What to do when an AI model gets a fact about your brand wrong

A wrong AI answer about your brand is easy to notice and, per every platform's own documentation, surprisingly hard to actually fix through the tool that shows it to you. This piece covers what each platform's feedback mechanism is documented to do, why none of them are the real fix, and what is.

What does each platform's feedback mechanism actually do?

ChatGPT. OpenAI's own developer community and product documentation are consistent on this: a thumbs-down doesn't change how ChatGPT responds in that conversation or in future conversations for that account. The signal is collected as training and evaluation data. It can shape a later model update, in aggregate, across every user, once enough signal accumulates. It is not a correction mechanism for the specific fact you just flagged.

Google AI Overviews. Google provides a "Report a problem" flow for an inaccurate or problematic Overview, where a user selects a category and adds detail. The stated mechanism is user feedback plus periodic human review, not an automated accuracy check, and not a ticket system with a case number or a guaranteed response. Google's own framing treats this as one input into ongoing quality improvement, not a per-fact correction request.

Perplexity. Users can flag an answer or file a report through Perplexity's support channel. Perplexity's own help documentation is direct about what this does: feedback feeds evaluation and safety pipelines rather than editing answers in place. The same documentation states the actual fix plainly: correct the cited source so the next retrieval returns the right fact.

All three converge on the same shape: feedback is real, and it does something, but "something" is aggregate model improvement over time, not a guaranteed or immediate correction of the specific claim about your brand.

Why isn't the feedback button the fix?

This isn't a gap in any one product. It follows from how these systems actually work. A thumbs-down or a report doesn't reach into a specific fact and rewrite it; it becomes one signal among many in a reinforcement-learning or human-review process that shapes future model behavior broadly. Expecting it to correct today's answer is expecting the wrong kind of fix from the wrong layer of the system, the same category of mistake as expecting a robots.txt change to fix a CDN-layer block, or expecting a one-time content edit to move a citation rate the underlying research says doesn't respond to it.

What's the real fix?

Perplexity's own guidance names the actual mechanism directly: correct the cited source so the next retrieval returns the right fact. That's the version of the fix available whenever the wrong answer came with a citation: a live retrieval pulled the wrong fact from a specific page, and correcting that page changes what the next retrieval finds.

Most brand-fact errors trace to one of a small number of sources: your own site (if it's stale or ambiguous), a third-party review or data platform (G2, Capterra, Crunchbase, a business directory), or Wikidata. All three are places you can actually edit, unlike the model's own weights. Entity consistency and Wikidata for brands already cover how to keep these aligned; this is the same work, framed as incident response instead of routine hygiene.

Is this retrieval-sourced or training-data-sourced?

The fix above only works when the error is retrieval-sourced, meaning a citation or source link is attached to the answer. Check for one. Perplexity attaches citations by default; ChatGPT and Claude only when their web-search capability is actively invoked; Google AI Overviews shows a source panel whenever an Overview appears. If a named source is there, that's the page to fix.

If there's no citation at all, the fact is more likely coming from training data the model memorized, with nothing fetched and nothing to link to. Mention vs. citation already covers why this category is harder: a mention sourced from training data has no retrieval event to trigger a refresh, and no source attached that a fact-checker (human or automated) could flag. Fixing every third-party profile today doesn't retroactively correct what a model already memorized from an earlier training snapshot; it only prevents the same stale fact from being pulled again on a future retrieval. That's a real, bounded limitation worth being honest about rather than promising a fix this category of error doesn't actually support yet.

What's a practical escalation order?

1. Your own site first. Fastest, fully in your control, and the most obvious place a model would retrieve from if it's citing anything at all.
2. Wikidata, with per-statement sourcing. Not just changing the fact: attaching a stated in reference (property P248) to the specific statement. Missing per-statement sourcing is the single most commonly cited reason a Wikidata edit itself gets reverted, which would leave the error both uncorrected and now carrying a visible revert in its history.
3. G2, Capterra, and other review or data platforms your entity-consistency footprint already touches: category accuracy and field completeness, the same hygiene G2/Capterra profile hygiene covers for other reasons.
4. The platform's own feedback report, last. Worth filing: it's real signal, and Perplexity's own guidance notes a specific, repeatable report with the exact wrong claim and the corrected source is more likely to get reviewed than a vague one, but as a secondary layer on top of the actual fix, not a substitute for it.

Frequently asked questions

Sources referenced

1. OpenAI Developer Community: What's the purpose of thumbs up/down feedback?
2. Perplexity Help Center: How can I report incorrect or inaccurate answers?
3. Google Search Central Community: Correcting false information in AI Overview

This piece is the incident-response version of the routine hygiene covered in what is entity consistency and Wikidata for brands. For why a citation and a mention need to be diagnosed differently in the first place, see mention vs. citation in AI search.