Here is the thing that makes automated review summaries much less mysterious than they look: they are overwhelmingly built out of repetition.
Not sentiment, not your rating, not recency alone. A summary surfaces the words that keep appearing. If thirty-one of your last ninety reviews contain the word loud, something is going to mention noise. If four of them do, nothing will, however strongly those four felt about it.
That is worth sitting with, because it changes what the summary is useful for. It is not a verdict on your restaurant. It is a word-frequency report on your reviews, rendered in a sentence — which makes it the one place you can read your own operation the way a stranger reads it.
Why repetition and not severity
A single furious review about a hair in the food is the worst thing in your review set and will not appear in any summary. Eleven mild reviews mentioning a wait will.
This feels wrong and is actually reasonable. A summary exists to tell a stranger what they can expect, and one bad night is not what they can expect. Eleven mentions of waiting is.
The practical consequence: the things that shape how you are described are mild, frequent and operational. Not your disasters. Your Tuesdays.
| Appears in summaries | Why | |
|---|---|---|
| Noise, temperature, seating comfort | Constantly | Mentioned casually by satisfied guests, so volume accumulates |
| Wait and pace language | Constantly | Every stage of service can generate it |
| One named dish | Often | Guests name specifics when they liked something |
| Parking, finding the door, the toilets | Often | Logistics are memorable and nobody fixes them |
| A severe one-off incident | Rarely | One mention, no matter how bad |
| Your star rating | Not as language | It is displayed separately; the summary reads text |
The noun inventory
You can do a rough version of the same counting yourself in half an hour, and the number it produces is more useful than the sentence Google shows you, because it is ranked.
Take your last ninety reviews, all star ratings. For each one, extract the nouns and adjectives that describe something concrete — loud, parking, cold, slow, friendly, cramped, short rib, queue, cocktails. Not nice, not lovely, not disappointing; those are verdicts and they carry no information about what happened.
Count them. You now have something like:
| Term | Count | Share of reviews |
|---|---|---|
| friendly / lovely staff | 34 | 38% |
| loud / noisy | 22 | 24% |
| wait / slow / queue | 19 | 21% |
| parking | 11 | 12% |
| short rib | 9 | 10% |
| cold | 6 | 7% |
Two things usually jump out of a list like that.
The top term is a positive you are not using. Thirty-eight per cent of reviews naming the staff is a genuine asset and almost nobody mentions it in their own description, their photos, or the way they brief new starters.
The second term is something nobody has ever been asked to fix. Noise at twenty-four per cent is not a complaint — it arrives inside otherwise glowing reviews — which means it has never generated a conversation, a work order, or a line in a manager's brief. It is simply true, repeatedly, and it is shaping how you are described to people who have not been yet.
Where the sentence actually comes from
Almost always, a repeated term traces back to one or two specific moments rather than a general condition.
Loud usually means a particular zone at a particular hour — the hard-surfaced section when the room passes about seventy per cent full, or Friday after nine. Not the whole restaurant.
Slow usually means a single gap in the sequence. "Slow" is the guest's word for a twenty-minute wait between ordering and the first drink arriving, and a different guest's word for a nine-minute wait for the card machine at the end. Those are five separate problems wearing the same label, and the surrounding wording tells you which.
Cold usually means one dish and one distance. Which is more often a plate and a route than a kitchen.
So the useful question is never "how do we get noise out of the summary". It is "which twenty square metres, at which hour, generates the word loud" — and that has a fix with a cost you can quote.
The lever, and its honest size
Your influence over the language runs through operations, and it runs slowly.
The arithmetic is unforgiving. If loud sits at twenty-four per cent of ninety reviews and you fix the acoustics perfectly tomorrow, the existing twenty-two mentions do not go anywhere. You need enough new reviews without the word to move the share, and at thirty reviews a month that is a visible shift in about two months and a convincing one in four.
Which produces two reasonable conclusions:
- Pick one term. Not three. The slowest-moving metric in your business does not reward a broad effort.
- Choose the one that is high-frequency and cheap. Acoustic panels in one section, a parking line added to your profile and your booking confirmation, or a card machine at the pass. Frequency, not severity, is what shifts a summary — so the unglamorous environmental things punch above their weight here in a way they never do in a complaint log.
What you cannot influence
The sentence itself. There is no edit box, no dispute flow, and no wording request. If it is stating something factually untrue about your hours or services, that is a different problem with a different route — capture it, correct the input, document it.
Which reviews get weighted. Recency and helpfulness factor in and you do not get to see the weights.
The pace. It is downstream of your review flow, which means a venue getting eight reviews a month is working with a much slower instrument than one getting sixty. If your volume is genuinely low, the ask itself is the lever first — and most venues' asking habit dies in week three.
Do not try to script the nouns
The tempting shortcut is to steer the words guests use: a card saying if you enjoyed the short rib, do mention it, or staff prompting for the word friendly.
Don't. Solicited phrasing is detectable, it reads as astroturf to the humans who notice it, and a cluster of reviews using suspiciously similar language is the pattern most likely to get filtered — which costs you the real reviews alongside the engineered ones. What actually crosses the line is worth knowing precisely.
The legitimate version of influencing language is making the thing more noticeable. A dish served so it is the first thing on the table gets named more. A host who introduces themselves by name gets named more. That is not scripting; that is giving guests something specific to remember.
The thing worth checking monthly
Not the summary. The ranked list.
The summary is a paraphrase and it moves for reasons you cannot see. The share of your recent reviews containing your top three terms is a number, it moves for reasons you can act on, and comparing this month's to last month's is the only way to know whether the acoustic panels worked.
OMMU keeps that count running — term frequency by month, with the review text behind each one — which is the version that survives a busy quarter. Ninety reviews and a tally chart gets you the same answer; the hard part is doing the tally again in October.