How Reevix's AI Learns

Audience: Technical stakeholders, growth analysts, and anyone who wants to understand the engine driving Reevix's decisions.

Reevix uses a combination of real-time behavioral classification, large language model generation, and a multi-armed bandit optimisation loop to continuously improve its interventions.

The Decision Pipeline

Every 50ms, for every active session, Reevix runs this pipeline:

Raw signals → Feature extraction → State classification → 
Intervention selection → Bandit scoring → Deploy or suppress

Stage 1 — Raw Signal Collection

The SDK streams over 30 raw signals to Reevix's edge service. These include:

Mouse position, velocity, and direction
Scroll depth and dwell time per page
Click patterns (including rage-click detection)
Idle time
Tab switch count
Form interaction events (focus, blur, error)
Product view history for the session
Cart contents and subtotal
UTM source
Prior purchase history (if identify has been called)
Intent survey answer (if taken)

Stage 2 — Feature Extraction

Raw signals are transformed into normalised features that feed the classifier. For example:

comparison_score = weighted sum of (product views × repeat views × category switches)
exit_intent_score = weighted sum of (cursor-to-top velocity + idle time + tab switches)
cart_hesitation_score = time since last cart mutation × cart value

Stage 3 — State Classification

A lightweight ML model (running on the edge, latency < 2ms) maps the feature vector to one of the five behavioral states: Exploring, Comparing, Evaluating, Ready, Buying.

The model is a gradient-boosted classifier trained on anonymized, aggregated behavioral data across all Reevix tenants. It is retrained monthly.

Stage 4 — Intervention Selection

Given the classified state and page type, Reevix filters the store's active message templates to those matching the current context. It then queries the bandit for each eligible template to get an estimated conversion lift score.

Stage 5 — Multi-Armed Bandit

Reevix uses a Thompson Sampling bandit to balance exploration (trying new or undertested templates) with exploitation (showing the template most likely to convert):

Each template maintains a Beta distribution over its conversion rate
On each selection decision, Reevix samples from each template's distribution and picks the highest sample
A conversion (purchase, add-to-cart, or extended session) updates the selected template's Beta distribution

Result: Over time, high-performing templates receive more traffic automatically, without you needing to manually analyze A/B test results or pause underperformers. The bandit handles this continuously.

LLM Message Generation

When you run the Setup Wizard or click Regenerate All, Reevix calls a large language model (Claude or GPT-4o, depending on your region) with:

Your wizard answers (brand, tone, goals, policies, trust signals)
A structured prompt enforcing factual accuracy (the LLM is explicitly instructed to never claim policies you haven't declared)
Output schema requiring: message_text, page_type, pain_point, intervention_type, max_length

The output is validated against your declared policies before being saved. Messages that reference undeclared policies (e.g., "free returns" when you haven't enabled that) are automatically filtered out.

How Long Before the AI Learns?

Sessions	What Happens
0 – 500	Pure exploration — all templates shown roughly equally. Bandit is gathering initial data.
500 – 2,000	Early signals. High-performing templates begin to edge ahead slightly.
2,000 – 10,000	Significant optimisation. 1–3 templates per pain point dominate traffic.
10,000+	Steady state. Bandit maintains ~5% exploration to continue detecting changes.

For stores with fewer than 2,000 monthly sessions, the learning period is proportionally longer. You can accelerate this by:

Adding more templates per pain point (more arms = faster differentiation)
Regenerating templates if early conversion rates are uniformly low

What Reevix Does NOT Do

No personally identifiable data is used for classification. All data is session-level.
No retargeting. Reevix only acts on visitors who are currently on your site.
No dark patterns. The LLM is explicitly prohibited from generating false urgency, countdown timers, or misleading claims.
No consent required for session signals (they are equivalent to server-side analytics). PII (email via identify) requires your store's existing consent flow.