The kind words that you see when you check what AI says about you are the answer to just one way of asking. Our guess is that business owners checking on themselves mostly type something like “recommend me a…”, while customers who are close to deciding may well ask “what are the downsides of this place?” or “any red flags?” Both questions are about you, yet the tone of the two answers can be very different, so seeing AI say nice things about you is no reason to relax yet.

Someone has now actually measured this. SE Ranking makes SEO tools, and on 4 November one of its products, SE Visible, published a study on its own blog. The team took the same point about the same brand and wrote it three ways: favourable, neutral and sceptical. For Samsung’s Frame TV, that meant asking for reasons to choose it, what to consider before choosing it, and its biggest drawbacks. They then classified ChatGPT’s 900 answers as positive, neutral or negative.

The study covered three categories: TVs, airline loyalty programmes and telehealth services for GLP-1 weight-loss drugs. This is a vendor’s own research, it tested only ChatGPT, the published sample questions are all in English and no Taiwanese brands were included, so it can’t tell us whether customers in Taiwan asking in Chinese would see the same thing, and we’re only treating it as a rough guide.

Sceptical questions have about twice the effect of favourable ones

The result: change the tone of the question and the tone of the answer tends to change with it, in all three categories. The two directions aren’t equal, though. Measured from the answers to neutral questions, sceptical questions hurt the tone about twice as much as favourable ones helped it.

What’s more, even answers to neutral questions lean slightly positive on average, so what you normally see happens to be the better-looking side. Celeste Gonzalez, who wrote the study, warns that if you only test neutral or favourable questions, you may come away with a much rosier picture than someone who arrives with a worry.

Say you run a gym in Yongkang, Tainan. Curious how AI sees you, you type “best gyms in Yongkang”. The answer lists you and says the place is pleasant, and you stop worrying. Meanwhile an office worker nearby is due to sign a one-year contract next week, and what he asks is “What are the downsides of this gym? Can I get a refund if I quit halfway through?” It’s the same gym, but he may well be reading a different tone, and you’ll never see that conversation.

How much the tone changes depends on how AI already sees you

There’s a limit to how much the wording can change the tone, though. Where it ends up also depends on the brand and the category. The study scored tone from -100, all negative, to +100, all positive, with 0 being neutral. With sceptical questions, TVs averaged about -11, still close to neutral, while the GLP-1 weight-loss telehealth services fell to about -53, with clearly more negative than positive.

The author’s reading is that when AI answers, it has more to go on than the tone of the question, including the information and reviews that it has read. So even when customers go straight for a weakness, some brands hold up: Sony’s TVs, asked about one particular weakness, still came out positive.

For a single business, we’d expect much the same: if what AI can read about you describes you clearly and your reviews are good, the tone won’t get as bad when a worried customer asks.

On refunds and warranties, the tone barely improves

The type of question that owners should watch most is what the study calls trust questions, such as whether you can get a refund, privacy, safety and warranties. Even when these were worded positively, the tone of the answers didn’t improve in any meaningful way, and they still came with caveats. The author’s reading is that answers on these subjects come with caveats anyway, so favourable wording doesn’t change them much.

Our guess is that these are the very things on a customer’s mind just before they pay. As for how customers ask, you have no say in that, which is where the study ends up too: people are going to ask these sceptical questions, and what a brand can influence is the information that ChatGPT has to work with when they do. So whether your website spells out how refunds are calculated and what the warranty covers is the part that you can actually change.

A few negative lines won’t tell you whether the wording caused them

You might think you could just ask “any red flags?” yourself. You’d get an answer, but when you see a few negative lines, you can’t tell whether the question caused them or whether that’s how AI sees you anyway. The author warns that a negative answer isn’t necessarily evidence that ChatGPT has a negative view of the brand. To tell the two apart you have to set different wordings of the same point side by side, and since AI’s answers vary from one run to the next, asking once isn’t reliable either.

Even after all that, you’ve only covered one point. What customers worry about differs from one line of business to the next, and owners can’t always guess it: you may think your refund terms are perfectly clear while the customer is wondering whether you’ll still be open when the contract ends. Working through them one at a time, then checking again later, isn’t something that you can finish with a few questions after closing time.

What you normally see is the favourable side, and the other side you can catch up on in a demo. A demo is time set aside with one of our consultants, who shows you live. Take the thing that customers in your line of business worry about most: the consultant asks ChatGPT about it there and then in both favourable and sceptical wording, and you can read the difference in tone for yourself. In the same session we’ll also take you through your website’s GEO audit score: geoweb.tw’s free website audit, which looks at how well AI can understand your site across 12 areas.

A demo shows you today’s tone; improving what AI reads about you takes long-term work. GEO stands for generative engine optimisation, and that’s what it does. It’s our core business, and handing it to us is what we call managed GEO: we rewrite the parts of your website that need to be clearer, bring what other websites and maps say about you into line, then regularly put the questions that customers ask to ChatGPT, Gemini and other leading AI engines to see whether they talk about you in a different tone. All you need to do is review the results.

Further reading