The Optimizer: how traffic is allocated

AXO decides how much traffic each variant gets with a multi-armed bandit rather than a traditional A/B test. In plain words: each variant's share of traffic is based on the evidence so far that it converts. Variants that are converting earn a growing share; variants that aren't are gradually starved of traffic. There is no fixed test duration, no waiting for a significance readout, and no “declare a winner” step – allocation is recomputed every hour from each variant's live impression and conversion counts, and it keeps adjusting for as long as the variants are live.

The specific method is Thompson sampling. The name “bandit” comes from casino slot machines (“one-armed bandits”): the method sends more pulls to the machines that are paying out, without ever fully abandoning the others while real uncertainty remains. Practically, that means early winners get traffic sooner, weak variants lose traffic without anyone having to switch them off, and a variant with thin evidence still gets enough exposure to prove itself either way.

Two boundaries worth knowing. First, each visitor's draw is deterministic per session – a visitor keeps the same variant across pageviews rather than seeing the page change mid-visit; the odds move between sessions. Second, the bandit only decides traffic shares: every variant it allocates between was written or approved by a human. You can inspect each arm's impressions, conversions, and current allocation on the Optimizer page, and view or reset bandit state in the AI Scorecard tab.

Real evidence wins

Sites can accumulate simulated traffic – sessions generated by the AXO simulator for demos and pre-launch dry runs. The optimizer guarantees that simulated traffic can never steer a site that has real traffic: the moment a real visitor is exposed to a variant, simulator sessions are excluded from the optimizer's evidence entirely, and allocation is computed from real impressions and real conversions only. A simulated session can never outvote a real conversion. This is automatic – there is no setting that lets simulated data back in.

The SIMULATED DATA badge

Before a site has any real exposures, the optimizer still runs on simulated traffic so you can watch the mechanics work – and the Optimizer page labels that state with a SIMULATED DATA badge. The badge means exactly one thing: every count on the page came from simulator runs, because no real visitor has been exposed to a variant yet. It clears on its own the moment real exposures arrive, when the optimizer switches to real traffic only. If you don't see the badge, the numbers you are looking at came from real visitors.

Lift Report & Holdout

AXO permanently withholds 10% of traffic from all personalization – this is the holdout group. By comparing conversion rates between the personalized 90% and the untouched 10%, AXO produces a statistically rigorous measurement of incremental revenue. The Launch Report shows your projected daily and monthly revenue lift, broken down by segment and variant type. This is the number that answers “what is AXO actually worth to us?”

Backtest

Before going live, use Backtest to run an offline A/B simulation against your historical session data. The engine replays past sessions, applies your current segment definitions and variant weights, and projects what CVR lift would have been – giving you a data-backed confidence score before committing real traffic.

Analytics Funnel

The Analytics tab provides a behavioral engagement funnel – from landing → scrolling past the intro → viewing a product → adding to cart → purchasing. Each step is broken down by segment, so you can see exactly where each behavioral cluster drops off and which placement interventions are moving them down the funnel.