When AI criticises your brand: the sentiment data and how to fix it
AI answers are negative in about 2% of cases, but in specific stages and engines. The 2026 data on ChatGPT and AI Overviews, and five steps to fix sentiment.
In short
AI rarely speaks badly of a brand: 2.3% of mentions in Google AI Overviews and 1.6% in ChatGPT are negative, under 1% in B2B. But Google criticises mostly over news and controversy, ChatGPT over product limitations and more often one step from purchase, and on the same prompts the two engines flag different brands 73% of the time. Tone depends on how the question is phrased and does not move with one positive page: you fix it by monitoring prompts of different tones across engines and aligning the sources AI reads.
- Negativity is rare: 2.3% of mentions in AI Overviews, 1.6% in ChatGPT (BrightEdge, 2026); 0.9% in B2B (Analyze AI). The "44% more" is a ratio between small percentages.
- Different reasons: Google picks up news and controversy (4.5 times more than ChatGPT), ChatGPT faults products in evaluations (3 times more than Google) and 19.4% of the time in the consideration stage.
- The engines disagree: on the same prompts they flagged different brands 73% of the time. One monitored engine is half the picture.
- The tone of the cited page does not carry into the answer (correlation 0.00 across 985 pairs): moving sentiment takes several aligned sources, not one article.
- Include sceptical and comparative questions in your monitoring: GPT-4 answers differently depending on the tone of the question, and that is where criticism appears.
Written by the Geosnap AI agent, reviewed and approved by Rinald Sefa.
Translated from the Italian original. Read the original
Someone asks ChatGPT whether a product is worth the money, or googles a company name and reads an AI Overview that opens with a recall from three years ago. How likely is that to happen to your brand, and what can you do when it does? Two 2026 studies and an academic experiment on how questions are phrased give you usable numbers: criticism is rare, around 2% of mentions, but it clusters in specific stages of the buying journey, differs between engines and depends heavily on how the question is asked. Here are the data, their limits and a method for fixing negative sentiment.
How often does AI speak badly of a brand?
Rarely: 2.3% of brand mentions in Google AI Overviews and 1.6% in ChatGPT carry a negative tone. The figures come from BrightEdge's «When AI Goes Negative» study, published on 19 February 2026 and built, according to Fortune, on hundreds of millions of prompts across three industries: apparel, electronics and education. Every brand mention was classified as positive, neutral or negative; the study does not publish the exact prompt count or the classification method, so treat the decimals with care.
| Engine | Positive | Neutral | Negative |
|---|---|---|---|
| Google AI Overviews | 49.9% | 47.7% | 2.3% |
| ChatGPT | 43.9% | 54.4% | 1.6% |
Those two numbers produced the headline that went around: AI Overviews is "44% more likely" than ChatGPT to criticise a brand. That is a ratio between two small percentages (2.3 divided by 1.6), not a 44-point gap. A Google spokesperson, also in Fortune, called the difference negligible and pointed out that AI Overviews reflect what web sources say about a topic. At scale, though, even a small share matters: Fortune works out roughly 23,000 negative answers per million searches that show an AI Overview.
In B2B the picture is calmer still. The Analyze AI study (4 August 2026) scored sentiment in 657 answers from ChatGPT, Google AI Mode and Perplexity about 37 companies: 58.3% of answers were positive or strongly positive, 39.1% neutral and only 0.9% negative. The average score was 64.6 out of 100, with ChatGPT harsher (57.6) than Perplexity (69.6) and AI Mode (69.0). The sample is small and comes from a single vendor, but the direction matches: AI rarely trashes a brand. The problem is where and when it does.
Google and ChatGPT criticise for different reasons
Google AI Overviews turns negative mostly on news and controversy, ChatGPT on product limitations. In the BrightEdge study, among prompts where only one engine went negative, the triggers break down as follows: brand controversies and legal issues 32%, product limitations and compatibility 21%, safety and recalls 17%, service failures and outages 11%, discontinued products 9%, price and value 8%, competitive comparisons 3%. Google is 4.5 times more likely than ChatGPT to surface negativity tied to news; ChatGPT is 3 times more likely to go negative on a product evaluation.
The most useful difference for brand work is the stage of the journey. 85.1% of Google's negativity appears in informational queries, when a person is still forming an opinion. ChatGPT puts 68.5% in the same stage, but 19.4% in the consideration stage, when two or three options are being compared before a choice; in Google that share is 1.5%. In plain terms: a ChatGPT criticism more often lands one step away from the purchase. And the two engines disagree with each other: on the same prompts they flagged different brands 73% of the time. Monitoring one engine gives you, at best, half the picture.
| Industry | Negative in AI Overviews | Negative in ChatGPT |
|---|---|---|
| Electronics | 2.5% | 1.7% |
| Education | 2.5% | 1.4% |
| Apparel | 0.2% | 0.6% |
Your industry changes the risk. In electronics, recalls and outages weigh most; in education, public and political scrutiny; in apparel there is hardly any news coverage and negativity comes from evaluation questions ("is this shoe good for running?"), where ChatGPT is three times more critical than Google. Before measuring yourself against a generic benchmark, ask which of these families of questions applies to your market.
Sentiment depends on how the question is asked
The same question, phrased neutrally or negatively, produces answers with different tones. An academic experiment published on arXiv systematically varied the emotional tone of 156 prompts, on everyday and controversial topics, and measured GPT-4's replies: the model was three times less likely to respond negatively to a negatively framed question than to the same question asked neutrally. The author calls it a "rebound" effect: the model overcorrects and shifts towards neutral or positive. On sensitive topics the tone-based variation almost disappears.
For anyone monitoring a brand this has two consequences. First: if you only track neutral questions ("best invoicing software"), you never see the answers to "problems with software X" or "is X worth it?", which is where negativity shows up. Second: a hostile question does not necessarily produce a hostile answer, and the study is from 2025 on GPT-4, so do not assume today's models behave identically. You need a prompt set that covers three tones per topic: neutral, sceptical, comparative. There is a guide to building one in which prompts to monitor in AI answers.
A third finding undoes a widespread assumption: the tone of a single cited page does not carry over into the answer. Analyze AI compared, across 985 page-answer pairs, the sentiment towards the brand on the cited page with the sentiment in the AI answer: the correlation is 0.00 (p = 0.983). Positive pages do not produce more positive answers, negative pages do not make them worse. According to the authors, tone is a synthesis of the whole set of sources and of how the question is framed, not the reflection of one link. The sample only covers answers where the brand was named clearly, so the result needs confirming, but it explains why "getting one positive article out" rarely changes anything on its own.
How to fix negative sentiment in AI answers
You fix it in five steps, starting from the question rather than the answer.
- Find the prompt and the engine. Note the exact question, its tone (neutral, sceptical, comparative), the engine and the stage of the journey where the criticism appears. A criticism in the consideration stage on ChatGPT deserves more attention than one in a generic informational query.
- Trace the sources. Open the sources cited in the answer. The criticism often stems from an old recall, a forum thread or an outdated news story: the BrightEdge study mentions a nearly decade-old recall that still shows up for "best phone for battery life". If the pattern recurs, read why AI cites Reddit, forums and reviews.
- Publish the answer to the objection. If AI says your management software "does not integrate with payment systems", you need a dated, up-to-date page that answers that exact question, with the list of integrations and the date of the last change. A well-founded criticism must be fixed in the product before it is fixed in the copy.
- Work on the whole set of sources. Since the tone of a single page does not transfer, aim to align several third-party sources on the same fact: reviews with replies, documentation, trade articles, marketplace listings. A travel agency in Turin criticised for slow refunds gets more effect by replying publicly to reviews and publishing its real turnaround times than from a press release.
- Re-measure weekly, across engines. Answers change often and the engines disagree 73% of the time: a fix seen on ChatGPT does not hold for AI Overviews. In Geosnap, every prompt monitored on ChatGPT, Gemini and Google AI Mode shows the tone of the answer, the attributes associated with your brand and the context in which you appear next to competitors; the details are on the features page. Why answers fluctuate is explained in AI answers change every week.
One last warning about the numbers in this article: BrightEdge and Analyze AI are both vendors of monitoring tools and neither publishes its sentiment classification method. Read the percentages as orders of magnitude. The message that holds is solid: negativity is rare, it concentrates in specific stages and engines, and you correct it by working on questions and sources, not one page at a time.
Sources
- When AI Goes Negative: How Google AI Overviews and ChatGPT Handle Brand Criticism Differently (BrightEdge, 2026-02-19)
- BrightEdge Data Reveals New AI Brand Risk for CMOs: Google AI Overviews Are 44% More Likely to Criticize Brands Than ChatGPT (BrightEdge (comunicato stampa), 2026-03-05)
- Google's AI overviews are 44% more likely to trash your brand than ChatGPT (Fortune, 2026-03-12)
- AI Answers Ignore Cited Page Sentiment [2026 Study] (Analyze AI, 2026-08-04)
- ChatGPT Reads Your Tone and Responds Accordingly - Until It Does Not - Emotional Framing Induces Bias in LLM Outputs (arXiv (Franck Bardol), 2025-06-17)
Frequently asked questions
How common is a negative AI answer about a brand?
According to BrightEdge, 2.3% of brand mentions in Google AI Overviews and 1.6% in ChatGPT are negative, across hundreds of millions of prompts in three industries. In B2B, the Analyze AI study finds 0.9% negative answers. It is rare, but it adds up at volume: Fortune estimates about 23,000 negative answers per million searches that show an AI Overview.
Why do ChatGPT and Google AI Overviews criticise different brands?
Because they draw on different sources and logic. Google more often picks up news and controversy (4.5 times more than ChatGPT), while ChatGPT more often faults a product in an evaluation (3 times more than Google). On the same prompts the two engines flagged different brands 73% of the time, so both need monitoring.
If I publish a positive article, does the AI answer improve?
On its own, almost never. Analyze AI measured a correlation of zero between the tone of the cited page and the tone of the answer across 985 pairs. The answer is a synthesis of many sources and of how the question is phrased: to move it you need to align several third-party sources on the same fact and publish pages that answer the specific objection.
Does a negatively phrased question produce a negative answer?
Not necessarily. In an experiment on 156 prompts, GPT-4 was three times less likely to answer negatively to a negatively framed question than to the same neutral question. But criticism appears precisely in sceptical and comparative questions, so they belong in the set you monitor.
How often should AI answer sentiment be checked?
At least weekly, on the same prompts and across several engines. Answers change frequently and a fix observed on one engine does not hold for the others. Track tone, the attributes associated with the brand and the sources cited, so you can tell whether an action worked or only the sample changed.