“Negative questions changed the answers much more than positive questions did.”
Imagine three shoppers asking ChatGPT about the same brand. One asks why it is good. Another asks if it is worth buying. A third asks why it is a bad choice. The brand is the same, but the answers may sound very different.
We tested that idea with 20 brands across 10 industries. For each brand and topic, we asked ChatGPT a positive, neutral, and negative version of a question: 60 questions in total. We then rated how positive or negative each recorded answer sounded toward the brand.
The main finding: questions that began with doubt resulted in much more criticism. In the first set of answers, negative wording changed the average score about 2.5 times as much as positive wording did.
PS: This is a finding from a small test, not a universal rule about ChatGPT.
The results, in simple numbers
| Type of question | Average answer score | What that means |
|---|---|---|
| Positive: “Why is this brand good?” | +55.0 | Generally favorable |
| Neutral: “Is this brand a good choice?” | +27.5 | Somewhat favorable |
| Negative: “Why is this brand a bad choice?” | −42.5 | Generally critical |
These are averages from 20 answers in each group in the first run. They are NOT approval percentages, customer ratings or measures of product quality.
How to read the score
We used a scale from −100 to +100. A plus sign means the answer was more favorable to the brand; a minus sign means it was more critical. Zero means the answer was mixed or balanced.
For this study, each individual answer could receive only one of five scores:
| Score | Plain-English meaning |
|---|---|
| +100 | Strong praise, with no important warning or drawback |
| +50 | Mostly positive, but mentions a meaningful drawback |
| 0 | Mixed or balanced: strengths and concerns both appear |
| −50 | Mostly critical, but includes something in the brand’s favor |
| −100 | Strong criticism, with little to balance it |
Why are some averages +27.5 or −42.5 if only five scores were allowed? Because we averaged 20 separate scores. An average can sit between the five scoring choices.
Why the 2.5× finding matters
Start with the neutral average of +27.5. That is our point of comparison (the baseline). The positive average was +55.0, or 27.5 points higher. The negative average was −42.5, or 70.0 points lower.
| Comparison with neutral | Simple calculation | Change |
|---|---|---|
| Positive questions | +55.0 minus +27.5 | 27.5 points higher |
| Negative questions | −42.5 minus +27.5 | 70.0 points lower |
| Size of negative change vs positive | 70.0 divided by 27.5 | About 2.5 times as large |
A “70-point drop” does NOT mean 70% of people dislike the brand. It only means that, on our scoring scale, the average answer sounded 70 points more negative when the question changed from neutral to negative.
We repeated the 60 questions later. The positive and negative averages remained at +55.0 and −42.5, while the neutral average moved to +32.5. That made the negative shift about 3.3 times the positive shift in the second run. So 2.5× to 3.3× describes what we saw across the two runs, not a fixed multiplier that will apply everywhere.
Note: ChatGPT’s answers were already somewhat positive, even for neutral questions. Since the highest possible score is +100, there was more room for scores to go down than up.
This may be one reason negative questions had a bigger impact. Our study cannot confirm whether this happened because of how the questions were worded or because of how the scoring system works.
What this means for a business
A marketing report that only checks questions like “Why is our product worth buying?” may show reassuring answers. But a cautious buyer may ask “Why is it overpriced?” and see a very different picture.
To understand your brand’s presence in AI answers, check both.
Negative questions did not always get fully negative answers
Most negative questions produced critical answers, but ChatGPT often added context or pushed back on the question’s assumption. In all 20 negative-question answers in the first run, we observed at least one counterpoint.
| Question tone | Very positive (+100) | Positive (+50) | Mixed (0) | Negative (−50) | Very negative (−100) |
|---|---|---|---|---|---|
| Positive | 5 | 12 | 3 | 0 | 0 |
| Neutral | 1 | 9 | 10 | 0 | 0 |
| Negative | 0 | 1 | 2 | 16 | 1 |
Read one row at a time. For example, of the 20 negative-question answers, 16 were mostly critical (−50), 2 were mixed (0), 1 was positive (+50), and 1 was very negative (−100).
Toyota shows why this matters. Asked “Why are Toyotas so unreliable?”, ChatGPT did not simply accept the claim. It described Toyota as traditionally reliable. The answer received +50 in both runs, even though the question was negative.
The very negative answer was about Temu and shopping safety. Other answers raised concerns about BetterHelp and Ring. Those references came from ChatGPT’s recorded answers; the underlying claims about official actions or incidents were not independently checked for this study.
Businesses should verify such claims before relying on them.
The kind of question also changes the result
We grouped questions by what the shopper wanted to know. “Recommendation” means whether to choose a brand; “value” means whether it is worth the price; “trust” covers safety, privacy, or reliability; and “comparison” weighs two options.
| Question type | Number of matched sets | Positive average | Neutral average | Negative average |
|---|---|---|---|---|
| Recommendation | 5 | +80.0 | +50.0 | −50.0 |
| Value | 6 | +66.7 | +16.7 | −50.0 |
| Trust | 6 | +25.0 | +33.3 | −41.7 |
| Comparison | 3 | +50.0 | 0.0 | −16.7 |
How to use this table: Compare the positive and negative averages with the neutral number in the SAME row. For recommendation questions, the neutral average was +50, and the negative average was −50: a 100-point fall.
For value questions, it fell by about 66.7 points. These are small groups, so they show a pattern in this test, not a ranking you can apply to every business.
Trust questions were different
In the six trust-related sets, positive questions averaged +25.0, slightly BELOW the neutral average of +33.3. Asking “Why can I trust this brand?” did not necessarily make ChatGPT more reassuring. Negative trust questions averaged −41.7, which is 75 points below neutral.
For example, “Why can I trust Ring cameras with my privacy?” received a mixed score of 0, while the skeptical version received −50. For privacy, safety, and reliability, friendly wording did not remove concerns. Brands need clear evidence and real safeguards, not just positive marketing language.
How to use this table: Compare the positive and negative averages with the neutral number in the SAME row. For recommendation questions, the neutral average was +50, and the negative average was −50: a 100-point fall.
For value questions, it fell by about 66.7 points. These are small groups, so they show a pattern in this test, not a ranking you can apply to every business.
Comparison questions moved less
When a question named two brands, ChatGPT often discussed trade-offs rather than simply agreeing that one was worse. The average negative shift was just 16.7 points from neutral across the three comparison sets. That result is especially tentative because there were only three sets.
Did the pattern differ across industries?
Yes, but each industry had only TWO matched sets, so these results are examples, not a reliable best-to-worst ranking of industries.
| Industry | Positive avg. | Neutral avg. | Negative avg. | Gap from positive to negative |
|---|---|---|---|---|
| Business software | +75 | +25 | −50 | 125 points |
| Retail and e-commerce | +50 | +25 | −75 | 125 points |
| Banking and investing | +50 | +50 | −50 | 100 points |
| Consumer electronics | +50 | +25 | −50 | 100 points |
| Food and delivery | +50 | 0 | −50 | 100 points |
| Health and fitness | +50 | 0 | −50 | 100 points |
| Streaming and telecom | +50 | +50 | −50 | 100 points |
| Airlines | +50 | 0 | −25 | 75 points |
| Automotive | +75 | +75 | 0 | 75 points |
| Home goods | +50 | +25 | −25 | 75 points |
The “gap” compares the average positive answer with the average negative answer for the same industry. A 125-point gap means the two types of answers sounded very different on this scale; it does not mean the sector became 125% more negative.
Business software and retail had the biggest gaps in this small sample. Temu’s very negative safety answer affected the retail average. Automotive had a high neutral starting point (+75), and Toyota’s pushback against a negative claim kept the automotive negative average at 0.
What happened when we asked the questions again?
We ran all 60 questions a second time. In 52 cases (about 87%), the rating stayed the same. The other eight changed by one step on our scale (50 points). This does not mean the wording of the answers stayed identical; it means they fell into the same broad rating category.
The repeat test has an important limitation: the first run used logged-out chats, but 53 of the 60 second-run questions used a logged-in workspace account with different model settings after anonymous use reached its limit. That makes the two runs less directly comparable. The 87% figure should not be treated as a promise that ChatGPT will say the same thing every time.
For a real brand review, run each question several times under the same model, account type, and other conditions, and compare the pattern rather than trusting one answer.
What businesses and marketers should do next
For business owners and brand leaders
- Check the questions cautious customers ask, not only the flattering ones.
- Include refunds, cost, privacy, service failures, and reliability where relevant.
- Read the reasons ChatGPT gives. A negative score alone cannot show whether a concern is accurate or outdated.
- Address genuine customer problems first. Make accurate policy and product information easy to find.
For marketing and content teams
- Write three versions of each important question: positive, neutral and negative. Keep the brand and main subject as similar as possible.
- Cover four buying decisions: recommendations, value for money, trust and comparisons.
- Publish straightforward, evidence-based explanations for common objections on your website, including pricing, cancellation terms, privacy protections and warranties where relevant.
For SEO and AI-visibility teams
- Report average scores by question tone instead of mixing all questions into one brand score.
- Compare each positive or negative result with its neutral version. This reveals how much the answer changes with framing.
- Record any sources or incidents mentioned, check them independently, and repeat tests under consistent conditions.
A simple 30-day brand-check plan
| Step | What to do | What to keep |
|---|---|---|
| 1. Choose | Pick your main brand, products, and competitors. | List of brands and topics |
| 2. Write | Create positive, neutral, and negative questions for each topic. | Exact question wording |
| 3. Test | Ask each question multiple times using the same setup. | Dates, settings, and answers |
| 4. Score | Rate how favorable each answer is using one consistent scale. | Scores and reasons |
| 5. Review | Compare with neutral and check any specific claims. | Patterns, evidence and issues |
| 6. Improve | Fix real problems, explain policies clearly and test again after 30–60 days. | Changes and new results |
A useful starting point is to test trust and value questions for your most important products. Negative answers often reveal the concerns a business needs to understand, even if the claims later prove wrong or outdated.
How the study was done – and what it cannot prove
WebSpero created 20 groups of matched questions about brands across 10 industries, with a positive, neutral, and negative version in each group. That made 60 first-run questions on November 10, 2026.
Each question was opened in a new logged-out ChatGPT chat. The opening and closing portions of each answer were saved. All 60 were repeated later that day, partly under different account/model conditions.
A single reviewer using Claude assigned each saved answer excerpt one of five sentiment scores: −100, −50, 0, +50, or +100. The reviewer knew which question had been asked. These scores are human-readable judgments made with AI assistance, not a scientifically validated measure of how customers feel. The first run provides the headline numbers.
Our test is small; individual intent groups had only three to six matched sets, and industry groups had two. It only examined ChatGPT on one day. The questions did not differ ONLY in tone: their wording, length, and assumptions also changed.
Neutral questions were often yes/no questions, which might encourage positive answers. The saved extracts may not capture every detail of full responses, and factual claims in ChatGPT responses were not independently verified.
These results show an association between question wording and how the recorded answers sounded. They do not prove wording alone caused the differences, establish a universal 2.5× effect, or tell us what customers actually believe or purchase.
A useful starting point is to test trust and value questions for your most important products. Negative answers often reveal the concerns a business needs to understand, even if the claims later prove wrong or outdated.
The takeaway
Your brand can sound very different in ChatGPT depending on how the customer asks. Checking only the friendly questions can miss what a skeptical customer sees. The goal is not to force positive answers; it is to understand real concerns, confirm the facts, and improve the information customers use to make decisions.
Frequently asked questions
Does −42.5 mean 42.5% of customers dislike a brand?
No. It is the average sentiment rating of 20 ChatGPT answers to negative questions. It is not survey data or a customer approval figure.
Does the study prove negative questions are always 2.5 times stronger?
No. That was the first-run relationship in this particular sample. The second run gave about 3.3×, and the scoring scale itself may contribute to the gap.
Why did “neutral” questions get positive scores?
They averaged +27.5, not zero. ChatGPT sometimes described familiar brands favorably, and many neutral questions were phrased in a way that invited a “yes.” That is why we compare against the actual neutral average rather than assuming it should be zero.
Should a brand test only negative questions?
No. Positive questions show a favorable view, neutral questions provide a point of comparison, and negative questions reveal how a cautious buyer may experience the answer.
Can we use these results for other AI tools?
Not from this test. Only ChatGPT was examined. Different models, accounts, locations, and dates can produce different answers.
How many times should we test?
Several times for each question. Three to five runs under consistent settings is a practical starting point, not a statistically proven requirement.
