Your Bad Reviews Are Now Training ChatGPT to Trash Talk Your Brand

When someone asks an AI tool about your brand, is the answer based on your positioning, or everyone else’s complaints?

Most brands still think of reputation management as a PR problem. Or a social media problem. Or a customer support problem.

That was already too narrow.

In AI search, reputation is also an SEO problem.

LLMs do not create your brand reputation from scratch. They compress what the web already says about you. That means reviews, Reddit threads, forums, social posts, comparison pages, affiliate content, customer complaints, support conversations, news coverage, and third-party product pages can all become part of how an AI system describes your company.

Not because the model “knows” your brand.

Because it has access to public signals around your brand, and it is designed to turn scattered information into a confident answer.

That does not mean LLMs are ruining your brand. That is the wrong framing. AI is not a shortcut, and it is not magic. It is a multiplier of existing quality or existing debt. That principle matters here because reputation debt does not stay isolated once AI systems start summarizing it at the moment of evaluation.

The uncomfortable part is that many of these inputs feel messy, informal, or unrepresentative.

A Reddit thread may not reflect your average customer. A review site may overrepresent unhappy users. An affiliate comparison may be shaped by incentives. A complaint from three years ago may still rank.

But all of it can still become part of the public evidence layer. And AI systems are very good at making messy public sentiment sound neat.

The dangerous assumption is this: If our owned content is accurate, AI systems will describe us accurately. That is not how this works.

Your website can say you are reliable, affordable, enterprise-ready, easy to use, and loved by customers. It can say those things clearly. It can use structured data. It can have polished product pages, strong case studies, and a well-maintained blog.

That helps. But it is only one input.

If review platforms repeatedly mention slow support, if Reddit threads describe painful onboarding, if comparison pages call out confusing pricing, and if social posts complain about cancellation, an AI response may reflect that broader pattern.

And in many cases, it should.

AI systems should not rely only on a brand’s own marketing copy. A buyer asking “Is this vendor trustworthy?” is not looking for a rewritten homepage. They are looking for evidence. The public web gives them that evidence, even when it is imperfect.

But the other side matters too.

Public sentiment can be distorted. A small group of angry customers can create an outsized footprint. Old problems can persist long after they are fixed. Competitor content can frame your weaknesses more aggressively than a neutral evaluator would. Affiliate pages can repeat stale claims because they convert. Forums can preserve outdated complaints because no one has the incentive to clean them up.

So the issue is not that owned content no longer matters. The issue is that owned content no longer controls the narrative by itself. Your brand narrative is now negotiated across the public web.

The obvious risk is that an AI response says something bad about your brand.

That is real. But it is not the full problem. The bigger risk is that the answer sounds plausible.

A buyer may ask: “Is Brand X good for enterprise teams?”

And the response may say: “Brand X is popular with smaller teams, but some users mention limitations around implementation, support responsiveness, and pricing transparency.”

That answer may be fair. It may be unfair. It may be outdated. It may be based on three review pages and one Reddit thread. The buyer may never know.

They may never click through to the original sources. That is what changes the risk profile.

Traditional search made the buyer do more synthesis. They saw the homepage, the review site, the Reddit result, the analyst page, the competitor comparison, and maybe a few “alternatives to” posts. They still had to interpret the pattern.

AI search does more of that synthesis for them.

OpenAI describes ChatGPT search as giving timely answers with links to relevant web sources, and Google describes AI Overviews as snapshots with links for exploring more on the web. In both cases, the experience is moving from “here are sources to evaluate” toward “here is an answer, with sources nearby.”

That does not remove the web. It compresses it. At small scale, a bad review is a customer support issue. At AI-search scale, repeated sentiment becomes a visibility and trust issue.

It can affect brand queries. It can affect comparison queries. It can affect “best solution for” queries. It can affect “who should not use” queries. It can affect sales enablement, pipeline quality, and executive confidence in organic visibility.

The answer may not block the visit. It may prevent the visit from ever happening.

Reputation debt becomes answer debt.

Traditional reputation management often asks: How do we suppress or respond to negative mentions? That question is too small for AI search.

The better question is: What evidence does the public web give an AI system to describe us accurately? That changes the work.

This is not about flooding the web with positive content. It is not about fake reviews. It is not about thin PR. It is not about trying to trick a model into saying nicer things.

That approach is risky on its own terms. The FTC’s rule on consumer reviews and testimonials prohibits fake or false reviews and testimonials, including buying or selling fake reviews, and authorizes civil penalties for knowing violations.

It is also bad SEO strategy. Google’s own guidance continues to emphasize helpful, reliable, people-first content rather than content created to manipulate search rankings. Its spam policies also call out manipulative behavior that can cause pages or sites to rank lower or be omitted from search results.

So the fix is not cosmetic. The fix is operational, editorial, and technical.

If customers are complaining because support is slow, the answer is not a blog post about your commitment to customers. It is better support.

If buyers are confused about pricing because your pricing page hides important constraints, the answer is not a thought leadership campaign. It is clearer pricing information.

If the market thinks you are not ready for enterprise because your public documentation does not explain implementation, governance, security, or support expectations, the answer is not to ask AI tools to “understand” you better. It is to publish better evidence.

Bad reputation signals should not be hidden when they reflect real customer problems. But outdated or unbalanced signals should not be allowed to define the brand indefinitely.

That is where in-house SEO teams have an important role. Not as reputation cleaners. As stewards of the public evidence layer.

The first place to look is review platforms. Review sites often contain repeated language around support, pricing, reliability, ease of use, implementation, cancellation, product quality, and value. A single bad review usually does not matter. A repeated pattern across multiple review sources does.

The SEO question is not “Do we have negative reviews?” Every brand does. The better question is: “Are the same complaints repeated across multiple platforms?”

If the same themes appear again and again, those themes are not just reputation issues. They are retrievable claims. They are language an AI system may find, compress, and repeat.

The second place to look is user-generated content. Reddit, forums, niche communities, Quora-style discussions, and public Slack or Discord archives can shape informal sentiment. These environments often contain the language buyers trust most because it does not sound like marketing.

That trust cuts both ways. A thoughtful customer answer can support your positioning better than your own website can. A detailed complaint can damage it more than a one-star review.

The third place is social media chatter. Social platforms can create fast-moving reputation narratives around launches, outages, pricing changes, layoffs, controversial policies, or support failures. Some of this is temporary noise. Some of it reinforces an existing pattern.

A one-day spike in frustration after an outage is different from years of posts saying your product is unreliable. One is an incident. The other is a market belief.

The fourth place is third-party editorial content. Reviews, comparisons, affiliate pages, analyst summaries, partner pages, and “alternatives” articles often define how a brand is positioned against competitors. These pages are especially important because they already use the same comparative language buyers use.

“Best for small teams.” “Not ideal for complex workflows.” “Cheaper alternative.” “More flexible but harder to implement.” “Strong feature set, mixed support.”

This is the language of evaluation, and AI answers often sound like evaluation.

The fifth place is your own content gaps. If your site does not answer important buyer questions, third parties will answer them for you. If you do not explain who your product is not for, someone else will. If you do not explain pricing tradeoffs, someone else will. If you do not explain migration, implementation, limitations, support, compliance, or cancellation, someone else will. Sometimes they will be accurate. Sometimes they will not. But silence creates space for other sources to define the answer.

Start by testing the questions a real buyer would ask. Not vanity prompts. Not “Tell me about [Brand].” Ask the uncomfortable questions.

Ask whether your brand is trustworthy. Ask what the biggest complaints are. Ask for pros and cons. Ask how you compare to a major competitor. Ask who should not use your product. Ask what users say about you on Reddit. Ask whether you are good for enterprise teams. Ask what the main risks are in choosing you. Run these prompts across the AI models your buyers are likely to use.

Do not overreact to one answer. LLM responses vary. They change by model, prompt, and available sources. Google tells users to check important information in more than one place and try multiple versions of a question when using AI Overviews. That same caution should apply to brands auditing their AI visibility.

Look for patterns. Track repeated negative claims. Track missing context. Track outdated information. Track competitor comparisons. Track unsupported generalizations. Track objections that appear again and again.

Then try to infer the likely source class. Support complaints may come from review sites, forums, and social posts. Pricing confusion may come from old plan pages, comparison content, and user threads. Feature-gap claims may come from competitor content or outdated reviews. Trust concerns may come from news coverage, customer complaints, or public social chatter.

You will not always know the exact source. That is frustrating. But you can often identify the type of evidence the model is reflecting. Then ask the hard question: “Is the AI response wrong, outdated, exaggerated, or simply reflecting a real issue?”

Those are different problems. A wrong answer needs correction and stronger authoritative evidence. An outdated answer needs fresher corroboration. An exaggerated answer needs balance and context. A true answer needs operational change.

SEO teams should be careful here. The goal is not to prove the model wrong. The goal is to understand what the answer reveals about the public web.

The first fix is to fix the real issue. This sounds obvious. It is often skipped.

If customers repeatedly complain about billing, onboarding, support, reliability, cancellation, or product quality, content will not solve the root problem. It may temporarily soften the narrative, but the underlying evidence will keep accumulating.

SEO should not be used to launder operational debt. That does not mean SEO has no role. It means SEO should bring the evidence back to the business. If recurring complaints are showing up in AI answers, those answers are a market research asset. Painful, but useful.

The second fix is to update owned content around known objections.

Most brand sites are strong on positioning and weak on buyer doubt. They explain the ideal use case. They under-explain the tradeoffs. That leaves AI systems and buyers to fill in the gaps from third-party sources. Create clear pages that address pricing, implementation, support expectations, security, compliance, product limitations, use-case fit, migration, cancellation, and who the product is not for.

That last one matters. A “who we are not for” page may feel uncomfortable. But it can create trust. It gives buyers and AI systems clearer language about fit. It prevents the market from defining every limitation as a failure.

The third fix is to strengthen third-party evidence.

Not by manufacturing positivity. By encouraging accurate, current, specific customer feedback. Recent reviews matter more than stale praise. Detailed case studies matter more than vague testimonials. Customer stories with measurable outcomes matter more than generic logos. Analyst pages, partner pages, implementation stories, community answers, and updated comparison content can all help create a more accurate public evidence layer.

The fourth fix is to respond where appropriate.

Not every negative mention deserves a response. Some threads should be left alone. Some posts are too old, too small, or too emotional to engage productively.

But recurring, visible, or inaccurate claims should be addressed with calm, specific correction.

Avoid defensiveness. The goal is not to win an argument. The goal is to improve the evidence layer.

A useful response explains what happened, what changed, where to find current information, and how customers can get help. That kind of response may help the original user. It may help future readers. It may also become part of what the public web says about your brand.

Do not try to manipulate LLMs with fake positivity. Fake reviews, spammed mentions, thin content, and low-quality PR may create short-term noise. They also create long-term trust risk. They can violate platform rules, search policies, and consumer protection expectations. They can also backfire if buyers discover that the evidence layer has been artificially inflated.

Do not assume one bad answer means a crisis. AI responses vary. One strange answer is not a strategy problem. A repeated pattern across tools, prompts, and source types is. Do not treat every negative mention as unfair.

Sometimes the model is surfacing a real market perception. That is not comfortable, but it is useful.

It tells you what buyers may already believe before they reach your site, your sales team, or your demo.

Do not leave this work only to PR. PR has a role. Social has a role. Support has a role. Product has a role. Customer success has a role. Legal may have a role.

But SEO should be in the room because this now affects discovery, trust, and conversion. Reputation is no longer separate from visibility. It is part of the answer.

LLMs do not just answer questions about your brand.

They summarize the reputation system around your brand.

That system includes your website. But it also includes reviews, social chatter, forums, competitors, customers, public complaints, third-party comparisons, and unresolved objections.

For SEO teams, the job is not to “optimize for AI answers” in isolation.

The job is to make sure the public web contains enough accurate, current, trustworthy evidence for AI systems to describe the brand fairly.

That is a harder job than publishing more content.

It requires operational honesty. It requires stronger owned pages. It requires better third-party evidence. It requires ongoing monitoring. It requires the company to care about what the market is actually saying, not just what the brand wants repeated.

Tokens: The chunks of text an AI model processes.

Learn More

One more thing: AI is only as good as it’s operator, and if you are reading this newsletter, you are better than most!

Till next time,

PS: Let me know what you think of this issue, or anything else here: [email protected]