Will Thumbs-Down Feedback Fix a Wrong Answer About My Brand?
Thumbs-down feedback alone is not a dependable fix for a wrong answer about your brand, and it comes with no timeline. Feedback feeds the signals AI companies use to improve their systems in aggregate. It does not edit a stored fact about your brand, because there is no stored answer to edit. Each response is generated fresh from the sources the assistant retrieves and what the model learned in training. Feedback is still worth sending, with a clear explanation and a link to a correct source, especially for serious errors. But the fix that actually changes answers is correcting the pages assistants rely on: your own site, profiles, directories and third-party articles. Asking staff to downvote an answer many times does not help. It can breach platform rules.
What does feedback actually do?
| Platform | Feedback option | What the company says |
|---|---|---|
| ChatGPT | Thumbs up or down on a response, with an optional comment | OpenAI says that if you provide feedback, the whole conversation associated with it may be used to train its models |
| Gemini | Thumbs down with details; a separate option to report a legal issue | Google says feedback helps Gemini improve, and feedback may be reviewed by people |
| Google AI Overviews | Feedback option on the Overview | Google says feedback helps it improve AI Overviews for everyone |
| Claude | Thumbs down on a response | Anthropic says it may use feedback for research and to train its models |
| Perplexity | Report option below the answer | Perplexity uses reports to improve answer quality |
OpenAI's own account of a 2025 update shows how feedback is used. In its review of a ChatGPT update that made the model overly flattering, OpenAI said the update had introduced an additional reward signal based on user feedback: thumbs-up and thumbs-down data. Feedback shapes model behaviour broadly, across millions of conversations. It is not a correction request for one brand.
Why doesn't feedback fix the answer?
- No stored answer. The next time someone asks, the assistant searches and writes again. If the sources still say the wrong thing, so will the answer.
- Aggregate use. Feedback is combined with huge volumes of other feedback to improve systems generally.
- Model memory lags. Answers from training data change only when newer models are trained.
- No individual reply. You usually will not hear back or see a direct change.
How long it takes to reverse negative sentiment more generally is covered in how long it takes to reverse negative brand sentiment in AI answers.
Is it still worth sending?
Yes, for three reasons. It records the error with the platform, which matters for serious or harmful claims. It adds to the signals that improve answers over time. And some platforms review serious reports by hand. Make each piece of feedback useful:
- Quote the wrong statement.
- Give the correct fact in one sentence.
- Link to a public source, such as your website, an official register or a licence lookup.
- Say who you are, such as "I work for [company]", where the platform asks.
Keep a record of what you sent and when.
What actually fixes wrong answers?
- Find the source. Open the citations in the wrong answer, and search for the wrong claim in quotation marks.
- Fix your own pages. State the correct facts plainly in text, near the top of the relevant page.
- Correct third-party sources: directories, company databases, review profiles and articles.
- Use formal routes for serious cases: legal reports for defamation, privacy requests for personal data about individuals.
- Retest over the following weeks with web search on and, where the assistant allows it, off, to see which answers have changed.
Platform-specific routes are covered in whether you can ask Google to correct a false Gemini answer and how to report an inaccurate Perplexity answer about your business.
Should staff downvote the answer repeatedly?
No. Repeated downvotes from colleagues add noise, not evidence, and they do not make an answer change for other users. Using multiple accounts or coordinated feedback to influence a system can also breach platform terms. One clear, well-evidenced report is worth more than a hundred thumbs-down clicks.
How do you know whether anything worked?
Track the answer, not the feedback:
- Run the same questions weekly for a month after fixing sources, then monthly.
- Record the answer and its sources each time, with search on and off.
- Note when the wrong claim stops appearing, and which source change preceded it.
If searched answers improve after a source fix and now cite the corrected page, that source was the likely cause. If only unsearched answers stay wrong, model memory is the issue; keep the sources consistent so newer models have the correct facts to learn from.
What might an illustrative case look like?
Consider an illustrative software company whose support team noticed ChatGPT telling users the product "has no API". Staff spent a week clicking thumbs down on the answer, and nothing changed. A check of the citations showed a two-year-old comparison article listing "API: no", and the company's own API documentation sat behind a login. The company published a public API overview page and asked the publisher to update the article, which it did. It also sent a single feedback note linking to the new page. Within three weeks, searched answers described the API correctly. The team stopped the downvoting and added the question to its monthly checks.
What do people ask about AI feedback buttons?
Does positive feedback on correct answers help?
It adds to aggregate signals, like negative feedback. It will not lock in a correct answer about your brand.
Is feedback private?
Platforms say feedback may be used to improve their systems and, in some cases, reviewed by people. Do not include confidential information.
Should customers be asked to leave feedback?
No. Put the effort into correcting sources, which changes answers for everyone.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources