Citino teamYou cannot edit an AI answer. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews repeat what they learned in training or what they read on the web, so a wrong fact about your brand has a source. Find that source, correct it, and the answer follows.
Why AI assistants get facts about your brand wrong
A wrong answer usually has one of four causes:
- An outdated page. An engine that searches the web reads whatever ranks for the question. If an old pricing page, a cached copy or a stale review-site listing still ranks, the engine repeats it. Perplexity, for example, searches the web in real time and cites the pages it used, so its answer is only as current as those pages.
- Training data. When an engine answers from memory rather than search, it uses what was in its training data. Every model has a training data cutoff; Anthropic, for example, lists one for each Claude model. A price you changed after that date is not in the model.
- Conflicting sources. If your site says one thing and three directories say another, the engine may side with the majority.
- Invention. Language models sometimes produce fluent statements that are not supported by any source. Researchers call this hallucination (Huang et al. survey the causes). Google is direct about it: its help page says AI Overviews “can and will make mistakes”.
The fix differs by cause, so diagnose before you correct.
Step 1: Write down the facts that matter
List the facts a buyer relies on when choosing you. Each one is a claim: a short statement you can check an answer against.
| Kind of fact | Example claim (illustrative) |
|---|---|
| Price | “Tallyfold’s entry plan has no setup fee and is billed monthly” |
| Plan or packaging | “Every plan includes multi-client billing” |
| Integration | “Tallyfold integrates with NetSuite and Shopify” |
| Limit | “No limit on the number of client accounts” |
| Company fact | “Tallyfold is based in Austin, Texas” |
Tallyfold is fictional. Write each claim in plain words, the way you would want an answer to say it, and note the date it became true. A fact that changed recently is the most likely to be wrong.
Step 2: Ask every engine, more than once
Ask each engine the questions where the fact matters, such as “how much does Tallyfold cost?” or “does Tallyfold work with NetSuite?”. Then compare the answer with your claim. Checking claims is one part of AI search analytics. There are four outcomes:
| Outcome | What it means |
|---|---|
| Correct | The engine states the claim as you wrote it |
| Differs | Close but not the same: an old plan name, the wrong billing period |
| Wrong | The engine contradicts your claim |
| Not stated | No answer mentions the fact |
Answers vary from run to run, so one check is not enough. Ask several times, on several days. A fact stated wrongly once may be noise; a fact stated wrongly every time has a source.
“Not stated” is a different problem: the engines do not know the fact at all. That is a content gap, and how to fix content gaps covers it.
Step 3: Find the page the wrong fact came from
Engines that search cite their sources. Open the cited pages and look for the wrong fact.
| What you find | Likely cause | Where to fix it |
|---|---|---|
| Your own old page still live | Outdated page | Update it, or redirect it to the current page |
| A review site or directory listing | Outdated third-party page | Claim and edit the listing |
| An article or comparison page | Outdated third-party page | Ask the publisher to correct it |
| No cited page states it | Training data or invention | Strengthen the pages that state the right fact |
If the wrong fact appears with no citation, the engine probably answered from training data. You cannot change that model, but you can make sure every page it might read next time says the right thing.
Step 4: Fix the source, not the answer
- Correct your own pages first. State the fact in plain text on the page a buyer would expect to find it, such as the pricing page for a price. Text inside an image or a PDF is harder for engines to read.
- Retire old pages properly. An old pricing page that still returns a page keeps getting read. Redirect it to the current page or update it.
- Use structured data where it fits. Google supports Product markup with price and availability and Organization markup for company details. Markup must match what the page says.
- Fix every copy. Update your profiles on review sites, directories and marketplaces. If sources disagree, the engine has to choose; make them agree. How to find and fix source gaps covers getting listed on the pages engines cite.
Arguing with the assistant in a chat does not fix the answer for anyone else. The pages it reads do.
Step 5: Get the corrected pages read again
A correction only helps once engines read the new version.
- Google: request a recrawl with the URL Inspection tool for a few pages, or resubmit your sitemap for many. Google says crawling can take anywhere from a few days to a few weeks.
- Bing and others: IndexNow lets you notify participating search engines that a URL has changed.
- Third-party pages: you depend on the site making the change. Note the date it went live.
You can also send feedback in the product. Google’s help page asks people to send feedback when an AI Overview does not look right. Feedback does not change the source pages, so do it as well as fixing them, not instead.
Step 6: Check each engine changed its answer
Ask the same questions again once the corrected pages have been recrawled. A claim is a fact, not a sampled share, so the result is per engine: correct now, still wrong, or changed to something else. Keep checking, because an old listing or a model update can bring the wrong version back.
If you also want to know whether the correction changed how often you are recommended, measure that separately. How to tell whether an AI search fix worked explains how.
Where Citino helps
In Citino you write each claim once, with the date it became true. Citino checks every engine’s answers against your wording and marks each one correct, differs, wrong or not stated, with the page the engine read it from. Marking an answer wrong opens a correction in Gaps with the evidence attached, and you are alerted when an engine changes what it says. See Claims.