Conversion as a system outcome: a design for behavioural optimisation under constraints

1Context and motivation

Conversion optimisation is commonly practised one surface at a time: a checkout button, a product page layout, a banner. Each change is tested in isolation and kept if it wins. Controlled experiments are the right instrument for that question¹, but the question is narrow. Whether a shopper buys also depends on things no page test varies: whether the item is in stock, whether the promised delivery date is credible, whether the payment method works, whether earlier experiences with the seller were good.

We propose treating conversion as an outcome of the system rather than a property of a page. In this view the storefront, pricing, inventory, delivery promise, payments and support each act on the shopper, and the task is to coordinate them under uncertainty and under explicit limits. Two strands of the literature motivate this. Trust in the seller is a documented determinant of online purchase intention, alongside the perceived usefulness and ease of use of the site². And the same intervention can help or hurt depending on where a shopper is in a decision: in a field experiment, ads showing exactly the products a person had viewed did worse than generic ads on average and better only once the person's preferences had become specific³.

This paper describes a design, which we call the Behavioural Optimisation Model (BOM), that follows from this view. We present it as a worked design argued from the cited literature, without measurements of our own.

2Architecture

The BOM has five layers.

1. Event layer. Behavioural events (views, scroll depth, searches and their reformulations, filter use, cart edits), commerce state (price, discount eligibility, stock, delivery windows, fees) and trust and support events (returns, chargebacks, support contacts), each with locale, device and referral context. Events are ordered per session.
2. Signal fabric. Feature pipelines that turn raw events into intent, friction and trust signals, each with a freshness timestamp, a decay function that down-weights stale values, and anomaly flags.
3. Decision core. Effect estimators and contextual bandits that choose treatments (ranking, offer, message, interface flow) under constraints.
4. Execution. Rendering and delivery to web, app and partner channels, with latency checks on the request path and rules-based fallbacks kept ready.
5. Learning and governance. An experiment registry, audit logs, outcome attribution, and monitors for exposure across groups and for latency.

2.1Signals

Intent. Query specificity, filter depth, cart persistence, revisit cadence and price-check loops. Intent scores are calibrated with a logistic model under monotonic constraints, so that more of a positive signal can never lower the score. This is our modelling choice, made for interpretability.

Friction. Slow-rendering components, repeated filter changes, stock-out encounters, price mismatches, unavailable variants, payment retries and backtracking.

Trust. Prior return experience, the historical accuracy of delivery promises, and support backlog. We include these because trust in the seller shapes purchase intention².

Time and context. Seasonality and event windows such as sales periods. Regional differences are modelled with hierarchical priors so that sparse regions borrow strength from related ones.

2.2Decision core

Effect estimation. For each candidate treatment the core estimates the change in outcome per customer group, with honest confidence intervals⁴. These estimates decide where to personalise and where to keep a standard experience: a treatment is suppressed for a group when its interval is wide or when it would breach an exposure constraint.

Contextual bandits. Treatments are allocated by contextual bandits⁵. The reward combines conversion and contribution margin; penalties apply for latency breaches and exposure imbalances. Confidence bounds are set per group, so that exploration is smaller where outcomes are volatile.

Conservative learning from logs. A new policy learned from logged data can do harm if unobserved factors drove both past treatment and outcome. Kallus and Zhou propose learning a policy that minimises worst-case regret against the current baseline policy over a bounded degree of such confounding⁶. The BOM adopts that rule for groups with sparse data: a new policy replaces the default only where it improves on it under the confounding allowance.

Exposure floors. A reward-maximising bandit can starve some options or groups of exposure. Following the formulation of a bandit in which each arm must receive a minimum share of pulls⁷, the BOM sets exposure floors per customer group and region.

Damping. Exposure rates change gradually, so that a short burst of noisy reward does not swing the allocation. The rate limit is a design parameter.

Safety checks. Before execution every policy passes consent and residency checks, a latency budget, an exposure check and content policy filters. Rollouts go from shadow mode through staged ramps to full traffic, with automatic rollback on latency, exposure or anomaly triggers.

2.3Objective and definitions

The decision core scores a treatment with a weighted objective:

score = w1 · conversion + w2 · average order value − w3 · logistics cost − w4 · return risk

The weights are set by the merchant's operator and recorded with each policy version; we propose no default values. Relative uplift is defined as (CVR treatment − CVR control) / CVR control. The exposure gap across groups is the largest minus the smallest group uplift; it is reported alongside, not folded into, the objective, because there are several incompatible definitions of fairness and the choice among them is not a technical one⁸.

3Questions the design raises

The design rests on hypotheses we have not tested. We state them as research questions.

1. Do intent, friction and trust signals together predict conversion better than behavioural signals alone, and is the gain stable across regions?
2. Which allocation policies keep conversion and margin within acceptable bounds while meeting exposure floors and latency budgets?
3. Does continuous experimentation with sequential tests give stable effect estimates across regions with different traffic levels?
4. Does damping reduce the variance of outcomes during traffic spikes, and at what cost in reaction speed?
5. What governance keeps iteration fast without eroding shoppers' and merchants' trust?

4Evaluation framework

Answering these questions would need layered experiments: bandit policies running inside A/B tests whose control arm is a fixed holdout, with hypotheses and metrics pre-registered per group¹. New policies run in shadow mode first. Because promotions move fast and teams will monitor results continuously, analysis uses always-valid sequential tests⁹, and pre-experiment covariates reduce variance¹⁰.

Metric	Purpose	Method	Source
Conversion rate, revenue per session	Primary outcome	Holdout comparison with CUPED and sequential tests	10, 9
Contribution margin	Guard against buying conversion with discount	Holdout comparison	1
p95 and p99 decision latency	Protect the shopper's experience	On-path measurement	1
Exposure and uplift by group	Equity across customer groups	Gap between group estimates with intervals	4, 8
Fallback rate	Resilience of the decision path	On-path counters	Design metric
Uplift variance over rolling windows	Stability of the policy	Variance of windowed estimates	Design metric
Returns, cancellations, support contacts	Trust	Holdout comparison	2

Source: methods as cited in each row; the table holds no measured values.

5Hypothetical scenarios

The following describe intended behaviour, not results.

Cart recovery during a flash sale. If payment retries rose sharply during a sale while cart edits indicated high intent, the core would throttle risky offers and the bandit would favour payment paths with lower decline rates. Whether timeouts and abandonment would fall is for an experiment to show.

Trust repair after late deliveries. If delivery delays in a region lowered trust signals, the core would shift messaging toward accurate status and revised delivery promises, while exposure floors kept the change from concentrating on one group. Its effect on cancellations would need a holdout.

Degraded mobile networks. If latency alerts fired for a mobile segment, the execution layer would switch to lighter models and cached fallbacks. The conversion cost of that switch is unknown until measured.

A new market with sparse data. The core would reduce exploration, lean on hierarchical priors and the conservative learning rule⁶, and run longer sequential tests.

6Data and ethical posture

Data collection is limited to signals that serve intent, friction or trust estimation, and each signal is mapped to the decisions that use it. Identifiers are hashed before storage; retention is shorter for sensitive events such as payments and support conversations; joining behavioural data with sensitive attributes needs an approval. For analytics on small groups, where released aggregates could reveal individuals, the design calls for differential privacy, which bounds what any release discloses about one person¹¹.

Interface copy and nudges are reviewed before testing, and manipulative patterns (false scarcity, hidden costs, obstructed opt-outs) are excluded by policy. Recommendations carry explanations for operators, merchants and shoppers can contest automated decisions, and high-impact changes need human approval.

7Limitations

• No measurements. We report no conversion effects, latencies or variance figures. Every statement about behaviour is an argument from design and literature.
• Sparse groups. Small groups give wide intervals; the design falls back to standard experiences for them, which avoids harm but also withholds gains.
• Non-stationarity. Abrupt changes such as policy changes or economic shocks can invalidate learned policies faster than damping and drift checks react.
• Attribution. When a layout, an offer and a delivery promise all touch one order, the design does not separate their contributions.
• Fairness definition. Exposure floors and reported uplift gaps are two of many fairness definitions, and they can conflict⁸.
• Confounding allowance. The conservative learning rule is only as safe as the confounding bound chosen, and that bound cannot be estimated from the data⁶.
• Human factors. Signals cannot replace qualitative research on why shoppers hesitate; the design gives that research no formal role yet.

8Conclusion

We have argued that conversion is an outcome of the whole commerce system and described a design that acts on that view: signals for intent, friction and trust; effect estimates per group; bandit allocation under exposure, latency and consent constraints; a conservative rule for learning from logs; and an evaluation framework set out in advance. Whether the design improves conversion, and at what cost to margin, equity or speed, is an empirical question this paper does not answer.