Bet365 Poker Bot Developer FAQ

Some commercial poker AI projects claim iPoker Network support, which by extension covers Bet365 Poker. Quality varies widely. The general pattern is that any bot that markets itself as targeting iPoker is built on a shared client-automation layer across all iPoker skins; the math underneath is whatever solver and opponent-model code the team has, mostly developed against deeper liquidity at GGPoker and ported. The economics are weaker than at higher-traffic operators — mid-stakes iPoker liquidity is thin and the regular population is well-defined enough to make the per-account bot lifetime shorter. Most teams that take poker AI seriously do not list Bet365 as a primary target.

No. The clients are different, the underlying protocols are different, the UI layouts are different, and the behavioural-telemetry surfaces are different. A PokerStars bot can share solver code and opponent-model code with a Bet365 / iPoker bot — those layers are platform-agnostic — but the UI automation layer (screen scraping, action emission, telemetry hygiene) is fully platform-specific and accounts for a large share of the engineering effort in any production poker AI.

Two-tier. Playtech runs the network-side anti-cheat across every iPoker skin: behavioural fingerprinting, statistical play-pattern analysis, anti-collusion graph models. Bet365 runs the operator-side fraud and KYC: sportsbook-grade payment monitoring, identity verification, source-of-funds checks. The combined surface is heavier on identity and payment than at a pure poker operator, lighter on bleeding-edge poker-specific behavioural modelling than at GGPoker. Catch-rate is real but slower; median lifetime to ban runs 2-4 months on typical implementations. Full picture in the detection note.

Marginal. The liquidity sits at mid-tier — fine for development, learning, and as a secondary site, not strong enough to be a primary income line. EV-per-engineering-hour curves at GGPoker materially above iPoker on most format / stake combinations. The reasons to point an existing bot stack at Bet365 are softer-population at low-mid stakes (sportsbook cross-pollination keeps the fish density reasonable) and as a diversification site to avoid concentrating account risk at one operator. The reason not to is that the network's cross-skin observation is genuinely effective and accounts have shorter expected lifetimes than the raw skin-level analysis suggests.

Because the poker product on Bet365 is a re-skinned Playtech iPoker client and the player pool is shared with other iPoker skins. When you play a NL10 6-max table on Bet365 Poker, your opponents are sourced from the combined iPoker pool, not just Bet365 customers. Network-side anti-cheat sees behaviour across every skin and joins on Playtech's internal player ID. From an engineering perspective, building "for Bet365" and building "for iPoker" are nearly the same project — the differences sit in operator-side risk policy and a handful of skin-specific client tweaks.

Network-side ToS prohibits third-party HUDs and overlays across all iPoker skins. Enforcement is variable but real — the client performs process detection at startup and during play. PokerTracker and Holdem Manager historically supported iPoker, and a portion of the playerbase has used them anyway; the operator does react when usage shows up in client telemetry. The realistic guidance is do not assume HUD use is safe even where it has worked previously. Most modern poker AI work assumes no HUD support and falls back to online opponent modelling for the same reason.

iPoker rake on NLH cash sits in the 5-7% range at low stakes with some rakeback through VIP programmes, dropping to roughly 3-4% at mid-stakes. Twister rake is in the 6-8% range built into the multiplier structure. Net EV after rake makes NL25 to NL100 the sweet spot for cash; below that the rake share eats too much, above that the regular population grows aware of solver play fast. Twister is mathematically positive at €5-10 buy-ins given enough volume, with very high variance. Higher stakes are too thin to be a primary target.

Technically yes, mathematically marginal. The Twister format starts at 10bb three-handed with a multiplier draw that can push the prize pool to 1000x the buy-in. Push-fold trees dominate from about 12bb down, and they are well-solved. The catch is that multiplier variance dominates the EV signal; a positive-EV strategy needs tens of thousands of games to converge. For most teams the engineering effort is better spent on cash-game lines. Some operators run Twister-specific stacks but the EV-per-hour after rake and variance is not compelling.

GGPoker, by a wide margin, for any project where bot revenue is the primary metric. Higher liquidity, more rake structure that supports mid-stakes solver play, a softer population at the recreational end, larger MTT prize pools. GGPoker also has the more sophisticated detection stack, which makes the engineering harder, but the EV-per-engineering-hour still comes out ahead. Bet365 / iPoker makes sense as a secondary site for diversification or as a development environment where account-loss cost is lower. The homepage of the parent project (Poker Bot AI) covers this comparison across operators.

The empirical floor for a useful exploitative deviation sits around 80-150 hands of joint observation per opponent — enough to estimate VPIP and PFR to within ±5 percentage points with overlay on c-betting and 3-betting response. A Bayesian prior conditioned on population distribution can in principle cut this to 40-60 hands. The harder constraint at iPoker is that mid-stakes table populations are smaller than at GGPoker — you see the same regulars across sessions, but at lower volume per session, so absolute time to convergence is longer despite the relatively defined population.

Each table runs its own state machine with its own decision queue. A naive implementation shares a CPU and clocks decisions sequentially, which creates observable cross-table action-latency correlations — table A's hard river decision delays table B's snap-fold, and the correlation is detectable at the network-side telemetry layer. The right architecture is preemptive: a tough decision on table A yields the CPU and rejoins the queue when easier decisions are pending, with each table's behavioural-timing distribution sampled independently. Production systems also add a per-table alertness parameter that drifts down over long sessions to mimic fatigue.

On a modern Android device (Snapdragon 8 Gen 2 class), 30-80ms for solver-table lookup, 5-15ms for opponent-model update, 10-30ms for UI inspection on the rendered client. Total compute budget around 100-150ms per action. The human-perceptible action latency is the behaviourally-shaped sample on top of that, typically 600ms to several seconds depending on decision difficulty. The constraint is rarely compute; it is power consumption and thermal management over long sessions, and the maintenance load on the UI automation layer.

Android is the production target. The application sandbox is permissive enough to support accessibility-service-driven input automation with relatively stable behavioural shaping, and the device-fingerprint surface is broad enough to be shaped credibly. iOS is significantly harder — the sandbox blocks the input-automation surfaces, the available alternatives (jailbreak, MDM-based exploitation) are unstable across iOS releases, and the device-fingerprint surface is narrower. Most production poker AI runs Android-only as a result, which has the side effect of concentrating detection-side analysis on Android-typical telemetry.

Hardware-level entropy. Mobile devices expose sensor data — accelerometer drift, gyroscope motion during play, magnetometer baseline, ambient light, CPU clock variance under load, battery discharge curve — that is hard to synthesise convincingly. A device sitting on a desk for 12 hours has measurably different sensor traces from a device being carried in a pocket. Some implementations place a real device on a powered cradle and inject low-level perturbations to the sensor stream; this defeats the cheapest version of the test but it is unclear whether it survives side-channel analysis on the full sensor vector.

Not an immediate ban-hammer. The visible sequence: quiet flag → soft restriction (withdrawal limits drop, KYC re-verification requested) → structured interview from Bet365 support → confiscation and account closure. The cycle takes 14 days to 9 months depending on review-queue capacity and triggering events; the most common accelerator is a first large withdrawal. The detection page covers the review pipeline in detail. The Independent Betting Adjudication Service (IBAS) recourse for UK customers can extend the closure-to-resolution timeline further.

A GTO solver — PioSolver, GTO+, MonkerSolver — computes a near-equilibrium strategy for a given game tree offline, iterating a CFR variant until convergence. The output is a large lookup table of action frequencies keyed on game state. A real-time bot uses that table as a baseline and queries it under tight latency, overlaying an online opponent model that biases the action toward exploitative deviation when the opponent estimate justifies it, with detection-aware noise on top. The solver is the theory book; the bot is the player.

In heads-up, every hand updates the same opponent's model — the model can be fine-grained because the sample comes in fast. In 6-max, five opponents each update their own model at one fifth the rate, so the model uses broader archetype buckets (loose-passive, tight-aggressive, station, nit) and refines the per-opponent fit over the session. The architectural difference is that heads-up engines can afford specific board-texture-keyed features per opponent; 6-max engines need archetype-conditioned features with per-opponent overrides only when the per-opponent sample is large enough to support them.

None for real-time decisions; some utility for adjacent tasks. Frontier LLMs hallucinate ranges, misapply ICM, lack frequency intuition, and produce sizing recommendations that match no recognisable equilibrium. A $5 heads-up bot built on solver output beats any LLM at live play. LLMs are useful at hand-history annotation, exploit-hypothesis generation worth solver-checking, and converting solver output into prose summaries for human study. The boundary between "useful annotation tool" and "useful in-the-loop player" is sharper than the field generally acknowledges.

Two compression layers. State abstraction collapses the game state into equivalence classes based on board texture, stack-to-pot ratio, position, and action history, reducing the lookup key from the full state space to an integer index. Action abstraction discretises bet sizes to a finite set (33%, 50%, 66%, 100%, 150%, all-in) and solves only at those discretisation points. Together the two typically compress by four to five orders of magnitude with bounded EV loss compared to the full CFR output. The art is in choosing abstractions that preserve as much EV-relevant structure as possible.

Three places. Online opponent-model convergence under thin volume and cross-skin observation (the network sees more than your bot does — design accordingly). Detection-aware policy combiners under stacked classifiers (Playtech network-side plus Bet365 operator-side — multi-classifier adversarial selection). Compression of multiway solver outputs to mobile inference budgets with acceptable EV loss. All three are open problems with enough surface area for a serious contributor to make real progress. The team chat is the right place to start the conversation.