When ChatGPT gets your company wrong: finding and fixing the facts

The wrong facts you will find

Run your customers' questions through ChatGPT, Perplexity, Gemini and Claude and the mistakes fall into a short list, each with a typical cause and fix.

• Old prices. The assistant quotes the plan you had two years ago, or a price in the wrong currency, or a per-month figure for a per-seat plan.
• A feature you do not have. Something that was on your roadmap page, in a partner's blog post, or in a competitor's description of you, stated as if it shipped.
• A feature you no longer have. You retired it, but the help article and the launch post are still live and still cited.
• The wrong founding year, founder or location. Usually from an old directory listing or a profile nobody maintains.
• A merged or renamed company. The assistant still uses the old name, or describes the company you acquired as a competitor, or merges you with a company that has a similar name.
• A competitor's feature attributed to you, or yours to them. Comparison pages written in a hurry do this often, and the assistant reads both companies' pages in the same search.
• Wrong integrations. "Does Fernwick integrate with HubSpot?" answered with the list from a marketplace page that was correct in 2023.
• "No free plan" when there is one, or the reverse. A common one, and costly, because it is the sentence a buyer acts on.

The pattern across all of them: the assistant did not make anything up, it believed something it read.

Why it happens

Four causes account for most of it, and each points to where the fix is.

Training data from years ago. Every model has a training cut-off, and the facts it absorbed about your company come from pages as they were then. With search off, or when the model does not bother to search, that is what it answers from. You cannot edit it. You can only make the current facts so widely repeated that the next model, and the search that supplements this one, picks them up.

Outdated third-party pages. Review sites, directories, "best X" lists, a partner's integration page, a Wikipedia article. These were right once, nobody owns keeping them right, and search-based assistants cite them anyway because they rank. Why AI answers lean on Reddit, reviews and forums explains how much weight they carry.

Your own old pages, still live. The 2022 pricing page that was never redirected. The launch post for the feature you retired. The help article for the old plan names. If it is on your domain, the assistant treats it as authoritative, and it should.

Inconsistent facts across your site. The pricing page says one thing, the comparison page says another, the footer says a third. An assistant that reads two of them picks one, and you do not get to choose which.

How to find them

You cannot fix what you have not listed, so the first job is a list.

Keep a fact sheet

Write down 20 to 40 facts about the company that are true today and that a buyer might ask about. One line each, plain, with a date. For Fernwick, a scheduling tool, the sheet would include lines like: "Free plan: yes, one user, unlimited bookings." "Team plan: X per seat per month, billed yearly." "Integrations: Google Calendar, Microsoft 365, HubSpot, Attio, Slack, Zapier." "Founded 2021 in Lisbon." "Not: a video calling tool; no built-in video." The negatives matter as much as the positives, because many wrong facts are things attributed to you. This sheet becomes the reference for everything below, and the source for the page you will publish in a moment.

Ask the questions your customers ask

Take the prompts that invite facts: "What does Fernwick cost?", "Does Fernwick have a free plan?", "Fernwick vs Slotwise", "Does Fernwick integrate with HubSpot?", "Who owns Fernwick?". Run each across the assistants, with search on where that is an option, and once more a few days later; answers vary between runs, so one pass is not a verdict. How to choose the prompts to track covers building the set.

Compare and log

Check each answer against the sheet, line by line, and record every mismatch. A spreadsheet with these columns does the job: date, assistant, prompt, the wrong claim, the true fact, the source the answer cited (if any), status. Echo's FactCheck does this automatically against the facts you keep, but the method is the same by hand. The cited source is the most valuable column: it tells you where to go next.

How to fix them, in order of effect

Work from the sources you control outward. Each step changes more answers for less effort than the one after it.

1. Fix your own pages and say the fact in text

Assistants extract facts from sentences, not from pricing toggles or images. Every fact on your sheet should appear on your site as a plain sentence somewhere a crawler can read it: "Fernwick has a free plan for one user." "Fernwick integrates with HubSpot and Attio." Check the pricing page, the integrations page, the About page and the home page, and make them agree with the sheet and with each other. A fact that lives only in a JavaScript-rendered table or a screenshot does not count. How to get your pages cited by ChatGPT has the page-level checklist.

2. Publish a "facts about us" page

One page, linked from the footer, that is the fact sheet in public: name, what the product does in one sentence, founding year and place, founders, plans and prices with a "last checked" date, integrations, what the product is not, and a contact for corrections. Journalists, partners and assistants all look for a page like this, and when one exists it tends to become the cited source. A press page does the same job for a company that gets written about.

3. Update the third-party pages that get cited

Go down the "source" column of your log, most-cited first.

• Review sites. Claim the listing, correct pricing, features and integrations, and keep them current. G2 and its peers are cited constantly for exactly the facts that go wrong.
• Directories and data sites. Crunchbase, LinkedIn, Product Hunt, app marketplaces, the local company register: the founding year and the "about" line usually come from one of these.
• Wikipedia, where justified. If the company has an article, request corrections on the talk page with sources; do not edit your own article. If it does not, do not create one for this purpose.
• Partner and integration pages. Ask the partner to update the listing. They want it right too.
• Lists and comparison pages. Write to the author with the specific correction and a link to the fact on your site. How AI uses "best X" lists and comparison pages covers the approach.

4. Remove or redirect the old pages

Take the old pricing page, the retired feature's launch post and the stale help articles, and either update them with a dated note at the top, redirect them to the current page, or return 410 so crawlers drop them. Do not leave them live and contradicting the sheet. A site map review once a quarter catches these.

5. Correct community threads as yourself

Where a Reddit or forum thread states the wrong fact and gets cited, reply as yourself, with your role in your profile, with the correct fact and a link. Short, factual, no pitch, and never from an anonymous account.

How long before the answers change

It depends on where the wrong fact came from.

Search-based answers update when the pages do. Once your page and the main cited source say the right thing and have been recrawled, the next search-grounded answer tends to reflect it. In practice that is days to a few weeks, depending on how often the engines fetch the pages involved. Perplexity and ChatGPT with search on move first; Google's AI Overviews follow Google's own index.

Training-based claims change slowly. If the assistant repeats the wrong fact with search off, or cites nothing, the fact is in the model's training data and nothing you publish changes the current model. It changes the next one, and only if the correct fact is now what most of the web says. The practical response is to make sure every search-grounded answer is right, keep the sources consistent, and accept that the ungrounded version catches up on the model's schedule, not yours.

Re-run the prompts that were wrong every two weeks after a fix and record the date each one came right. That gives you your own evidence of how long each kind of correction takes, which is worth more than any general rule.

Keeping it from coming back

Wrong facts return because the world changes and the sheet does not. Three habits keep that in check.

1. A monthly check. Run the fact-inviting prompts across the assistants, compare against the sheet, log the misses. An hour a month for a small team.
2. Update the sheet when the facts change. A price change, a new integration, a retired feature or a rename triggers a sheet update, a site review and a note to the review sites, on the same day. Put it in the launch checklist.
3. Own the sources. Assign one person the listings on the review sites, the directories and the partner pages, with a quarterly reminder. Unowned listings are where next year's wrong facts come from.