How to tell whether your chess is improving - measuring instead of guessing
Rating is a slow, noisy instrument, and it is the only one most players own. Three faster measurements exist: errors per reviewed game, where in the game those errors live, and per-concept mastery from clean solves. All three move before rating does, and all three can be counted rather than felt.
Rating is a slow, noisy instrument
Rating answers one question - how you are doing against people near your strength - and answers it with a long lag and a lot of noise. Thirty blitz games can move a rating a hundred points in either direction without anything changing about your chess. Over a year it is a fair summary. Over a fortnight it is mostly weather.
That would be fine if rating were one instrument among several. The problem is that for most players it is the only one, so a bad week reads as "I am getting worse" and a good week reads as proof that whatever they tried is working. Both readings are usually wrong, and both change what people train next.
Measure the thing you are trying to change
If the plan is "stop hanging pieces", the measurement is not rating. It is how often a piece goes en prise in your reviewed games, this month against last. If the plan is "handle endgames better", the measurement is how many of your errors now fall in the endgame phase - not what your rating did while you were working on it.
This is the whole trick, and it is not a chess idea. Pick the specific thing, count it before, count it after, keep the conditions comparable. Anything you can count you can steer by, and rating is the last number to react rather than the first.
Three measurements that move before rating does
Errors per reviewed game. Take the counted errors from games that actually went through an engine pass and divide by the number of games. It moves within weeks and it responds to the work directly - fewer real errors is close to the definition of playing better. Mikhail charts it by week, which is enough resolution to see a trend and not so much that noise looks like one.
Where the errors live. The same counted errors split by phase - opening, middlegame, endgame - earlier games against your most recent. A total that stays flat while its distribution shifts is real progress: you have moved the problem, which is the step before removing it.
Per-concept mastery. A drill ledger in which a concept's score rises only on a clean, unassisted solve and falls on a miss. Because revealing an answer is recorded and never rewarded, the number is slow and honest, and its change since your first attempt on that concept is the part worth reading.
Compare like with like
Most apparent improvements are changes of conditions. Errors per game drop because you switched from blitz to rapid. The endgame column empties because your last ten games ended in the middlegame. Session accuracy rises because the drills got easier. Before believing a number, check that the games behind it resemble the games behind the previous one.
The cheapest protection is to keep one comparable pool - the same time control, reviewed the same way - and let everything else be extra. This is also why review coverage matters: a measurement built from the six games you felt like analysing is a measurement of your mood.
What a real improvement curve looks like
Not a line. Flat for weeks, a step down, flat again, a bump upward when you start meeting stronger opponents or a harder time control. If your chart is smooth, it is probably too short or too smoothed. Judge it the way you would judge a running pace: by the level of the flat stretches, not by the slope of any particular fortnight.
Give any change six to eight weeks of comparable games before deciding it worked. That is slower than anyone wants, and it is exactly why the counted measurements matter - they at least move inside that window, while rating often has not begun to.
How Mikhail reports it
The statistics page reports measured change only. Game figures come from games that went through the engine, training figures from server-graded attempts, and a section without enough behind it says "not enough data yet" rather than drawing a flattering line through three points. Sections appear as the data appears.
That is a deliberately unexciting design. A progress page that always shows progress is a comfort object; one that sometimes says nothing is an instrument. If the number has not moved, the useful response is to check that the work matched the measurement - and then to give it more games.
Questions
How long before I should expect to see anything?
Counted errors per reviewed game can move inside a month if you are reviewing consistently. Rating is slower and noisier. If you need an answer sooner than that, you need a narrower measurement rather than a faster one.
My rating is flat but my errors per game are falling - which do I believe?
Both, in that order. Falling errors usually arrive first; rating follows once the improvement survives contact with opponents. If errors keep falling for months and rating never moves, look at what is ending your games - the clock, conversion, or the opponents you are choosing.
Does solving more drills prove I am improving?
It proves you solved more drills. Mastery rises only on clean unassisted solves, which makes it a fair training measurement, but the measurement that settles it is what happens in your next reviewed games.
Why does the statistics page sometimes show nothing?
Because there is not enough measured data behind that section yet. An empty section is a true statement; a trend drawn through two points is not.
Try it on your games
The part of Mikhail this guide describes is Statistics, which needs a free account — your games and your counts have to live somewhere.