Measuring consumer decisions mediated by recommender systems and AI assistants: a design

1Context and motivation

A shopper on a large online store rarely sees a neutral list of products. A ranking model orders the feed, a recommender fills the "similar items" rail, a model may summarise the reviews, a price may be personalised, and a conversational assistant may answer questions and assemble a basket. Each of these layers changes what the shopper sees and how much effort a decision takes. Each can also change how far the shopper relies on the store's judgement rather than their own.

Most stores evaluate these layers one at a time, against click-through or conversion. That leaves three questions unanswered. Does an intervention help the shopper reach a better decision, or only a faster one? Does it shift reliance and trust in ways that later reverse? Does it narrow what shoppers are shown? This paper describes a measurement architecture designed to answer those questions. We describe the design and argue each part from published work. We have not deployed it and report no outcomes.

2Related work

2.1Bounded rationality and choice overload

Consumers choose with limited information and limited capacity to compute, so their choice processes are simplified relative to the economist's model of global rationality¹. Large assortments make this limit visible: in field and laboratory experiments, people were more likely to buy jams or chocolates when offered six options than when offered 24 or 30, and reported greater satisfaction with what they chose². An AI layer that filters a large assortment acts directly on this constraint.

2.2Interactive decision aids and recommenders

In a controlled experiment with a simulated online store, a recommendation agent built from a shopper's stated attribute weights reduced search effort, produced smaller but better consideration sets, and improved decision quality; a comparison matrix had similar effects on consideration sets³. Recommenders also have system-level effects. Analytical models and simulations show that common recommenders, which rely on past sales and ratings, can reduce sales diversity through a rich-get-richer effect for products that are already popular⁴. A measurement design therefore has to track exposure diversity as well as individual decision outcomes.

2.3Trust in automation and in algorithms

Online purchase intention depends both on the perceived usefulness and ease of use of the site and on trust in the e-vendor⁵. When the assistant is an algorithm, trust governs reliance: people rely on automation according to their trust in it, especially when the system is too complex to understand fully, and the design goal is appropriate reliance rather than maximal reliance⁶. Evidence on how people respond to algorithmic advice points both ways. People lose confidence in an algorithmic forecaster faster than in a human one after seeing it make the same mistake⁷. Yet lay people in six experiments adhered more to advice they believed came from an algorithm than from a person, while experienced forecasters relied on it less⁸. Neither aversion nor appreciation can be assumed; reliance has to be measured, and measured over time, because it responds to observed errors.

Transparency is not a simple remedy. In a field experiment on an algorithmic grading interface, people whose expectations were violated trusted the system less unless the algorithm was explained, but too much information also eroded trust⁹.

2.4Choice architecture, digital nudges and dark patterns

Interface design elements guide choices in digital environments, which makes every ranking and default a form of digital nudge¹⁰. The same levers can be used against the user. An automated crawl of about 11,000 shopping websites found 1,818 instances of dark patterns across 15 types, designs that steer users into purchases or disclosures they would not otherwise make¹¹. In the European Union, providers of online platforms may not design interfaces that deceive or manipulate users or materially impair their ability to make free and informed decisions, and platforms that use recommender systems must set out the main parameters of those systems and any options users have to modify them¹².

2.5Privacy behaviour

Privacy concern is uncertain, context-dependent and malleable, and it can be shaped by the commercial interests that collect the data¹³. A system that measures behaviour at fine grain changes the context in which that behaviour happens, so the measurement design itself is part of what must be governed.

3Design: a behaviour measurement architecture

We call the design a consumer behaviour measurement architecture. It has four parts: an intervention log, a state representation, a set of estimators, and a reporting and policy layer.

3.1Intervention log

Every algorithmic intervention is logged as a first-class event, not inferred afterwards. A record holds the intervention type (ranking variant, recommendation slot, review summary shown, assistant turn, price rule applied, explanation shown), the model and policy version, the candidate set and the propensity with which the served option was chosen. Logging propensities is a design requirement because the off-policy estimators below cannot be computed without them¹⁴. Behavioural events (views, searches, impressions, clicks, carts, checkouts, returns) and context (device, channel, region) are joined to interventions by session and request identifiers. Optional self-report items (a one-question trust or relevance prompt) are sampled sparingly and stored with their sampling rule.

3.2State representation

The architecture keeps three levels of state.

• Session state: search depth, number of products compared, whether the shopper read raw reviews or a summary, whether the assistant was used, and time from first view to decision. These operationalise effort and the screening and comparison stages that decision aids act on³.
• User state: a reliance index (the share of decisions that start from an algorithmic surface), its trajectory after an observed recommendation error, and self-reported trust where collected. Reliance after error is tracked because aversion is triggered by seeing the algorithm err⁷.
• Cohort state: distributions of the user states by cohort and region, used for reporting rather than for targeting.

3.3Estimators

• Experiment analysis. Interventions are evaluated in randomised controlled experiments with pre-declared target and guardrail metrics¹⁵.
• Off-policy evaluation. Before a new ranking or assistant policy is exposed, its value is estimated from logged data with a doubly robust estimator, which is accurate when either the reward model or the logging-policy model is good¹⁴.
• Reliance dynamics. A model of how the reliance index moves after correct and incorrect recommendations, so that a short-run gain in reliance can be checked against later decline^{6, 7}.
• Exposure diversity. Concentration of impressions and sales across the catalogue, compared between treatment and control, because recommenders can narrow sales diversity⁴.

3.4Reporting and policy layer

The outputs are dashboards of effort, reliance, trust and diversity per intervention; a register of interventions with their parameters in plain language, which also serves the recommender transparency duty where it applies¹²; and policy rules. The policy rules are our design choices: every personalised surface offers a way to see why an item appears and a non-personalised alternative; price personalisation is off by default; and any interface change is screened against a dark-pattern checklist before testing¹¹.

Table 1 maps each component to the problem it addresses and its evidence.

Component	Problem it addresses	Evidence
Effort and comparison-depth metrics	Conversion alone hides whether a decision aid improved decision quality	3
Assortment-size covariate	Large choice sets can reduce purchase and satisfaction	2
Reliance index with post-error trajectory	Reliance falls after visible algorithm errors	7, 6
Source-of-advice experiments	People may over- or under-weight algorithmic advice	8
Trust measures beside usefulness	Purchase intention depends on trust in the vendor	5
Explanation-depth variants	Too little or too much transparency can both lower trust	9
Exposure diversity metric	Recommenders can concentrate sales on popular items	4
Dark-pattern screening	Interface design can steer users against their interest	11, 10
Propensity logging and doubly robust evaluation	New policies cannot be evaluated from biased logs	14
Plain-language parameter register	Recommender parameters must be disclosed in some jurisdictions	12

Source: our design, with the works cited in the table.

4Evaluation protocol

Before any claim about an intervention's effect on shoppers could be made, the following protocol would apply.

• Pre-registration. Hypothesis, target metric, guardrails (return rate, complaint rate, opt-out rate from personalisation) and minimum detectable effect are recorded before launch¹⁵.
• Offline first. A candidate policy is scored with the doubly robust estimator on logged data; only policies that pass offline enter an online test¹⁴.
• Decision quality, not only conversion. Where a quality proxy exists (later return, repeat purchase of the same item, self-reported satisfaction), it is a co-primary metric, following the distinction between decision effort and decision quality³.
• Longitudinal readout. Reliance and trust are read out at the end of the test and again after a holdback period, since trust is dynamic⁶.
• Diversity guardrail. A treatment that raises conversion but concentrates exposure beyond a set threshold is flagged for review⁴. We leave the threshold as a parameter; it needs a measured baseline.

5A worked power model

Reporting effects separately by region multiplies the traffic an experiment needs. Table 2 makes the cost concrete with the standard two-sample approximation for a difference in proportions: n per arm is about 2(z(1 minus alpha/2) + z(power))^2 p(1 minus p) / d^2, where p is the baseline conversion rate and d is the absolute effect to detect. With a two-sided alpha of 0.05 and power of 0.8, the constant 2(1.96 + 0.84)^2 is about 15.7. The baseline rates and relative effects are illustrative inputs, not measurements.

Baseline conversion p (input)	Relative effect (input)	Absolute effect d	Users per arm, pooled	Users per arm, four regions each powered
3%	5%	0.15 percentage points	203,000	812,000
3%	10%	0.30 percentage points	51,000	203,000
1%	10%	0.10 percentage points	155,000	622,000

Source: Worked model; inputs in the table.

The arithmetic has two consequences for the design. First, region-level claims about behaviour need four times the traffic of a pooled claim at the same sensitivity, so the architecture reports pooled effects by default and treats regional differences as hypotheses to be tested where traffic allows. Second, small relative effects on low baseline rates are expensive to detect at all, which argues for variance-reducing designs and for choosing metrics with higher base rates where they are valid proxies¹⁵.

6Regional variation as a hypothesis

The original version of this paper asserted how shoppers in different regions respond to AI features. We have no measurements to support those statements and found no cited source that does, so we removed them. What the design does assume is narrower. Legal duties differ by jurisdiction: in the European Union the transparency and interface-design duties above apply to online platforms¹². Privacy concern varies with context¹³, so consent and data use settings are configured per jurisdiction. Any difference in reliance, trust or response to explanations between regions is a quantity the architecture is built to estimate, subject to the power limits in Table 2.

7Ethics, privacy and governance

• Assistance versus manipulation. The same ranking and default levers that reduce effort can steer¹⁰. The design therefore treats guardrails on returns, complaints and personalisation opt-outs as stopping rules, not secondary metrics.
• Appropriate, not maximal, reliance. Raising reliance is not a goal in itself. The target is reliance that tracks the system's actual accuracy⁶.
• Minimisation. Fine-grained behavioural logging changes the context it measures¹³. The design collects only fields an estimator uses, keeps records in their jurisdiction, and samples self-report prompts at low rates.
• Contestability. Shoppers can see why an item is shown and switch to a non-personalised view; merchants can see the parameters that rank their products.

8Illustrative scenarios

These scenarios are hypothetical and show how the parts would interact. They are not results.

Review summaries. A store testing model-written review summaries against raw reviews would read out time to decision, return rate and the share of shoppers who still open raw reviews. If the summary shortened decisions but raised returns, the protocol would treat that as a failed guardrail rather than a win.

An assistant that errs. If an assistant recommended an out-of-stock or incompatible item, the reliance index for affected users would be tracked over the following weeks. A sharp and lasting drop would be consistent with algorithm aversion⁷ and would argue for showing uncertainty in the assistant's answers.

Explanations. An experiment with no explanation, a short "why this item" note, and a detailed breakdown would test whether the middle option yields the most trust, as the transparency literature suggests it might⁹.

9Limitations

• No measurements. We have not deployed this architecture. Every effect it is designed to detect is unmeasured.
• Proxies. Decision quality is observed only through proxies such as returns and repeat purchase, which have other causes.
• Self-report. Trust items are sparse and subject to response bias; low response rates may not represent all shoppers.
• Power. As Table 2 shows, regional and cohort-level effects need large samples, and many stores will not have the traffic to estimate them.
• Off-policy limits. Doubly robust estimates depend on logged propensities and on overlap between the old and new policies; where overlap is poor, offline estimates are unreliable.
• External validity. Most cited behavioural findings come from specific laboratory or field settings, not from AI shopping assistants; they motivate what to measure but do not predict the result.
• Legal scope. The legal duties cited apply to some platforms in some jurisdictions; the design's policy rules go beyond them by choice.

10Conclusion

We have described an architecture for measuring how recommenders, summaries, assistants and personalised prices change consumer decisions: an intervention log with propensities, session and user states for effort and reliance, experimental and off-policy estimators, and a reporting layer that tracks trust and exposure diversity alongside conversion. Each component is tied to published work in Table 1. The design's purpose is to replace assertions about how AI changes shopping with estimates. The next step is to run the protocol on one surface, such as review summaries, and report what is measured.