We actually use the twitter detections to launch analyses of the seismic data in order to get confirmed results for events that aren’t reported yet [1] but there are some statistics for the twitter detections in the supplementary material of that article [2]. Basically, in 2016-2017 (wow, so long ago), we detected 893 earthquakes via twitter, with a median delay of 67s and a median separation from the published epicentre of 94km. Note that estimating earthquake epicentres is nontrivial anyway and so, for comparison, 10km accuracy would often be considered ok. So the twitter, I mean X, method isn’t optimal but it gets you down to the right region. Partly it’s because geocoding the tweets is inaccurate and partly it’s because people live clumped together in cities rather than smoothly spread over the surface of the earth.
There is also the Earthquake Network (EQN) app that works on very similar principals to Google’s system - phones monitor their MEMs accelerometers, when they are left charging with their screen off, and when enough neighbouring phones detect vibrations simultaneously an earthquake is detected and apps nearby are alerted. It’s been running since 2012.
The USGS created a system to do exactly this about 15 years ago. I’m not sure whether they’re still running it but at the EMSC, we've been running a similar system for many years to highlight earthquakes important to the public and improve our messaging. Twitter doesn’t give access to geotags anymore but we do manage to roughly estimate an earthquake’s location by analysis of the tweets. Estimating magnitude is much more difficult. Naturally there are some false positives but it works well overall.