AI Answers Show Wrong Information About Our Brand: How to Fix It

The short answer. AI answers show wrong information about a brand when the pages the engines read are outdated or contradict each other, when facts are missing or hard to find on those pages, or when an engine is repeating what it remembers from training rather than looking anything up. Fixing it means correcting the source, then checking each engine again to see whether the corrected fact now appears.

This page is a workflow. It does not promise that an engine will adopt a correction by a given date, and it separates what moves in days from what moves slowly.

Where do AI assistants get wrong details about a brand from?

An answer engine builds an answer from three things, and a wrong fact can come from any of them (see Why):

• Pages it reads. Engines search the web and cite what they find: your pricing, product and About pages, and third-party pages such as lists, reviews, directories and profiles. If those pages are stale, or say different things, the engine has stale or conflicting facts to quote. Two founding years, two prices or two headcounts across pages lower a model’s confidence in every statement about that entity (entities and claims). A 2024 survey of knowledge conflicts in language models reports that when retrieved texts contradict each other, models are sensitive to the order of the evidence, favour evidence that matches their training-time knowledge or is corroborated by more documents, and are susceptible to misinformation (the evidence, grade C).
• The searches it runs. Engines turn a question into their own searches, often “best …” lists, and answer from what ranks. A wrong claim on a page that ranks for those searches travels.
• What it remembers. Asked without search, an engine answers from training alone, and it can confuse your brand with another company. That is not hypothetical: on 2 October 2026, asked without search, ChatGPT took Vercatus for a pension fintech and Gemini took it for an Amazon PPC agency (our own measurement).

The distinction matters because the fix and the speed differ. A gap in what an engine remembers moves slowly and closes through presence across the web over time; a gap in what it finds by searching can close in days (Why: memory).

How to see what the engines currently say

Counting is not the same as checking accuracy, so use two tools together.

• The poll and Answer Share show how often the engines name or cite you on your topics, engine by engine, with a range (the poll, how Answer Share is measured). A drop on one engine is a cue to look at what that engine is saying.
• Engine Briefs ask each engine directly about you once a week, with search on and, where the engine allows it, with search off: what it knows about the brand, whether it would recommend it, and the weaknesses it gives. Each card lists the facts the engine states and what changed since the previous brief, and Show the probes opens the full answers behind it (Briefs).

A wrong claim that appears in the facts of the same engine’s brief week after week is a recurring one. A claim that appears once may be noise. Read the briefs with the label they carry: this is what the engine says, which is perception, not evidence of what moves its answers. The pages it read and the searches it ran, in Why, are the evidence for where a claim came from.

The correction process

1. Find the claim and where it came from. Take the wrong statement from a brief or from the probes behind it. Then look at the pages the engine read for that topic and the searches it ran; the source is usually among them. Vercatus turns a fact an engine gets wrong, one that contradicts your site such as an old price or a wrong founder, into a Correct a fact play (plays).
2. Correct the authoritative page, and every page that repeats it. State the fact plainly in visible text. Check your site, your documentation, press and About pages, and the profiles you control, so they agree. If the page shows a date, make the visible date and the structured dateModified match; do not change a date to look fresh when nothing changed (the evidence, grade A).
3. Let the engines see the change. Update the sitemap’s last-modified date. For engines that support it, notify them: IndexNow shares a submission with all participating engines, including Bing; Google is not among them, so use Search Console for Google (IndexNow, grade A). Vercatus’ own check reads the page again when a play’s check runs.
4. Check each engine again. A correction is confirmed when a newer brief states your version. A brief that says nothing does not count: the check is strict on purpose. Re-read the brief for each engine that was wrong, across the ten engines Vercatus reads. Expect a search-driven answer to change before a memory-driven one.
5. Know what proves what. A correction play shows no expected gain, because fixing a fact is hygiene rather than a bet on share. The brief is the proof the fact changed. A shipped play also gets a verdict on Answer Share, read on days 14, 30 and 45 against topics left alone (verdicts); that answers a different question, whether your share moved, and it is not evidence that a particular claim was fixed.

To draft the corrected passage, each play comes with a guide and a draft that you edit and publish yourself (working a play). Vercatus does not publish for you.

Specific cases

Outdated pricing quoted by AI assistants

Pricing is the claim most often out of date, because it changes and gets copied. Put each plan’s price, currency, billing period and inclusions in visible text, ideally a real HTML table with the plan name in every row, mirrored in Offer markup. In Vercatus’ own retrieval simulations, pricing tables that flatten into passages without plan attribution answered pricing questions only partially (pricing page citability, grade C). Keep one pricing page as the source of truth, redirect old pricing URLs to it with a permanent redirect, and update the third-party listings that still carry the old figures. A recurring subscription price has no single billing-period field in schema.org, so express the unit explicitly (subscription price markup, grade A). Limited-time prices should carry a validity window (sale price validity, grade A).

A discontinued product still listed

Say on the product page that it is discontinued, with the date and what replaces it, and make availability data match the text (offer state as data, grade A). If you remove the page, redirect it with a permanent redirect to the closest replacement, and keep the redirect for at least a year; do not send many old URLs to one unrelated page such as the home page, which may be treated as a soft 404 (site moves and redirects, grade A). If nothing replaces it, return a real 404 rather than a page that says “not found” with a success code (soft 404, grade A). Then update the third-party pages that still list it.

Wrong funding figures in answers

Funding figures usually come from third-party pages, news items and company databases rather than your own site, so the fix starts outside it. Publish the correct figure and its date once, plainly, on a page you control, such as About or a press page, and publish only a figure you are ready to stand behind. Correct the profiles you can edit. Google’s knowledge panels update as the information they draw on changes, and an official representative can claim a panel and suggest changes; Wikidata accepts an item for any clearly identifiable entity described by serious public references, so a brand can maintain its own record (consistency across the web, grade A). For pages you do not control, the Sources view shows which third-party pages the engines read for your topics, so you can ask those editors to correct them.

Related

• Entities and claims: how Vercatus finds the contradictions that make engines doubt a fact.
• The knowledge base: about 180 entries, each graded A to D with its sources and a date to re-check.
• How plays work and verdicts: what a shipped change is, and how it is judged.