Every NFL pick since 2017, against the player they passed on

All 32 teams, measured four different ways against the players who were still on the board. Buffalo comes out on top, and it is the only team whose margin survives a test against chance.

The board, best to worst

Order comes from a team's average rank across the four comparisons below. Gold is a team whose picks gained career value over the alternatives; red is a team that lost it.

The rankings

Each number is career value gained or lost per pick, in wAV, against the alternative. Records read as team better–alternative better–even. Click a column to sort by it.

What this table is not. It is not a draft success analysis and it does not measure how much talent a team accumulated. It measures one thing: whether a pick outperformed the specific player taken right after him. A team can rank well here and still have a bare roster, because beating your neighbor on the board is not the same as landing a star. Washington is the clearest case, and most of the gaps in the middle of this table are not distinguishable from chance. Readers have since found real flaws in the method, and what happens when you fix them is set out at the end.

2017–2023 drafts for the first two comparisons, 2019–2023 for the two ESPN comparisons. Read the order with care: most of the gaps in the middle of this table are not distinguishable from chance. See how much of this is noise.

How this was measured

For every pick I asked the same four questions, then totaled the answers by team.

1. The next pick
Was this player better than the one taken immediately after him?
2. The next player at the same position
Better than the next player drafted at his position, even when that player went a round or two later?
3. ESPN's best available
Better than the highest-ranked player still sitting on ESPN's Scouts Inc. board at that moment?
4. ESPN's best available at the same position
The same question, limited to the position the team was drafting. This is the fairest of the four, because teams draft for need rather than straight down a board.

Better means a higher career wAV, Pro Football Reference's weighted Approximate Value, which condenses a career into one number that works across positions. Pairs within 2 wAV count as even. Career stats run through the 2025 season.

What the numbers say

Buffalo isn't close

The Bills went 28–11–12 against the next pick and 30–9–11 against the next player at the position, a surplus of +571 wAV in career value over the players other teams took in the same spots. It comes from hitting repeatedly rather than from one star: Tre'Davious White, Matt Milano, Dawson Knox, Ed Oliver, Christian Benford, James Cook and Terrel Bernard all outplayed comparable players drafted after them.

Kansas City took the right player off the board

The Chiefs rank first against ESPN's best available at +4.94 per pick. Patrick Mahomes at tenth in 2017, one pick ahead of Marshon Lattimore, is worth +72 on his own, and Creed Humphrey and Trey Smith are two of the largest late-round wins anywhere in the data.

Detroit's recent run is the best in football

The Lions lead everyone against the best available at the position, +5.85 per pick, almost entirely on the strength of 2021 through 2023: Penei Sewell, Amon-Ra St. Brown, Aidan Hutchinson, Jahmyr Gibbs and Brian Branch.

Miami is the most lopsided team here

The Dolphins finish last in both position comparisons at −5.48 per pick, going 10–26–6 against the next player at the position. Both of their marquee offensive picks were followed closely by better players at the same spot: Tua Tagovailoa by −22 and Jaylen Waddle by −23.

New England and Las Vegas bring up the rear

The Patriots went 18–27–18 against the next pick, −256 wAV, with Sony Michel one selection ahead of Lamar Jackson as the single worst pick of the group. The Raiders finish last overall at 17–28–11; Maxx Crosby is the one big hit, while Henry Ruggs at twelfth in 2020, immediately before Justin Jefferson, costs them −56.

Both of those are descriptions of what happened, not findings about how these teams draft. Neither margin survives the test in the next section: the Raiders' last-place figure is the single least significant result in the table, and would be matched by chance about half the time.

The largest single swings

Picks that beat the very next selection

  1. T.J. Watt, Pittsburgh, 2017 at 30, over Reuben Foster: +88
  2. Lamar Jackson, Baltimore, 2018 at 32, over Austin Corbett: +81
  3. Christian McCaffrey, Carolina, 2017 at 8, over John Ross: +79
  4. Patrick Mahomes, Kansas City, 2017 at 10, over Marshon Lattimore: +72
  5. Fred Warner, San Francisco, 2018 at 70, over Royce Freeman: +69

Picks the very next selection beat

  1. John Ross, Cincinnati, 2017 at 9, one pick before Mahomes: −108
  2. Sony Michel, New England, 2018 at 31, one pick before Lamar Jackson: −89
  3. David Njoku, Cleveland, 2017 at 29, one pick before T.J. Watt: −69
  4. Dylan Donahue, New York Jets, 2017 at 181, one pick before Aaron Jones: −67
  5. Cam Akers, Los Angeles Rams, 2020 at 52, one pick before Jalen Hurts: −66

Team by team: click your team

All 32 teams are below, in alphabetical order. Click any one of them to see its record in all four comparisons, its largest swings against the very next selection, and a read on where its drafting has gone right and wrong.

Every one of the 2,312 picks behind these numbers is in the full spreadsheet, alongside each comparison and the method.

A note on why I only looked at the very next pick

The counterfactual problem, and what each of the four comparisons does about it.

A draft is a chain reaction. The moment a team takes a different player, every pick after it changes. Different players come off the board, teams' needs shift, and the guy you wanted eight picks later might have been snapped up at pick three instead. So any "they should have taken X" that reaches deep into the round is quietly assuming that 30 other front offices would have behaved exactly the same way in a draft that no longer exists.

Comparing against the very next pick is the smallest possible change to history: one swap, and the rest of the draft is largely undisturbed. It's the cleanest counterfactual available.

It's also the honest one. If I went looking 8 or 9 picks down, I could find a good player after almost any selection ever made and call it a miss.

That's hindsight with a search function, not analysis.

Relaxing it, carefully

The same-position comparison

My attempt to loosen the rule a little without losing the plot. It lets the alternative be a round or two later, but only at the position the team was actually drafting, so it stays a decision they were plausibly weighing in the room.

Asking it the right way around

The ESPN best-available comparisons

These handle the "who else was there?" question from the correct direction. They use a ranking written before the draft, so the alternative is a player the league itself had rated higher at that moment, not one we know turned out well.

How much of this is noise?

The obvious objection to all of the above, tested rather than argued.

A reader called this junk science, on the grounds that comparing a pick to whoever happened to be taken next introduces so much randomness that the results stop meaning anything. That is a fair challenge and it has an answer, so here it is.

The raw noise is enormous, exactly as the objection says. The margin on a single pick has a standard deviation of 17.6 wAV, rising to 26.4 in the first round, and individual pairs run from −108 to +88. Nobody should read a single pick comparison as a judgment on a front office.

The question is whether that noise cancels out across the 41 to 63 picks each team made. To find out I ran a permutation test: shuffle which picks belong to which team 10,000 times, holding each team's round-by-round profile fixed so draft capital stays controlled, and count how often chance alone produces a result as extreme as the real one.

The spread across all 32

p = 0.26

Not distinguishable from chance. If team identity carried no information at all, you would see a table this spread out about a quarter of the time. The ordered list is a description, not a ranking anyone should defend position by position.

Buffalo at the top

p = 0.056

Borderline, and just short of the conventional line, after accounting for the fact that we are looking at the best of 32 teams. On its own 51 picks Buffalo's margin has a 95% interval of +2.8 to +11.8, which clears zero. The strongest result here, and still not a comfortable one.

Las Vegas at the bottom

p = 0.51

Nothing at all. A last-place finish this deep would happen by chance about half the time. This is why the claim that Las Vegas drafts worse than anyone has been removed from this piece.

Bootstrapping each team against its own picks says the same thing from another angle: only two of the 32 teams, Buffalo and Las Vegas, have a 95% interval that excludes zero, and the Las Vegas one only barely. Everything between second and thirty-first is statistically indistinguishable from everything else between second and thirty-first.

The bigger limit is structural, and it is not about sample size. This comparison is zero-sum: if my pick beats the next pick, that team loses by construction, so the league average is pinned near zero. It measures how a team did against its immediate neighbors on the board, not how much talent it came away with. Those are different questions, and they can point in opposite directions.

Washington is the cleanest example in the dataset, and it is worth spelling out because it cuts against my own table. The Commanders rank 9th against the next pick. On the absolute production of those same 61 picks they rank 29th of 32, at 11.4 wAV per pick against a league average of 13.6, with 7 Pro Bowl selections total, 28th in the league. That gap of 20 places is the largest of any team. Both numbers are correct. They answer different questions, and the second one is closer to what fans mean by drafting well.

Across all 32 teams the two rankings correlate at rho = 0.63, so they are related but far from the same measure. If what you want to know is which teams stocked their roster, rank by absolute value produced, not by this.

So what is left standing? The per-pick records are facts, and so are the individual swings: T.J. Watt really was taken one pick before Reuben Foster, and John Ross really did go one pick before Patrick Mahomes. Those need no inference. Buffalo's run is the one team-level result that holds up to scrutiny, and even that is borderline. The rest of the table is best read as a description of what happened rather than a measurement of how well each front office drafts.

That is a narrower claim than the one I started with, and the objection deserves the credit for it.

Four objections, and what happens when you fix them

Readers named specific flaws in the method. Each one is testable, so I tested them.

After this went up, a reader worked through the methodology and sent a list of concrete problems with it rather than a general complaint. Three of the four could be checked directly against the data. All three were right, and one of them changes the answer enough that it belongs on this page rather than in a comment thread.

1. The method cannot see the last pick of any draft

Every pick here is measured against the player taken immediately after him. The final pick of each draft has no one after him, so he is dropped. That is seven picks across the 2017–2023 window, and six of the seven are inconsequential: Chad Kelly, Trey Quinn, Caleb Wilson, Tae Crowder, Grant Stuard, Desjuan Johnson.

The seventh is Brock Purdy, pick 262 in 2022, 44 wAV. San Francisco receives no credit for him in the next-pick comparison. The one pick this method structurally cannot see is the best value selection in the dataset.

2. Older draft classes quietly dominate the average

I had assumed career length cancels out, since a pick and the player taken after him have had exactly the same number of seasons to accumulate value. Within a pair it does cancel. Across classes it does not, because an eight-year-old class has far more room to spread out than a two-year-old one.

The 2017 class carries about three times the variance of the 2023 class. A team average across seven drafts is therefore closer to a verdict on 2017 and 2018 wearing a seven-year label. Dividing each margin by seasons available corrects for it.

Standard deviation of the per-pick margin, by draft year (wAV)

201726.5
201821.0
201918.2
202016.5
202114.6
202211.4
20238.9

3. Nothing adjusted for where a team was picking

A selection at 5 and a selection at 205 were being scored on the same scale. The suggested fix was a draft-position value curve. Rather than import one, I fit the expectation directly from these picks:

expected wAV per season = 9.658 − 1.605 × ln(pick)

That values the first pick at about 9.7 wAV per season and pick 250 at about 0.8. Scoring every selection as actual minus expected-for-slot produces a ranking that correlates with the published one at rho = 0.77, so it is a genuinely different measure rather than a rescaling of the same one.

4. Dwayne Haskins

He remains in the dataset, because removing a row to make a number look better is not something I am willing to do. But he had been used as a headline example of a bad pick, and he died in 2022. That framing is gone from this piece.

All three methods, side by side

Each team scored three ways: the published margin against the next pick, that same margin normalized per season available, and value over the expectation for the draft slot. The final column is how far a team moves between the first method and the third.

2017–2023 drafts. Sorted by the published ranking. Rank columns run 1 (best) to 32 (worst).
Team vs next pick# Per season# Vs slot expectation# Move
Bills+7.241+1.1481+1.08910
Ravens+4.772+0.7102+0.6583+1
Saints+4.463+0.5353+0.17412+9
Chiefs+2.874+0.3925+0.6842-2
Chargers+2.845+0.4774+0.3046+1
Bears+2.196+0.2839+0.13214+8
Cowboys+1.527-0.00619+0.27870
Eagles+1.518+0.23710+0.20810+2
Commanders+1.209+0.00717-0.38226+17
Lions+1.0210+0.3886+0.13713+3
Packers+1.0111+0.3537+0.2388-3
Rams+0.9112+0.19912+0.4804-8
Steelers+0.7413+0.09515+0.2239-4
Falcons+0.7314+0.12314-0.16220+6
Dolphins+0.6415-0.00118-0.09218+3
Jaguars+0.6016+0.3068-0.20922+6
Buccaneers+0.5917+0.13013+0.3275-12
Cardinals+0.5518+0.22811-0.37825+7
49ers-0.1219-0.08720+0.19411-8
Vikings-0.2620-0.13223-0.21623+3
Colts-0.4021-0.23525+0.04416-5
Giants-1.0022-0.31227-0.39328+6
Texans-1.1823+0.06916+0.04715-8
Titans-1.2124-0.67431-0.39127+3
Bengals-1.4425-0.09921-0.13919-6
Browns-1.7226-0.23224-0.51430+4
Broncos-2.0927-0.29426-0.16321-6
Seahawks-2.3928-0.11322+0.02117-11
Panthers-2.7829-0.59830-0.484290
Jets-3.5030-0.47628-0.63831+1
Patriots-4.0631-0.55729-0.29324-7
Raiders-4.6132-0.73832-0.678320

What survives all three. Buffalo ranks 1st and Las Vegas 32nd under every one of them. The tails are robust to the choice of method, which is more than can be said for the middle of the table: Washington falls 17 places, Tampa Bay climbs 12, Seattle climbs 11.

Washington is worth dwelling on, because three independent routes converge there. The absolute production figures put them 29th. A thread of Commanders fans said the 9th-place finish was self-evidently wrong. And this slot model, built with no knowledge of either, drops them to 26th. When the eye test, the raw output and a corrected method all point the same way against my published table, the published table is the thing that is wrong.

One caution against over-reading the stability. Robustness to specification is not robustness to chance. Buffalo and Las Vegas hold their positions across every way of computing this, and the permutation test in the previous section still puts the Las Vegas result at p = 0.51. A number can be stable across every method you try and still be something randomness produces half the time. Both statements are true, and neither one rescues the other.

The reader who supplied these objections is not named here because I have not asked permission. The analysis is better for them.