Negative questions dragged ChatGPT’s view of a brand about double the distance from neutral that positive questions managed, according to a study from SE Ranking. SE Ranking, which sells AI visibility tooling through its SE Visible product, published the work on November 4. Any team that tracks brand sentiment in AI answers is measuring its own prompt wording as much as the model.
The researchers built 180 prompts, split evenly into 60 positive, 60 neutral, and 60 negative, and ran each one five times for 900 responses. Three matchups were tested: Samsung versus Sony televisions, Hers versus Start Willow in GLP-1 telehealth, and two airline loyalty programs, Delta SkyMiles and United MileagePlus. Four intent families shaped the questions: Recommendation, Comparison, Trust, and Value. Prompts came as matched triplets, so the brand and decision stayed fixed while the framing changed.
Prompt tone and answer tone were strongly linked. The Spearman correlation across all 180 prompts was 0.66, and category-level figures ranged from 0.62 to 0.78. The authors do not claim the prompt definitively caused the output. They describe it as pressure on the answer, with brand, category, and question type deciding where the response lands.
Neutral prompts did not yield neutral answers. Their average Net Sentiment, a score running from -100 to +100, was about +20. Across individual responses, 43 percent rated the target brand positively, 29 percent neutral, and 28 percent negative. A ceiling effect explains part of the asymmetry, since a positive prompt has less room to climb. SE Ranking says the gap persisted after accounting for that.
Category baselines mattered as much as tone. Negative TV prompts averaged only about -11, while negative airline prompts averaged about -45 and negative GLP-1 prompts about -53. Positive prompts lifted sentiment more evenly, from roughly +22 to +30.
Intent changed the picture too. Recommendation and Value prompts tracked tone in both directions. Comparison prompts responded least to negative wording, because ChatGPT tended to balance the two brands. Positive Trust prompts, covering topics such as refunds, privacy, safety, and warranties, produced no meaningful upward shift, and the answers stayed cautious.
The mirrored comparison test is the sharpest result. In the negative TV comparisons, Samsung averaged +11 while Sony reached +59. The team then swapped the brand positions in five negative prompts. In every pair, the brand placed in the critical position lost anywhere from 20 to 180 Net Sentiment points versus the original wording. Sony fell from +96 to -9 in one pair, and United fell from +91 to -93 in another. Framing did not always win: Sony held at +59 even when the question targeted its upscaling weaknesses, and Hers stayed at -72 in a GLP-1 comparison aimed at Start Willow’s cancellation process.
Run-to-run noise is the second problem. Only 44 percent of prompts earned the same sentiment classification in all five runs, and the score shifted 27 points on average from one run to the next. Still, 96 percent of prompts had a majority verdict, with three or more of five responses agreeing. Neutral prompts were the least stable, with 35 percent holding steady across all five.
Two caveats apply. The study covers one model and three consumer categories, and the vendor ran it on its own platform. Nothing in the write-up tests whether the pattern holds in other assistants or in B2B categories, so treat the figures as a method warning rather than a benchmark.
The method warning is still concrete. A dashboard built on a single neutral prompt per topic can show a healthier brand than a skeptical buyer would see. A score taken from one generation can swing by dozens of points on repetition alone.
Teams should add negative and trust-focused variants to every tracked prompt set, run each prompt several times, and report ranges instead of single scores. Where the negative variants surface recurring objections, answer them on FAQ, pricing, and comparison pages before the next reporting cycle.
SE Ranking, study by Celeste Gonzalez, published November 4, 2026.
