AI does not read your star rating. It reads your reviews.
For most of the last two decades, reputation management meant moving one number. You wanted the average up, you wanted it above the shop next door, and everything else was detail. That worked because customers saw the number first and the sentences second, if at all.
That order is inverting.
More and more of the first impression a customer forms is not a list of businesses with numbers beside them. It is a paragraph. Someone asks an assistant where to eat, which plumber to call, which clinic takes their insurance, and gets prose back. That paragraph was not assembled by ranking averages. It was assembled by reading the reviews and describing what they say.
An average throws away the part that now matters
A rating is a summary statistic that keeps magnitude and discards everything else. It tells you people were broadly pleased. It cannot tell you what they were pleased about, what they forgave, or what they kept mentioning on the way out.
A language model has neither that limitation nor that habit. Give it forty reviews and it will not compute a mean. It will notice that eleven of them mention waiting, that the ones mentioning waiting still gave four stars, and that the word worth keeps appearing next to the word wait. Then it writes a sentence carrying all of that: busy at lunch, but people say it is worth it.
Two businesses can hold an identical rating and be described in opposite terms. One gets consistently good, staff remember regulars. The other gets opinions vary, several mentions of slow service. Same number. Completely different sentence. The sentence is what the customer acts on.
What a summary keys on
Three things do the heavy lifting, and none of them is the average.
Repetition. One person mentioning a cold room is noise. Nine people mentioning it is a fact about your business, and it survives into the summary even when every one of those reviews was positive overall.
Recency. A theme that stopped appearing four months ago reads as solved. A theme that started three weeks ago reads as current. Your rating cannot express the difference, because a two year old five star weighs exactly the same as this morning's.
Specificity. Great place carries almost nothing into a summary. The chef came out to check on the table carries a great deal, because it is concrete enough to reuse. Vague praise lifts your average and contributes nothing to how you get described.
That last one is genuinely counterintuitive. A wall of thin five star reviews can raise your number while leaving a summary engine with nothing good to say about you.
Read them as a transcript, not a scoreboard
Find the sentence that repeats. Every business has one, and most owners already suspect what it is. The useful question is not whether it is fair. It is how often it appears, whether it appears more than it used to, and whether it sits alongside complaints or alongside compliments.
Then treat it as an operations problem rather than a marketing one. If eleven people mention waiting at lunch, that is a rota question. The review is where you found it, not where you fix it. This is the part most reputation tooling misses. The value is not in replying faster. It is in noticing that your reviews have been quietly reporting a staffing gap since June.
And when you ask for reviews, ask better. Please leave us a review produces vague praise. Asking someone what specifically they enjoyed produces a sentence with a detail in it, and details are what survive summarisation.
The number is not going away
None of this means the rating stops mattering. It remains a filter, a threshold, a first cut. Fall far enough and nothing else gets a chance to be read.
But once you are past that threshold, another decimal place buys you less than it used to, and the sentences buy you more. The businesses that work this out early will spend the next few years fixing the thing everyone keeps mentioning. Their competitors will spend it chasing an extra tenth of a star.
One of those is reputation work. The other is arithmetic.