> I wonder how far apart railway signals usually are
It varies substantially across the network, dependent on the mode of operation of the signalling, the desired headway, the maximum operating speed, the service braking distance of the rolling stock at line speed, factors in the layout that might influence safety (e.g. junctions), the number of colour lights used on signal heads (i.e. 2, 3 or 4 colour lights), signal sighting distance, whether signal visibility might be affected by sun glare, etc. And of course this assumes that the area is actually using line-side signals – most of the network does, but there's notable exceptions using in-cab signalling or computer-based train control.
> suggests the data mostly comes from railway signalling information, plus a bit of "AI" in some way
I'm in no way affiliated with the website, but in areas that follow track circuit block principles (much of the heavily-used main GB railway network, excluding many more lightly-used outlying lines), the "signalling information" that article indicates the site derives from is almost certainly Network Rail's Train Describer (TD) feed. This feed reports the headcodes for signalling berths across the layout. The signalling system will normally step headcodes automatically as the train activates successive track circuits or operates axle counters in the track.
The Train Describer can only report headcodes in berths, and berths might be quite long! It certainly doesn't provide second-by-second progress of trains, so this site's mapping engine is likely doing some proprietary interpolation to make the train position indicators appear to "move" in real time to give the illusion of trains making progress. (Whatever the inputs to their algorithm/model are, their calling it "AI" loses all technical specificity.) This may be based on line speed or perhaps the observed average time a headcode normally remains in a given berth, allowing them to derive a typical 'average speed' for that berth or section.
In other areas of the country, train position may be reported by GPS on equipped units, or in some areas, the passage of trains is dependent on manual reports by the signaller, so the map may not have much real-time data to infer train position from. I find it highly unlikely that they're deriving much accurate data in real time from smartphone apps, but it could be a (noisy, incomplete) set of inputs to improve their model of how trains typically make progress through particular berths.