01 · The headline numbers
3,433
reviews read
across 168 completed audits
60
Google listings
holding 80,083 reviews between them
144
could be filed
out of the 3,433 we read
100
audits found nothing
out of 168, so 3 in 5
The number worth sitting with is the last one. Three audits in five turned up nothing that could be filed, and we say so on the results screen rather than inventing a case to sell. The industry number you have probably read instead is some version of “most businesses have fake reviews”, and it does not survive contact with a count.
02 · How the numbers were produced
There are two datasets on this page and it matters which one a figure comes from.
Dataset A · the audit funnel
Every free audit that ran to completion between 18 April and 21 September 2026: 168 runs on 60 distinct Google listings. The audit reads the lowest-rated slice of a listing rather than all of it, which is why 3,433 reviews were read on listings that hold 80,083 between them. A review counts as flagged when it can be argued under a specific published Google clause, which is a candidate for filing, not a removal.
Dataset B · the scored set
569 reviews on 14 listings that have been through full scoring, 554 of them carrying a removability score from 0 to 100. This is the set behind every split by star rating, by strength and by clause.
Nothing on this page is modelled or extrapolated, and no listing, business or reviewer is named. Counts move as more audits run, so quote the date with the number.
03 · What a typical audit finds
Of the 168 completed audits, 66 turned up at least one review that could be filed and 100 turned up none. Among the ones that found something, most found one or two:
| Reviews flagged on one listing | Audits |
|---|---|
| None | 100 |
| One | 29 |
| Two | 22 |
| Three | 1 |
| Four | 4 |
| Five | 8 |
| Six | 2 |
The listings themselves ranged from one with no reviews at all to one with 4,047, and averaged 477. Two completed runs recorded no flag count and are left out of the table rather than counted as zero.
04 · What the reviews themselves look like
The scored set skews negative on purpose: the audit goes looking at the bottom of a listing. Read it as a picture of the reviews an owner is worried about, not of Google.
| Rating | Reviews, out of 569 |
|---|---|
| 1 star | 180 |
| 2 stars | 38 |
| 3 stars | 24 |
| 4 stars | 25 |
| 5 stars | 302 |
Now the part that decides whether anything can be done. Every scored review carries a removability score, and the distribution is lopsided:
A clause to name and evidence sitting on Google's own platform.
Something is off, but the argument needs more than the review text to stand up.
Negative, sometimes brutal, and inside Google's rules. Nothing to file.
One-star reviews were 180 of the 569. Reviews that scored as a strong case were 10 of the 554 carrying a score. The gap between those two counts is the whole business: being harsh, unfair or flatly wrong about what happened is not a policy violation, and no amount of wanting it to be makes it one.
05 · Which clauses actually show up
Among the reviews that were flagged at all, here is what they were flagged for. A review can carry two at once, so the column does not add up to the flag count:
| What it was flagged for | Reviews |
|---|---|
| Rating only, no text | 36 |
| Spam or bot behaviour | 8 |
| Never visited the business | 6 |
| Off topic | 2 |
| Advertising another business | 1 |
| Repeated content | 1 |
| Incentivised | 1 |
Rating-only reviews dominate, and they are the hardest thing in this category to win: a star with no words gives Google almost nothing to read, so the case has to be built from the account around it. Everything people expect to see near the top, defamation, competitor sabotage, extortion, is rare in a count even though it is loud in the inbox.
06 · Three things we did not expect
55 reviews disappeared on their own, and 23 of them were five stars
Between one scrape and the next, 55 reviews we were tracking stopped being on the listing. 26 were one or two stars, 23 were five. Google’s own sweeps take down positive reviews at close to the same rate as negative ones, which is worth knowing before you read a disappearance as someone getting away with something.
A quarter of reviews had no text at all
147 of 569 were a rating and nothing else, and among the one and two star reviews it was 37 of 218. Owners tend to assume the silent ones are the fakes. Sometimes. Mostly they are people who could not be bothered to type.
A second, independent read disagreed with almost every case
We run a second scoring pass that has no sight of the first one, as a check on ourselves. On the 215 reviews it has read, it came back “nothing to file” on 196, “hold” on 18 and “file” on one. Two honest readers of the same policy disagree this much, which is a useful thing to know about anybody, us included, who sounds certain about your reviews.
One more, less surprising but worth a number: owners had replied to 136 of the 218 negative reviews in the set, nearly two in three. Replying is common. It is also a separate decision from filing, and the cases where a reply makes things worse are real.
07 · What these numbers are not
Anyone quoting this page should quote this section with it. The limits are not small print, they are the difference between a statistic and a claim.
- Not a sample of Google. These are businesses that came to us, which means they already suspected they had a problem. A random listing would almost certainly look cleaner, not dirtier.
- Not the whole listing. The audit reads the lowest-rated slice. Read “144 of the 3,433 reviews we read” and never “144 in every 3,433 reviews on Google”.
- Flagged is not removed. A flag says a clause can be named and evidence exists. Google decides, and Google makes the final call on every removal.
- Small, and we will say so. 60 listings is enough to be worth publishing and not enough to be a benchmark. We would rather show the sample than round it into a headline.
- No removal rate here. The number of filings that have run all the way to a Google decision is still too small for a percentage to mean anything, so we do not publish one. When one good month would move a figure by a lot, that figure is marketing.
08 · Using this page
Quote any figure here, in a post, a deck or an article, with a link to this page and the date on it. The counts change as more audits run, and a number without its date goes stale quietly.
If you want the exact query behind a figure, or a cut we have not published, ask on the contact page and we will send it. The method notes in this page’s source name every table and column the numbers came from.
FAQ
How many Google reviews break Google's rules?
In our own data, 144 of 3,433 reviews read across 60 listings could be filed under a published Google clause, which is one in twenty-four. That is a figure about the reviews our audit reads, not about Google as a whole: the audit reads the lowest-rated slice of a listing rather than all of it, and the businesses running it already suspected a problem.
What share of businesses have a fake Google review?
Of 168 completed audits, 66 turned up at least one review that could be filed and 100 turned up nothing at all. So on this sample it is closer to two in five than to the nine in ten you see quoted around this industry, and our sample is skewed towards businesses that came looking for a problem.
Do most one-star reviews qualify for removal?
No, and it is not close. One-star reviews were 180 of the 569 in our scored set, while only 10 of the 554 that carry a score came back as a strong case. Being harsh, unfair or wrong about the facts is not a policy violation, and most one-stars are exactly that.
Can I quote these figures?
Yes. Use them with a link to this page and the date, because the counts move as more audits run. If you want the query behind a specific number, ask and we will send it.
Why don't you publish a removal rate?
Because the number of filings that have run all the way to a Google decision is still too small for a percentage to mean anything. One good month would move it by a lot, which is exactly the kind of figure that misleads. We will publish it when the sample is big enough to survive the next month unchanged.