I suspect the algorithm is just the frontloaded questions they ask you -- of course they can shrink all that down into a handful of bytes and unpack it on the other end as "car crash, 3 people, injuries reported, lat/long".

So the percentage "works", in a way, until you get to the freeform text.

Also, a 75% reduction (which I assume was meant by their 300 percent) is not that impressive when it comes to text compression. I'd guess that should be easily reachable with zstd by just creating a pre-shared dictionary generated from a bunch of typical emergency messages. Especially when those messages are partially auto-generated by a wizard-style questionnaire and will thus adhere to a previously-known structure and contain a lot of known elements and words.