This is the whole dataset, published because nobody else has counted it. Every figure below is a count on live Google listings read in August 2026 — not a survey, not an estimate, and not an industry average borrowed from somewhere else. If you want to cite any of it, please do; a link back is enough.
What was measured
Pure cafés and coffee shops only — no restaurants that happen to serve coffee. Read from each business's public Google listing, sorted newest first, with the ordering verified rather than assumed.
Most London cafés answer nothing
Of those 328 London cafés, one hundred and eighty-seven had replied to none of their recent reviews. That is 57 per cent. At the other end, 34 cafés — one in ten — reply to almost everything, at 95 per cent or above. Between those two groups sits everybody else, replying occasionally and without a pattern.
The median café in this set carries 1,170 reviews. Silence at that volume is not an oversight; it is a standing decision, usually made by default.
The replies that do get written are short
The median reply is 31 words. Ninety per cent are 77 words or fewer. More than half are a single paragraph. This is worth knowing if you are evaluating software that writes replies for you: most of it produces three paragraphs and around 150 words, which is longer than nine out of ten replies a real owner writes.
Thirty-seven per cent of replies to unhappy customers engage nothing
There are 395 replies in the data written to reviews of three stars or fewer. We measured how much of each one touches the specific content of the review it sits under — the dish, the wait, the person, the price. One hundred and forty-eight of them score zero. They answer the star rating, not the review.
“Thanks for your brilliant rating – it means a lot to our team.”
664 sentences appear at more than one business
Taking every sentence of five words or more, 156 of them appear word for word at two or more separate cafés. One turns up at 29. These are not similar sentences; they are the same sentence, at businesses in different postcodes under different owners.
Counted across the 2,907 replies from 328 London cafés read for this piece in August 2026 — the first count, before the corpus was widened. Sentences of five words or more; where a shorter sentence was contained inside a longer one, only the longer was counted, so nothing is double-reported.
One reply in ten sends the customer somewhere else
We ran the whole set through the checks we apply to our own writing. The most common thing we found was not rudeness or bad grammar. It was cafés moving the conversation off the page.
Two hundred and ninety-seven replies — one in ten — ask the reviewer to get in touch privately. It reads as helpful and it costs the business the argument: the complaint stays public and the answer moves somewhere nobody can see. Every future reader sees a problem and no response to it.
One hundred and forty-three replies use “I” rather than “we”. Fourteen concede fault outright, in writing, under their own name — which is a statement that outlives the shift it was written on.
What it adds up to
Fifty-seven per cent of London cafés say nothing. Of the ones that do speak, more than a third say nothing about what happened, one in ten sends the customer away from the page, and 664 sentences are shared property. Answering is not the hard part. Sounding like the business that was actually there is.
Since this was written
This piece was written from the London count. We have since read the same thing in 18 cities: 14,130 replies from 914 cafés across 90 neighbourhoods. The London figures above are left as they were measured — the answer rate in particular cannot be recomputed, because outside London we only read reviews that already carried a reply.
What the wider count changed, and what it did not: the shape of the problem held. Across all 18 cities, 51 per cent of replies still share no distinctive word with the review above them, and 8 per cent still break at least one check. What grew was the repetition figure — 664 sentences now appear at two or more separate businesses, with the most repeated closing line turning up at 68 cafés spread across every city we read.
In other words: London was not unusual. It was representative.
What we would measure differently now
Two things. First, we would separate the checks that describe a real risk from the ones that describe our own house style. Telling a business it is at risk because it wrote "I" instead of "we" is not measurement, it is preference, and we have since stopped counting it that way.
Second, we would be clearer that a word-overlap test cannot see across languages. 922 reviews in the wider corpus are not in English, and running an English test on them produces a zero that means nothing.
Both corrections made the headline numbers smaller. That is usually a sign the corrections were the right ones.
Method, and how to use this
328 pure cafés across 24 London neighbourhoods, read from live Google listings in August 2026. Sentence repetition counts sentences of five words or more, and where a shorter sentence sits inside a longer one only the longer is counted, so nothing is double-reported. Engagement is measured by how much of a review's distinctive vocabulary the reply picks up, weighted so rare words count and common ones do not.
No individual café is named anywhere in this report, and none will be. The pattern is the point. If you are a journalist, an analyst or an operator and you want a figure checked or a cut of the data we have not published, write to us and we will run it.