Field notes: the afternoon one query tile turned up sites on three continents

There's a particular kind of fatigue that comes from looking at the same kind of structure for six hours straight. Not boredom exactly. More like the visual version of saying a word so many times it stops meaning anything. You've been staring at rooflines, berms, the way a compound's perimeter track makes a specific shape when a specific kind of work happens inside it, and by hour five your eyes start seeing that shape in cloud shadows.
Every analyst who has done broad area search by hand knows this afternoon. You have one site. You know what it is, or you have a strong read on it, and the job is to find out if there are others like it somewhere in a region the size of a small country. So you pull up the archive and you start panning.
Tile, tile, tile. You build a mental template of the thing you're hunting and you hold it in your head while you scroll, which is exhausting in a way that's hard to explain to someone who hasn't done it. The template degrades. You start second-guessing a match you were sure of an hour ago.
What's changed the shape of that afternoon for a few analysts I've talked to isn't a new sensor or a new archive. It's a different way of asking the question. Instead of carrying the template in your head, you mark the one site you already know and let a model carry it for you, comparing against everything it's seen, flagging visual similarity rather than you flagging it. The archive stops being something you browse and starts being something you query.
The version of this that sticks with me is the one where the known site wasn't rare. It was a mid-sized processing layout, the kind of footprint that could plausibly belong to a dozen different industries. The analyst ran it as a similarity query across a search area that spanned three continents, mostly out of curiosity about how the tool would handle a shape that common. It came back with a shortlist that included two sites nobody on the team had flagged before, built years apart, in countries that had no obvious reason to share a construction pattern. One turned out to be a coincidence of layout, a facility doing something unrelated that happened to need the same geometry. The other wasn't a coincidence at all.
That's the part worth sitting with. A tile-by-tile search is bounded by what the analyst already suspects is worth looking at, because searching everywhere by eye isn't possible, and you search where you expect to find something. A similarity query carries no such expectation. It compares shapes against the whole archive at once, which is exactly why the shape sometimes turns up in a country that was never on anyone's list to check.
None of this replaces judgment. The false positive in that story still needed a human to look at the imagery, check the context, rule it out. What changed was where the human's time went: instead of spending the afternoon generating candidates, the analyst spent it evaluating them, which is a much better use of six hours.
I bring this up because calling a tool like this "faster search" undersells what changes on the ground. The first pass gets wider on its own, without the analyst deciding in advance where to look. That frees the eye for judgment calls: weighing context, ruling things in or out, instead of scanning hundreds of tiles for a shape it has half-forgotten. Broad Area Search's early-access version of this approach is built around exactly that one-tile, wide-area query, for analysts who are tired of being the only thing standing between a known site and every place on earth that looks like it.
If an afternoon like that sounds familiar, it's worth seeing what one query tile turns up in your own search area.