How to Get Better at Chess: What the Data Actually Shows

Type "how to get better at chess" into a search engine and the same five-item list comes back, in some order, on page after page: play more games, review your own games, solve puzzles, learn basic endgames, stop memorizing openings. Chess.com's own top-ranking page for this exact query opens with a close variant of that list. None of the five instructions is wrong. None of them is measured against what a rating actually does when a player follows it, and one instruction on that list, skip the openings, is flatly contradicted by every structured study plan sitting next to it on the same search results page.

Two things break the pattern open. A 2026 study tracked roughly 44,000 Chess.com players and measured rating gained per hour of each major activity, not just whether players who did more of something improved. And our own data, built from a Maia3 human move model and Stockfish scored against real opening theory, can answer a question none of those five-item lists even attempts: not what should a player do, but what should you do, given your actual rating.

What actually moves your rating, measured

Southwick and colleagues followed close to 44,000 players and logged rating gained per hour spent on each of four activities: gameplay, puzzles, reviewing your own games, and watching video lessons. Deliberate-practice activities carried a 3.61x learning-efficiency advantage over plain gameplay overall, and the per-activity breakdown is where the real surprise sits. Puzzles returned 0.73 rating points per hour, the least efficient of the four, with plain gameplay sitting just above them. Game review returned 4.41. Video lessons returned 5.21, roughly six to seven times what puzzles returned for the same hour spent.

Read that carefully before drawing the wrong conclusion. It does not say puzzles don't work. Puzzles are how tactical pattern recognition actually gets built, and the mechanism behind that training is real. What the study says is narrower and more useful: per hour spent, puzzles bought players the least rating of the four activities measured, landing just below plain gameplay rather than above it. If your goal is the fastest path to a higher number and you have one fixed hour, the same dataset says spend it reviewing a game you actually played or watching someone explain a position, not solving one more puzzle set. The same researchers found something worth keeping in view for later in this piece: a human coach was not required for these gains. Structured, self-directed activities produced them on their own.

A separate, older line of research points the same direction from a different angle. A study of adolescent tournament players found that serious study and serious competitive play each independently predicted performance over a player's career, using the kind of linear mixed-models analysis built to separate the two effects rather than blend them. The more useful detail sits in what happened to the players who quit competing: they had benefited from their deliberate practice to the same extent as the players who stayed active. The hours you already put in do not stop counting the day you stop entering tournaments.

The best move is closer to a coin flip than the opening pages admit

The "don't memorize openings" advice sits on the same search page as three separate structured study plans, each assigning openings real study time. ChessMood's plan for players under 1500 gives openings 15 percent of study time, alongside 70 percent on tactics and 15 percent on middlegame and endgame combined. A chess.com forum thread on the same 1000 to 1600 band proposes roughly 80 percent split between tactics and endgames combined against 20 percent for opening understanding, not memorization. Duolingo's blog publishes a third breakdown entirely: 25 percent tactics, 25 percent endgames, 20 percent middlegame, 20 percent general improvement, and only 10 percent openings. Three plans, three different numbers, each presented as a settled answer to the same question, ranking next to each other for the same search.

Our own data suggests why the openings argument keeps splitting three ways: the "best" opening move is very often not much better than the second choice, which makes memorizing it a low return on the hours it costs. We ran Stockfish 18 to depth 20 across 28 real opening positions, scoring every legal move, and found a median gap of just 9.5 centipawns between the best move and the second best. Fourteen of those 28 positions, exactly half, had a gap under 10 centipawns, closer to noise than to a real advantage. Which openings are actually findable by rating covers the full picture; the part that matters here is what happens as players get stronger. Across those same 28 positions, the rate at which a player's most likely move actually matches Stockfish's top choice falls from 67.86 percent at 1000 rating to 57.14 percent at 1800, and stays there through 2200. Stronger players agree with the engine's precise pick less often, not more, and it barely costs them: the count of positions where the human move gives up at least 50 centipawns drops from 3 at 1000 rating to 0 at both 2000 and 2200. Drilling a database line to find the exact top engine choice is chasing a gap that, by 2000 rating, no longer cost anything serious to miss in any of the 28 positions we scored.

What our own data says about "just do more puzzles"

Findability, our name for the probability a player at a given rating actually plays a specific move, lets us test the puzzle question a different way than a rating-per-hour study can. We ran Maia3 against the finish of the Opera Game, Morphy against the Duke of Brunswick and Count Isouard, Paris 1858, at seven rating bands from 1000 to 2200. Averaged across the real decision points in that famous forced sequence, findability barely moves with rating at all: 16.83 percent at 1000, 16.37 percent at 1500, only 17.65 percent by 2200. A well-known, heavily drilled combination like this one does not reliably get easier to spot just because a player's rating climbs. Be careful how far you stretch that, because our own opening data runs the other way. Across the 28 opening positions from the section above, mean findability climbs steadily with strength, from 35.23 percent at 1000 rating to 49.95 percent at 2200. So the honest reading is narrower than "recognition doesn't improve." Ordinary positional choices clearly do get easier to find as players improve. A queen sacrifice that has to be calculated all the way to mate stays hard at every rating we tested. That is worth knowing before you read a puzzle streak as evidence your tactics are solved, and it is a separate question from whether puzzle training pays per hour, which is the one Southwick's data answers and ours does not.

The same underlying model is what our repertoire scheduler runs on, and it is the most concrete answer we can give to "how do I actually retain what I study." Every opening line in Chessdrive is scheduled by ts-fsrs, and each card's own target retention is computed, not chosen, from priority = reachProb * (1 - humanProb), where humanProb is Maia's probability that a player at your rating finds that move on their own. A move you would likely find anyway sits near the bottom of our retention range, 0.80. A move that is common in your lines but unlikely for someone at your rating to find without review sits at the top, 0.97, and stays there. At 0.80 retention, three correct reviews in a row stretch the gap between them to 8 days, then 75, then 521. At 0.97, the same three reviews barely move: 3 days, then 3, then 6. That gap is deliberate. The moves you would find anyway earn long rest, and the moves you specifically wouldn't get reviewed on a much shorter leash for as long as they stay in your repertoire.

Where a coach, or a different tool, beats us

Southwick's data found a human coach was not required for the rating gains it measured; structured self-directed practice, game review and video lessons specifically, worked without one. That is not the same claim as "a coach adds nothing." A coach watching you play live can ask why you considered a move in the moment, and follow the answer wherever it leads. No automated review, ours included, can do that yet: Chessdrive can show you the move you missed and the moment a game turned, but it cannot ask you a follow-up question about what you were thinking and adjust based on the answer.

There is a second honest gap. Video lessons were the single most efficient activity in Southwick's data, at 5.21 rating points per hour, ahead of game review's 4.41. Chessdrive's own centerpiece leans on the activities the same study ranked lower, tuned puzzles and play at your level, alongside game review, which does land in the efficient half of the same dataset. If structured video instruction from a strong player is specifically what you're after, a dedicated course or channel built around that is the better tool for that slice of your study time than an app built around review, puzzles and repertoire drilling.

One more honest note, this one about the goalposts rather than the tools. Writing on rating plateaus argues that a stalled rating is often partly illusory: skills can improve well before a player reliably applies them under the pressure of a real game, and the rating pool itself gets stronger every year, quietly raising the bar under a player who is not actually standing still. The same source cites playing at least 50 blitz games a month, a little under two a day, as a volume threshold associated with real long-run gains. That threshold sits on the low-efficiency side of Southwick's ledger, with plain gameplay rather than with game review or video lessons, which is the actual tension worth naming out loud: raw game volume helps over a long enough horizon, but hour for hour, it is not where the fastest gains sit.

There's a related piece of advice worth stress-testing rather than repeating. Several sources recommend deliberately seeking out stronger opponents, the reasoning being that you can't coast past your own mistakes against someone better. That's reasonable on its own, but it sits awkwardly next to the volume finding above: the matchmaking pool that produces 50 blitz games a month is mostly players near your own strength, not stronger ones you went looking for. Nothing in the data reconciles the two, so treat "play up" as advice for specific, deliberately arranged games, not as a description of what your normal rated queue already gives you.

So what do you actually do

Not a fourth percentage split. The data doesn't support inventing one, and three incompatible splits already crowd the search results for this exact query. What it supports is a different ordering question: which hour buys the most rating, not which topic deserves what share of a plan. Southwick's numbers put game review and video lessons well ahead of puzzles and plain gameplay for that specific question. What a chess engine's evaluation number does and does not tell you matters here too, because reviewing a game only pays off if you're reading the right thing in it, not just an evaluation bar swinging between two numbers.

One rough map is worth keeping precisely because it isn't a percentage split. Improvement writing describes chess skill in coarse stages instead of numbers: first not hanging your own pieces, then tactics, fork, pin, discovered attack and the rest, then strategic play once blunders stop deciding most of your games. No rating number is attached to any of the three stages in that writing, and that omission is honest rather than lazy. Which stage is actually limiting you is a question about your own games, not one a ranking page can answer for you.

Endgames are the one place the SERP and our own numbers agree without much argument. A chess.com forum commenter called endgame weakness "quite common under 1800," and that lines up with a real, teachable gap rather than a preference. The five endgame ideas worth learning before 1500 covers exactly that ground. And tactics remain worth training, just not because doing more of them is the fastest lever available: they're how puzzle rating and game rating end up measuring two different things, and understanding that gap is part of what makes a puzzle habit worth keeping even at 0.73 rating points an hour.

None of this names the one thing that will move your rating starting today, because the honest answer depends on your own rating and your own games, not a universal list. What it does say is that the five-item list you already found before opening this page is missing its most important qualifier: how much of your next hour it will actually buy you. That number was never on any of those pages to begin with.

Published 22 September 2026. Last verified: 22 September 2026, against the sources cited above and against Chessdrive's own generated Stockfish, Maia3 and FSRS output described in this article. Forum sentiment and third party pages can change; confirm current details on each source's own page before relying on them.