GenGrowthGEO Agent · AI answers
See whether AI answers cite your site
Eight buyer questions against real ChatGPT answers: five retrieval probes asked three times each, three natural buyer questions asked once. You confirm every fact and every question before anything runs, and the report keeps the retrieval numbers and the natural-demand numbers apart.
Run GEO AgentAccount access gate
A verified Supabase session is required to run. Each run makes 18 billed provider calls, so it is not available anonymously.
Report contents are not stored
The report is returned to this page and never written down: no account history, no server copy of the assistant's answer. Minimal billing and quota metadata is kept by the existing credits system. Leaving the page ends the report, so copy anything you need before you go.
Confirm what this site is
These words go into the questions, so use your buyer's vocabulary rather than your internal naming. Nothing is billed until you confirm the run.
A short noun phrase. It is substituted into calibrated wording, so an instruction or a sentence here will be refused rather than trimmed.
Optional, comma separated. The first one or two are used in the measured alternatives question; without any, that question falls back to its measured rival-free wording.
Passed to the provider as the search country. There is no default: an unset market stops the run rather than quietly becoming US. EU and UK are not country codes and are not offered.
en
English only in this version. The wording was calibrated in English, and a translated calibration-locked question is a different, unmeasured question.
How it works
From buyer questions to observed citations
Every number in the report is recomputed from the samples underneath it, and a sample that observed nothing is reported as exactly that rather than counted as a zero.
01
Confirm the facts
Tell the Agent what the site sells, who buys it and which market to ask in. Brand names and the category are proposals until you tick them; nothing is inferred on your behalf.
02
Confirm the questions
Eight questions are generated from calibrated wording. Edit anything that does not sound like your buyer — an edited retrieval probe is visibly demoted, because its citation counts only meant something for the exact string that was measured.
03
Sample by mode, not uniformly
Retrieval probes are asked three times, because one sample cannot tell a site that is not cited apart from a run that happened not to cite it. Natural buyer questions are asked once: repeating a question the assistant answers from memory mostly repeats the same non-observation.
04
Read it with its denominator
Retrieval and natural demand never share a number, prompted and unprompted mentions never share a number, and a probe that failed to trigger a search is excluded as an instrumentation failure rather than counted against you. Copy what you need before you leave — the report is not saved.
This report observes what one assistant answered at one moment. It is not a ranking, not a forecast, and not a measure of traffic.