Field notes: the week the archive refresh reshuffled every match we'd called

Tuesday morning, I pulled the same AOI I'd been working for three weeks. Same query tile, same radius, same everything I thought I'd locked down. The top of the ranked list had moved. A site I'd called a confident match two weeks earlier, logged it, briefed it, had slid from rank four to rank nineteen. Nothing about the site had changed. The archive had.
This is the part of similarity search nobody warns you about until it happens to your own AOI. A ranked list of matches is a snapshot of how one tile compares to everything else in the archive right now. Add a new collection pass overnight, backfill a gap with reprocessed imagery, swap in a sharper look at a region that used to sit under cloud shadow, and the whole neighborhood around your query tile shifts. The population it's being ranked against moved; the thing you were looking for stayed exactly where it was.
I've started thinking of it the way a fisheries biologist thinks about a tagged population survey. You don't recount every fish every week and expect the same numbers. New individuals enter the frame, old ones drop out of range, and your estimate of "how many look like this one" is only ever a statement about the sample you had when you ran it. Treating a similarity export as a one-time deliverable, something you run once, print, and file, is the mistake. It's a living list, and it needs a date stamp as much as a confidence rank.
The refresh that caused my particular reshuffle was mundane: a new acquisition pass over a sector that had been thin on recent coverage, plus what looked like a reprocessing pass that cleaned up some older haze-affected tiles. Six new candidates entered the top twenty that hadn't existed in the archive two weeks prior. Three previously-ranked sites dropped out because better-matching tiles had simply arrived. That's the system working as designed. It's just not what a one-and-done workflow is built to absorb.
There's a seasonal version of this too, and it bites harder because it's slower to notice. A query tile pulled from a July scene, full canopy, strong shadow contrast, won't rank the same set of candidates in January as it did in August, even against an unchanged archive, because the appearance of the thing you're matching against shifts with the season. If your query site itself gets re-collected in a different month, you're quietly changing what "similar" means.
None of this is an argument against the approach. It's an argument for logging dates on calls the way you'd log a collection date on any other source, and for re-running a standing AOI on a cadence instead of trusting a single export from three weeks back. The alternative, which most of us still do by default, is scrolling tile by tile across a region too large to review by eye, hoping the one site that matters hasn't already drifted out of the window you last checked. That's the gap Broad Area Search is built around: mark the one site you already know, and let the matching run again against whatever the archive looks like today, rather than whatever it looked like the week you happened to pull the list.
My rank-four site is back in the top ten now, by the way. A later pass picked up a clearer look at it. The list will probably move again before the month is out, and that's fine, as long as somebody's checking it on a schedule instead of trusting the one printout in the folder.
If a standing list of sites feels like it needs rechecking more often than it's getting checked, that's worth a look at what Broad Area Search does with the next refresh.