Do Poker Training Apps Actually Work? An Honest Answer

You've had a trainer on your phone for three weeks. Maybe 400 spots done, a streak you're weirdly proud of, and a little accuracy number that has crept up from 68 to 74. Then you sat down Saturday, ran into a cooler on hand nine, punted a stack on hand forty, and closed the session down two buy-ins wondering whether any of it did anything at all.

Poker training apps work, but only when four conditions are met, and most of the apps on your phone meet none of them. The condition that matters most is the one nobody markets: the app has to make you commit to a decision before it shows you the answer, then grade the decision rather than the outcome of the hand. Everything else, the spot count, the solver depth, the ELO badge, is secondary to whether that loop closes.

Short answer: Poker training apps work when they force a decision before revealing the answer, grade the decision rather than the result, keep the spot narrow, and spread reps across days. Apps that fail those conditions, including play-money apps and video libraries, produce no measurable improvement no matter how long you use them.

Key takeaways

• Four conditions decide whether a poker training app does anything: decision before answer, feedback on the decision rather than the result, one narrow spot, reps spread across days.
• Poker punishes learning because outcomes are noisy. A trainer works by removing the noise, not by adding information you couldn't have read in a book.
• In Roediger and Karpicke's 2006 testing-effect experiments, the group that reread material recalled 83% after five minutes and 40% after a week, while the group that tested itself recalled 71% and then 61%. Rereading feels better and retains worse.
• You cannot see a training app's effect in your win rate. Confirming a 2 bb/100 improvement at 95% confidence takes roughly 500,000 to 1,000,000 hands.
• You can see it in scored decision accuracy in about 600 reps per period, which is three weeks of daily drilling.

Why do poker training apps work at all?

A poker training app works by converting poker from a learning environment that punishes you into one that teaches you. That sounds abstract. It's the most practical idea in this article, so stay with it for two paragraphs.

Robin Hogarth and colleagues drew a distinction in their 2015 paper on kind and wicked learning environments in Current Directions in Psychological Science. A kind learning environment is one where the feedback you get matches the thing you're trying to learn. A wicked learning environment is one where it doesn't, so experience actively teaches you the wrong lesson.

Poker at the table is about as wicked as it gets. You make a correct fold and lose the pot you would have won. You make a spew call and bink a two-outer, and your brain, which is a pattern-matching machine that does not read solver output, files that under "worked." Feedback arrives minutes later, attached to the wrong variable, and it arrives roughly a thousand times a session. That's why some players have 400,000 hands of experience and the same leak they had at hand 5,000.

A drill app makes the environment kind. The result of the hand is deleted. The only thing you're graded on is the decision, and the grade shows up two seconds after you make it. That single change, feedback on the decision instead of the outcome, is the entire mechanism, and it's why "I'll just learn from playing more" doesn't work for most people.

What are the four conditions a poker training app has to meet?

The four conditions are: you decide before you see the answer, the feedback grades the decision rather than the result, the spot is narrow enough to repeat, and the reps are spread across days. Miss one and the app becomes entertainment with a progress bar.

Condition one has the strongest evidence behind it. Retrieval practice is the finding that trying to produce an answer from memory improves later retention more than restudying the same material does. In Roediger and Karpicke's 2006 experiments in Psychological Science, students who reread a passage four times recalled 83% of it five minutes later, against 71% for students who read it once and then tested themselves three times. A week later the ordering flipped hard: 40% for the rereaders, 61% for the testers.

Read that twice, because it's the reason your video backlog isn't working. The method that felt like it was working was worse. Watching a coach explain a river spot is rereading. Being handed the river spot cold and made to pick a size is testing.

Condition four is the same story in a different coat, and we've written the daily version of it up as a 15-minute poker study routine with a literal 7-day calendar. Same total minutes, different shape, much better retention.

Which types of poker apps actually meet the conditions?

Drill-style trainers meet all four conditions, hand replayers meet one, and play-money apps meet none. Here's the honest sort, by category rather than by brand, because brands ship features and change.

Two things follow from that table. First, "is this app any good" is the wrong question, and "does this app make me decide before it shows me the answer" is the right one. Second, most people don't need a better tool, they need a tool from a different row. Our comparison of six poker training apps sorts specific products into these rows, and the full feature-by-feature comparison covers current pricing, which moves too often to quote in a blog post. If you already have a desktop solver subscription and you're wondering why nothing's changed, the GTO Wizard alternative breakdown is the same argument applied to one product.

How would you even know if a poker training app worked?

You cannot detect a training app's effect in your win rate, and anyone telling you they did is reading noise. This is the part every article on this query skips, and it's the part that will save you the most self-deception.

The margin of error on a cash-game win rate is brutal. PrimeDope's variance calculator puts a reasonable standard deviation for 6-max no-limit hold'em at 75 to 100 bb/100. Take the pessimistic end, 100 bb/100, and the 95% margin of error on your measured win rate over n hands is 1.96 × 100 ÷ √(n/100) in bb/100. That produces this:

At a standard deviation of 75 bb/100 those margins shrink by about a quarter, so the honest statement is a range: confirming a 2 bb/100 improvement at 95% confidence takes somewhere between 500,000 and 1,000,000 hands, and comparing a before period against an after period needs roughly double that in each period. At 1,000 hands a week, a million hands is about nineteen years. Your trainer subscription will not outlive the measurement.

So measure the thing that resolves fast instead. If your app scores each decision right or wrong, decision accuracy is a proportion, and proportions converge orders of magnitude quicker than win rates. Around 600 scored decisions in each period is enough to see an 8-point accuracy change at 95% confidence, which at 30 reps a day is about three weeks. That's the number to watch. Not your graph.

The catch, and it's a real one: accuracy inside an app is a proxy. It proves you learned the baseline. It doesn't prove you executed it at 11pm on a Saturday with two buy-ins already gone. Nobody has published data closing that gap, so treat in-app accuracy as necessary rather than sufficient, and confirm it the old way, by noticing whether the spot still makes you tank at the table.

When do poker training apps not work?

Training apps fail against pool-specific mistakes, because a solver baseline assumes your opponent plays well and yours doesn't. This is the strongest argument against trainers and it deserves a fair hearing rather than a straw man.

Steve Blay of Advanced Poker Training makes it directly in his piece on whether a GTO trainer will make you a poker legend, arguing that "a solver only guarantees that it will teach you how to beat other optimal opponents" and that exploitative play is where the money is. He sells a product built on lifelike opponents rather than solvers, so read him with that in mind. He's also not wrong. A trainer will never tell you the nit in seat four folds every river, and at $1/$2 live that single read is worth more than a month of c-bet drills.

The counter is that these fix different halves of the problem. Unforced errors are the larger half below mid-stakes: calling too wide from the blinds, c-betting every flop out of habit, folding the turn after calling the flop. A trainer removes those. Exploitation is a second layer you add on top, by hand, at the table, and it works better once you're not simultaneously leaking three big blinds a hundred to your own defaults. Work down our ranked list of common poker leaks and you'll notice how few of the expensive ones are about your opponent at all.

Four other failure modes, all of them yours rather than the app's:

1. Drilling a spot you don't play. Beautiful MTT ICM accuracy is worth nothing to a $1/$2 cash player. Pick spots by frequency in your own game.
2. Random mode. It feels like practice and it's closer to scrolling. Lock the parameters.
3. Streak theatre. Opening the app to preserve a number is not a rep. Duolingo taught an entire generation this reflex and it does not transfer.
4. Skipping diagnosis. If you don't know which leak costs you most, you'll drill the one you find interesting. Our framework for fixing a poker leak starts with the diagnosis step for exactly this reason.

Honest opinion, and it's not a flattering one for our own category: most people who quit a training app didn't quit because the app was bad. They quit because they were drilling a spot at random with no way to know it was working, which is a fair reason to stop doing anything.

How to actually drill this

Knowing the four conditions changes nothing on its own. Here's how to test whether your app clears them, using your own game as the sample.

1. Audit your current app against the four conditions tonight. Open it and check: does it take your answer before showing the baseline, does it grade the decision or the pot, can you lock position and stack depth, does it work in five minutes. If it fails condition one or two, no amount of discipline rescues it. Change rows in the table above.
2. Pick one spot with every parameter named. Not "big blind defence". Something like: big blind versus a button open, 100 big blinds effective, 6-max, facing 2.5x. If it doesn't fit on one line, it's still too broad.
3. Run 30 reps a day for 20 days, same spot, same time slot. That's roughly 600 decisions, which is the sample you need for the accuracy read to mean something. Write down your accuracy on day one before you start.
4. Compare day-one accuracy to day-twenty accuracy on the same spot. An 8-point move or better is a real change. A 3-point move is noise, and the honest response is to keep going, not to declare victory.
5. You're fixed when the pause is gone at the table. The observable outcome isn't the number in the app, it's that the spot comes up mid-session and you act without a beat of thought. Still tanking after 600 reps? You're drilling a slightly different spot from the one costing you money. Narrow it and restart.

That loop is what LeakSeek is built around: solver-backed spots, a decision required before the baseline appears, feedback on that decision, and a leak report that ranks your mistakes by EV loss so step two takes thirty seconds instead of an evening. Its real limits, so you can judge it fairly: it's phone-only, there's no desktop solver, and after the 7-day trial you get one free session a day unless you subscribe at $14.99/month. This worked hand breakdown shows exactly what the feedback on one decision looks like, so you can check condition two yourself before downloading anything.

The accuracy number was never the point. Three weeks from now the button opens, you're in the big blind with K7 offsuit, and you just know. That's the whole product, and it's the only receipt that counts.