A Correction Tracker for Wrong AI Answers About Your Brand
Log each wrong statement once, as a claim rather than a screenshot: what the assistant said, what's true, the evidence for what's true, which surfaces said it in how many of how many runs, and the source it cited, if any. Then trace the source and pick a route. A page you control gets fixed or redirected. A third-party page gets a correction request to its owner, with your evidence. An answer with no citation gets the product's feedback control and a re-test schedule. A claim that persists after all its sources are correct gets recorded and re-tested, not chased. Close a row only when the claim stops appearing across repeated runs, never after one clean answer.
Someone on the sales team pastes an answer: an assistant says your product starts at a price you retired last year and lacks a feature you shipped in the spring. It's tempting to fire off a fix in every direction. A tracker makes it one row, one route and a date to check again.
The tracker
| Field | Example (illustrative) | Why |
|---|---|---|
| Claim as said | "Lumina Desk starts at $49 a month" | The exact words you're correcting |
| What's true, with evidence | Starts at $19 a seat; pricing page, checked 2026-09-29 | A correction request needs proof |
| Where and how often | Surfaces S-01 and S-03; 4 of 9 runs | Separates a pattern from a one-off |
| Cited source | A 2025 review article, or none | Decides the route |
| Source owner | Us, a third party, or nobody | Who can change it |
| Route | Fix own page, ask owner, feedback, or record | The action taken |
| Status | Open, actioned, re-testing, closed | Every row has to end somewhere |
| Re-test dates | 2026-10-06, 2026-10-20 | Closing needs repeated clean runs |
Lumina Desk is a fictional product; the row is there for shape.
The four routes
Your own page. Old pricing pages, archived posts and stale changelog entries are the usual culprits. Update the page, or redirect it if it shouldn't exist, and state the correct fact with a visible date. This is the route you can act on fastest, because it's entirely in your hands.
A third party's page. Send the owner the specific sentence, the correct fact and your evidence, and record the date you asked. If the site publishes a corrections policy or contact, use that channel. Don't ask them to change opinions, and don't send the same request twice in a week.
No citation. Use the product's own feedback control. For AI Overviews, Google's help page describes a thumbs-down with a 'Report a problem' option, and says the feedback includes your most recent search query and its results. Other assistants have their own controls. Treat all of these as feedback, not a correction request with a guaranteed outcome, and keep re-testing.
Everything is already correct. Sometimes the sources are fixed and the claim persists. Record it, keep the re-test dates, and resist publishing a pile of pages that repeat the correct fact; one clear, dated statement on a page you own is the input you control.
When to close a row
Close it when the claim is absent across at least as many runs as it appeared in when you opened the row, on two consecutive re-tests. Opened at 4 of 9, close after two re-tests at 0 of 9 on the same surfaces. A single clean answer proves only that one run went your way.
What doesn't belong here
Opinions. 'Expensive for small teams' or 'hard to set up' aren't factual errors, even when they sting. Put them in a sentiment log. A tracker full of rows that can never close because they aren't wrong stops getting read, and then the real errors sit in it unnoticed.
FAQ
Should we contact the company behind the assistant? — Use the feedback control the product offers, with your evidence ready. Fixing the sources you can reach is usually a better use of the time than escalating further.
What if the wrong claim is about a competitor? — Out of scope for this tracker. Our competitor-audit guide covers what's worth acting on there, and it's less than most people expect.
How long before a correction shows up? — We don't know of a published timeline from any assistant vendor, and it likely varies by source and surface. Your re-test dates will build your own answer over a few months.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.