Because each position is calculated from public orbital data with a mathematical model, not reported by the satellite itself, so it is a prediction rather than a measurement.
Every position label on this site says estimated, and the footer of each tracker page says Positions are estimates, not live GPS. That is not modesty — it is an exact description of how the number is made.
Satellites do not broadcast their positions to the public. What is public is the orbital data that ground radars and telescopes produce: the US Space Surveillance Network tracks objects, the 18th Space Defense Squadron fits orbits to those observations and publishes them as element sets, and CelesTrak republishes them. This site downloads those TLEs and runs the SGP4 model in your browser to work out where each satellite should be right now.
A kilometre or two is tiny compared with what you need. From the ground in Delhi, the ISS at around 420 km height is still a single bright point; a position error of a few kilometres moves it by well under a degree in your sky. For seeing a pass, finding which way to face or watching the ground track cross India, an SGP4 estimate is more than good enough.
Where precision does matter — docking, collision avoidance, steering imaging satellites — operators use their own tracking data and more detailed models. This site is honest about which kind of number it shows.
The page footer reads ‘Data: CelesTrak TLE / SGP4 propagation · Positions are estimates, not live GPS’. Open the ISS tracker
Error in a TLE-based prediction has two sources: the error in the elements at epoch (the orbit-determination fit) and the error the model adds as it runs forward. Vallado and co-authors, revisiting Spacetrack Report No. 3, note that propagator accuracy is ‘generally in the kilometer-level range’, and also that TLEs do not say which version of SGP4 produced them, so mixing versions can occasionally make results differ far more.
The BSTAR drag term illustrates the second point. The same paper explains that least-squares orbit fitting often gives a BSTAR that has ‘soaked up’ other force-model errors, and that negative BSTAR values sometimes appear, which would mean energy being added — physically wrong, but a by-product of SGP4's limited force model. BSTAR is a fitting parameter, not a measured property of the satellite.
How much does a kilometre matter for a sky-watcher? Angular error ≈ position error / range. For the ISS seen at 1,000 km range, a 2 km along-track error is 2 / 1,000 rad ≈ 0.11° in your sky. In time, at 7.66 km/s, 2 km means the station arrives about a quarter of a second early or late. Both are far below what the eye can notice, which is why pass times on this site are given to the minute.
The station's own navigation data is not broadcast to the public. Public trackers calculate the position from published orbital elements, so they show an estimate.
About 2 / 7.6 ≈ 0.26 seconds — the satellite arrives a quarter of a second later than predicted.