The State of McKinsey Solve 2026: What 26,622 Practice Runs Reveal

1. First run to best run: scores move, and Red Rock moves most

For every user with at least two runs of the same game, we compared their first completed run with their best. The median Red Rock user gained 20.3 percentage points (95% CI 18.9–21.6), moving from 61.2% to 81.5%. Sea Wolf users gained least — 6.7 points (95% CI 5.8–7.6), from 71.4% to 78.1% — which fits a game with one scored decision per site and far less room to recover a bad start.

How much of that gain is real?

"Best minus first" flatters itself: the maximum of k noisy draws rises with k even when nothing is learned. So we ran two counterweights. The first is the plain first→second gain, which has no max-of-k selection in it. The second is an order-permutation placebo: each user's own runs are shuffled 10,000 times and the same statistic recomputed. Shuffling does not move a user's maximum, so what the placebo measures is best minus a randomly positioned run — the gain you would see from within-user variance alone, with no learning in it. The difference between the headline and the placebo is the part that is not explained by noise.

The net column is a plain difference of two medians and is reported as descriptive: unlike every other estimate on this page it carries no bootstrap interval, because the difference of two medians computed on overlapping resamples needs its own resampling loop (query block 10) rather than the one behind the headline CIs. Read it as a direction and a rough magnitude, not a precise estimate. Roughly half of the headline gain in every game is regression to the mean. The surviving movement — +11.4 pp in Red Rock, +5.7 in the SFL, +3.3 in Sea Wolf — is the number worth quoting, and it is closely tracked by the independent first→second measure, which is what you would expect if it is real. Even so, this is not proof that practice causes improvement: candidates who come back are self-selected, and part of the remainder is familiarity with our scenario generator rather than transferable skill. Treat the headline figures as an upper bound and the net column as a floor.

2. The most-failed component in Solve is consistency, not maths

Aggregating every scored sub-component across all runs, one number dwarfs the rest. In the Sustainable Futures Lab, candidates lose 60.7% (95% CI 59.4–62.0) of the available consistency points — the score for whether your later decisions stay coherent with your earlier ones. Their judgment on any individual scenario is far stronger (33.5% lost), and their stakeholder management stronger still (24.5%). It is holding a line across thirteen linked questions that breaks them.

The Red Rock Study is much flatter: report (41.2%), analysis (38.5%), chart building (30.4%) and cases (27.9%) sit inside a thirteen-point band, and the two heaviest components are also the two worst. There is no single weak phase to drill — the losses are spread, which in practice means timing, not topic, is the binding constraint.

Sea Wolf has no phases — it is three site submissions — so we counted errors per submitted site instead:

• 33.5%An attribute average outside the site's target range
• 12.6%An undesirable trait included in the selection
• 9.8%The required desirable trait missing entirely

Trait errors are the rarer failure because they are binary and candidates learn the rule fast. Range errors are roughly three times as common, because they require averaging three microbes across three attributes under a shared 30-minute clock. Full component list:

Red Rock's Investigation phase is not separately scored — it feeds the Analysis and Report phases — so it is excluded rather than reported as zero. The SFL list is complete: its four components (judgment 65, consistency 20, ranking 18, stakeholder 15) sum to the full 118-point maximum.

Does the component table reconcile with the scores?

Component losses and headline scores are two views of the same points, so they have to agree. Weighting each component by its share of the available points gives an implied mean total per game; here it is against the pooled median of every run in that game.

The implied mean sits one to two points below the median in each game, which is the signature of a left tail: abandoned and heavily timed-out runs drag the mean down without moving the middle. Sea Wolf's implied mean is derived from its per-site error rates and the 100-point-per-site deduction model.

3. Most candidates practise two or three times, then stop

Across all three simulators, 4,261 of 9,004 game-user pairs are a single run — one attempt, never repeated. Returning users cluster at two or three runs. Only a tail (681 pairs) reaches seven or more, and that tail generates a disproportionate share of all sessions: 283 Red Rock users account for 3,538 of its 11,063 runs.

Median score by volume bucket tells a more careful story than the improvement finding does. All three games trend mildly upward with volume — Sea Wolf 68.9% → 80.1%, Red Rock 63.5% → 74.6%, SFL 57.9% → 65.4% — but the gap between the one-run and seven-plus buckets is smaller than the first-to-best gains, because the high-volume bucket is also where the candidates who found the game hard end up.

These bucket medians are unweighted descriptive medians of every run in the bucket, on the same user counts shown in the table above; they carry no confidence interval because they are not used to support an inferential claim. This is a survivorship pattern, not a dose-response curve. Read it as a description of who practises how much, not as evidence that the fourth run buys you points.

4. Practice is overwhelmingly last-minute

3,622 of the 5,174 candidates in the study (70%) stated a test date, and 3,401 of them completed at least one run before that date. The remaining 221 first practised after their stated date — a rescheduled or mistyped date, most likely — and are excluded from the buckets rather than folded into the last one. All shares below are therefore on a base of 3,401. Grouping each candidate by when they first practised, 2,392 (70.3%) did not start until the final 48 hours — 66.0% of everyone who stated a date at all. Another 19.0% started in the two-to-four-day window, and just 10.7% began more than four days out. The median lead time between a candidate's very first run and their test day is one day.

Buckets are mutually exclusive and sum to 3,401 candidates and 20,402 runs. The unit is the candidate's first run: a candidate who starts four days out and practises again the night before counts once, in the 4+ days row. The runs column counts every run by the candidates in that row, including runs after their test date — it is a measure of how much that cohort practises in total, not of pre-test volume.

Note the structural bias here: we ask for a test date at signup, and people typically sign up because a test is imminent. So this finding describes when our candidates practise, not the ideal preparation window. Read by candidate, only 363 of the 3,401 (10.7%) get started more than four days out — so beginning a week ahead already puts you ahead of roughly nine candidates in ten. Our prep plan is built around that reality.

References

External sources used for the structure and naming of the assessment. None of them supplied scores: McKinsey publishes neither its scoring model nor candidate results, so every score in this study is a SolvePrep simulator score.

1. McKinsey & Company — Our application process (Solve assessment)The firm's own description of where Solve sits in the application process. Used for the game names and the position of the assessment in the funnel — not for scoring.
2. McKinsey & Company — Solve, our problem-solving gamePublic description of the assessment format. McKinsey publishes no scoring model or candidate results, which is why every score in this study is a SolvePrep simulator score.
3. SolvePrep — McKinsey Solve guideOur own structural documentation of each game, reviewed by an ex-McKinsey subject-matter expert. Defines the components used in Finding 2.