Why one slow client can drag everyone down. Clients take turns on the same channel, and a slow client's turns take much longer.
Change these to see how much they matter. The notes below explain where the defaults come from.
Wi-Fi clients take turns on the channel, and standard contention gives each one roughly the same chance to win a turn. It doesn't give them the same time. A client at a low rate needs far longer to send the same data, so while it transmits, everyone else waits. With equal chances, every client ends up with about the same throughput, pulled down towards the slowest one. This is known as the 802.11 performance anomaly.
With equal chances, the fast clients only get their next turn after the slow client's turn ends, so what hurts them is how long each slow turn lasts. A distant Wi-Fi 4/5/6 client at 6.5 Mbps bundles frames and keeps transmitting for up to the longest transmission time (5.484 ms), while an 802.11a/g client at 6 Mbps sends a single 1,500-byte frame and hands the channel back after about 2 ms. On its own the bundling client is the more efficient of the two, but next to fast clients it holds them back more. Lower the "Longest transmission" assumption to see the difference shrink.
With airtime fairness the AP shares out time instead of turns: each client gets the same share of the channel's time, and fast clients fit many turns into theirs. Fast clients get most of their speed back and the slow client gets what its share of time allows. The total goes up a lot. Many APs offer a setting like this for the traffic they send; how it works varies between vendors.
Each preset is a PHY rate for a stated configuration, calculated as in the MCS Index: for example Wi-Fi 6 with 2 streams at 80 MHz and MCS 11 is 1,201 Mbps. A client far from the AP drops to a lower MCS, so a slow client is often simply a distant one.
Every turn costs time before and after the data: waiting DIFS (34 µs), the average random backoff (7.5 slots × 9 µs = 67.5 µs), the preamble, SIFS (16 µs) and the acknowledgement (44 µs at 6 Mbps). These are the 5 GHz values used in the 802.11 Medium Access Simulator. Together with a preamble of a few tens of microseconds they come to about 200 µs, which is an approximation you can change above.
802.11n, ac and ax clients bundle many frames into one transmission (aggregation), here up to the data-per-turn setting. One transmission can last at most 5.484 ms in 802.11n/ac/ax, a limit set by a field in the legacy preamble, so a slow client sends less per turn. 802.11a/g clients can't aggregate at all: they send one 1,500-byte frame per turn, which makes them especially costly. 802.11b uses different timing and isn't modelled; see this article on 1 Mbps basic rates.
The throughput here is what the channel can carry for these clients with everything else perfect: no collisions (which grow with more clients), no retries, no other networks and no TCP or application overhead. Real throughput is lower, but the effect of a slow client is the same. For designing around airtime, see this article on airtime utilisation.