AI brand monitoring: find and fix what ChatGPT gets wrong

AI brand monitoring means asking ChatGPT, Gemini, Perplexity and Google's AI answers the same questions about your brand every month, saving each answer with its sources, and fixing the claims that are wrong. Start with five prompts about what you are, what you cost, your alternatives, whether you fit a buyer's use case and what reviewers say. When an answer is wrong, find the page it came from. Fix your own pages first, then the third-party pages the engine cites, then your entity data, and use the feedback buttons last.

Search Console's Generative AI report counts impressions only, and no engine shows you the words it used about you. To know what buyers read, you have to ask.

Which prompts to check for AI brand monitoring

Check five prompts, worded the way a buyer types them, and reuse the exact text every month. A reworded prompt is a different question and breaks the comparison.

What is Acme Analytics?
How much does Acme Analytics cost?
What are the best alternatives to Acme Analytics?
Is Acme Analytics good for a 10-person marketing agency?
What do reviews say about Acme Analytics?

The "what is" prompt shows whether the engine knows you and how it categorizes you. Pricing goes stale easily, because review sites and old posts keep quoting the old price. Alternatives shows who you get grouped with. For "good for X", use your real best-fit customer. Reviews shows which complaints have stuck.

If another company shares your name, also run "What is Acme (acme.io)?". Buyers type the plain version. The second tells you whether the engine can tell you apart once it has a hint.

Shortlist questions that don't name you, such as "best analytics tool for agencies", need a different check. AI Answer Visibility asks ChatGPT, Perplexity and Gemini such a question and shows whether each answer names or cites you.

How to monitor your brand in ChatGPT and other AI engines

Check the engines your buyers use. For most businesses that means ChatGPT, Google's AI Overviews and AI Mode, Gemini and Perplexity, plus Claude or Microsoft Copilot if customers mention them. Asking five recent customers which assistant they used beats any market-share chart.

Run the full set once a month, and again within a week of a price change, rebrand or launch, since that's when answers fall out of date. After a fix, rerun that one prompt weekly until the answer changes.

Use a logged-out browser or ChatGPT's Temporary Chat, a fresh conversation per prompt, and three to five runs each. Answers vary, for reasons covered in why ChatGPT gives different people different answers. A claim that shows up in most runs is the one buyers see.

How to record what ChatGPT says about your brand

Log every answer in a spreadsheet, one row per prompt, engine and run, with the full text pasted in. Screenshots can't be searched or compared.

The mode column matters most. OpenAI's article on searching the web with ChatGPT says answers that use web search may include citations and a Sources list. An answer with no citations most likely came from training data, and that changes how you trace it.

How to classify wrong AI answers about your company

Sort each wrong claim into one of five classes. Each has a usual source and a usual first fix.

Then rank them by cost. A wrong price or a mix-up with another company loses deals this week. A missing minor feature can wait. A fair complaint isn't an error, and its fix is in the product.

How to trace a wrong AI answer to its source

Trace an error through its citations when the answer searched, and through a rerun with search when it didn't.

1. Open each cited source and find the sentence carrying the wrong claim. Note its date. OpenAI warns that citations "can be incomplete, outdated, or incorrect."
2. If no cited page contains the claim, the model blended sources or filled a gap. Your own pages probably say too little about that fact.
3. If there were no citations, regenerate with search. ChatGPT's refresh control offers Search the web when available. A correct searched answer means the error lives in training data. A wrong one means a live page feeds it, and the citations name it.
4. Search the wrong phrase in quotes on Google and Bing, such as "Acme Analytics" "$29". Pages that repeat it can be cited next month.

Gemini needs care. Google's help article on Gemini's sources describes them as links to content related to parts of the response, which stops short of saying Gemini used them. Treat them as leads and confirm the claim on the page itself.

Nobody can edit a model that has shipped. Fix the web anyway, since searched answers read it today and the next model trains on it.

How to correct wrong AI information about your company

Correct the record where the engines read it, starting with the pages you control and moving outward.

Fix your own pages first

Engines that search fetch your site directly, and you can change it in an hour. Put prices in plain text, not in an image or a calculator. Open the home page with one sentence that says what you are and who it's for. If you renamed, say so on the About page, as in "Acme Analytics was called Metricly until March 2025." Old posts that quote retired prices need an update note or a redirect.

Then check the engines can reach those pages. OpenAI says a site must allow OAI-SearchBot to be eligible for ChatGPT search, and its search article names Microsoft among ChatGPT's search partners, so confirm Bing has indexed your current pricing page. The AI crawler list has the user agents to allow.

Then the third-party pages being cited

Work down the cited URLs in your log, most-cited first. Claim and update your profiles on the review and directory sites that appear. Send the author of a stale comparison article the correct figure and a link to your pricing page. In forum threads, reply with current facts and say who you are.

Wikipedia's conflict of interest guideline strongly discourages editing an article about your own company. Propose the change on the talk page with the {{edit COI}} template and a source.

Old complaints need newer reviews, not arguments. Ask recent customers to review you on the sites the engines cited.

Then your entity data

Entity data tells an engine that "Acme" on a review site, on LinkedIn and on acme.io is one company. Add Organization markup with your exact name, logo and sameAs links to your official profiles. Brand Entity Checker checks that your name, logo, markup, sameAs profiles and Wikidata item tie back to your site. If Google shows a knowledge panel for you, claim it and suggest edits. This layer does the most for the "confused with another company" class.

Use the feedback buttons, with low expectations

Every assistant lets you flag a bad answer. Their help pages offer brands no channel for submitting corrected facts, and none promises a report will change a specific answer.

Write what the answer claimed, what's true and the URL that proves it. Perplexity's article on reporting inaccurate answers also asks for the query URL.

OpenAI's content reporting article says its Model Quality team may review reported domains and apply filters so ChatGPT stops relying on unreliable sources. That route covers content that may break OpenAI's terms or the law, such as a defamatory page. An out-of-date review site doesn't qualify.

Check what ChatGPT, Perplexity and Gemini say about your brand

Brand Narrative Check reads your home page, then asks ChatGPT, Perplexity and Gemini, with web search on, one question that folds the what-is, pricing, alternatives and reviews prompts together. The question includes your domain, so a namesake doesn't get mixed in.

For each engine, the report shows whether it recognized your brand, described a different company or found nothing, plus its description, category, pricing, tone, strengths, weaknesses, alternatives and cited sources. That covers the error classes above, as far as your home page states the facts. It also lists potential discrepancies with your home page, quoted word for word from both texts so you can judge each one.

A run costs 150 credits and takes one answer per engine. It compares answers with your home page only and doesn't track them over time, so keep your log and rerun it monthly.