A buyer opens ChatGPT and asks for the best tools in your category. The answer names four companies, links to two of them and puts one first. If you aren't on that list, you may never hear from that buyer, and nothing in your analytics will show the conversation happened.

More purchases start this way now. In G2's March 2026 survey of 1,076 software buyers, 51% said they begin research in an AI chatbot more often than in Google, up from 29% in April 2025. Sixty-nine percent chose a different vendor than they'd planned based on a chatbot's guidance, and a third bought from a vendor they'd never heard of. For a young company, that last number is an opening. It also raises a question founders keep asking us: what is AI saying about us? We built Overheard AI to answer it, and we've released it free and open source.

One answer tells you very little

The obvious test is to ask ChatGPT yourself. The trouble is that the answer changes each time you ask. SparkToro and Gumshoe had 600 volunteers run the same prompts nearly 3,000 times. They found less than a 1 in 100 chance that ChatGPT or Google's AI returns the same list of brands twice in 100 answers, and the order repeated even less often.

What held up was how often a brand appeared across many answers. That matched what we'd seen building visibility trackers for client work. One screenshot of an AI answer can make you feel great or terrible, and neither feeling survives asking again. So Overheard AI never reports a single ranking. Every number it shows is a share of answers.

What it measures

You describe your brand, pick the questions your buyers ask and choose the assistants. Overheard AI asks each question several times of each assistant, with web search on for every answer, and keeps the full text. A second, cheaper call reads each answer back and records the brands it named, their place in any ranked list and the sites it linked.

The dashboard shows how often you're mentioned, how often you make the top three and how often your site is cited, next to your competitors. Brands you didn't list but the assistants keep naming show up as discovered competitors. The app also asks each assistant what it knows about you and sums that up in four parts: what you do, who buys from you, what you're known for and what people criticize.

A few counting rules keep the numbers honest. An answer counts at most one mention per brand, so a chatty model can't outscore a decisive one. A question that names your own brand is asked and shown but never counted, because "Is Acme any good?" will always mention Acme. Failed calls are shown separately and never counted as misses.

It runs on your machine

Overheard AI runs on your own computer with your own API keys for OpenAI, Anthropic or Google. One key is enough. Every answer is stored in one SQLite file that you own. There's no account and no hosted service, and nothing leaves your machine except the calls to the providers you have keys for.

Each answer takes two calls on your keys, one to ask and one to read the answer back. Before a run starts, the app shows the planned calls and an estimated cost. Treat that as a rough figure, since the number of web searches varies, and treat your provider's billing page as the real one.

One limit is worth knowing. It asks the assistants through their APIs, with search on. That's close to what people see in the ChatGPT, Claude and Gemini apps, but it isn't identical, and the apps can answer differently.

Try it before you add a key

You can see it working without spending anything. The first launch opens a tutorial that builds a demo project with six months of history. The demo's numbers are generated, and every demo page says so.

When you're ready, you need Node 22.13 or newer. Install it by hand with five commands, or paste the setup prompt from the README into a coding agent such as Claude Code or Cursor. The agent never handles your keys. You add those yourself.

Overheard AI began as the measurement engine inside a hosted product we were building. We rewrote it to run on one machine and released it under the MIT license. If the results raise a bigger question about how buyers find you, we're happy to talk it through.