on the Companies House register, active, in or near this town? A great many legitimate UK businesses fail that test for reasons that have nothing to do with honesty. Sole traders are not on the register at all. Trading names differ from registered names. A real one-person plumbing business fails a name match every time.

So the name-match family of figures, 49% to 58% unmatched across four studies, measures how far AI recommendations sit outside the company register. It does not measure how many recommended businesses are fake.

The 5.2% figure comes from the same 728 recommendations put through four passes, including a town-aware re-rank, successor-company checks and a public trading footprint test. It measures something much stricter: after looking properly, how many could not be shown to be a real business trading in the town the engine named. That answer is 38.

Whito publishes 5.2% as a floor. Thirty-one rows were coded as close calls and every close call was resolved towards verifiable. Websites were not fetched row by row.

Two further findings sit alongside this and neither is a matching artefact:

• Five of 60 Google AI Mode answers named a business whose only Companies House registration was dissolved, one of them since 2017 (29 July 2026).
• Two of the ten roofing review profiles behind AI’s most trusted recommendations belong to dissolved companies. Eight of the ten score 9.9 out of 10 or better (July 2026).
• 29 of 100 limited-company listings on five UK business directories could not be resolved to an active company. Hotfrog failed on 11 of 20, Yell passed all 20 (July 2026).

3. What puts a business into the answer?

Every AI answer is assembled from sources. Whito logged which ones.

Sources: AI citation sources study, Whito AI Visibility Index, what sources AI trusts. Two different engines and two different panels, listed separately for that reason.

One test looked at whether AI favours household names. It did not. Of 45 businesses named by Google AI Mode and Perplexity across seven local buying questions, 40 were independents, 88.9%, and only 5 were household names (August 2026).

4. Can the engines see you at all?

An engine cannot recommend what it cannot read. Almost no UK business blocks AI crawlers. Most have published nothing for one to read.

Sources: AI readiness audit, AI Door Test, trades structured data study, CQC register study, price transparency study. These are five separate samples and must not be added together.

The register that the NHS, councils, comparison sites and AI assistants all read is missing the website field for nearly half of England’s registered care services. When Whito spot-checked ten blank home care entries, three had a working website the register did not hold. A business can have a working website and still be invisible at the point where the decision is made.

5. What is an AI citation worth?

Whito sells research on AI visibility, so here is what AI visibility has been worth to Whito.

Source: Do AI citations drive traffic?, Bing Webmaster Tools, read 29 August 2026. Both sides of the ratio cover the same window.

Anyone selling AI visibility as a traffic channel should publish this number for their own site. Citations are not clicks. They may still be worth having, and a business should know which of the two it is paying for.

Method

Every figure on this page was collected by Whito and is published in full on its own study page, with that study’s own method section. The pattern across all of them:

• One verbatim prompt per test, run once, with no retries and no cherry-picking. Every answer logged, including the unhelpful ones.
• Fixed panels. Twenty UK towns from Reading to Belfast for the Index, six UK cities for the sector series, fifteen towns for the July trade studies.
• Verification against the public record. Companies House for company status, and the statutory regulator where one exists, such as the Care Quality Commission for care services and the General Dental Council for dentistry.
• Named businesses are reported as visibility facts only. Whito makes no claim about the quality of any named firm’s work, and has no commercial relationship with any of them. Where a study found a failure, the failure is the engine’s, not the business’s, and firms are not named unflatteringly.
• Collection dates are stated everywhere, because a different day produces a different list. That variability is itself one of the findings.

Known limits

• These are separate samples and cannot be added together. 120 sites in July, 49 sites in June, 200 sites in August, 159 answers in August, 728 recommendations in August. They are different panels answering different questions.
• The agreement metric is not consistent across rows. The June test measured top-pick agreement. The sector tests measured whether a name appeared at all. Do not read the two as one trend line.
• Name matching is not a fraud test. Sole traders are not on the Companies House register, and trading names differ from registered names. The 49% to 58% unmatched family measures distance from the register, nothing more.
• 5.2% is a floor. Close calls were resolved towards verifiable and websites were not fetched row by row.
• One run per prompt. AI answers vary between runs. These studies measure what the engines said on the dates stated, not how repeatable those answers are, with one deliberate exception: the plumber repeat test, which asked the same question thirty times and got 51 different businesses.
• Engine coverage is uneven. The Index is Google AI Mode alone. The sector series is three engines. The Reddit finding is ChatGPT alone. Each row says which.
• None of this proves causation. Nothing here shows that any property of a website caused an engine to name a business.
• One website is not a study. The citations-per-click figure is Whito’s own traffic and should not be generalised to anyone else’s.

How to cite this page

This research is free to reuse with attribution and a link. No permission needed.

If you are quoting a single figure, cite the study it came from rather than this page, and take the collection date with it. Every row above links to its source. Journalists checking a number before publication can write to hello@whito.co.uk and we will point you at the raw log.

Correction log

A figure on this page may be corrected. It will never be quietly smoothed. Any correction changes the number and the date and is listed here.

• 3 September 2026, Edition 1. First publication. No corrections yet.

Common questions

How accurate are AI business recommendations in the UK?

On Whito’s measurements, the businesses named are usually real but the choice is close to arbitrary. Of 728 businesses Google AI Mode recommended across 20 UK towns on 22 August 2026, 38, or 5.2%, could not be confirmed as trading in the town named. But three engines asked the same question rarely agree: across six cities they named 109 different accountancy firms and agreed on exactly one.

Do ChatGPT, Gemini and Google AI Mode recommend the same businesses?

Rarely. Two of three engines agreed on the best UK tradesperson in 1 of 12 identical questions on 30 June 2026, which is 8%. In the August sector series, 85% of accountancy firms, 74% of law firms and 66% of dental practices were named by only one engine. Vets had the highest overlap tested, at 45% named by more than one engine.

Does AI recommend businesses that no longer exist?

Sometimes. Five of 60 Google AI Mode answers collected on 29 July 2026 named a business whose only Companies House registration was dissolved, one of them since 2017. Two of the ten roofing review profiles behind AI’s most trusted recommendations belong to dissolved companies, and eight of those ten score 9.9 out of 10 or better.

What sources do AI engines use to recommend a local business?

The business’s own website above everything else. It was cited in 143 of 159 Google AI Mode answers logged on 22 August 2026, which is 90%, with Google reviews or Business Profiles in 128, or 81%. Trustpilot appeared in 2 answers and no news outlet appeared in any. Checkatrade sat behind 50 of the 60 trade answers.

Should a small business block AI crawlers?

Almost nobody is, and almost nobody has decided either way. Only 1 of 49 UK business websites actively blocks any AI crawler and 92% give AI crawlers no instruction at all, measured in June 2026. The more common problem is having nothing readable to offer: 92% of 120 independent sites had no AI crawler policy, about 6% used correct business structured data, and 35% of top-rated tradespeople had no website at all.

Do AI citations bring traffic?

Not in any proportion to their number. Microsoft Copilot and its partners cited whito.co.uk 17,400 times across 127 pages in the three months to 28 August 2026, while Bing sent 167 clicks in the identical window. That is 104 citations for every click. One website is not a study. That figure is Whito’s own traffic and nobody else’s.

The takeaway

AI engines are not lying about UK businesses. They are guessing between them, and they are guessing from a very thin record: your own website, your Google profile, and whichever review platform dominates your trade. Three engines asked the same question name three different firms because there is not enough public evidence to distinguish them.

That is a structural problem, and it has a structural fix. Publish the things a machine can check: your registered company details, your regulator number, your prices, your service area, in plain readable HTML on your own site. Nearly half of England’s registered care services do not have a website recorded on the register that decides who gets recommended. Fixing that beats any tactic.

Whito in Numbers holds the operating figures behind this site, including the unflattering ones.

Jacob Whitmore Founder and Lead Analyst

Jacob Whitmore founded Whito. He works on what UK businesses are actually charged for marketing, and on getting those numbers published with the date they were checked.

See Full Bio