ServeTell

Scout a matchup

Pick a server and a returner. Every number is shrunk toward the tour norm by how much data backs it, so a hot streak in a few matches can't fool you.

Serve forecast

Where to aim the first serve

Seen from 's side of the net. Big number: chance wins the point aiming there against this returner (model, ±90% range). "Raw vs all" is their plain win rate against every opponent, the kind of split a broadcast shows. Bar: how often they go there now.

The returner's read

Where 's next first serve to the same court tends to go, given the last one. Shaded cells move at least 4 points away from their usual mix.

Under pressure

First-serve mix on break points against , vs every other point.

Second serve check

What 2.5 million serves say

A player's "best spot" is mostly noise

We took every player's first-serve win rate going T vs. going wide in matches through 2024, then checked it against what happened in 2025–26. If those splits were real, the dots would sit on the dashed diagonal. The best-fit line is nearly flat: about of an ATP split carries forward, and for WTA. Simply assuming "no difference" forecast 2025–26 better than the raw splits did.

That is what game theory predicts. If one spot paid better, returners would cheat toward it until it stopped paying. Walker & Wooders found pros sit at this balance in 10 Wimbledon finals. Testing T vs. wide inside each of player-court splits, come out "significant", about the 5% pure chance would flag. After correcting for testing that many players at once, not one split survives.

For coaches: don't rebuild a serve plan around a split from a handful of matches. ServeTell's numbers are shrunk by how much evidence backs them.

Each dot is one player on one court (deuce or ad) with 30+ T and 30+ wide first serves in 2025–26. Lines: what you'd expect if splits were all signal (dashed) vs. what happened (solid).

One pattern is real: body first serves lose

Across both tours, first serves aimed at the body win about fewer points per 100 than serves to the T or out wide, on both courts. It is the one target gap that shows up the same way on both tours and both courts, with hundreds of thousands of serves behind it.

For servers: keep the body serve as a surprise, not a staple. Caveat: charters log where the ball lands, so a mis-hit wide or T serve can be recorded as "body". Part of this gap is mis-hits.

Share of points won by the server when the first serve is aimed at each spot (faults count, so a missed first serve leads to a second-serve point).

The real tell is timing: pros switch too much

A truly random server would hit the same spot twice in a row at a predictable rate. Real pros repeat less. of WTA and of ATP servers switch significantly more than chance (permutation test inside each match). Across the tours that is repeats vs. expected (WTA) and vs. (ATP).

The tell holds up out of sample. Learning each server's habits from matches through 2024 and guessing their 2025–26 serves, using the last serve lifts the hit rate from to (ATP) and from to (WTA).

Not everyone falls for it. The ATP's strongest repeaters are , and across ATP servers, repeating more goes with a stronger first serve (r = ). The WTA shows no such link (r = ), so treat it as a pattern worth testing, not proof.

For returners: after a serve to one spot, shade toward the other. For servers: going back to the same spot is less predictable than it feels.

Switching score for every server with 300+ back-to-back first serves to the same court. A server who picks at random lands under the dashed bell curve. Real servers sit to its left: they switch more than chance. Below −2 is significant.

Break points don't change the mix

We compared each server's mix on break points to their own mix on every other point. On break points they go to their favorite spot of the time vs. expected (ATP), and vs. (WTA). No tour-wide pressure tell. What does change is the result: servers win about fewer break points than other points, whichever way they aim.

For returners: don't guess "they'll go to their favorite" on big points. Check the individual player in the scout above, but samples there are small.

First-serve points won by the server, break points vs. all other points (tiebreaks excluded).

Keep the second serve a second serve

Should a player just hit two first serves? That pays when (first serves in) × (points won when in) beats their second-serve win rate. For the median player it costs (ATP) and (WTA). Only of regular servers would gain, and just clears statistical significance.

For coaches: the safe second serve is right for almost everyone. If you coach a big server, check them in the scout before trying it.

Each dot is a player with 400+ charted second serves. Above the line = would win more by going big on the second serve.

How it works

Data

The Match Charting Project: volunteer, shot-by-shot charts of pro matches ( ATP, WTA, ). We parsed every point into serve attempts. For each one we know first or second serve, the target (wide, body, T), the court (deuce or ad, worked out from the score), break-point status, surface, both players, and who won the point.

Model

A penalized logistic regression predicts the chance the server wins the point. It combines a tour baseline (serve × court × target × surface) with each server's and each returner's overall strength, their strength by serve, and their target-specific quirks. The penalty pulls every player effect toward zero unless the data insists, the same as a random-effects model. Penalties were tuned on 2024 matches. Target-specific quirks needed the heaviest penalty (λ = ), which is finding 01 again.

Validation

We trained on matches through 2024 and scored on ATP and WTA matches from 2025–26 that the model never saw.

Point outcomes are close to coin flips, so log-loss gains look small. They are steady across tours and splits. Hold % comes from the standard game formula applied to the forecast serve-point win rate.

Calibration on 2025–26 matches

Serves grouped into tenths by forecast. Points on the diagonal mean "a 60% forecast wins 60% of the time".

Limitations

Charted matches lean toward big events and well-known players. Lower-ranked players have fewer matches, so their numbers are pulled harder toward the tour norm.
Targets are recorded where the ball lands, not where it was aimed.
The switching test shuffles serves within each match. A server who changes plan mid-match (say, attacking one weakness for a set) shows up as a "repeater", so repeat counts mix habit with in-match adjustment.
The direction tests are observational. Players pick targets based on things we can't see (the returner's position, wind, fatigue).
Our "two first serves" check assumes a first serve keeps its usual accuracy when hit as a second serve. Under pressure it likely won't, so the real gain is even smaller.