How to set a similarity score threshold for a site search

You mark one site, run a similarity search against the rest of your archive, and get back a few thousand tiles with scores attached. Now you have to decide where the real matches stop and the noise starts. Picking that number feels arbitrary the first time you do it, and it stays a little arbitrary every time after, but there's a workable way to approach it.
Start with the known matches, not the score
Before you touch a slider, find two or three locations you already know resemble your query site. Maybe it's a second facility you confirmed last year, or a training example from an earlier tasking. Run the search and see where those known matches land in the ranked list and what score they get. That score is your anchor. If your two confirmed look-alikes come back at 0.81 and 0.76, a threshold of 0.9 will cut them out before you ever see them. Set the floor below your weakest known match, not above it.
Similarity scores from embedding models aren't calibrated the way a percentage match would be. A 0.7 on one query tile and a 0.7 on another aren't directly comparable if the two sites have different roof materials, shadow angles, or surrounding clutter. The score only means something relative to that specific query, so calibrate it fresh against that query's own known matches each time you run it.
Recall versus clutter
A loose threshold pulls in everything that's plausibly similar, including a lot of built-up areas, sports courts, and parking structures that share geometry with your site but nothing else. A tight threshold cuts the noise but risks dropping a real match that happens to sit at an odd angle, has partial cloud cover, or was built with a slightly different material. There's no setting that avoids both problems at once.
Running the search twice gets you further than guessing at one number up front. First pass, set the cutoff loose enough that you're confident nothing real gets excluded, then skim the list fast for obvious non-matches and discard them by eye. Second pass, once you've got a feel for where the true positives cluster on that run, tighten the retrieval cutoff for any follow-up query on the same site type. Over a few searches against the same target category, you build a working sense of where the signal actually sits, which is more useful than any fixed number you'd read off a chart.
Treat the cutoff as a sort order, not a gate
The more durable habit is to stop treating the threshold as a hard line at all. Rank the results by match confidence and work down the list in order, rather than drawing one cutoff and only looking above it. The tiles at 0.95 get a quick glance and a confirm or reject. The tiles between 0.7 and 0.85 get the closest look, because that's where most of the judgment calls sit. Below that band, confidence drops fast enough that it's usually faster to skim the remainder in bulk than to set a threshold and hope it lands in the right place.
This also protects you against a specific failure mode: a threshold set once and reused across different query sites. A compound with a distinctive roof shape and strong shadow contrast will produce a tight, high-confidence cluster of matches. A site that's smaller or visually closer to its surroundings will produce a flatter curve where good matches and false positives sit at similar scores. Carry the first threshold over to the second search and you'll either miss real sites or drown in clutter. Set it fresh, against the known matches for that query, every time.
Broad Area Search is built around exactly this workflow: mark one known site, get back every location across the search area that looks like it, ranked so the real work is reviewing a short ordered list instead of browsing an archive tile by tile.
If that sounds like the search you keep doing by eye, it's worth getting on the early access list.