Responding to Negative AI Answers About Your Brand

A negative AI answer falls into one of three types. Each needs a different response, so decide which one you have first.

Tone alone is not a fact. An answer can sound cool or hedged while every claim in it is correct. Read how sentiment in AI answers is defined and measured before you treat a flat tone as a problem. Can't tell whether an error was invented or just old? This comparison of hallucinated and outdated answers walks through the diagnosis.

An AI answer comes from one of two places. The model answers from what it learned in training, or it searches the web and uses pages it retrieved. Each engine decides per question whether to search. ChatGPT search, for example, runs at the model's discretion.

Answers with citations

Open every citation. You'll usually find one of four things:

• A third-party page that states the negative claim.
• Your own page, stating the fact badly or out of date.
• A page that doesn't support the claim at all.
• A page that supports it accurately, which means the criticism is real.

Cited does not always mean used. OpenAI's API returns the URLs a model consulted separately from the ones it cited, and the consulted list is often longer.

Answers with no citations

These usually come from training data, which is frozen at a cutoff. OpenAI's model documentation lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for ChatGPT's Instant path. Nothing you publish today reaches those answers until a newer model ships.

Can the engines reach you at all?

OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. A firewall or CDN can also block a crawler your robots.txt allows. If your site is unreachable, the engine can only work from other people's pages. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

Work through these in order. Skipping the tracing steps is how teams end up rewriting pages the engine never read.

1. Write down the exact buyer questions that produced the negative answer. Use the phrasing a prospect would use.
2. Run each question on each engine you care about, logged out and with memory off. Saved memories and custom instructions change answers.
3. Save every answer with the date, the engine, and whether it showed citations.
4. Classify each negative claim as wrong, stale or fair. One answer can contain all three.
5. Trace each claim to its cited page and copy the exact sentence the engine relied on.
6. Check access: robots.txt rules for each search crawler, firewall responses, and whether your key facts are in the HTML before JavaScript runs.
7. Fix wrong and stale facts at the source (next section).
8. Address fair criticism at its root (the section after).
9. Re-run the same questions on the same engines and compare claim by claim.

For a fuller method on step 2, see how to find out what AI says about your company. The free checker on this site gives you a quick snapshot for up to three buyer questions if you want a starting point.

The fix depends on where the wrong fact lives.

• Your own page is cited and wrong. Update it. Put the correct fact in plain text in the HTML. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so facts that only appear after scripts run may be invisible to them.
• The fact is missing from your site. State it plainly on a page that can be fetched. A specific sentence (a price, a date, a one-line description) gives the engine something it can quote. That is not a guarantee it will.
• A third-party page is wrong. Ask the publisher to correct it and send evidence. Update directory listings and profiles you control. When engines cite other sites instead of yours, those sites are the route into the answer.
• Your own pages disagree with each other. Make the pricing page, about page and footer say the same thing. See what AI does when sources disagree.
• The answer comes from training data. Fix the live sources anyway, then wait. The correction reaches retrieval answers first and trained answers only after a model update.

Arguing with the bot does not help. Correcting ChatGPT in a chat changes that conversation, not the answer other users get. For the platform routes that do exist, read fixing ChatGPT errors in trained knowledge vs search.

If the cited sources are genuine complaints, the engine is doing its job. No page you publish will reliably talk it out of summarizing them. A July 2026 critical survey of 45 GEO studies found no technique with a stable, cross-platform causal effect on discoverability. There is no content shortcut here.

What you can do:

• Group the complaints into themes: support response times, billing surprises, a missing feature.
• Fix the underlying issue. This is the only change that removes the cause.
• Respond publicly, under your real company name, where the complaints live.
• Publish a dated, specific note of what changed, so later pages and answers can reflect it.
• Ask real customers for reviews through your normal channels, with no incentive tied to a rating.

Old criticism can keep appearing after you've fixed the problem, especially in training-data answers. Read whether you can control brand sentiment in AI answers for an honest view of what's in your hands.

Under pressure, teams reach for shortcuts. Each of these is illegitimate and adds risk rather than removing it.

• Fake reviews or sock-puppet posts. If they're exposed, the exposure becomes a new negative source that engines can cite.
• Hidden text or cloaking for AI crawlers. Google's spam policies cover both. Showing crawlers different content than people see puts your search presence at risk.
• Prompt injection. Some pages embed instructions aimed at AI agents. A June 2026 study of LLM search agents found they can be swayed by manipulated content once it is in their results, and vulnerability varied by model and setup. This is an attack, not a tactic.
• Mass-publishing rebuttal content. In Otterly's 2026 experiment, Google deindexed two new sites carrying roughly 1,000 AI-generated posts each, with no manual action shown in Search Console. Google's scaled content abuse policy describes that pattern.

The same research is a reason to watch defensively. Third-party pages can carry injected or misleading content about you. Note them when you trace citations.

Checklist

• Buyer questions written down, in buyer language
• Answers saved per engine, with date and citation status
• Each negative claim labelled wrong, stale or fair
• Every citation opened and the source sentence copied
• robots.txt checked for OAI-SearchBot, PerplexityBot and Claude-SearchBot
• Firewall and CDN checked for blocked AI crawlers
• Key facts visible in the HTML before JavaScript runs
• Your own pages consistent on prices, dates and product names
• Correction requests sent to third-party publishers
• Real fixes shipped for fair criticism, with a dated public note
• No fake reviews, hidden text or injected instructions anywhere

Good results

• The wrong fact is gone across several repeat runs, not just one.
• The cited pages now state the correct fact.
• Criticism, where it appears, references your fix or is framed as past.

Bad results (or misleading ones)

• One clean run that you treat as a win. Answers vary run to run.
• A blended score that improved while one engine still states the error.
• A training-data answer still repeating the old fact. That's expected until a new model ships, not a sign your fix failed.

Treat this as a routine, not a one-off.

• Monthly: re-run the same question set on each engine and compare claim by claim. Keep the questions fixed so changes mean something.
• After each fix: re-check once the corrected page is live and has had time to be recrawled. Judge retrieval answers first.
• After a major model launch: re-run everything. A new model can bring a newer training cutoff, which can retire stale facts or surface new ones.
• Quarterly: re-sort your list. Move resolved items out, and check whether fair criticism has shifted to a new theme.

Score each engine on its own. One engine can repeat a claim that the others have dropped, and an average hides that.