How our GTO coaching is built

Last updated: September 3, 2026

Every poker app says its advice is "GTO-backed." Almost none say what that means or where the numbers came from. This page is the long answer, written so that someone who has never opened a solver can follow it. The short answer is in the box.

1. What a solver actually does

Poker is a game of hidden information. You never know the other player's cards, so there is no single "right move." There is only the right move most of the time, given every hand the other player could have.

A solver finds that by playing one situation against itself, over and over. Player A tries a plan. Player B finds the best answer to it. A adjusts, B adjusts, and each round the plans get a little harder to beat. Eventually neither side can improve by changing anything, and that is the answer. Two kids playing rock-paper-scissors thousands of times would arrive at the same idea: throw each one exactly often enough that the other kid can't win on purpose.

Nobody teaches the solver poker strategy. It knows the rules, the money, and the cards each side could hold. Good strategy falls out of the math.

2. Two kinds of data, built two ways

Preflop charts are curated, not solved. Free public sources cover preflop well, so our open-raise and defense ranges are synthesized from published charts. Each range carries provenance: which sources it draws on, in what priority, and how disagreements were resolved. These are the ranges the coaching features grade against, and the Ranges tab shows them at 6-max for each rake setting; its 8-max and 9-max charts are a simpler, older approximation and say so. Where sources disagree, a priority-weighted average decides, never a coin flip. Where only one source covers a spot, the range is marked low-confidence in its own file.

Flop and turn strategy is solved by us. No free source publishes postflop solutions at the depth a hand replay needs, so we run the solver ourselves. The rest of this page is about that.

3. What we feed the solver

We use TexasSolver, an open-source solver that runs from a terminal. It has no buttons and no opinions. For every flop we hand it a small text file describing one situation:

• The line. Who raised before the flop, who called, and which of them acts first on the flop. Today every line is a single-raised pot where the big blind called an open from one of the five other seats.
• The money. The pot after the preflop action and the stacks behind, at 100 big blinds.
• The hands. Exactly which hands each player would have arrived with. These are our own curated preflop ranges at the rakeless setting. If you have the app set to a live rake structure, the charts it shows you are tighter than the ranges behind these solves — at $1/$3 the five opening seats run about 23% fewer hands, and on the button 83 against 117. Section 7 says what we know, and do not yet know, about what that costs.
• The allowed moves. Check, bet a third of the pot, or move all in, with a raise or shove available in reply. Fewer options mean a faster, more accurate solve and a strategy a human can actually learn.
• The target. How close to unbeatable the answer must be before the solver is allowed to stop. We set it to three tenths of a percent of the pot.

4. Which flops, and why 49

There are 22,100 possible flops, but suits are interchangeable, so strategically there are 1,755 distinct ones. Solving all of them for every line would take months on the hardware we use. Instead we solve 49 per line, chosen to cover the kinds of flop that change how the hand plays.

The selection is our own, and it is deterministic. We sort every flop into a bucket by four features: suit texture (rainbow, two-tone, monotone), pairing, the highest card, and how connected the cards are. Each bucket gets a share of the 49 slots in proportion to how often it is actually dealt, and within a bucket the most common flops are picked first. Anyone with the method can regenerate the same 49 boards, which is what lets us call the board set our own data rather than a copy of somebody's product.

One solve runs the full hand through the river and takes a laptop with every core busy somewhere between ten and forty minutes, depending on the board. A line of 49 flops is an overnight batch.

5. How we check the answers

A solver will confidently solve the wrong game if you describe the wrong game. So every batch has to pass three checks before a single number reaches the app, and a batch lands whole or not at all.

• The solver's own score. It reports how beatable its answer still is. Anything above our three-tenths-of-a-percent target is refused, including solves that merely ran out of time.
• Known ballparks, per position. Poker theory knows roughly how often an early-position raiser should bet the flop, how often the big blind should fold to it, and how rarely the big blind should lead out. Each batch is compared to those figures for its own seat. A batch that lands far outside them is treated as a setup error, not a discovery.
• Who was sitting where. The solver's output is read back and the hands present at each decision are compared to the range that was supposed to be there. The right numbers in the wrong seat fail this check by name.

6. When your flop isn't one of the 49

Most of the time it won't be. When you replay a hand, the app first looks for your exact flop. If it isn't solved, it looks for the closest solved flop in the same bucket: same suit texture, same pairing, then the nearest high card and connectivity. Whatever answers, the app tells you which board it was. It will never silently show you a rainbow solve for a monotone board, or an unpaired one for a paired board. If nothing honest matches, it says so and shows nothing.

The turn works the same way, one level down. If the flop we served is your own, your turn card is solved too and you get the exact answer. If the flop was a stand-in, the turn is matched by what the card did to the board — paired the top card, completed a flush, brought an overcard, and so on — and the app says when that is what happened. We measured how much that grouping actually explains before shipping it: for three of the four turn decisions it explains two and a half to three times the variation left inside a group, which is enough to stand in. For the fourth, the big blind betting after the small blind checks the turn, it explains barely half that, so we do not stand in at all — that decision is shown only when your own turn card is solved, and withheld otherwise.

7. What's covered today

Rake matters postflop, and it matters more than we first measured: a solve that ignores the drop is a genuinely different game from the one at a live table. Our flop and turn solves are rakeless on purpose, because the solver we use does not model rake and because the rakeless solution is the right baseline to measure adjustments against. The app labels them that way. Rake reaches a solve through two doors, and we have measured one of them so far. The first is the pot. Rake makes every pot smaller, and a smaller pot makes calling worse: eight boards re-solved with the pot cut by the rake a $1/$3 game takes, ranges held fixed, moved the average hand by roughly one to two percentage points, and the one movement that showed up on every board was folding slightly more against a bet — which is exactly what a smaller pot should do. Part of what looked like a rake effect turned out to be the solver landing in a slightly different place on a second run, so we measured that too and subtracted it. The second door is the ranges, and it turned out to be five to ten times wider. A raked game is played tighter before the flop — at $1/$3 about 23% fewer opening hands, at $2/$5 about 7% — and these solves are fed the rakeless ranges. We re-solved the same eight boards with the $1/$3 ranges instead: the average hand moved seven to twelve percentage points, against one to two for the pot. And that understates it, because it can only compare hands that appear in both: about a quarter of the grid is missing from a $1/$3 range entirely, so those hands are not shifted, they are absent. The clearest single consequence is that folding against a flop bet rose about seven points. If you play $1/$3 and fold a little more than these pages say, that is not a leak, it is your game.

8. Credits

TexasSolver is open-source software, and we are grateful for it. We run it as a tool; the app ships only the strategies we computed with it and the checks we wrote around it. Our preflop charts draw on freely published material from GTO Wizard's free tier, PokerCoaching, Upswing Poker, Red Chip Poker, and Range Converter, each cited per range in our provenance files.

9. Questions

If something here is unclear, or you think a number in the app is wrong, email support@chiptrove.com. A wrong number is a bug and we treat it like one.