Last Updated on July 10, 2026
When an AI calls a tradesperson “trusted”, it is quoting a review score. So we went underneath the scores.
We took the roofing firms that ChatGPT, Gemini and Google AI Mode presented as most trustworthy in our 15-town study, pulled their full review profiles on the trade platforms the AI quotes, and cross-referenced every one against Companies House. Ten firms got the deep treatment: total reviews, ratings, membership dates, and review velocity over the last 12 months.
We are not naming any of them, because the point is not ten roofers. The point is what “trusted” is actually made of.
Pattern one: review profiles outlive the companies behind them
Two of the ten profiles belong to firms whose registered companies are dissolved. Not dormant. Dissolved, in one case since 2017.
Both profiles are very much alive. One holds a perfect 10/10 from 137 reviews and collected 11 of them in the last 12 months, the newest six days before we looked. The other holds 10/10 from 223 reviews, 30 of them in the last year, on a platform membership that started nine months before the company at its own advertised address was struck off.
There can be innocent explanations. Plenty of tradespeople close a limited company and carry on as sole traders. But the customer reading that profile cannot tell, the platform does not flag it, and the AI quoting the score certainly does not check. A review profile is not a business. It just looks like one.
Pattern two: perfection is the baseline
Of the ten profiles, eight score the equivalent of 9.9 out of 10 or better. Two are literal 10/10s. One is a flat 5.0 out of 5, another a 4.99.
Think about what a 10/10 from 137 customers means: not one person, ever, knocked off a single point. Across an entire trade, that is not a distribution of quality. It is a distribution of who asks whom for reviews.
When nearly everyone is perfect, the score contains almost no information. The AI tools treat 9.9 versus 9.5 as a ranking signal. The honest reading is that platform review scores in the trades have ceased to discriminate at the top, and the real signal moved elsewhere.
Pattern three: perfect scores, tiny denominators
One firm was described by Google’s AI as fully vetted with a “4.99/5.0 rating”. The 4.99 is real. It is built on twelve reviews.
Another was praised for a “perfect 5.0/5.0 across multiple categories”. Fourteen reviews.
Neither AI answer mentioned the sample size. And in the first case, Google’s own map card was simultaneously showing the same firm at 4.0 from its 8 Google reviews. Which brings us to the next pattern.
Pattern four: the same firm has a different score on every platform
One Coventry roofer is 4.9 on Google and the equivalent of 4.6 on Checkatrade. A Reading roofer is 4.0 on Google and 4.99 on TrustATrader. The AI answers blend these into a single confident “highly rated” verdict, and in every case we checked, the number quoted was the more flattering one.
That is not the AI lying. It is the AI doing what it does: repeating whichever source it happened to read. Every count we could check matched the platform exactly, including the dissolved firms’ counts. The numbers are accurate. The verification is absent. Those are different failures, and the second one is worse.
Pattern five: stale totals dressed as current form
One firm was recommended with “over 200 reviews and high recent scores”. The total is right: 204 reviews accumulated since 2015. The recent part is not: six reviews in the last 12 months. A profile that averaged twenty reviews a year for a decade has slowed to one every two months, and the AI read the odometer, not the speed.
Across our ten firms, review velocity in the last 12 months ranged from 6 to 40. These firms are presented to consumers as equivalently “trusted”. Their current trading activity differs by a factor of nearly seven.
Pattern six: platform trust and legal existence are strangers
Two of the ten profiles, with 176 and 204 reviews respectively and platform memberships dating to 2020 and 2015, belong to firms we could not match to any registered company at all. Again: possibly sole traders, entirely legal. But it means the two trust systems, review platforms and the companies register, do not talk to each other. A firm can be nine years deep in five-star reviews and legally untraceable, or legally dissolved and still winning “most trusted” this week.
What the law does and does not fix
Since April 2025, fake reviews are illegal in the UK under the Digital Markets, Competition and Consumers Act, and platforms must take reasonable steps to prevent them.
Here is the uncomfortable finding of this study: nothing we found requires a single fake review. Dead-company profiles, twelve-review perfection, cross-platform score shopping and stale totals are all achievable with 100% genuine reviews. The law bans fabrication. It does not require the resulting numbers to mean what consumers think they mean.
How to actually read a review profile
Four checks that take two minutes and would have caught everything above. Check the date of the newest reviews and how many arrived in the last year, not the total. Check the sample size behind any “perfect” score. Check the same firm on a second platform and see if the story holds. And put the name into Companies House; if it is there, check it is active. None of the AI tools did any of these.
The takeaway
Review scores are the raw material of every AI trust recommendation, and at the top of the trades they no longer separate anyone: nearly everybody is perfect, some of the perfect are dissolved, and some of the busiest profiles belong to firms no register has heard of.
Trust is not a score. It is a score, a date, a denominator and a company number, read together. Until the AI tools do that, someone has to.
Method: from the 173 firms in our July 2026 roofer study, we sampled the firms AI tools cited with specific ratings or review counts and collected full platform profiles for ten of them (Checkatrade and TrustATrader): overall rating, total reviews, membership start date, reviews in the past 12 months and most recent review date, cross-referenced against Companies House records collected for the original study. Small deep-dive sample: these are patterns, not prevalence estimates. Firms are deliberately unnamed; the underlying data is held on file. Collected 10 July 2026.
