How to Track What AI Says About Your Brand

You check your Google reviews. You'd notice a one-star takedown within a day. But do you know what ChatGPT tells people who ask about your company?

Most owners have never looked. That's a gap worth closing, because AI assistants now answer a meaningful share of "who should I hire / what should I buy" questions, they speak with total confidence, and they are sometimes flat wrong. The errors are ordinary ones: outdated pricing, a defunct product line, a competitor described as you. Unlike a bad review, a bad AI answer is invisible until you go looking.

The good news: monitoring this yourself costs nothing but an hour a month. Here's the system.

Step 1: Build your prompt set

Don't ask "what do you think of Acme?" Customers don't phrase it that way. Write 8–12 prompts that mirror real buying questions:

Direct entity checks: "What does [company] do?" / "Is [company] legit?" / "How much does [company] cost?"
Category shortlists: "Best [your category] for [your customer type]" / "Best [category] in [city]". These are the money prompts, because they're where you're competing for inclusion, not description.
Comparisons: "[You] vs [main competitor]" and "[competitor] alternatives."
Problem-first questions: the plain-language version of the problem you solve, such as "how do I [job your product does]", to see whether you appear when nobody names any brand.

Write them down verbatim. Consistency is what makes month-over-month comparison meaningful.

Step 2: Run them across the engines that matter

Minimum viable coverage in 2026: ChatGPT, Google Gemini (and a normal Google search to catch AI Overviews), and Perplexity. Add Claude and Copilot if your buyers skew technical or corporate. Two habits that make results honest:

Use fresh chats, ideally logged out or in a private window. Assistants personalize, so you want the answer a stranger gets, not one colored by your own history of asking about your company.
Run the important prompts twice. These systems are probabilistic, so the same question can produce different shortlists. If you appear in one run of two, you're marginal, not established. That's worth knowing.

Step 3: Log it like a scoreboard

A spreadsheet is fine. One row per prompt-engine pair, once a month:

Mentioned? yes / no.
What was said. One line, verbatim if it's wrong.
Accurate? flag anything outdated or false.
Who else appeared. The competitor set the machine has decided you belong to (or that you don't).
Sources cited. Perplexity and AI Overviews link them, and ChatGPT does when it browses. This column is the actionable one: it tells you which pages are feeding the answers.

The trend matters more than any single snapshot. Noise between runs is normal, but a mention that persists across engines and months is real presence.

Step 4: When the answer is wrong

There's no support line where you correct a language model, so the honest framing is: you can't edit the answer, but you can outweigh its sources.

Trace it. Wrong answers usually have traceable causes: an outdated pricing page, a stale directory entry, an old article, a name collision with a similarly-named company. The cited sources (or a search for the wrong claim) usually reveal the culprit.
Fix what you control. Correct your own pages, kill zombie listings, update profiles. Retrieval-driven answers can shift within weeks of the source changing.
Publish the correction where machines read. For claims you can't edit at the source, put the current facts on the record somewhere dated, third-party-hosted, and unambiguous. A distributed press release announcing the new pricing, the rebrand, the relocation creates exactly the kind of recent, syndicated document that retrieval systems surface above stale pages. Recency wins ties.
Use the feedback buttons, without expectations. Thumbs-down with a correction on ChatGPT or Gemini may help in aggregate. Treat it as a lottery ticket, not a strategy.
Escalate only for defamation-grade errors. If an assistant asserts something damaging and false (regulatory trouble you don't have, a scam accusation), document it with screenshots and dates and use the provider's formal reporting channels. This is rare, and most errors are staleness, not malice.

Step 5: When you're missing entirely

Absence is the more common finding, and it's information: the public record about you is too thin for the machine to know you belong in the answer. The fix is the whole GEO playbook in miniature: machine-readable facts on your site, consistent entity info everywhere, and above all more third-party documentation such as coverage, roundups, reviews, and a steady cadence of announcements on the record. Expect lag. Retrieval answers can pick you up in weeks, but answers drawn from trained knowledge move on training-cycle time. This is a quarterly game, not a weekly one.

The emerging tool landscape

Purpose-built AI-visibility trackers have arrived fast. Profound, Peec AI, Otterly, and features bolted onto Semrush, Ahrefs, and the enterprise SEO suites all now run prompts across engines on a schedule and chart your "share of voice." Honest guidance for a small business:

They automate exactly what the spreadsheet does, at $50–$500+/month. The value is scale and trend charts, not secret access. Nobody has an API into the model's opinions.
The category is young and churning, so expect tools to merge, pivot, and die through 2026–2027. Don't sign annual contracts for a moving target.
Do the manual version for two or three months first. You'll learn which prompts matter for your business, which makes you a much smarter buyer if you later pay for automation. And for many SMBs, an hour a month in a spreadsheet is genuinely enough.

Bottom line

What AI says about you is now part of your reputation, and it's checkable in an hour a month: fixed prompt set, three engines, one spreadsheet, quarterly review. Wrong answers get fixed by outweighing their sources. Missing answers get fixed by building a record worth citing. The owners who check are the ones who'll notice, and correct, before their customers do.

The reliable way to put fresh, correct facts where AI systems read them: [distribute the news](/pricing).