Reviewing Culture

REVIEWING CULTURE

What a star actually measures, and who gets to write one

Written 15 September 2026 – Northampton

The notification comes in at about twenty past nine on a Sunday evening.

You are not working. You are on the sofa with the telly on and a cup of tea going cold, and the phone buzzes on the arm, and you pick it up because everybody picks it up.

One star.

And then you read it, and there is a problem, because you do not recognise the job. Not the address, not the date, not the description of what supposedly happened. You go back through the bookings anyway, because you always do, and you check the year before that in case the name has changed.

There was no job. This person has never been a customer.

Your partner is in the kitchen. You read it out to her, and she comes and stands in the doorway with a tea towel, and the two of you look at a paragraph written by somebody neither of you has ever met, about work neither of you has ever done, which will be there in five years when somebody in Milton Keynes is deciding whether to let a stranger into their back garden, or not.

That is the whole subject of this article, and everything that follows is an attempt to understand what actually happened on that sofa, and what happened for the other party.

FIRST, THE UNCOMFORTABLE PART

We are not going to write six thousand words about unfair reviews without saying the obvious thing first.

We have had bad reviews we deserved.

There have been stretches, set out honestly in the second article in this series, where the van was off the road, where health and personal circumstances stacked up, and where people who had paid us waited longer to hear from us than they should have. Some of them said so publicly. They were entitled to. A customer who has been let down and says so in a review is doing exactly what the system is for.

So this article is not an argument that criticism should be harder to publish. It is not a plea for businesses to be protected from their own mistakes. If we had wanted that we would not have built a website that publishes every job we do in forensic detail, including the difficult ones.

The question is narrower and it is genuinely hard. When a business looks at its own reviews, it is looking at a mixture of things. Customers who were let down. Customers who were satisfied and bothered to say so. Customers who were satisfied and never said anything at all. And, in an amount nobody can quantify, people who were never customers.

How do you tell those apart? What is the number at the top of the page actually measuring when it adds them all together? And why repeating the process of apologising and repeating data to a customer becomes more difficult than most people are aware of.

We went and found out. The answer is better evidenced than we expected and worse than we hoped.

ONE. THE MIDDLE IS MISSING

Open any review page for any business anywhere. Count the fives. Count the ones. Then go looking for the threes.

They are not there.

This shape has a name. It is called a J shaped distribution. A large mass at five stars, a smaller mass at one star, and almost nothing in between. And it is not a quirk of your business or a sign that your customers are unusually polarised. It appears on essentially every product and service on every platform.

Hu, Pavlou and Zhang established why, in M I S Quarterly, one of the highest ranked journals in information systems. Two biases produce it, and both happen before anybody types a word.

Acquisition bias. The people who buy something already had a favourable view of it. Nobody hires a window cleaner they expect to be terrible. So the pool of people who could possibly leave a review is skewed before the transaction even happens.

Underreporting bias. People with extreme views, in either direction, are far more likely to write than people with moderate ones. The person who is quietly satisfied does not sit down on a Sunday evening to record that the windows are clean. There is nothing to say.

And the authors state the consequence flatly.

The mean rating is a biased estimator of quality.

Then they prove it, which is the part that ends the argument.

When all consumers are asked to write a review, the distribution becomes normal.

Remove the self selection and the J shape disappears. It was never a property of the thing being reviewed. It was a property of who bothered.

Which means the middle of every review page in existence, the ordinary, satisfied, unremarkable majority, is structurally absent. Not suppressed. Never written. Those people exist, they got what they paid for, and they are invisible.

TWO. AND THEN IT WEIGHS MORE

That explains what gets written. It does not explain what it does to the person reading it, or, for that matter, to the person it is about, sitting on a sofa at twenty past nine.

For that, the reference is Baumeister, Bratslavsky, Finkenauer and Vohs. Bad Is Stronger Than Good, published in Review of General Psychology in 2001 and one of the most cited papers in the discipline.

Their finding is that the greater power of bad over good is not confined to one domain. It runs through everyday events, major life events including trauma, relationships, social networks and learning. Bad feedback has more impact than good. Bad information is processed more thoroughly.

We should be careful here, because this is exactly the kind of place where a business with a grievance starts overclaiming. That paper is a review of general psychological findings. It is not a study of online reviews and it does not mention them. What it licenses is an inference, and we are stating it as one. This is what the psychology of negativity bias implies for a page of star ratings.

But the inference is sound, and it produces the shape of the problem.

The sample is skewed by who writes. Then the skew is weighted again by how it is read. Two distortions, stacked, and neither of them has anything whatever to do with how the work was done.

THREE. AND IT CHANGES WHAT PEOPLE BUY

Mechanism is one thing. Consequence needs measuring.

Which question mark measured it, and the study is better than anything else we found on this subject.

They ran a controlled behavioural experiment on nine thousand nine hundred and eighty eight people. Each participant was given an online shopping task. Choose your favourite from five products, all at the same price. One of the five was a Which Don't Buy. A product Which had independently tested and rated as poor.

Participants were randomly allocated into six groups. One control group saw only real review content. Five treatment groups saw varying amounts of fake review content.

Price held constant. Product set held constant. Only the reviews changed.

People were more than twice as likely to buy a poor quality product that had been boosted by fake reviews.

And.

All the fake review treatments significantly increased the likelihood of a consumer choosing the Don't Buy product.

Not most of them. Every single one.

That completes the chain, and it is worth stating as three findings from three independent sources because that is what it is.

The sample is skewed. Hu, Pavlou and Zhang. The skew is weighted again in the reading. Baumeister and colleagues. And the result changes what people actually buy. Which, on nearly ten thousand participants, with a control group.

The money involved has been measured too. Which cites Competition and Markets Authority research suggesting twenty three billion pounds a year of United Kingdom consumer spending is influenced by online reviews.

Twenty three billion pounds, steered by a number that its own literature describes as a biased estimator.

FOUR. HOW MUCH OF IT IS FAKE

The platforms publish their own enforcement figures. A company reporting on its own policing is not a neutral source, so everything in this section is given by each firm's own account.

Google blocked or removed more than two hundred and ninety two million policy violating reviews in 2025, and more than two hundred and forty million in 2024. It removed over thirteen million fake Business Profiles in 2025, and over twelve million the year before. It states that the vast majority were removed before anybody saw them.

Trustpilot removed four and a half million fake reviews in 2024. Seven per cent of everything posted on the platform that year. Ninety per cent were taken down automatically, a fifty three per cent increase in automated removals on 2023.

Tripadvisor flagged and removed two hundred and fourteen thousand artificial intelligence generated reviews in 2024.

Sit with the first one. Two hundred and ninety two million.

That is not a fringe problem being handled at the edges of an otherwise sound system. That is the routine operating volume of contamination in the mechanism the public uses to decide which tradesman to let into the house.

And note carefully what those numbers are. They are counts of what was caught. Neither Google nor Trustpilot nor Tripadvisor publishes an estimate of what was missed, and there is no reason to assume the answer is small.

There is a word for what happened on the sofa

Here is the thing we did not expect to find, and it changed how we thought about that Sunday evening.

Tripadvisor sorts fake submissions into four categories. Boosting, vandalism, member fraud, and paid reviews. Boosting accounts for fifty four per cent of what it finds. Member fraud, thirty nine.

Vandalism means malicious negative reviews.

A major global platform formally classifies malicious negative reviewing as a distinct category of fraud, tracks it as an operational metric, and has given it a name.

We had spent a long time looking for academic work on coordinated negative reviewing and finding almost nothing. It exists. It just was not in a journal. It was in a transparency report, because the people who deal with it every day are not researchers. They are moderators.

We are not going to overstate how common it is. Boosting and member fraud together account for the large majority of what Tripadvisor catches, and vandalism is a minority of a minority. The finding is not that it is widespread. The finding is that it is a recognised, named, tracked category. And that almost every public conversation about fake reviews concerns businesses inflating their own ratings, which is the opposite direction.

The other direction has a name, and the name is vandalism.

FIVE. AND SOME OF IT IS NOT A REVIEW AT ALL

There is a version of this that does not go through a review platform, and it is the one that has done us the most damage.

Local groups. Community pages. The place where somebody asks for a recommendation and forty people answer.

This business has been written about in those places by people who have never bought a service from us, never met us, and in all likelihood never will. Not customers with a complaint. A customer with a complaint is owed an answer and gets one. Strangers. Repeating something half heard. Occasionally using the word criminal.

We are not going to name anybody, quote anything, or link to it, because doing so would be doing to them precisely what we are describing. And we are not going to dress it up as a tragedy, because it is not one. It is an ordinary feature of running a visible small business in 2026, and almost everybody in the trades has a version of the same story.

What is worth noticing is the structure, and it is genuinely strange when you set it out.

And there is nothing at all holding the other direction. A person with no connection to a business whatsoever can publish an accusation about it, permanently, to an audience of thousands of its neighbours, with no obligation of accuracy, no register, no licence and no consequence.

That is not a complaint about free speech and we want to be very clear about it. People are entitled to talk about businesses. A trade that wanted that stopped would be a trade with something to hide, and the entire architecture of this website, the case log, the published methods, the dilution rates, exists so that people have something factual to talk about instead of something vague.

It is an observation about where obligation has been placed. Every duty in the relationship runs one way.

SIX. WHAT THE PLATFORMS SAY ABOUT TRUTH

So you have a review you know to be false. You assume there is somewhere to take it.

There is not, and both major platforms say so in writing, on pages anybody can open.

Google's published policy states.

Negative reviews should not be reported just because you disagree with them or dislike them.

And that Google doesn't get involved in conflict between businesses and customers.

Trustpilot's guidelines for businesses, in the version live since February 2026, state.

Reviews are not removed just because a business thinks they are unfair or critical.

Set them side by side.

Google tests breach of content policy. Trustpilot tests breach of guidelines. Neither tests whether the review is true. Where a breach is unproven, on both platforms, the review stands. Neither will let the business remove it. Both will let the reviewer remove it, at any time.

Two independent platforms, in their own published words, both stating that truth is not the criterion.

And look hard at that last pair, because that is the part that sits with you afterwards. The person who wrote the statement can withdraw it at any moment, for any reason, with two taps. The person it is about can never remove it, and cannot have its accuracy examined by anybody.

That is not a platform failing. It is the platform working exactly as designed, and saying so in public.

And the obvious response is the prohibited one

There is a trap here and the natural small business instinct walks straight into it.

Google's rating manipulation policy prohibits incentivised reviews, expressly including content posted following a request to revise or remove a review in exchange for an incentive. Trustpilot prohibits pressuring a reviewer or offering an incentive to change or delete one.

Offering a customer anything at all to take a bad review down is a policy breach on both platforms. And, under the fake review provisions of the Digital Markets, Competition and Consumers Act 2024, potentially an infringement too.

The thing a worried business most wants to do is the one thing it definitely must not.

Where the platforms are not passive

We should be fair, because platforms do nothing is not what the evidence says.

Trustpilot has won a landmark United Kingdom legal case against websites selling fake reviews, and states it has brought ten legal cases in two years against businesses abusing its rules. It brought the landmark action before the new legislation required it to, at its own expense.

So the honest position is this. The platforms police fabrication at industrial scale, through courts and automated detection, and decline to adjudicate accuracy in the individual case. Both of those are true at the same time, and a business that only says the second half is not being straight.

SEVEN. AND THEN A BANK USED THEM

This is the part we would not have believed before it happened to us.

In 2025, following a disputed transaction of twenty two pounds and fifty pence, our business bank account was closed. The full account of that is in the second article in this series, and the detail that belongs here is one sentence.

We were told that part of what informed the decision was a review of our online reviews.

A regulated financial institution. Not a customer choosing a window cleaner, not a neighbour asking on a community page, but a firm authorised by the Prudential Regulation Authority and regulated by the Financial Conduct Authority. Reached a conclusion about our business using a data source that its own sector's regulator has published guidance on, and that the peer reviewed literature describes as a biased estimator of quality.

Read section one of this article again with that in mind.

The J curve means the middle is absent. Baumeister means the negative entries carry disproportionate weight in the reading. Which demonstrated, on nearly ten thousand people, that review content changes decisions even when the underlying thing is identical and the price is held constant. Google removed two hundred and ninety two million policy violating reviews in a single year. Tripadvisor has a formal category called vandalism.

Every one of those was true of the page that was read about us.

We are not going to allege bad faith, because we do not think there was any. We think somebody opened a browser, looked at a rating, and formed an impression. Which is exactly what everybody does, and exactly what the research says happens when they do.

That is the point. The bias is not a defect that careful people avoid. It is in the instrument. And an instrument with a published, measured, peer reviewed bias was used to make a decision that removed a business's ability to trade.

If you want a single reason why this article exists at eight thousand words rather than five hundred, that is it. We assumed reviews were reputation. It turns out they are also evidence. And nobody is checking whether they are any good at being evidence.

EIGHT. THE LAW, AS IT NOW STANDS

The ground moved in 2025 and most people have not noticed.

The Digital Markets, Competition and Consumers Act 2024 brought its consumer provisions into force on 6 April 2025, and the fake review ban sits at Schedule twenty, paragraph thirteen.

That reference is worth getting right, because it is widely miscited. The Act is chapter thirteen of 2024. The provision is Schedule twenty, paragraph thirteen. They are not the same thing and quoting the wrong one is the fastest way to lose an argument with somebody who has read it.

What it bans. Submitting, or commissioning somebody else to write or submit, a fake consumer review. Publishing a review that conceals that it was incentivised. Publishing reviews in a misleading way. Failing to take reasonable and proportionate steps to prevent fake or misleading reviews appearing. And offering services that facilitate any of it.

A fake consumer review is defined as one that purports to be, but is not, based on a person's genuine experience.

And the placement matters more than the wording. Because it sits on the list of practices unfair in all circumstances, the Authority does not have to prove the practice affected the transactional decision of the average consumer. The act itself is the infringement. The sanction is not criminal. The Authority can determine a breach without going to court and impose fines of up to ten per cent of annual global turnover.

What reasonable and proportionate steps actually means

The Authority published C M A two zero eight, Fake Reviews Guidance, in April 2025, with a three month adjustment period before enforcement began.

Publishers must have published policies on fake and incentivised reviews. They must complete a risk assessment and build processes to detect, investigate and act. They must remove banned content once published.

And then this, which explains something otherwise puzzling.

A business cannot rely on having outsourced review moderation to a third party as a defence. It remains individually responsible.

The five open investigations

On 26 March 2026 the Competition and Markets Authority opened five investigations into fake and misleading online reviews. We read one of the case pages directly on 14 September 2026 rather than relying on the reporting.

The allegation concerning Autotrader and Feefo, in the Authority's own words.

The C M A is investigating whether a number of one star reviews, which were moderated by Feefo Limited, were not published on Autotrader's platform, and were not counted towards star ratings, therefore denying consumers a fully rounded picture of other customers' experiences.

Not published, and not counted. Two mechanisms in one sentence. And a review that exists but is excluded from the arithmetic is invisible to everybody, including the person who wrote it.

The other three concern Dignity, over whether staff were asked to write positive reviews of its crematoria services. Just Eat, over whether its ratings methodology inflated particular establishments' star ratings. And Pasta Evangelists, over whether customers were offered undisclosed discounts for five star reviews.

The Authority's own caveat travels with every one of those names.

At this initial stage, it should not be assumed that the business has infringed consumer protection law and no finding has been made.

When we checked on 14 September 2026, the Autotrader case was open, it had been opened on 26 March 2026, the page had last been amended on 3 August 2026, and a case update was due that month and had not yet appeared. The twenty seventh is the press release date, which secondary reporting routinely conflates with the opening.

Note that both Autotrader and Feefo are named subjects. Under the guidance, outsourcing the moderation does not move the duty.

NINE. THE DOOR THAT IS ALMOST ALWAYS SHUT

Suppose a review is not merely unfair but false, and you can prove it. What then?

The Defamation Act 2013, section one, provides that a statement is not defamatory unless it has caused or is likely to cause serious harm to reputation.

And section one, subsection two, adds a condition that applies only to businesses.

For a body that trades for profit, harm is not serious harm unless it has caused or is likely to cause the body serious financial loss.

Parliament's explanatory notes give the reasoning. Bodies trading for profit cannot claim for injury to feelings and are in practice likely to have to show actual or likely financial loss, and the requirement that the loss be serious is set to match the test in subsection one.

So a sole trader who knows a review is false must demonstrate serious financial loss. And must do so before a court will treat the statement as defamatory at all.

Try to quantify that. One review. No control period. No enterprise analytics. A round where the phone rings less some months than others for fourteen reasons, of which this might be one. It cannot practically be evidenced, and the cost of attempting it would exceed the value of the work in dispute by an order of magnitude.

The law recognises the harm. It sets a threshold for it. And it sets the threshold at a level a business of two people cannot reach.

The right exists. Exercising it does not.

This is information rather than legal advice. We are stating the statutory position and what it means in practice. We are not advising anybody to sue and we are not advising anybody not to.

TEN. AND THE OTHER SIDE OF RECOGNITION

There is a part of this economy that is neither review nor complaint, and it is worth a short section because we were in it for eight years.

We were listed by ThreeBestRated for eight consecutive years, from 2017 to 2024.

The listing is generated from businesses ranking well in public search. Each time it arrives, it is followed by an invitation to purchase certificates and badges to display.

We are recording that without adjectives, because the facts do the work. A business is identified as good by an automated process, told it has been awarded something, and then offered the opportunity to buy the physical evidence of it.

We are not suggesting anything improper and we are not naming it to attack it. We are naming it because a reader trying to work out what any of these signals mean deserves to know how they are produced. And because the honest answer is that eight years of being listed told us almost nothing about the quality of our work that we did not already know from the work.

ELEVEN. SO WHY WE ANSWER EVERY ONE

Put all of it together and the shape is unavoidable.

The distribution is an artefact of who wrote. The negative entries weigh more in the reading than the positive ones. Fake reviews demonstrably shift what people buy. One platform removed two hundred and ninety two million policy violating reviews in a single year. A second has a formal category for malicious negatives. Neither of the two largest tests whether a review is true. The reviewer can delete at any moment. The business never can. The legal remedy requires proving a loss the business cannot measure. And a regulated bank used the resulting number as an input to a decision.

There is exactly one instrument left.

The reply.

Not to argue. Not to win, because there is nothing to win. To put the other half of the record in the same place, under our own name, where anybody reading can see both and make up their own mind.

That is the whole of our policy on reviews and it is the only policy available to us. Every review gets an answer. The fair ones get an apology and an account of what we changed. The unfair ones get the facts, set out plainly, without heat.

And the thing we have to say about our own

Here is the part that costs us something to write.

The J curve cuts against us in both directions. Our five star reviews come from exactly the same skewed pool as everybody else's. The people who already liked us, who had an extreme enough reaction to sit down and type. The satisfied middle never wrote about us either.

Which means our own rating is not a measurement of our work. It is a measurement of who bothered, about us, same as everyone.

We would be fools to treat it as anything else, and we would be hypocrites to spend eight thousand words explaining why a number is unreliable and then quote our own.

That is precisely why we publish the case log instead. Photographs, dates, substrates, products, dilution rates, weather conditions, what worked, what did not, and what we would do differently. Every job. Including the ones that went badly.

Not because it is more flattering. It frequently is not. Because it is evidence, and a star is not.

TWELVE. WHAT WE ACTUALLY THINK

Matthew started on the markets, getting to the pitch at four in the morning, before any of this existed.

Your reputation, then, was the person standing in front of you. If you sold somebody something and it was wrong, they came back the following week and told you to your face, and you sorted it out, and that was the end of it. The transaction and the accountability happened in the same place, between the same two people, and both of them were present.

None of that is true any more, and we are not going to pretend it was a golden age. Plenty of people got away with plenty, and a market stall had no case log either.

But something specific has been lost and it is worth naming. Not politeness, and not deference. Proximity. The person making the judgement is no longer in the room, is frequently not a customer, is sometimes not a person, and never has to be right.

A star rating is not a lie. It is not a conspiracy and the platforms are not villains. It is a measuring instrument with a known and published bias, which almost everybody using it believes to be neutral. Including, it turns out, banks.

That is all we are asking anybody to understand. Not that reviews should be ignored. That they should be read for what they actually are.

And if you are ever standing in a kitchen doorway with a tea towel at twenty past nine on a Sunday, reading something about yourself that never happened, that is not a failure of your business.

It is a J curve, with a stranger on the end of it.

SOURCES

Peer reviewed. Hu, Pavlou and Zhang, two thousand and seventeen, On Self Selection Biases in Online Product Reviews, M I S Quarterly, volume forty one, issue two. Baumeister, Bratslavsky, Finkenauer and Vohs, two thousand and one, Bad Is Stronger Than Good, Review of General Psychology, volume five, issue four, pages three two three to three seven zero.

Statute and regulator. Digital Markets, Competition and Consumers Act 2024, Schedule twenty. Defamation Act 2013, section one, and its Explanatory Notes. Both at legislation dot gov dot uk. C M A two zero eight Fake Reviews Guidance, the Autotrader consumer protection enforcement case page, and the press release of 27 March 2026, all at gov dot uk.

Which. A consumer champion which campaigns for legislative change. The real impact of fake reviews, at which dot co dot uk.

Platforms, by their own account. Google Business Profile help pages on reporting inappropriate reviews and on prohibited and restricted content. The Google blog on protecting businesses on Maps. Trustpilot Guidelines for Businesses, February 2026 version, Action We Take, March 2026 version, and the Trustpilot Trust Centre. And the Tripadvisor Transparency Report 2025.

Sources and licensing. Contains public sector information licensed under the Open Government Licence version three point zero. Quotations from copyright works are used under the quotation exception, section thirty, Copyright, Designs and Patents Act 1988, and each is attributed in the text. Figures are reported as published by the named body. Where credible sources disagree, all are given. Nothing in this article is legal advice.

Matthew Kenneth McDaid, trading as Shining Windows