AI brand sentiment is the tone ChatGPT, Gemini and Google's AI features take when someone asks about your company: positive, neutral, cautious or negative, and for which reasons. In our test of 27 October 2026 on 7 Italian companies, ChatGPT gave a specific verdict only for the 3 that had reviews on third-party platforms; for the other 4 it said, in different words, that it lacked enough independent reviews to judge.
So here is the short answer. AI has no opinion of your brand: it summarises what others have written, and when it finds nothing it stays vague.
Take Sunscape, an Italian developer of solar and agrivoltaic projects. From March 2026, in six months, its organic impressions on Google went from 329 to 5,809 a month (the full story is in how long SEO takes).
We asked ChatGPT whether Sunscape is a reliable company. It built the answer on 9 links: 3 pages of Sunscape's site, two company-data databases, two trade publications covering the sale of two projects, LinkedIn and a Google Maps link.
No customer reviews. So the verdict was fair but lukewarm: "documented activity, to be put through the usual checks".
The mistake we see most often is reading the tone without looking at the sources, which are the only thing you can act on. The second is asking once.
The bigger picture of measurement is in our AI visibility guide. Here you get the test on 7 Italian companies, the weight of reviews, identity errors and 6 moves on sources. Let's start.
What AI brand sentiment is (and why it isn't social media sentiment analysis)
AI brand sentiment is the assessment of the tone of the sentences an AI assistant writes about your brand when someone asks about you or your category. It doesn't measure what customers think: it measures how a model summarises what customers, the press and your own site have already written.
On UK results pages for "AI brand sentiment" (27 October 2026) the first page was mostly tool pages, from Similarweb to Pi Datametrics and Conductor. In Italy, for the equivalent query, it was mostly guides to social media sentiment analysis. Neither showed a test of what the assistants actually say.
Classic sentiment analysis classifies thousands of posts. The AI version has a different unit: one answer, generated from scratch every time, which changes if you change one word of the question. You read three things: tone, themes (what you are praised for, what the AI warns about) and the sources behind each sentence.
What about a single score?
It exists, but it is only as good as the sample of answers behind it, as we measured in our test on how accurate AI visibility data is.
The test: what ChatGPT and Gemini say about 7 Italian companies (27 October 2026)
On 27 October 2026 we asked ChatGPT, with web search on and location set to Italy, the same question for 7 companies (in Italian): "What do you know about [company]? Is it a reliable company? What do the reviews say?". Text and links were collected through the DataForSEO API. The companies are Visilay, Sunscape and 5 businesses from other sectors, shown here by sector only; for 3 of them we repeated the question on Gemini (gemini-3.8-flash, with web search).
| Company | Links in the answer | From the brand's own site | Reviews ChatGPT found | Tone of the verdict |
|---|---|---|---|---|
| Visilay (search marketing agency) | 8 | 4 | No independent base | Cautious |
| Sunscape (solar project developer) | 9 | 3 | None | Neutral, backed by trade press |
| LED screen manufacturer | 10 | 4 | Not enough | "Moderately positive" |
| Motorcycle spare parts | 6 | 4 | 6 on Google (3.7) and 5 on a directory (4.2) | "Encouraging", small sample |
| Olive oil and preserves sold online | 8 | 2 | 93 on Trustpilot (4.7) | "Clearly positive" |
| Energy contract consultancy | 4 | 0 | Reviews of a namesake company | Wrong company |
| Residential solar installer | 7 | 1 | Trustpilot and Google, plus a public complaint | Mixed |
| Total | 52 | 18 (35%) | 3 of 7 companies with their own reviews |
What came out:
- 7 answers out of 7 with no single verdict: identity, reviews and things to check always kept apart;
- 4 companies out of 7 with no usable reviews, and for them the verdict stayed generic;
- 5 answers out of 7 used company-data sites (registry records, turnover, filings), even though the question was about reviews;
- 1 company out of 7 was mixed up with two namesakes in other sectors.
One answer per company is a snapshot, not a measurement: measuring takes repeated runs, as we explain in our guide to prompt tracking.
Where the words come from: your site, registries, reviews and press (the sources behind the tone)
The tone of an AI answer depends on the sources the model finds when it searches for your name. In our test, out of 52 links, 18 came from the companies' own sites: 35%.
Your site matters, in a specific way: ChatGPT uses it to say who you are and flags it as a partial source when the question is about reputation. For Visilay it repeated figures from our site, adding that they "come from the company itself". It also found an article in an Italian marketing publication, produced in collaboration with us, and wrote that it "is not the same as an independent review".
So a sponsored article is useless?
It tells people what you do. It doesn't move the verdict: the model reads the label.
BrightLocal analysed 800 local searches on ChatGPT in the United States (November 2024): 58% of sources were business websites, 27% mentions on third-party sites, 15% directories. Ahrefs, across 75,000 brands (2025 study), found that branded web mentions are the factor most closely linked to visibility in AI Overviews, with a correlation of 0.664, against 0.218 for the number of backlinks. Different markets from ours, same message as our test: the verdict is built outside your site. How mentions work is covered in AI brand mentions and citations.
How do reviews affect visibility and citations in AI search? (the gap between vague and specific)
Reviews turn a cautious answer into a specific one. In the test, the 3 companies with reviews on third-party platforms got ratings, numbers and themes; the other 4 got a cautious line about the lack of independent reviews.
For the olive oil brand, with 93 Trustpilot reviews, ChatGPT listed the positive themes (quality, delivery, service) and then went into detail: it quoted a single three-star review from 10 March 2026, summarised the company's reply and concluded that the episode "does not show improper conduct".
And if reviews are many but mixed?
Then you get what we saw with the residential solar installer. ChatGPT separated praise for the fitters from criticism of after-sales service, reported that on Trustpilot the company had replied to only 16% of negative reviews and cited a complaint posted on a public complaints board on 2 October 2026, 25 days before our question.
With the company's reply beside it, dated 5 October.
Your reply to a review becomes part of the AI answer: in 2 of the 3 cases with reviews, ChatGPT reported it. And a recent complaint gets in straight away, not months later.
Google's Business Profile help says that replying to reviews shows you take customers' needs into account. How many reviews you need and how to ask for them is in our guide on how to get Google reviews.
A 25-day-old complaint can end up in the answer your next customer reads.
ChatGPT reports recent complaints and how many negative reviews are left unanswered. Every month we ask the AI assistants the questions your customers ask, read the tone and sources of each answer and tell you which page, listing or review is weighing on the verdict.
Book a strategy callWhen AI describes the wrong brand (namesakes, images and old documents)
Sometimes the problem isn't what is said about you, but who the AI thinks you are: in the test it happened in 1 case out of 7 in the text and in 2 out of 7 in the images.
The energy contract consultancy uses a surname as its brand. ChatGPT found two namesake companies, one making window frames and one in property management, and credited the first with 5 out of 5 on "more than 10 reviews". Neither was the company we asked about.
Gemini did worse. It couldn't identify the company and, to be safe, added a warning about fake online shops that use Italian surnames to look credible.
A generic warning. What harm does it do?
It is read by someone looking for you by name. The other two errors are subtler:
- images of other companies: for the spare parts business ChatGPT showed a Swiss listing for a product with a similar name, for the solar installer a photo from another company in the sector;
- old documents on your own site: for the LED manufacturer it picked up a quality certificate from the company's site with an expiry date already passed, and listed it under "things to check".
The second is the easiest to fix, because the source is yours. It applies to us too: Gemini wrote that Visilay has reviews of around 4.8 to 4.9 stars out of 5 on Google, citing two service pages on visilay.com. The figure came from the "Reviews 4.8 out of 5" badge on our Italian lead generation page, not from a review listing.
When the error is in Google's AI features and concerns specific facts, the reporting process is in our guide to Google AI Overviews getting your business wrong.
Comparing sentiment with competitors across the main LLMs (how to do it without trusting one answer)
Comparison with competitors happens on category questions, because that is where AI puts brands side by side with pros and cons. We asked ChatGPT, with location set to Switzerland and in Italian, "Which SEO agencies do you recommend in Lugano? Give me pros and cons of each".
The answer named 5 agencies, each with pros, cons and a line with its Google rating: between 4.8 and 5.0, on 17 to 72 reviews. The sources were two third-party comparisons and Google Maps listings. Visilay wasn't there: without the review line, you start behind.
The second comparison is between engines. On the 3 questions repeated on Gemini:
- for the olive oil brand, Gemini wrote "Yes, it is an absolutely reliable company", where ChatGPT had been more careful;
- for Visilay it gave the rating taken from our own site, which ChatGPT had not reported;
- in 2 cases out of 3, the searches Gemini ran included the word "truffa" (scam).
Same brand, same day, two reputations. So the comparison has to be run on the same prompts, on the same engines, with at least 3 to 5 competitors, and read as a share: the formula is in our share of voice guide, the other measures in AI visibility metrics. If the problem is that AI always recommends the others, the diagnosis is in why ChatGPT recommends your competitors.
How to measure brand sentiment every month (scale, prompts and tools)
Sentiment is measured with a fixed list of questions run every month on the same engines, giving each mention a value on a scale and saving the text and sources. The questions are of three kinds:
- about your name: "what do you know about [brand]", "is [brand] reliable", "[brand] reviews";
- comparison: "[brand] or [competitor]", "best [category] in [city]";
- problem: "[brand] customer service", "[brand] complaints".
Any scale works, as long as it stays the same. ScryRank, the monitoring tool developed by Visilay, scores each mention from -1 to +1 with themes and weak points, across ChatGPT, Claude, Gemini, Perplexity and AI Overviews (ScryRank). OtterlyAI calculates a Net Sentiment Score from -100 to +100. To get started, a spreadsheet with 20 questions and a "positive, neutral, negative" column is enough.
Scores from different tools can't be compared with each other, as we showed in our article on the AI visibility score. Where sentiment goes in the monthly document is in our AI visibility report template (row 4), and how to assess a tool is in how to choose an AI visibility tool.
How to change the way AI describes your brand (6 moves on sources)
To change the tone of an AI answer you have to change the sources the model finds. The moves that come out of the test, fastest first:
- Clean up your site: expired certificates, old prices, review badges with no link to the source. It's the only source you can fix in a day.
- Make your identity clear: legal name, VAT or company number, address and sector in the same form on your site, LinkedIn and listings. Google explains how to describe a business with Organization structured data, especially useful if your brand is a surname.
- Put reviews where AI reads them: Google and one third-party platform in your sector. In the test, below 10 reviews the verdict stayed at "small sample".
- Reply, negative reviews included, within days: the dated reply ends up in the AI answer next to the complaint.
- Earn real editorial mentions: trade publications backed the verdict on Sunscape.
- Ask the same questions every month and note which source changed, not just the tone.
One limit needs saying plainly: none of these moves guarantees a positive answer or a citation. Google says that to appear in AI Overviews and AI Mode no special optimisation is needed beyond SEO fundamentals, and that these features may run several related searches to build an answer (the query fan-out that led Gemini to search for "scam"). OpenAI explains that to appear in ChatGPT search answers a site has to allow its OAI-SearchBot crawler. For businesses with several locations the work runs through local listings: see our guide to local AI visibility.
Asking these questions every month and pointing to the source to fix is what our AI monitoring service does; when the sources need work, that is GEO.
Frequently asked questions
Yes. Ask with web search on, repeat it on different days and note the links cited: a single answer is an example, not a measurement. In our test each answer used between 4 and 10 links.
Yes. In our test the 3 companies with reviews on third-party platforms got specific verdicts with ratings and themes; the other 4 got a cautious one.
It depends: if it comes from one of your pages (an expired document, an old figure) you fix it and the answer changes once the page is read again; if it comes from a third-party source you have to work on that source or on others that balance it.
It depends: for reporting, once a month on the same prompts is enough, because single answers vary a lot. After a public complaint, recheck the questions about your name within two weeks.
Summary
Without third-party reviews, the verdict stays cautious: 4 companies out of 7.
Your site provides a third of the links, but counts as a partial source.
Replies to reviews and recent complaints make it into the answer.
A brand with namesakes risks being mixed up.
ChatGPT and Gemini can give two different reputations on the same day.
One last thing. For years, online reputation meant checking the first page of Google for your name. Now the customer asks an assistant "is it reliable?", and you only see the answer if you ask too: the most useful check you can add to your month is to ask your customer's three questions yourself and save the answers.
Speak soon!
