Reviews and AI recommendations for local businesses
What we track when a local client asks how AI engines talk about their locations: per-response sentiment, the accuracy of claims in answers, and which review platforms show up in the citations.
When a local business owner asks about reviews and AI, the question underneath is usually "will better reviews get ChatGPT to recommend us?" We can't answer that with a yes. We haven't seen a source that tells us how any AI engine weighs reviews when it picks a local business, and we're not going to invent one.
What we can do is measure the part that's visible. What do AI answers say about each location? Is the tone positive or negative? Are the claims true? And which review sites, if any, are the answers citing? That's the program we run, and this post explains how.
Keep running the review program anyway
First, the part that doesn't depend on AI at all. Asking happy customers for reviews, never gating or buying them, and replying to every review (especially the bad ones) is good practice for the people reading those reviews before they book. Our local SEO playbook has the details. We keep doing it for clients because customers read reviews, and we don't need an AI ranking factor to justify that.
The AI side is a separate question, and we measure it separately.
Sentiment per response, per location
We run client programs on Promptwatch, with one monitor per location: a coordinate monitor for walk-in sites, a city monitor for service areas. The prompts in each monitor are identical, so locations can be compared.
Promptwatch scores the sentiment of every response from 1 to 100, and writes a weekly AI summary of how sentiment moved. For a local client, that per-response score is the useful unit. A chain-wide average hides the one branch that every engine describes as "mixed reviews, slow service." Split by location monitor, it stands out.
We read the weekly summary first, then open the responses behind any drop. The score tells us where to look. The response text tells us what the engine actually said, and that's what the client will want to see.
A couple of settings make the numbers cleaner. Brand aliases make sure "Brand Riverside," "Brand" and the old trading name all count as the same client. The Brand Book's analysis instructions let us spell out what counts as a mention, which matters when a chain and its individual branches get named in different ways.
Accuracy: are the claims in the answer true?
Sentiment is only half of it. A perfectly positive answer that says the location is open until 9pm when it closes at 6 is still a problem. For local clients we read responses with a short list of claims to check:
- hours and days open
- services offered at that location (not just somewhere in the chain)
- prices, when an answer quotes one
- address, parking, accessibility
- whether the answer attributes a competitor's complaint or praise to the client
Every wrong claim goes on a fix list with the response it came from. Then we look at what that response cited, because the error usually came from somewhere: an old listing, a stale page on the client's site, a review that describes a service the location no longer offers. Our checklist for profile and site facts is in Google Business Profile and AI answers: what we check.
Which review platforms show up in the citations
This is where reviews and AI answers meet in data we can actually see. Promptwatch's citation analytics show the domains and specific pages each answer cited. Filtered to one location's monitor, we can see whether review platforms appear in the citations at all, which ones, and which pages on them.
We don't decide in advance which review sites matter. Whatever turns up in the citation report for that location is what we look at. If a review platform keeps appearing behind answers in one city, we check the client's profile on it: is it claimed, are the details right, what do the recent reviews say? If a platform never appears, we don't tell the client to chase it for AI reasons.
We also watch Reddit, YouTube and LinkedIn citations, which Promptwatch reports on separately. For some local trades, a Reddit thread where people discuss providers in their city gets cited more than any review site. That's not something to manipulate, but it's worth knowing about and responding to honestly when a real customer complaint is sitting there.
When sentiment drops
When one location's sentiment drops, we check the responses for a few causes before anything else. Is a cited source carrying a run of negative reviews? Is the answer repeating an old problem that's since been fixed? Or is it mixing up the client with a similarly named competitor? Each has a different fix.
Real negative reviews are an operations problem first. We pass them to the client with the responses that quoted them, and the reply goes on the review platform, written by someone who can actually address it.
Old problems that are fixed need a newer, citable source that says so. That might be an updated location page, a reply on the review that explains what changed, or a local listing with current information.
Mix-ups with a competitor go back to the facts check: name consistency, aliases, address and phone on every listing.
We log these as tasks in Promptwatch's Action Board so they stay with the location they belong to, and the monitor keeps running so we can see whether the answers change.
What we tell clients
We tell them we can show what AI engines say about each location, how the tone moves week to week, which claims are wrong and where those claims came from. We don't tell them that more reviews will move them up in ChatGPT, because we can't back that up. The sentiment trend for a branch, read alongside the citations, is the evidence we can stand behind. Our wider notes on sentiment tracking are in brand sentiment analysis across AI engines.
If you want to see what AI engines say about your locations, email hello@1001seomedia.com.