Know when AI cites you. Prove it.

The observables

Mentioned, cited, or neither

Every rendered answer is read against your target, and the run writes one observation per answer. A brand mention in the answer text and a citation in the source list are recorded as separate things. The run never counts a warm paragraph as a reference.

A run that cannot render an answer stays unknown. The CLI keeps it out of the rate instead of counting it as a miss. A blocked page check works the same way: it never proves a link disappeared.

Your panel

Prompts you choose

A citation panel is a small JSON file you own. It names your domain or page, your brand and its aliases, the prompts your buyers actually ask, and how many answers each prompt gets per run. No vendor prompt database, no keyword quota and nothing to rent back.

The sample count is committed before the run starts, so an epoch cannot quietly stop early when the numbers look convenient. The file is yours to edit, version and carry to another machine.

Your access

Engines you authorize

ChatGPT answers through your own browser session today. A mock engine costs nothing and exists to test the whole pipeline before any spend. More engines are under validation in the package and land as they prove out.

Engine access stays with you. The CLI holds no engine accounts, buys no quota and keeps no answer index. Every observation names the engine, provider and model it came from. Every evidence snapshot is content-addressed, so you can trace a number to the answer behind it.

The statistics

One check is noise

Ask the same question twice and the citation set can come back different. Published measurement research on generative search treats a single-run visibility number as one sample of a moving distribution. The CLI treats it the same way. It samples in epochs, with a fixed number of answers per prompt per run. Every prompt-and-engine pair gets a citation rate with a Wilson interval and its sample size.

Each epoch is compared with the one before it, and the CLI calls a change only when the interval supports it. The tool makes exactly four calls: declined, grown, not distinguishable and insufficient data. A citation is never called removed. Rates move; the words follow the evidence.

Example change call. Citation rate declined from 0.80 (0.62 to 0.90, n=30) to 0.23 (0.11 to 0.41, n=30) between two weekly epochs.

The evidence

What one observation looks like

This is the first live observation from our own panel, kept the way the run wrote it. Our site was not cited in this answer, the sample is one answer deep, and the verdict says so instead of dressing the number up.

The CLI smooths the rate so a single miss never reads as a hard zero. The interval and the n carry the real story, and every number on a card like this traces to a kept snapshot of the answer.

Proof assets

See the code, run the tool

The citation watcher ships inside the open-source package. The repository holds the engine code, the statistics and the compatibility fixes; the docs hold the arguments, verdicts and exit codes.

Continue

Keep watching