How do I know if ChatGPT or Gemini is giving accurate information about my company?

Start with an audit, not a tool

Clients often come to us after a customer or a salesperson spots something odd. The wrong founder name. A pricing tier that was retired two years ago. A feature the company never had. The first question is always the same: how bad is it, and how would we know?

Our answer is a structured audit. It needs a spreadsheet more than software, though software makes it repeatable.

The fact sheet comes first

We ask every client for a short, dated source of truth before testing a single prompt. It lists the legal name, location, what the company does, the current products and plans, leadership, certifications they can prove, and what they do not offer. That last item matters more than people expect. Many wrong answers are false additions rather than omissions.

If the client cannot state a fact plainly, we fix that first. An engine cannot be expected to find a clear answer where the company itself has none.

The prompt panel

We write the questions a customer, a journalist or a competitor would type, in several phrasings. Identity ("what is X"), pricing, founders, comparison against named rivals, suitability for a use case, and reputation ("is X any good"). We run them on the engines the client's buyers use, and we keep the full response text and the cited sources.

For repeatability we run the panel in Promptwatch. It stores complete responses with citations, runs prompts daily per model, country and language, and lets us group prompts by topic, persona and intent. We set up a Brand Book first. Its analysis instructions decide what counts as a mention, and brand aliases catch name variants, which stops a misspelling from hiding a response.

Scoring

We mark each claim in each answer: correct, wrong, outdated, missing or unsupported. We also note position and tone. Promptwatch scores sentiment per response from 1 to 100 and writes a weekly AI summary, which is useful for the tone question, but the claim-level scoring we still do by comparison with the fact sheet. A tool does not know your pricing is wrong unless you tell it what is right.

Where we usually find trouble:

• Pricing and plans. Old figures live on in third-party articles long after a site update.
• People. Founders, titles and company history.
• Capabilities. Features borrowed from a competitor with a similar name.
• Comparisons. Competitor pages that frame the client unfairly.

Finding where a wrong claim comes from

The citations are the lead. If a Gemini or ChatGPT answer states a wrong price, we look at what it cited. The pattern is usually an outdated third-party page, a thin directory listing, or a comparison article. Promptwatch shows page and domain citations and reports offsite mentions, where the brand is named on third-party pages without a link. A Page Tracker lets us follow specific URLs, including a competitor's page that keeps getting cited.

Then the fix is practical: update our own page so the fact is unmistakable, contact the owner of a wrong third-party page, publish a clear answer to the exact question, and track whether the answer changes.

One note on the engines. Promptwatch monitors ChatGPT Search and Gemini through the live interface, along with Google AI Overviews, AI Mode, Copilot and Perplexity. Models such as Claude, DeepSeek and Mistral are tracked through the API, and we say that in client reports so nobody assumes otherwise.

Making it continuous

A one-time audit goes stale. Models update and pages change. We keep a fixed panel on a daily run and read the weekly sentiment summary. We use usage alerts so we do not run out of quota unnoticed, and a weekly report so the client sees the same view each time.

What we tell clients honestly

We cannot make ChatGPT or Gemini correct itself. We can show what was said, show which sources it leaned on, and help fix those sources. Some corrections take weeks and some never fully land. When the topic touches legal, medical or financial claims about the company, we tell clients to involve their own counsel before sending takedown requests or publishing corrections. We are an agency, not a law firm.

A simple first pass you can do today

1. Write the fact sheet, one page.
2. Write ten questions a customer would ask.
3. Run each in ChatGPT and Gemini, in a fresh session, and paste the answers into a sheet.
4. Mark each claim correct, wrong or outdated.
5. Check the cited sources for each wrong claim.

That is an afternoon, and it usually produces a to-do list. When the list needs to be rerun every week, Promptwatch is the tool we use.