The question you ask ChatGPT changes the answer it gives about your brand — and a negative question changes it about twice as much as a positive one. A study published on November 4 by Celeste Gonzalez of Lastmile Retail and the AI visibility platform SE Ranking tested 180 prompts across 900 ChatGPT responses and found that negative wording shifted brand sentiment by 56 points from the neutral baseline, while positive wording moved it by only 26 points. The finding lands as a warning for any brand tracking how AI assistants describe it: if your monitoring only uses neutral or flattering prompts, you are probably seeing a rosier version of your brand than your customers see.
Gonzalez and SE Ranking built 180 prompts — 60 positive, 60 neutral and 60 negative — across three product categories: Samsung versus Sony televisions, Hers versus Start Willow GLP-1 telehealth services, and Delta SkyMiles versus United MileagePlus airline loyalty programs. The questions spanned four intents: recommendations, comparisons, trust and value. Each prompt was submitted to ChatGPT five times, producing 900 responses in total, and brand sentiment was scored on a scale from -100 to +100 without the classifiers seeing the original prompts.
On that scale, neutral prompts averaged a net sentiment of about +20, meaning ChatGPT's answers to supposedly neutral questions were already mildly positive. Negative prompts averaged -36 and positive prompts averaged +46. The gap between the two directions is the headline: negative framing pulled sentiment 56 points below the neutral baseline, while positive framing pushed it only 26 points above it. Across all 180 prompts, the correlation between prompt sentiment and response sentiment was a Spearman rho of 0.66 — a strong relationship, but one that left plenty of room for the context of the question to matter.
The effect was not uniform. Brand comparisons produced the biggest sentiment swings, which makes sense: when a question pits two brands against each other, the framing of the question effectively picks a winner before the model starts reasoning. Trust questions behaved very differently. Asked about reliability, refunds or privacy, ChatGPT's answers resisted positive framing — a glowing question did not buy a glowing answer about a brand's trustworthiness. That asymmetry matters for reputation work, because the questions where a brand most wants a positive spin are the ones where positive spin works least.
There is a second problem buried in the data: repeatability. Only 44% of prompts received the same sentiment classification across all five runs, although 96% produced a majority classification. In other words, a single run of a single prompt is not a measurement. At the individual response level, 43% of the 900 responses rated the target brand positively, 29% neutral and 28% negative — a distribution that looks far less flattering than the mildly positive average for neutral prompts alone suggests.
For anyone doing AI visibility work, the practical lesson is to stop measuring brand sentiment with one prompt and one run. A defensible tracking setup now needs matched prompt sets — the same question asked positively, neutrally and negatively — repeated several times, with the results averaged. Anything less will flatter the brand and hide the exact questions where customers are most likely to hear something negative: comparisons and trust queries, the two intents closest to a purchase decision.
The study also reframes what link building and PR are for in the AI era. ChatGPT's answers draw on what the web says about a brand, and the framing of the user's question decides how much of the negative side of that record surfaces. Earning strong, independent coverage and genuine reviews is no longer just a ranking play; it is the raw material the model weighs when a shopper asks a loaded question. A brand with a thin public record has little positive material for the model to draw on when the question turns hostile — and hostile questions, this study shows, move the needle twice as far.
Sources: Search Engine Land, November 4, 2026, reporting a study by Celeste Gonzalez of Lastmile Retail and SE Ranking.
Related reading on this site: Google Organic Click Share Hits a 15-Month High as Zero-Click Searches Fall, Reddit Begins Testing Ads in Its Own Search Results, 83% of ChatGPT Shopping Picks Also Rank in Google's Top 40, Semrush Finds, Google Rewrites Its 'Crawled – Currently Not Indexed' Guidance, Search Console's Device Filter Is Now Multi-Select, and Email and Paid Search Are Growing as Organic Search Shrinks.
Linkbaits helps you create content that earns citations and backlinks. Start with the toolkit and build your first linkable asset this week.
Start free trial