Tools that query a model directly through an API, without a browsing or search step attached, can only observe what the model says, not what it retrieved, because nothing was retrieved. Any source names in that output are model-recalled: plausible, sometimes accurate, but drawn from training-data patterns rather than a verified fetch of the current page.
This is a real measurement limitation, and ClerAEO AI reports it openly rather than blurring the line: API-only queries surface model-recalled citations, while browsing-enabled or search-integrated modes reflect observed retrieval. The two are not interchangeable evidence of visibility, and any AEO tool that treats them as identical is overstating what it actually measured.