reach
An assistant can confidently describe outdated copy
The failure that looks most like success. The page was found, fetched and read correctly — just not this month's version of it.
Three answers about one site, collected from the same provider in a single afternoon. None of them had read the live page. Their copy dated to eighteen days, nineteen days, and current — and the nineteen-day answer cited the site's live URL while doing it. Every fact in all three was accurate. All three were accurate about a version that no longer existed.
Nothing in the answers marked it. Present tense throughout, a link to the current address, and no hedge — a citation reads as currency and is not one. It is a link the model attached, not a page it just read.
What it costs is not that an answer is old. It is that the answer is confidently about a site you no longer run, and the difference lands exactly where you changed something. One of those three valued the collection in six figures, because the copy it read said the works had been acquired. That sentence was deleted three weeks earlier. Every answer reading the copy that replaced it describes the same works as belonging to other people. The deletion was deliberate and defensible; the point is that both answers were live at once, and only one of them was about the current site.
This is a reach failure rather than a credence one, and it is worth keeping them apart. The system did not fail to find you and did not disbelieve you. It found a copy of you, and the copy was fine. There is no fix on the page, because the page is already right — the only available move is to know it is happening, which takes dated snapshots and a deliberate check rather than reading the answer.