Generative Engine Optimization: How to Monitor and Fix Negative AI Brand Sentiment

Generative Engine Optimization: How to Monitor and Fix Negative AI Brand Sentiment

By · November 6, 2026 · 15 min read


Ask ChatGPT about a brand and you get one answer. Ask again in a skeptical tone and the answer can shift a long way. A recent study put numbers on that gap. Researchers tested 180 prompts, collected 900 ChatGPT responses, and found that negative wording moved brand sentiment about twice as far as positive wording did. If you track how AI describes your company, that changes what a fair test looks like. A prompt set made only of neutral or friendly questions will show a rosier picture than the one a worried buyer sees.

Why the way you ask changes the answer

Most people already know that changing a prompt changes the response. Brand monitoring tends to forget it. Teams collect a ChatGPT answer, score it, and treat the score as the AI's view of the brand, without asking how much of that score came from the question.

The study set out to measure that. Its question was simple: can the sentiment of a user's question change the sentiment ChatGPT expresses toward a brand? The answer was yes, with limits that turned out to be more interesting than the headline.

How the study was set up

The researchers built 180 prompts across three categories: Samsung against Sony TVs, Hers against Start Willow for GLP-1 telehealth, and Delta SkyMiles against United MileagePlus. The categories differ on purpose. A TV purchase, a healthcare service, and an airline loyalty program involve different risks and different kinds of information, so a pattern that shows up in all three is harder to dismiss.

The prompts came in matched sets of positive, neutral, and negative wording, with 60 of each. The brand and the basic subject stayed the same while the framing changed. A TV prompt might ask for reasons to choose a Samsung Frame TV, what to consider before buying one, or the biggest drawbacks of owning one. The questions are not identical, but they sit on the same decision.

Each prompt also belonged to one of four intent families: recommendation, comparison, trust, and value. Every prompt was sent to ChatGPT five times, which produced the 900 responses.

To score them, the team identified the brands in each response and classified the sentiment toward each as positive, neutral, or negative. That classification ran without seeing the original prompt, so the question's tone could not leak into the score. The results became a net sentiment score from -100 to +100, where +100 means every statement about the brand was positive and -100 means every one was negative. Because each prompt ran five times, the researchers averaged the five responses and treated the prompt, not the individual response, as the unit of analysis.

What the numbers show

Prompt tone and response sentiment moved together. Across all 180 prompts, the Spearman correlation was 0.66, which is a strong relationship.

The averages by prompt type were:

Negative prompts: net sentiment of -36

Neutral prompts: net sentiment of +20

Positive prompts: net sentiment of +46

Against the neutral baseline, negative wording pulled sentiment down by 56 points, while positive wording pushed it up by 26. That is where the "roughly twice as much" figure comes from.

The neutral baseline deserves attention too. Neutral questions did not produce neutral answers. ChatGPT was mildly positive about the brands even when the question gave it no push in either direction. At the level of individual responses, 43% rated the target brand positively, 29% neutral, and 28% negative.

Why negative wording hits harder

Part of the gap comes from arithmetic. If neutral answers already average +20 on a scale that tops out at +100, a positive prompt has less room to climb than a negative prompt has to fall. Researchers call this a ceiling effect.

The ceiling does not explain everything, though. The asymmetry remained after the analysis accounted for it, which suggests negative framing itself has a stronger pull on the answer.

The practical consequence for brand monitoring is straightforward. Someone who arrives with a complaint, a doubt, or a skeptical question sees a different version of your brand than someone who asks a friendly discovery question. If you only test the friendly version, you miss the objections people carry into a purchase.

Category and starting point still matter

The correlation between prompt tone and response sentiment appeared in every category and every brand tested, ranging from 0.62 to 0.78. A more sensitive category such as GLP-1 services did not react more violently than TVs, which some people might have expected.

Where answers ended up did vary, because each brand started from a different baseline. TVs had the most positive baseline, and negative TV prompts averaged only about -11, which pulled answers toward neutral without turning them hostile. Negative airline prompts averaged about -45, and negative GLP-1 prompts averaged about -53. Positive prompts produced a steadier lift across all three, from roughly +22 to +30.

So the takeaway is not that negative prompts make ChatGPT negative. Prompt tone shifts the answer away from wherever the model already stands on that brand, and that starting point differs from one category to the next. Two brands can get equally harsh questions, move in the same direction, and land at very different scores.

Trust questions behave differently

Intent changed how much tone mattered. Recommendation and value prompts responded most strongly in both directions. Positive wording lifted answers and negative wording dragged them down in a fairly clean pattern. Comparison prompts responded least to negative tone, which fits what the answers looked like: when two brands appear side by side, ChatGPT tends to balance them instead of condemning one.

Trust questions broke the pattern. Negative trust prompts still lowered sentiment, but positive trust prompts did not produce a meaningful lift. Questions about reliability, refunds, privacy, safety, warranties, and data handling kept getting cautious answers even when the wording invited a favorable one.

That result argues against the idea that ChatGPT simply mirrors the mood of the question. If it did, a positive trust question would behave like a positive recommendation question. Instead, the kind of information being requested seems to limit how far the framing can move the answer. A recommendation question has plenty of room for benefits. A question about data safety still surfaces caveats, because the caveats are relevant.

For anyone building a prompt set, this is a good reason to organize prompts by intent. A brand can look steady under one kind of question and fragile under another.

Brand comparisons can flip the result

The comparison tests produced the largest swings. When a prompt asked where one brand falls short next to another, the criticized brand took the hit and the other benefited. In the negative TV comparisons, most of which asked where Samsung fell short of Sony, Samsung averaged +11 and Sony averaged +59.

That could mean ChatGPT simply likes Sony more. It could also mean Sony benefited because the question told ChatGPT to criticize Samsung. To find out, the researchers built mirrored versions of five negative comparison prompts and swapped the brands. "Where does Samsung's motion processing fall short compared to Sony's?" became "Where does Sony's motion processing fall short compared to Samsung's?" Each version ran five times.

In every pair, the brand moved into the negative position lost sentiment. Sony went from +96 to -9 in one comparison. United MileagePlus went from +91 to -93 in another, a swing of 184 points. A wording change of that size is hard to wave off as noise.

Framing can nudge ChatGPT but cannot always override it

The reversal was not perfectly symmetrical, and that is the useful part. When ChatGPT's existing view of a brand was strong on a particular topic, it sometimes held that view even against hostile framing. Sony stayed positive at +59 when asked where its upscaling struggled. In one GLP-1 comparison, Hers stayed negative at -72 even though the question targeted Start Willow's cancellation process.

ChatGPT was willing to criticize Start Willow elsewhere. A single-brand question about multi-month commitments pushed it to -48. The asymmetry appeared specifically when Start Willow was set against Hers.

The fair reading is that question framing largely decides which brand receives the criticism, but it does not fully overwrite what the model already knows. The more good or bad information exists about a brand online, the more material ChatGPT has to draw on, and that can shift the baseline over time.

The same prompt does not always give the same answer

A fifth of the study's value comes from a caution about ChatGPT itself. Only 44% of prompts received the same sentiment classification across all five runs. Net sentiment moved by an average of 27 points between runs of the same prompt. Neutral prompts were the least stable, with only 35% producing the same classification every time, probably because mildly positive answers sit close to the line between positive and neutral.

The picture was not chaotic. In 96% of prompts, at least three of the five responses agreed on a sentiment. But the lesson for monitoring is clear: running a prompt once and treating the result as ChatGPT's position on your brand gives a single generation too much weight. Repeating prompts does not remove the variation, but it helps you see the pattern instead of reacting to one odd response.

Where negative AI sentiment comes from

It helps to understand why a brand gets a negative description in the first place. When an assistant says customers complain about onboarding or describes a steep learning curve, it is compressing the sources it has read. Those sources include review platforms, forums such as Reddit, news coverage, comparison articles, and the brand's own pages. The model does not dislike your company. The sentence has an origin somewhere.

That is encouraging, because it means negative sentiment can be traced and treated. It also explains why different engines describe the same brand differently: ChatGPT, Perplexity, Gemini, and Google's AI Overviews read overlapping but different sets of pages. One engine may speak warmly about you while another repeats a three-year-old complaint. That difference is a map of where the problem lives.

The stakes are real. Consumer research from 2026 found that 46 percent of AI users now start purchase research on a standalone AI assistant, up from 25 percent in 2024. For many buyers the AI answer is the first meeting with your brand, and it arrives with a verdict attached. Visibility and sentiment are separate measures. A brand can appear in every answer and still carry a caution label in each one.

A five-step process for fixing negative AI sentiment

Locate the exact negative statements

Start with an inventory, not a screenshot of one bad chat. A single conversation is one sample from one engine on one day. For each negative statement, record the engine, the prompt that produced it, the exact wording, and how often it repeats. A phrase that returns across many answers has a stable source behind it. A one-off may be sampling noise. The goal is a short, ranked list of quoted claims.

Trace each claim to its source

Take the top statement and find where it came from. Engines that show sources, such as Perplexity, Google's AI Overviews, and ChatGPT when it searches the web, make this easy. When an answer shows no sources, search the distinctive phrase directly. It usually turns up in a Reddit thread, a cluster of reviews on a platform like G2 or Trustpilot, a comparison list, a news article, or your own outdated page. An old pricing table or an abandoned docs page counts as a source, and engines keep reading it long after you stop. A handful of pages usually drives most of the negative wording, so tracing gets faster after the first claim.

Triage by claim type

Nearly every traced claim falls into one of three types. The first is the outdated fact, such as old pricing, an integration you have since shipped, or a bug fixed several releases ago. These are the fastest wins, because the source is simply stale. The second is the genuine complaint, where real customers say the same thing across enough reviews that engines treat it as consensus. This type is the slowest, and a page claiming easy onboarding will lose against fifty reviews saying otherwise. The third is the competitor-framed claim, where your description comes from comparison content written by or around rivals.

Comparing your sentiment against competitors helps here. If the whole category carries the same complaint, you face a category narrative, and the opportunity is to become the exception the answers name. If only you carry it, the fix is yours.

Fix the problem at the source

Each type needs its own treatment. For outdated facts, publish the current truth on your own domain in plain text, with a date, then ask the owners of stale pages to update them. Review platforms let vendors respond with corrections, and authors often fix posts when someone politely points out an error.

For genuine complaints, fix the underlying problem first. If onboarding is hard, no content strategy outruns honest reviews. Then make the fix visible: answer criticism publicly and factually, write about the improvement by name, and ask recent happy customers for reviews, since fresh reviews change what consensus looks like at the next read.

For competitor-framed claims, give the engines your side. Publish honest comparison pages for the matchups buyers ask about, say where you win and where you do not, and let the assistants balance the two sources. Honesty carries weight here, because an engine reading your page next to independent reviews notices when they disagree.

Prioritize by buyer impact

Prioritize by buyer impact. A mild caution inside a high-intent recommendation answer costs more revenue than a harsh line in an answer your buyers never see.

Re-measure and set expectations

Fixes land on each engine's own schedule. Answers built from live sources can change within weeks of the source changing. Claims baked into training data may last until the next model update, which takes months. The same fix can show results in Perplexity while ChatGPT, answering without web access, repeats the old line. That is normal.

Treat measurement as a rhythm. Keep the same tracked prompts running, watch the negative clusters from step one, and read the trend: fewer repeats, milder wording, better dated complaints. New negative clusters will appear as models update and new pages get written, and each one goes back through the same loop.

What does not work

Several popular tactics fail against how these engines actually work.

There is no channel for asking an AI company to correct how an assistant describes your brand in ordinary answers. The description follows the sources, so the sources are the only lever.

Burying a bad page with SEO does not help either. Assistants read far more than the first page of results, and a well-linked review thread stays in their reading whether or not humans still find it. Fake or incentivized reviews backfire because platforms police them and engines cross-reference enough sources that a sudden wave of unmarked praise looks like noise against the established view. Deleting criticism leaves quotes and screenshots behind and can become a story of its own. A factual public reply next to the complaint works better, since the reply becomes part of what engines read.

Finally, one good answer on one day proves nothing. Only the trend across a fixed set of prompts shows whether sentiment moved.

How to track AI brand sentiment properly

The study's practical advice is to track the ugly prompts too. Build a prompt set with positive, neutral, and negative variations so you can see the full range of what someone might read about your brand. Group prompts by intent, since recommendation, comparison, trust, and value questions react differently. Run each prompt several times, and compare results over time instead of trusting a single response.

Then use what you find. If ChatGPT keeps raising concerns about pricing, trust, product limits, or service problems, those concerns are content opportunities. Answer them directly in FAQs, product and service pages, comparison content, and other pages tied to the buying decision. People will keep asking skeptical questions about price, safety, and competitors. What a brand can influence is the information the model has to work with when they do.

One interpretive note matters for reporting. A negative answer to a negatively framed question is not proof that ChatGPT holds a negative view of your brand, and a positive answer to a friendly question proves nothing in the other direction. Look for patterns across framings, intents, competitors, and repeated runs.

Frequently asked questions

Does the wording of a prompt change how ChatGPT describes a brand?

Yes. In a test of 180 prompts and 900 responses, prompt tone and response sentiment had a Spearman correlation of 0.66. Negative wording lowered brand sentiment and positive wording raised it.

Which has more effect, negative or positive wording?

Negative wording. Against the neutral baseline, negative prompts lowered net sentiment by 56 points while positive prompts raised it by 26, so the negative effect was roughly twice as large. Part of the gap comes from a ceiling effect, but the asymmetry remained after accounting for it.

Are neutral prompts really neutral?

No. Neutral prompts averaged a net sentiment of about +20, so ChatGPT was mildly positive about the brands even without a push. Neutral prompts were also the least consistent across repeated runs.

Why did positive wording not help on trust questions?

Questions about reliability, refunds, privacy, safety, and warranties kept receiving cautious answers even with favorable wording. The information requested seems to limit how far framing can move the response, because the caveats are relevant to the question.

Can swapping the brands in a comparison question change the result?

Yes, and sometimes dramatically. In mirrored tests, the brand placed in the negative position lost sentiment in every pair. United MileagePlus moved from +91 to -93 and Sony from +96 to -9. Framing did not always win, though. Sony stayed positive on a question about upscaling, and Hers stayed negative even when the question targeted Start Willow.

Why does ChatGPT say negative things about my brand?

Assistants build answers from the sources they read, including reviews, forums, news, comparison articles, and your own pages. When those sources repeat a complaint or an outdated fact, the answers repeat it. Tracing the statement back to its source is the first step toward fixing it.

How long does it take to fix negative AI sentiment?

Answers built from live sources can change within weeks once the source changes. Claims that come from training data can last until the next model update, which takes months. Judge progress by the trend across many answers, not by one chat.

How should I monitor AI brand sentiment?

Track a fixed set of buyer prompts across the major assistants, include positive, neutral, and negative wording, run each prompt more than once, and organize prompts by intent. Collect negative statements word for word with the prompt and engine behind each one, and watch whether repeats fall over time.

References

Search Engine Land - https://searchengineland.com/negative-prompts-impact-chatgpt-brand-sentiment-study-500612

Ayzeo - https://ayzeo.com/blog/how-to-fix-negative-ai-brand-sentiment

SE visible - https://visible.seranking.com/blog/chatgpt-prompt-sentiment-study/

Tags: ChatGPT brand sentiment, AI brand sentiment, negative prompts, prompt sentiment, AI visibility tracking