The number
Through four weeks, the recruiting gap between two FBS teams explains 32% of the final margin. From week 5 on, it explains 5%.
That's 2,501 completed FBS-vs-FBS games from 2023 through week 3 of 2026, every one where we hold a talent composite for both sides. Correlation of the talent gap with the actual margin runs 0.569 in weeks 1–4 (811 games), falls to 0.233 in weeks 5–9 (820 games), and sits at 0.264 from week 10 on (870 games). Square those and you get the share of the scoreboard recruiting can account for: 32.4%, then 5.4%, then 7.0%.
The usual objection is range restriction — conference play pits similar rosters against each other, so a thinner spread of talent gaps would drag the correlation down on its own. Fair, and it isn't the whole story. The slope moves too. A 100-point edge in the composite is worth 5.33 points of margin in September and 2.81 in October. The relationship doesn't just get harder to see. It gets smaller.
Ask the November question instead
Margin in a single game is the wrong thing to ask about anyway. The question that actually comes up on a broadcast in November is about the rest of the season: this team out-recruited that one, so back them the rest of the way.
So we measured that directly. After each week of the season, take everything a team has done so far, and separately take its recruiting composite, and ask which one better predicts the win rate it goes on to post over its remaining games.
Recruiting never wins. Not once, at any week. After week 7, recruiting explains 1.7% of a team's remaining win rate across 414 team-seasons; the record it has already put on the field explains 20.2%. After week 10 the gap is 0.9% against 16.6%, over 405 team-seasons. Twelve-to-eighteen times more signal in the games than in the stars, and that ratio is roughly stable from the moment a team has three results to its name.
0.9% is the number worth carrying into a playoff argument
By week 10 the committee is meeting and the rankings talk starts. This is where the recruiting argument gets made most often and where it holds up worst.
After week 10, a team's recruiting composite accounts for 0.9% of the variance in how it finishes. Not zero — we're not claiming zero, and a 405-team sample over three and a half seasons can't support a claim that clean. But 0.9% is indistinguishable from noise in any practical sense. If two teams are being compared in November and the tiebreaker offered is that one of them signed a better class, that tiebreaker is carrying less than one percent of the available information while the other ninety-nine sits right there in the results.
The honest version of the recruiting argument is a September argument. It's a real one then — 32% of a game's margin is a lot of explanatory power for a single number. It has simply expired by the time anyone reaches for it.
Why it decays: the games answer the same question, better
Recruiting is a projection of how good a roster will be. Results are a measurement of how good it turned out to be. Early in the year you only have the projection, so it does real work. Every week after that, the measurement gets sharper and the projection gets more redundant.
That's not a theory we like — it's what happened when we tried to use recruiting in the model and couldn't. We ingest the composite for all 643 FBS team-seasons we cover and deliberately keep it out of the rating fit, because every weight we tested made the projections worse, monotonically. On a full-season walk-forward across 2024 and 2025, the correlation between our rating gap and actual margin ran 0.6713 with no recruiting prior at all, 0.6632 with a light one, 0.6459 at a moderate weight, and 0.5833 when recruiting was given real say.
What made that especially worth pinning down is that the regression which motivated the whole exercise looked great. Talent showed +0.0128 R² on margin in weeks 1–4, which is exactly the window where the measurement above says it should. Two things flattered it. A free additive regressor in an ordinary regression never has to outcompete the evidence already in the model — it just gets handed a coefficient. And the subset moved with the feature: a composite only exists for teams the source rates, which quietly dropped every FBS-vs-FCS game from the sample, the easiest blowouts on the board. The apparent edge didn't survive being made to compete.
So what do you do with it in November
Use the classes to set your priors in August and drop them the moment the games start arriving. Three results is enough for a team's own record to carry more information about its next month than its recruiting ever did.
In practice that means the November arguments worth having are about what a team has done — who it beat, by how much, on the road, against what. That's also how our college board is built: ratings fit on actual results, blended across two seasons, with the recruiting composite sitting on the team pages as context and nowhere near the projection. A team that signed the best class in the country and then went 5-4 is a 5-4 team, and the board prices it as one.
The exception is genuinely new information. A roster that turned over through the portal, a quarterback change, an injury that reshapes an offense — those move the measurement because they change what the games are telling you. A signing-day ranking from twenty months ago doesn't.
What this doesn't say
It doesn't say recruiting is worthless. Programs that sign top-ten classes year after year are the programs whose in-season results are good, which is precisely why the results carry the signal — recruiting is upstream of the thing we're measuring, not unrelated to it. Cause and information content are different questions and this only answers the second one.
It doesn't say a week-1 preview should ignore the classes. In September it's the best single number available, and our own board leans on last season's results the same way for the same reason.
And it's three and a half seasons: 2023 through week 3 of 2026, 2,501 rated FBS-vs-FBS games, one composite provider. The week-12 row in our season-ahead test has 34 teams in it and we've ignored it here, because a sample that size will say anything you want. The transfer portal is also actively rewriting what a recruiting class even means, and if that relationship moves, this measurement moves with it. We'll re-run it every year and publish whatever it says.
The measurement script is in the repo, the composites are ingested and on the team pages, and every game in the sample is one you can look up.
Sources
Every number here comes from the graded ledger the boards publish. Research from a statistical model, not betting advice; no outcome is guaranteed. Method: how MatchWiz works.