Skip to content
Vouch

All posts

AI reputation management: fixing what AI says about your brand

By Vouch7 min read

Vouch blog cover with the title "AI Reputation Management for Your Brand" beside a line drawing of a shield with a tick inside it and a small red dot at its top right.

A buyer asks an AI assistant about your brand and hears last year's price, a discontinued product, or a flaw you fixed two releases ago. That is an AI reputation problem, and it behaves differently from a bad review: there is no comment thread to reply to, and the answer is assembled from whatever the assistant can read.

AI reputation management is the work of finding those wrong answers and correcting what feeds them. This guide covers where assistants get their facts, why wrong ones persist, and how to fix them at the source. Documented platform behaviour carries links and dates; the working advice is labelled as editorial recommendation.

Where assistants get their facts

Assistants answer from two pools: what they learned during training, and what they retrieve live from the web when they answer. The split varies by product and by question, and the vendors covered here do not publish exact splits, so treat any precise claim about it with caution.

What you can verify is the control surface. Google's AI features documentation (as of November 2026) says there are no additional requirements to appear in AI Overviews or AI Mode, that robots.txt for Googlebot manages how sites are crawled for Search, and that preview controls (nosnippet, data-nosnippet, max-snippet, noindex) limit what is shown from your pages. In other words, for Google's AI features, your indexed pages are the material the answers are built from.

Why wrong facts persist

Three mechanisms explain most stale answers, two documented and one an editorial inference:

  • Duplicate and near-duplicate pages. Bing's webmaster team explains in Does Duplicate Content Hurt SEO and AI Search Visibility? (December 19, 2025) that LLMs group near-duplicate URLs into a cluster and choose one page to represent the set, and that the chosen page can be an outdated version. Your archived 2024 pricing page can win the cluster and feed the answer.
  • Recrawl lag. Google's troubleshooting notes for preview controls state that recrawling can take anywhere from several days to several months. A corrected page does not correct the answer the same afternoon.
  • Third-party sources. Review sites, directories and forum threads also feed answers, and you do not control their update cycles. This point is editorial inference, not vendor documentation, but it follows from assistants reading the open web.

How to find what AI says about you

You cannot fix what you have not heard. A practical routine:

  1. Write down the ten to twenty questions a buyer would ask about your brand: pricing, comparisons, downsides, trust, alternatives.
  2. Ask them in each assistant your customers use, from a clean session, and record the exact wording of each answer with the date.
  3. Repeat weekly. Answers vary run to run, so treat a single answer as a sample, not a verdict.

Our post on tracking brand mentions in AI covers the monitoring routine in more detail. The goal here is a dated record of specific wrong claims, in the assistant's own words, because every fix in the next section starts from one.

How to fix wrong answers at the source

The recommendations below are editorial suggestions, grounded in the documented behaviour above.

Fix your own pages first

For each wrong claim, find the page on your site that states the outdated fact and correct it, or retire the page. One fact should have one canonical home: consolidate duplicates with canonical tags or redirects, per Bing's guidance, so the current version is the one a cluster picks.

Make the corrected fact easy to read

State the fact plainly in the page's HTML text, with a date ("as of November 2026"). A clear, dated sentence is both easier for a crawler to use and easier for you to keep honest.

Correct third-party profiles you control

Update listings, directories and review-site profiles you own. Where a third party states something factually wrong about a current product, their own correction or update process is the route; keep a record of what you asked and when.

Think twice before blocking

Blocking a crawler removes your corrected information along with everything else. Robots.txt and preview controls are blunt instruments: they limit visibility, they do not edit answers. The trade-offs of blocking specific AI crawlers are covered in Should you block GPTBot?.

A short worked example

Demo Cycles is an invented brand for illustration. Say an assistant tells buyers that the Demo Cycles City model costs 900, when the current price is 750. The dated record from the monitoring routine shows the wrong figure appearing in two assistants across three weeks. The search for the source finds three candidates: an archived spring promotion page still indexed with the 900 price, a directory listing the team forgot, and a review written when 900 was correct.

The fixes follow the order above. The promotion page is redirected to the current product page, so only one price page remains. The product page states the 750 price in plain text with "as of November 2026". The directory profile is updated through its own edit flow. The old review cannot be edited, and that is fine: it was accurate when written, and the goal is that current sources outnumber and outdate it. Weekly re-asks then show the new price appearing in answers over the following weeks. No part of this required access to the assistant itself.

How long fixes take, and how to verify

Set expectations with the documented lag: Google's own notes put recrawl processing at days to months, and the other documentation reviewed for this post publishes no faster guarantee. None of the documentation reviewed for this post offers a way to edit an answer directly, so the source is the reliable lever.

Verification is the monitoring routine from earlier, run against your dated record:

  1. Re-ask the same questions weekly in the same assistants.
  2. Compare wording against your record, not against memory.
  3. Count a fix as landed only when the corrected fact appears consistently across several runs, not once.

What this cannot fix

Two honest limits. First, facts absorbed during training can linger after the web has moved on, and the documentation reviewed here does not say how or when those fade. Second, reputation work on AI answers is gardening, not a one-off repair: pages drift, old versions resurface, and new wrong claims can appear as your products change. A steady routine beats a heroic cleanup.

How Vouch helps

Vouch gives you the dated record without the manual runs. Threads are six-question modelled buyer conversations, from discovery to decision, run on 4 of the 7 assistants (ChatGPT, Claude, Gemini and Perplexity) three times a week. They record how assistants describe your brand at each stage, including the trust questions where wrong facts tend to surface, and record where a rival is named instead. Single prompts run daily across all seven tracked engines, and the weekly Vouch Rate (0 to 100) combines stage survival with the final recommendation, so a persistent wrong claim shows up as a pattern, not an anecdote.

Vouch is invite-only through a waitlist, every invitation starts with a 14-day free trial, and plans start at $39 a month (as of November 2026). Join the Vouch waitlist, or check which AI crawlers can read your pages today with the free AI crawler checker.