What a mobile commerce team should do, and what I would do differently today.
Results matter only when someone can act on them. This page turns the findings into four recommendations, says plainly how far they can be trusted, and then looks back at the study with a few more years of practice behind me.
Four recommendations
Each one rests only on results that held at p < .05, and follows the same line: what the study found, what it means, what to do.
Design for discovery, not only for search.
Browsing to discover predicts intent to return and recommend about three times as strongly as searching to decide. β = .45 against .16, both p < .05
The shopper who wanders is more valuable over time than the shopper who looks one thing up and leaves. Discovery features are retention features, not decoration.
- Give discovery real space: new arrivals, editorial collections, "you might also like" and category browsing.
- Track breadth of browsing (categories or products viewed per session) as a leading indicator of return visits.
- Do not judge discovery features by same-session conversion alone.
Put personalization and navigation ahead of "anywhere, anytime".
Usefulness is the only kind of value that raises exploration directly. Within it, customization and navigation carry the weight, and ubiquity adds nothing significant. weights .60, .42 and .09 (n.s.)
Constant access is now taken for granted. What shoppers still notice is whether the app knows them and whether they can find their way around it.
- Invest in the relevance of recommendations and offers, based on past behaviour and location.
- Treat search filters, categories and menus as core product, and test them with real tasks.
- Stop positioning convenience of access as a differentiator.
Use enjoyment and social features to build immersion, then protect it.
Enjoyment and social value do not move exploration directly. Both raise immersion, and immersion raises both kinds of exploration. Within social value, interactivity dominates. indirect effects p < .05; interactivity weight .85
A beautiful interface or a review section pays off only when it keeps the shopper absorbed. Anything that breaks concentration cancels the benefit.
- Make the app fast to respond and easy to talk back to: chat, questions and answers, feedback.
- Place reviews and ratings inside the browsing path, not on a separate screen.
- Audit interruptions: pop-ups, forced sign-ins and notifications that pull the shopper out.
Serve the two modes differently.
Immersion deepens focused searching more than twice as much as it widens discovery browsing, and the two modes have different starting points and goals. β = .54 against .24
A shopper who is narrowing down wants clarity. A shopper who is looking around wants variety. One layout rarely serves both well.
- For searchers: clean product pages, specifications, comparison tools and uncluttered design.
- For discoverers: richer visuals, variety and suggestions.
- Use entry point and behaviour (search box against home feed) to infer the mode.
How far to trust these results
What supports them
- Scales adapted from published research, pretested with experts and piloted before fieldwork.
- Every reflective scale met its reliability and validity thresholds.
- Bootstrapped confidence intervals, six control variables, and an out-of-sample prediction check.
- The main pattern is consistent across tests: usefulness acts directly, enjoyment and social value act through immersion, discovery drives return.
What calls for caution
- One questionnaire, one moment: the results show association, not cause and effect.
- An online panel in one country, so the findings are not population estimates.
- Several constructs are closely related, and two validity checks passed only narrowly.
- Intentions were measured, not actual return visits.
For a product team, that adds up to this: treat the recommendations as well-founded hypotheses to test in your own app, not as settled rules.
What I would do differently today
I completed this study in 2021. Reviewing it now, these are the six changes I would make, and they are the standards I apply to my work today.
Test for common method bias
Why. Every measure came from one questionnaire at one moment, and almost all constructs correlate between .60 and .88, including a control variable. Some of that is shared method, not substance. Today. Separate predictors and outcomes in the questionnaire, add a marker variable, and report a full collinearity test.
Report collinearity among the predictors
Why. The three kinds of value correlate at about .85 with each other and with immersion. That can make individual paths unstable: the negative, non-significant path from enjoyment to searching, next to a positive correlation of .73, is the classic sign. Today. Report inner VIFs and interpret such paths as suppression, not as findings.
Hold one decision rule and report effect sizes
Why. The thesis mixes p < .05 and p < .10 when calling hypotheses supported. Today. One threshold throughout, as on the results page of this case study, with f² effect sizes so that readers can judge practical importance and not only significance.
Sharpen the measure of social value
Why. Two interactivity items are about speed of response, which overlaps with usefulness, and interactivity carries most of the weight in social value. Today. Separate responsiveness from two-way communication and human presence, so that "social" means social.
Measure behaviour, not only perceptions of it
Why. The exploration items ask how shoppers feel about exploring, not what they did. Today. Pair the survey with app analytics (session depth, category breadth, search use) or run an experiment that varies one feature at a time.
Broaden and document the sample
Why. One online panel in one country, with no demographic profile reported. Today. Quotas or weighting against known population figures, data quality checks reported, and at least a second market before generalizing.
What this project demonstrates
- Design
Framing a question
Turning an open business curiosity into a model with testable paths, grounded in existing theory.
- Measure
Building instruments
Adapting and validating multi-item scales, with deliberate reflective and formative choices, a pretest and a pilot.
- Analyze
Advanced modelling
PLS-SEM with higher-order constructs, mediation, bootstrapping and predictive assessment. In commercial terms, a key driver analysis with the mechanism made visible.
- Communicate
Making it usable
Translating coefficients into recommendations, stating limits honestly, and reviewing my own work critically.