Gemini can answer with a live search attached or from what the model already holds, and the difference matters more than anything else on this page. An ungrounded answer describes what a model absorbed during training: not what a buyer sees today, not influenced by anything you ship, and traceable to no source.
We require grounding. When an answer comes back without it — which does happen, because the model can decide not to search — that is recorded as its own case rather than counted as a normal result. Mixing the two silently would blend 'what the web says' with 'what the model memorised' inside one published percentage.
Two things our own runs found
- A dead model id looks exactly like an unconfigured oneThe default model in an integration expires. Because failed calls are excluded from every denominator, an engine that has gone dark keeps every dashboard looking healthy while measuring nothing. We record the failure as a stored fixture rather than quietly repointing.
- Citations can point at the middlemanGrounded answers sometimes return redirect URLs on the vendor's own domain with the real publisher in the title. Classified naively, every citation in the product becomes third-party forever and your own site can never be credited. We read the publisher rather than the redirect.
What we can honestly tell you: whether this engine named you, in what order, alongside whom, and citing what — in the answers we bought, on the days we asked. What nobody can tell you is why it chose you, or that a change on the site will move it. We report the measurement and leave the causal story alone until we have measured enough weeks to know what its own noise looks like.
Why not just drop the grounding requirement?
Because it makes every error disappear and every call succeed, while changing what is being measured into something a buyer never sees. It is written into the integration so nobody makes that change on a quiet afternoon.
What happens to an ungrounded answer?
It is stored and flagged. Whether it should count in the visibility denominator at all is an open question we have not settled unilaterally, because the answer changes what every published percentage means.
Try it on your own site, free
No card, and the free tier does not expire — the first verdict takes about six weeks to arrive.