Say you run a scooter rental shop on Kinmen, one of Taiwan’s outlying islands, and most of your customers are tourists who arrive by plane or ferry. From last summer through this spring, every so often someone would complain on Google Maps: “The battery died halfway through the ride. I called the shop to swap the scooter and waited almost an hour.” In May you replaced that whole batch of old scooters and took on an extra person just to handle swaps. All summer the reviews were almost all praise, and your rating sat steady at 4.6.

Then in autumn, someone planning a trip to Kinmen asks ChatGPT: “Which scooter rental on Kinmen is best?” You’re in the answer, but with a line attached: “Generally well reviewed, though some travellers mention the scooters can be unreliable and replacements take a while.”

That line is describing you before May. AI isn’t making it up; it just got hold of the old version of you. To get it to change its story, deleting reviews or buying good ones won’t work. The only way is to get the fixed version written down yourself, then go back and ask it again a while later to see whether its tone has come back.

AI picks out what people keep mentioning

The AI tools that read reviews for customers all work in much the same way: they read through a pile of reviews, pick out what lots of people mention, and condense it into a sentence or two. OpenAI and Google say so themselves in their help pages.

When customers look for things to buy in ChatGPT, it may show a review summary next to a product, and OpenAI says that summary draws on reviews from public websites and highlights the likes and dislikes people commonly mention. Google Maps goes into more detail. At the end of October 2024 it announced AI-written review summaries for places, starting in the US, and so far there’s been no word from Google on when they’ll reach Taiwan. Its help page says the summary only uses reviews from the past year, and if there are too few reviews, or nothing people keep saying in common, there may be no summary at all.

So what ends up in those one or two sentences is whatever comes up often. A star rating averages everyone’s scores, but a summary has to pick out a few specific things to say, so a 4.6-star shop can still get one complaint written in next to the praise if enough people have made it. It’s the same when a customer just asks ChatGPT which place is best: it may go online and look things up before answering, and travellers’ reviews and bloggers’ trip write-ups could all be part of what it reads.

Once things are fixed, hardly anyone writes about it

If the new scooters arrived in May, why is AI still talking about waiting for a replacement in autumn? Because people rarely go out of their way to write about the side that’s been fixed.

Happy customers mostly write “the scooters are new”, “the owner was really helpful” or “they even lent us raincoats when it rained”. Hardly anyone bothers to write “the battery was fine this time”, because when things go smoothly there isn’t much to say. So the reviews that mention swapping scooters still come down to those few complaints. Several people told the old story, almost nobody tells the new one, and it’s no surprise AI picks the old one. With a summary like Google’s, which draws on a full year of reviews, this spring’s complaints won’t start dropping out until next spring.

Beyond reviews, there are articles that stay online indefinitely. Suppose a blogger wrote up a three-day trip to Kinmen last summer, with a passage about how the scooter he rented from you died halfway and he waited an hour. That post won’t take itself down, and when ChatGPT goes online to look things up, it may well be what it reads. OpenAI itself warns on the help page for ChatGPT search that cited sources can be incomplete, outdated or incorrect.

Not a word of “replacements take a while” is made up; all of it happened. It’s just that the reviews and articles telling the old story are all still up, while hardly any tell the new one, and AI goes with the majority, so what it describes is last year’s you.

You watch the 4.6; customers read that one line

That’s also why owners are slow to notice. What you see every day is the 4.6 next to your shop’s name, so you figure your reputation is fine; what customers read in ChatGPT is “replacements take a while”. Once AI has already boiled a pile of reviews down to one sentence for them, are they going to click through and read them one by one? If they don’t, they never see that long run of praise from the summer.

Whether AI says good things or bad things when it talks about you is what we call AI sentiment, and it’s one of the things we record when we regularly put questions to AI for our clients. It’s scored separately from whether AI gets things right: getting it right is about whether facts like your address, opening hours and prices are correct, while sentiment is about tone. You can’t fault the line about the Kinmen shop on the facts, but a customer who reads it starts worrying about another long wait before they’ve even arrived.

Customers tap out your stars one at a time, and Google adds them up every day and shows them next to your name. Nobody adds up whether the line AI says about you is good or bad, and unless you ask, you won’t know how it’s describing you right now.

You can’t touch the reviews, but you can write replies and your website

Once you know an old complaint will keep getting repeated, your first thought is probably to make those reviews disappear, or to get good ones piling up fast. Under Google’s review rules, almost none of that is allowed. Offering money, discounts or freebies for reviews: not allowed. Offering them to get customers to change or remove a negative review: not allowed either. Inviting only the customers you expect to be happy, or discouraging people from leaving negative reviews: not allowed. Pressing customers to leave a rating while they’re on your premises, or telling them what to write: also not allowed.

What Google does allow is asking customers to write about their genuine experience, without offering anything in return or trying to influence how many stars they give or what they say. In shop terms, you can invite every customer, and once you have, whether they write something good or bad is up to them. Reporting won’t help either. What Google lets businesses report is reviews that break its content rules, and a customer honestly writing up a complaint isn’t breaking any.

The old ones can’t be taken down, so what’s left is getting the fixed version written down. The most direct place is your reply under those old complaints: say, as the business, what you’ve changed and from which month, and everyone who sees that review from then on sees your explanation too. But we went through Google’s and OpenAI’s official documentation and found nothing on whether AI reads business replies, so replies alone aren’t enough.

The other place is your own website. The new scooters, the extra person handling swaps, the fact that it started in May: put all of it there. When ChatGPT looks things up online, your website is one of the pages it may read, and the only one where you decide everything it says. Bring the business descriptions you can edit on rental and travel sites up to date too. Then keep inviting every customer to leave a review as usual; as reviews from after May build up, the old complaints make up a smaller and smaller share.

Whether the tone has come back, you have to ask again later

Once that’s done, whether AI has actually changed its story is the thing you’re least able to judge yourself. Ask ChatGPT once and all you see is this one question, this one time: ask “Which scooter rental on Kinmen is best?” and it may just say your scooters are new; ask “What should I watch out for when renting a scooter on Kinmen with my parents?” and the wait for a replacement may crop up again; switch to Google’s Gemini and it may pick out something else entirely.

So sentiment has to be measured with a whole set of questions customers would ask, judged good or bad against the same standard, and asked again after a while, before you can tell whether May’s new scooters have made it into what AI reads. Which old issue gets mentioned most, and how clearly your website has to spell things out before AI picks up that this place has changed, is different for every business.

That’s what our managed GEO service does: we fix your website content, bring what’s said about you elsewhere into line, then regularly put the same set of questions to each AI tool to see how it describes you and whether its tone has come back. You just look at the results.

If you’d like to know how AI describes you right now, you can book a demo with one of our consultants. Bring the old complaint that worries you most; the consultant comes up with a few questions a customer would ask about it, puts them to ChatGPT there and then, and goes through the answer with you to pick out the lines about you and see whether the thing you fixed long ago is still there. The same session also covers your website’s GEO audit score. The audit is free, and it looks at whether AI can understand your website.

Further reading