Reward mode: purchase conversion

How the bandit allocates traffic when Purchase conversion is selected.

---

TL;DR

Only one thing matters: what share of article visits ended in a purchase.

Visit with a purchase = success (1). Without = 0.
Order amount does not matter. Reaching checkout without paying does not matter.

Article effectiveness ≈ purchases / journeys. Higher → more traffic (via Thompson Sampling), but not 100% to one article: others keep being tested.

---

Details

Per-visit formula

$$\text{reward} = \begin{cases} 1 & \text{if purchased} \\ 0 & \text{otherwise} \end{cases}$$

Funnel stage and purchase_value are ignored.

Article score

$$\begin{aligned} \text{mean\_reward} &= \mathrm{AVG}(\text{reward}) = \frac{\text{purchases}}{\text{journeys}} \\ \text{reward\_sum} &= \mathrm{SUM}(\text{reward}) \approx \text{purchase count} \\ \alpha &= 1 + \text{reward\_sum} \\ \beta &= 1 + \text{journeys} - \text{reward\_sum} \end{aligned}$$

Dashboard Effectiveness = (\text{mean_reward}) (conversion (0\ldots 1)).

How it differs

Traffic

Same Thompson / selected algorithm. Below min_journeys — uniform.

A first purchase on one article does not lock 100% traffic forever: low-N competitors stay uncertain and keep getting explored.

---

Example

Typical picture: leader 30–45%, other converters share the rest — not a single winner takes all.

---

When to use it

• Goal is purchase conversion, not revenue or “reached the store”
• Tracking reliably sends the purchase event (purchased)
• With very few purchases the metric is noisy — keep a sensible min_journeys

Switch in bandit settings. Effectiveness recomputes immediately after save.