Almost every 2048 guide gives the same handful of tips, and almost none of them show any evidence. That is a fixable problem: the tips are specific enough to be written as algorithms, and an algorithm can be made to play thousands of games and report what happened.
So that is what this page is. We turned the common advice into policies, ran 168 complete games under the same rules the board on this site uses, and compared the outcomes. One of the famous tips holds up extremely well. Another one, repeated everywhere, turns out to be a reliable way to lose.
| Policy | Games | Reached 1024 | Reached 2048 | Reached 4096 | Median score |
|---|---|---|---|---|---|
| Expectimax search | 4 | 0.0% | 100.0% | 25.0% | 38,788 |
| Same search, corner term off | 4 | 100.0% | 75.0% | 50.0% | 48,100 |
| Fixed priority: up → left → right → down | 80 | 0.0% | 0.0% | 0.0% | 2,504 |
| Random legal move | 80 | 0.0% | 0.0% | 0.0% | 1,066 |
Every policy plays by identical rules — same 4×4 board, same 90/10 split between spawned 2s and 4s, same win condition. Nothing separates those rows except the quality of the decisions, which is the useful thing to know: the spread between the top and bottom of that table is entirely skill, not luck.
Verdict: true, and it is the single most valuable habit you can build.
This one is testable directly. The solver scores every board it considers partly on whether the biggest tile sits in a corner. Switch that one term off, change nothing else — same search depth, same starting positions, same random seeds — and the win rate falls from 100.0% to 75.0%. Its median score falls from 38,788 to 48,100.
That is a clean measurement of one idea in isolation, and it is a large effect for a single heuristic. The reason is structural rather than mystical: a tile in a corner has two walls behind it, so only two of the four directions can dislodge it. A tile against an edge has one wall. A tile in the middle has none and will be shoved somewhere by every move you make. Corners are simply the only cells on the board where you can park something permanently.
The corresponding measurement in play: across its games the solver kept the largest tile in a corner on 88.7% of moves. A policy moving at random manages 11.5%, which is roughly what the board gives you by accident.
There is no preferred corner; the board is symmetric. What costs games is switching. Every corner implies a direction you must stop using, and rebuilding that structure in a different orientation takes many moves during which nothing is protected. Pick one at the start and treat it as permanent.
Verdict: false, and actively harmful past the mid-game.
The most widely repeated 2048 tip is some version of: press up and left over and over, and when neither works, press right and immediately go back to your vertical move. It is stated as a complete strategy. Written out as an algorithm — try up, else left, else right, else down — it is completely precise, so we ran it 80 times.
It reached 2048 in 0.0% of games. It reached 1024 in 0.0% and 512 in 6.3%. Its median game ended after 224 moves with a score of 2,504.
This is worth being precise about, because the tip is not worthless — it is worth a great deal early. Compare it to moving at random, which reaches 1024 in 0.0% of games, and the pattern is obviously doing real work. It builds a gradient, it keeps the big tile pinned, and it will carry a beginner to 256 or 512 without any thought at all.
What it cannot do is finish. The failure is specific: the fallback move. When up and left are both illegal, the pattern says press right — and right is the one direction that pulls your largest tile away from the corner you spent the whole game protecting. Every time the board tightens it takes that move, and eventually it takes it at a moment when a new tile lands in the vacated corner. From there the structure cannot be rebuilt, and the run is over even though the board still looks healthy.
The practical correction is not to abandon the pattern but to stop treating the fallback as automatic. When you run out of your two directions, that is the moment the game is asking you an actual question. It is also exactly the moment to press Hint and see whether a search agrees with your instinct.
Empty cells matter more than corners.
The solver rates every board on four things, and the weights are not equal. Free space carries a weight of 2.7. Corner control carries 1. Ordering carries 1. Those weights were tuned to win games, and they say plainly that the most valuable quantity on the board is room to manoeuvre — ahead of the corner rule that every guide leads with.
This is the correction most intermediate players need. Space is optional value: an empty cell is a move you have not been forced to make yet. Once you are down to one or two free cells, the board decides for you, and it does not matter how beautifully ordered your tiles are if you have no legal move that preserves the arrangement.
In practice this means merging is not automatically good. A merge in the middle of the board frees a cell, which is why it is tempting — but if taking it drags your structure out of order, you have traded a permanent problem for a temporary one. And it means the moment to be most careful is not when the board looks messy but when it looks nearly full, however tidy.
Verdict: true, and it is a symptom worth reading.
A 2 sitting between two 16s cannot merge with anything. To clear it you must first build another 2 next to it, then a 4, and so on — several moves of work, during which everything else must hold together. Left alone it occupies a cell permanently, which by the argument above is the most expensive thing that can happen to you.
We did not measure this one as a separate policy, and it would be dishonest to claim we did — it describes a board state rather than a rule for choosing moves. But it follows directly from the effect that was measured. A trapped tile is a permanently lost cell, and lost cells are what the heaviest term in the evaluation is counting. Clear the anomaly while it still has a reachable neighbour; every move you wait makes it more expensive.
Even the search loses sometimes: it failed to reach 2048 in 0.0% of its games, and it never once reached 8,192. There is no strategy that removes the possibility of a 4 landing in the one cell that breaks your board. Playing well moves you from losing most of the time to losing occasionally, which is the entire game — anyone promising a method that always wins has not measured it.
The numbers come from scripts/simulate.js in this site’s repository, run on August 5, 2026. It is worth stating the method plainly so you can judge it:
/js/ai.js rather than reimplemented, so the simulation cannot drift from the real game.1), and each game’s seed is derived from its index, so the whole run reproduces exactly and the corner ablation faces the same starting positions as the unmodified solver.Two limitations you should weigh. First, these policies are formalisations of the advice, not the advice itself — a person following “keep to two directions” also exercises judgement that no fixed priority order captures, so the 0.0% figure is a floor for the idea, not a verdict on everyone who uses it. Second, 4 games gives roughly a percentage point of uncertainty on the search win rates, so treat small differences between them as noise. The gaps described on this page are much larger than that.
The one thing worth taking from all of this: the tips are not equally good, and the difference between them is measurable rather than a matter of taste. Corner discipline and protecting space are worth building habits around. Any fixed sequence of keys is worth outgrowing.
Ready to put it into practice?
Play 2048