How to Spot and Fix Mistakes in AI Answers

Throughout this article, we’ll use Ahrefs Brand Radar to collect AI answers and see which pages they cite. We’ll also use a custom app we built in Letaido, which connects to the data from Brand Radar to automate the fact-checking process. This step is optional, but it can save you a huge amount of time.

The basis of the fact-checking process is two kinds of prompts. Together, they reveal both known problems and errors you did not think to look for.

A fact checker can’t catch an error unless you tell it what the right answer is. That’s the source-of-truth document: a plain list of checkable claims about your product, each with a link that proves it.

It does double duty. The fact checker compares AI answers against it, and the human reviewer uses the same list to decide whether a flagged response is real before anyone emails an author.

Once you’ve found an incorrect AI answer, the next question is: where did the bad information come from?

You can trace these errors back to their sources manually, but it quickly becomes time-consuming when you’re checking hundreds of answers. For larger audits, we recommend using AI to do most of the heavy lifting.

Not every incorrect page is equally worth chasing. Prioritize the ones where a correction is most likely to improve AI answers or reach potential buyers directly.

Here are the main signals to look at:

• Severity: Fix errors that could affect a buying decision first, such as incorrect pricing, plan access, services offered, product limits, or missing features.
• Citation frequency: Pages cited often across answers or platforms have more influence than one-off sources.
• Citation stability: Sources that appear consistently across repeated checks are more valuable than those that show up once.
• Organic traffic: Pages with meaningful search traffic can reach buyers through both traditional search and AI answers.
• Authority: Referring domains and Domain Rating can strengthen the case, especially when a page is also frequently cited and relevant.
• Relationship: Do you or a colleague already know the author? Could someone introduce you? Has the publication responded to you before? For example, for our campaign, Glen was the right person to approach one site because he already had a relationship with them. In another situation, the citation came from an influencer we already had relationships with, so our Partnerships team was the best fit to handle the outreach.

You can find some of that information in Brand Radar:

And if you’re using our Letaido app, the data is pulled from Brand Radar and served right next to the affected page:

Once you’ve found a page you want to reach out to, the next step is figuring out who is actually responsible for it.

For a small number of pages, the standard web-fetching tools available in AI assistants are usually enough to pull the page and look for author or contact information.

At a larger scale, though, checking URLs one by one quickly becomes impractical. You’ll typically want to use a scraping API like Apify or another source of page HTML.

Our app first tries Ahrefs-stored snapshots. If that fails, it tries Apify, and if the contact can’t be found on the cited page, it follows the link to the author’s page (if linked) and tries again there.

Start with the person named as the author. Check the article byline and author page first, then look at their LinkedIn or X profile if you need a way to contact them.

If you can’t find the author or they’re no longer there, move one level up. For a publication, that might be an editor or content lead. At a small company, the founder or marketing lead may be the best contact.

As a last resort, use the company’s general contact or support channel and ask for your message to be forwarded to whoever manages the page or content.

Tools such as HeyReach or BuzzSumo can help fill in missing contact details and identify likely people to approach. Many of them offer an API or MCP, so you can wire their functionality into an AI-assisted process.

Keep your outreach helpful, not demanding. You can thank them for covering your product, point out one or two specific errors, link to a reliable source for the correct information, and make it clear there’s no pressure to update the page.

To illustrate, here’s one of the messages we’ve sent:

This lowers the barrier to starting a conversation, although it may be less effective because you’re asking the recipient to reply before they can see the value of the correction.

If you don’t hear back, follow up once. After that, move on to the next page on your list. Some of our follow-up messages actually worked.

A few more practical tips for outreach:

• Be friendly. Do not accuse the author of spreading misinformation or threaten them with reputational harm.

• Assume the page became outdated rather than that the author acted in bad faith.

• Keep the request specific. One verified correction is easier to act on than a long list of complaints.

• Consider LinkedIn Premium or Sales Navigator if you need to send more direct messages or follow up without an existing connection. Sales Navigator allows you to send follow-up messages even if the first message wasn’t read by the recipient.

• Track at least the URL, contact, outreach date, channel, claim to fix, response, follow-up, and confirmed change.

• And just in case, you can take a screenshot of the page before outreach. If you use a contact form, save proof that the message was submitted. If you’re using Ahrefs, you can use Site Explorer to document the changes on the page between dates.

We contacted 26 authors about inaccurate information in their articles. 14 replied (53.8%), and 6 made corrections we could verify on their pages (23%).

This is still ongoing, and of course, authors aren’t obligated to make any changes. Even so, getting nearly a quarter of the articles corrected is a promising result.

Here’s one of the outreach targets, before and after (diff view via Ahrefs’ Site Explorer).

And yes, updating pages, even though we didn’t catch them all, did introduce correct facts in AI answers. Here’s an example (a big thank you to Joonas from thatmarketingbuddy.com):

Two more authors said they would make changes, but those were still pending when we reviewed the results. We kept them out of the confirmed correction count.

For 12 of the 26 contacts, we received no reply.

We never received an outright refusal to update an article. Some authors, though, were willing to make a correction only if we mentioned or promoted their product in return. We declined. We felt that keeping published information accurate should be worthwhile on its own, and offering something in exchange would be unfair to the authors who made corrections without asking for anything.

The outreach also exposed a problem with our own explanations. Some of the confusion came from unclear descriptions of Brand Radar’s pricing and feature limits. That gave us something to work on ourselves: making it easier for someone writing about the product to understand what was included.

Getting a reply was easier than getting a change made. That’s worth remembering when you plan this work. You’ll need time to check whether promised updates happen, and you should expect some messages to go unanswered.

You can’t control what AI assistants say, but you can make the correct information easier for them—and everyone else—to find.

This means monitoring brand mentions across the web and reacting quickly when you spot inaccuracies. In my experience, content creators care most about their work right after publishing, so reaching out early gives you a better chance of getting them to fix the page.

Of course, this doesn’t guarantee anything. You don’t know whether a particular page will ever be cited in an AI answer. But acting fast may stop misinformation from making its way into some AI responses.

You can also look for signals that a page is more likely to get cited. Its DR, reputation, and relevance to the topic you care about can all be useful proxies—especially if you’re operating in a narrow niche with only a handful of relevant pages on the web. Freshness can matter too, particularly when a new page is more up to date than the sources AI systems are already citing.

You can set up brand mention monitoring with a tool like Firehose.

You could also connect your agent via API and have it automatically fact-check each new source as soon as it’s found, giving you near-real-time monitoring.

Web mentions have become a currency. Getting your name onto certain pages costs, and apparently, getting your brand represented accurately can cost, too. That matters because where your brand is mentioned online—and what those mentions say—can have a huge impact on how visible you are in AI answers.

That’s why an app like ours, which automates the process of finding and fixing inaccurate information, can be so useful. But as we found out, unless you’re prepared to offer something in return for someone updating information about your brand, you may not get very far.

The bigger problem, though, is something AI developers need to address: how AI systems choose which sources to trust when answering questions about brands. Right now, that process doesn’t really serve anyone.

You can’t expect every content creator or website owner to constantly check their facts and keep information about every company up to date. But you can expect a business to care about keeping its own information accurate. It has every reason to—wrong or outdated information can directly hurt its business.

AI systems should be better at recognizing that. When it comes to basic facts about a company, accurate and up-to-date information from the business itself should carry more weight than an old mention on a third-party website.

Thanks for reading! Come and say Hi to Glen and Mateusz.