How to Fix Incorrect Brand Facts in LLM Citations: Report, Correct, Recheck

This post is the quick path for one wrong fact, with a correction log and message templates you can copy. For the full recovery blueprint (tracing the seed source, building a canonical facts page with JSON-LD, and speeding up recrawls), read our complete guide to fixing incorrect brand facts in AI answers.

A citation is good news, even when the sentence next to it is wrong. It tells you which page the engine read. That matters, because LLM citations are often shaky. When the Tow Center asked eight AI search tools to identify the source of 1,600 news excerpts, they gave incorrect answers to more than 60% of them. In a 2025 study by the EBU and BBC, sourcing problems caused 31% of significant issues, including answers that pinned a wrong claim on the outlet they cited.

Before You Report: Capture the Answer and Its LLM Citations

Answers change on every run, so save the wrong one before you do anything else.

1. Note the engine and mode. Write down whether web search ran. In ChatGPT, answers that used search can show citations and a Sources button; an answer with neither probably came from memory.
2. Copy the exact prompt and the date and time you ran it.
3. Save a share link and a screenshot of the full answer, sources included.
4. Quote the wrong sentence word for word.
5. List every cited URL next to the sentence it supports.
6. Write the correct fact and the URL that proves it, ideally a dated page on your own site.

Does the cited page say it?

Open each of the LLM citations and search the page for the wrong number or phrase. Each of the three possible results sends you down a different path.

OpenAI's own help page says ChatGPT's search results and citations "can be incomplete, outdated, or incorrect", and asks users to check that a source supports the answer. The EBU and BBC study counted claims the cited source didn't support among its sourcing problems, the largest category of error.

Report It Through Each Engine's Feedback Channel

Every major engine lets you flag an answer, and your note can point at the LLM citations under it. Write the same three sentences each time: the wrong claim, the correct fact, and the link that proves it. Here's what each vendor says the channel does, as of September 2026.

ChatGPT

Click the thumbs-down icon under the answer and describe the problem. OpenAI says the whole conversation behind a thumbs-up or thumbs-down may be used to train its models, even if you've opted out of training. That's the whole promise: no reply, and no fix for one answer.

Formal reports are a separate path for content that may break OpenAI's policies or the law, through the thumbs-down menu's "Safety or Legal concern" option or a web form. OpenAI says reported domains may be reviewed by its Model Quality team, which "may apply filters" so ChatGPT relies less on unreliable sources. Save it for fraud, such as a scam number shown as your support line, not a stale price. For product cards in ChatGPT shopping results, OpenAI offers a menu option to report products with incorrect information, and notes that price changes may show up with a delay.

Perplexity

Perplexity asks you to use the flag icon below the answer or open a support ticket. It wants the link to the thread, a description of the error and the answer you expected. Misinformation and outdated information are both on its list of issues it looks for. The article doesn't say what happens after you report.

Google AI Overviews and Gemini

Under an AI Overview, click the thumbs-down icon, then Report a problem, and pick a category. Google says the report includes your query and its search results and helps it improve AI Overviews. In the Gemini app, click Bad response under the answer, choose a reason and add details. Google says that feedback helps make Gemini "more helpful, accurate, and safe". Neither page promises a correction.

If the wrong fact sits in a Google knowledge panel, the path is different and stronger. A verified representative can suggest edits, and Google says it reviews that feedback "within a few days" and emails a resolution. Not every panel can be claimed, and descriptions can't be edited at all; Google points you to the source they came from.

Microsoft Copilot

Use the thumbs-down icon and explain what went wrong. Microsoft's transparency note says Copilot's feedback tools send reports that are reviewed by Microsoft's operations teams. Because Copilot draws on Bing, a page fix plus a Bing recrawl does more than the report alone.

Correct the Page Behind the Citation

The report is a note to the vendor. The page is what the engine reads. This is the step that actually moves LLM citations.

If it's your page

Fix the sentence, add a visible "Updated" date, and add a short dated note where an old fact used to be: "Until June 2025, the free plan covered 3 users. It now covers 5." Then ask for a recrawl so the LLM citations that point to the page pick up the new text: Request indexing in Google Search Console, and an IndexNow ping or URL Submission for Bing. Our hub guide explains which recrawl tool reaches which engine.

If someone else owns it

Write to the page owner. Engines can't edit other people's sites; as Bing puts it, you must contact the site owner to correct content at the source. Keep it short and specific. Three templates follow.

Template 1: a directory or review site

Subject: Correction request for [Brand] on [page title] Hi [Name or team], your page at [URL] says "[wrong sentence]". That was true until [date], but it's no longer correct. [Correct fact.] Our pricing page, updated [date], confirms it: [proof URL]. Could you update the listing? Thanks, [Name, role, Brand]

Template 2: a journalist or blog author

Subject: Small factual fix in your [month year] article on [topic] Hi [Name], thanks for covering [Brand] in "[article title]". One detail is out of date: the article says "[wrong sentence]". [Correct fact], as shown here: [proof URL]. If you're able to add a correction or an update note, we'd be grateful. [Name, role, Brand]

Template 3: a Wikipedia talk page

{{edit COI}} I work for [Brand] and have a conflict of interest. The article says "[wrong sentence]" in the [section] section. Please change it to "[correct sentence]". Source: [independent source, or the company's own dated page for an uncontroversial fact]. Thank you. ~~~~

Wikipedia's conflict-of-interest guideline strongly discourages editing your own company's article directly and points you to the {{edit COI}} request instead. On Wikidata, the community's self-promotion essay says you may correct clear factual errors on an existing item but shouldn't remove sourced claims; use the item's talk page for anything more.

When the owner fixes it but the old version lingers

If a site updated its page but "Bing and Copilot still show the old version," anyone can ask for a refresh with Bing's Content Removal Tool. Google's Refresh Outdated Content tool does the same for pages you don't own, but only after content was removed or changed significantly.

Recheck on a Schedule

A fix isn't done until the answers and their LLM citations change. Rerun the exact prompt you captured, three times in each engine where the error appeared, on this schedule.

Call the fact fixed only when it's right in every run at two checks in a row. One clean run is luck. For the metric behind this, Accuracy Rate, see our GEO metrics framework; for why single checks mislead, see the technical reality of tracking AI answers.

The One-Fact Correction Log

Keep one log entry per wrong fact; it's the record you'll want if the error comes back. Here it is filled in for Plannora, a made-up project management tool at plannora.io; the review site stackreview.co is made up too.

Two columns do most of the work. "Does the source say it?" tells you whether to chase the page or the engine, and the dated checks show whether the fix stuck.

If the wrong fact won't budge

If an answer still states the old fact at day 56 and cites nothing, it's probably coming from the model's training data. A page fix can't reach that until the vendor releases a new model, so keep your sources correct and recheck after each major model update. If the LLM citations still point to a page that's now correct, the engine may not have recrawled it yet; request a recrawl again and check the page is reachable by the engine's crawler. Our guide on how to track brand mentions in AI search shows how to fold these checks into a regular routine.

Where Rankbox Fits

Rankbox doesn't track LLM citations or monitor what AI says about your brand, so the capture, the reports and the rechecks in this guide are manual. The free AI Visibility Prompt Kit helps you write the prompts. Rankbox helps with the pages that fix the problem: its Citation-Ready Writer writes source-backed articles, such as a pricing explainer or a company facts page, that reach your site through Rankbox's API. The Business plan is $49.50 a month with a 7-day trial; see pricing.