Reviews are the one body of text about your business that you did not write. A model treats them accordingly: as the closest thing to independent evidence it can find about a place it has never visited.
Models read sentences, not stars
When an assistant builds an answer to "where should I eat in Greenpoint tonight", it needs a reason. It will not write "this one has 4.8", because that tells the reader nothing. It will write "praised for the handmade pasta and quick service", if it can find sentences that support it.
Hence the practical difference. A hundred ratings with no text give a model zero material. Fifteen reviews where customers describe specifics give it fifteen possible reasons.
Recency beats volume
Models lean heavily towards fresh information about places, because they are built to be careful with stale facts. A venue may have changed owner, menu or hours. A review from four years ago is weak evidence of anything.
A realistic target for a small business is a few new reviews a month, steadily. That works better than a sprint where you collect thirty in a fortnight and then nothing for a year.
Owner replies, or free space for facts
A reply is text you write, sitting next to text a customer wrote. The model reads both. It is your one chance to add information the customer did not mention.
| Instead of | Write |
|---|---|
| Thanks for the review! | Thank you. We bake the sourdough daily from 6am, and on Saturdays until it runs out. |
| Glad you enjoyed it, see you soon. | Glad the short rib worked for you. It is on the menu Thursday through Sunday. |
| Sorry to hear that, please get in touch. | Apologies for the wait on Saturday. Since May we have an extra person on the weekend shift. |
Notice what happens in the right-hand column. Every reply adds a fact: hours, a product name, a change in the business. Those are the things a model can repeat.
Negative reviews hurt less than you think
A profile of nothing but five-star ratings looks suspicious both to a person and to the systems watching for abuse. A handful of three and four-star reviews with a level-headed owner reply makes the whole profile more credible.
What matters is how you answer. A calm reply that states what changed reads well, including when the reader is a machine.
Why buying reviews moves you backwards
- Google removes them. Not always immediately, but it does, and the average you paid for goes with them.
- They read alike. Vague, short, no product names. A model extracts no reason from them, so even if they survive they do nothing for you.
- They arrive in a wave. Twenty reviews in three days after a year of silence is a pattern both Google and any human visitor can spot.
- You risk suspension. A suspended profile disappears from maps and from assistant answers at the same time.
Reviews beyond Google
A model searches in several places, so reviews elsewhere count too, if less. For most businesses it is worth covering Apple Maps, Facebook and one vertical site, such as a booking platform or a trade directory. The goal is less about collecting reviews everywhere and more about leaving no place where something outdated is written about you.
How to check whether reviews are doing their job
A simple test: ask an assistant about your trade and town, and if your business comes up, ask it why. If the model echoes a phrase that appears in your reviews, you know it read them.
The systematic version of that test is measuring share of answers, the same method described in the post on generative engine optimization. You can also start with the free scan, where we ask ChatGPT three questions and show you the full answers straight away.
Reviews are only one source, though. Alongside them it is worth sorting out your Google Business Profile, since that is where the model gets your category, hours and services.