How to fix wrong information about your brand in AI answers
You cannot edit an AI assistant's answer directly, but you can fix the pages it is built from. Identify the exact wrong claim, check whether it is incorrect or just outdated, find the pages that carry it (often including your own), correct your pages first, then third-party sources, and re-measure over the following weeks.
Step 1: Pin down the exact claim
"ChatGPT gets us wrong" is not actionable. Write the claim down in a structured way: subject (your product), property (price per month), value stated (€35), correct value (€29, valid since March). Note which engines make it, in which languages, and how often: one answer in twenty is a different problem from every answer.
Step 2: Classify it
- Incorrect: it was never true. Often a confusion with a namesake or a competitor, or a detail invented to fill a gap.
- Outdated: it used to be true. An old price, a discontinued product, a former CEO. This is the most common case, and the easiest to trace.
- Unverifiable: you have no public, citable source that says otherwise. The fix is to publish one.
Keeping a fact sheet (your verified facts, each with a date from which it is valid) makes this step fast and consistent across the team.
Step 3: Find the pages that carry it
Start with the citations in the answers that make the claim: the cited page often contains the wrong value. Then search the web for the question the way an engine would, and read the top results; search-grounded assistants draw from that pool. Look in particular at:
- your own pages: old pricing pages, archived press releases, PDF brochures, help-centre articles, localized pages that were never updated;
- comparison articles and reviews that quote your specifications;
- encyclopedia and directory entries, including business listings and app store pages;
- discussion threads where the claim is repeated by users.
A claim repeated across many pages is harder to shift than one carried by a single outdated page, so count how widespread it is.
Step 4: Fix your own pages first
Your own pages are the only sources you fully control, and an outdated owned page that an engine keeps retrieving is one of the most frequent causes of wrong answers.
- Update or redirect outdated pages; do not leave old versions reachable and indexable.
- State the correct fact in plain text, in a sentence an engine can lift: "Plan X costs €29 per month, VAT excluded, since March 2026."
- Add dates to facts that change, and a "last updated" date to the page.
- Publish a page that answers the question directly if none exists, for example an FAQ entry for "Is product X discontinued?".
- Use structured data (Product, Offer, Organization, FAQPage) so the facts are machine-readable.
- Keep every language version consistent; a stale translation can keep an old value alive in one language.
Step 5: Correct third-party sources
- Publishers and review sites: ask for a correction, politely, with a link to your source. Many update product specifications on request.
- Directories and listings: claim and update your business profiles.
- Encyclopedias: follow their rules. On Wikipedia, propose changes on the article's talk page, disclose your conflict of interest, and cite independent sources.
- Discussions: reply transparently as the company, with the correct information and a source. Do not use fake accounts.
Step 6: Use the engines' own feedback tools, with realistic expectations
Most assistants let users flag a bad answer. It is worth doing for serious errors, but it is not a reliable correction channel: feedback is not guaranteed to change future answers. Some providers also have processes for legal or safety issues, such as defamation or personal data; use them for those cases.
How long does a correction take to show up?
It depends on where the wrong claim comes from:
- Answers built from live web search can change within days of the source pages changing and being re-crawled.
- Search-feature answers generally follow the search engine's own re-indexing pace, often a matter of weeks.
- Claims coming from the model's training data can persist until the provider releases an updated model, which can take months. Strong, current, citable sources help the assistant override them when it searches.
Step 7: Re-measure and watch for recurrence
Re-run the same prompts, several times each, on each engine, and compare the share of answers that still contain the wrong claim. Keep watching: a corrected claim can return if an old page resurfaces or a new article copies an outdated source.
What should you avoid?
- Planting reviews, posts or articles under false identities. It violates platform rules and, if exposed, becomes the new story.
- Publishing large volumes of thin pages aimed at AI. They rarely help and can damage your search visibility.
- Fixing only the answer you saw. The same claim usually appears in other engines and languages.
How Tonecast helps
Tonecast's claim check compares every claim in every sampled answer from ChatGPT, Claude, Gemini and Perplexity with your fact sheet, classifies it as consistent, incorrect, outdated or unverifiable, and links it to the pages that carry it, including your own. The gap & action list collects wrong claims and information gaps to work through, and claim prevalence shows whether a claim is spreading in the sources before it reaches the answers. Tonecast does not write the correction for you; it shows where the problem is.