Historical Context
The One-Click Purchase concept was patented by Amazon in 1999 and marked a turning point in the development of e-commerce by eliminating key friction in the conversion funnel—the need for repeated data entry. In a mobile environment, where each additional screen reduces conversion by 20-30%, this feature is critical for impulse purchases. However, its implementation creates a methodological trap: users with saved cards systematically differ from other audiences in terms of higher trust in the platform and a history of repeat purchases, making simple group comparisons meaningless.
Problem Statement
During a phased rollout (e.g., first on iOS, then on Android), we face double endogeneity. Firstly, self-selection based on the presence of payment tokens correlates with loyalty and payment capacity. Secondly, different user base growth dynamics between platforms distort temporal trends. Directly comparing conversion rates between "one-click" and regular users shows a difference of 2-3 percentage points, but this reflects audience quality rather than the feature effect. It is necessary to separate the true incremental effect from the self-selection bias.
Detailed Solution
The optimal method is Difference-in-Differences (DiD) combined with Propensity Score Matching (PSM) or the Synthetic Control Method. The steps are as follows.
First, we form a cohort of "treated" users: those who have access to the feature at the time of rollout for their platform. For the control group, we use users with saved cards on the platform without the feature (e.g., Android users during the iOS experiment), matched through PSM based on pre-processed characteristics: purchase frequency, average check, session depth, and tenure.
Then we apply DiD: we compare the change in conversion (before and after) in the test group with the analogous change in the control group. This removes time-invariant user characteristics and platform effects. To strengthen causal inference, we use Instrumental Variables (IV): the very fact of feature availability (determined by the operating system update date, not user choice) acts as an instrument for actual use of One-Click.
Additionally, we apply Regression Discontinuity Design (RDD) around the time threshold since the card was saved (excluding users who added a card <14 days ago) to eliminate anticipatory behavior before major purchases. The result is an estimate of the Local Average Treatment Effect (LATE) for users ready for impulse purchases with reduced friction.
To isolate the effect of One-Click Purchase, a quasi-experimental design should be used, combining Difference-in-Differences and Propensity Score Matching. The key step is leveraging the phased rollout across platforms as a natural experiment, where the temporary availability of the feature serves as an instrumental variable.
First, we perform matching of users with saved cards across platforms based on historical metrics (LTV, session frequency, category preferences). Then we calculate the difference in conversion differentials before and after the feature activation. To adjust for varied propensity for use, we employ a two-stage least squares method (2SLS), where the first equation predicts the likelihood of feature use based on its availability, and the second predicts conversion based on the predicted use.
It is important to stratify the analysis by product categories: for impulse items (accessories, cosmetics), a high effect on reducing cart abandonment is expected, while for considered purchases (electronics), the effect is minimal. The final metric is the incremental increase in conversion precisely due to the reduction in checkout time, rather than the self-selection of the audience.
In the marketplace "Bystruta", a plan was made to implement One-Click Purchase to reduce the share of abandoned carts on mobile devices. The feature required prior saving of the card and address. The product manager insisted on a phased rollout: first iOS (65% of the audience), and six weeks later Android, as integration with Apple Pay was technically simpler.
Considered Option 1: Direct Comparison of Conversion An analyst suggested simply comparing the purchase conversion between users with active One-Click and without it for a month after release. Pros: immediate result, clear business metric. Cons: catastrophic self-selection bias—users with saved cards had already made 3+ purchases previously and demonstrated a baseline conversion of 4.2% compared to 1.8% for others. The 2.4 p.p. difference reflected loyalty, not the effect of the feature. The option was rejected due to invalidity.
Considered Option 2: Classic A/B Test with Forced Disablement A technical lead suggested conducting a clean experiment: randomly disabling One-Click for half of the eligible users on iOS. Pros: gold standard for causal inference. Cons: legal risks (breaching expectations of users with saved data), ethical problems (deliberately worsening UX for loyal customers), and technical difficulty of "forgetting" saved tokens at the frontend level. The product committee rejected the option as unacceptable for the business.
Chosen Solution: DiD Quasi-Experiment with Geographic Stratification The analytics team chose the approach with Propensity Score Matching and Difference-in-Differences. For each iOS user granted access to the feature in week 1, a "twin" was matched from Android users with similar history (±10% on LTV, ±1 on number of orders in 90 days) from a region with the same product availability. Windows of 4 weeks before and after release were analyzed.
Result: Naive comparisons showed an increase in conversion by +2.1 p.p., but the cleaned DiD estimate revealed a true effect of +0.7 p.p. for overall conversion and +1.4 p.p. for the "accessories and household goods" category (impulse purchases). The average check did not change statistically significantly. Based on the data, a decision was made to expand to Android and launch a campaign to encourage card saving for new users, increasing the share of eligible audience from 30% to 55% over the quarter.
How to handle anticipatory behavior when users save their card just before a planned major purchase, creating endogeneity of activation time?
Answer: This is an example of the Ashenfelter's dip effect in retail. Users often add payment methods in anticipation of well-known events (Black Friday, birthdays), so the observed high conversion after "saving a card" reflects pre-existing intentional demand rather than the effect of convenience. To neutralize, it is necessary to use a narrow window design: exclude from analysis users who saved a card in the ±7-day window around a purchase, or apply Regression Discontinuity regarding the threshold of the minimum check for saving a card. Alternatively, only historical users with saved cards (>30 days of payment method tenure) can be used as an eligible group, excluding "fresh" savers.
What is the difference between ITT (Intention-to-Treat) and ToT (Treatment-on-the-Treated) evaluations in the context of One-Click, and why is compliance important?
Answer: ITT measures the effect of the availability of the feature for all eligible users, including those who do not use it (dilution effect). ToT measures the effect directly on users who clicked the "Buy Now" button. Candidates often confuse these metrics, suggesting analyzing only actual users, which leads to selection bias—active users are already predisposed to purchasing. The correct approach is to assess LATE (Local Average Treatment Effect) through instrumental variables, where the fact of feature availability (platform rollout) instruments for actual use. This provides an effect for "compliers"—users who use One-Click precisely because it is available, not due to special preferences. It is important to understand that LATE does not generalize to the entire population if compliance correlates with characteristics (e.g., younger users are more likely to use express purchases).
Why can the implementation of One-Click artificially reduce the effectiveness of the organic channel in last-click attribution, and how to diagnose it?
Answer: One-Click compresses the time window between recognizing the need and making a purchase (consideration window). Without friction, a user who sees a product on Instagram buys immediately within the session, not returning through a search engine the next day. In last-click attribution models, this order is assigned to the paid channel (social), whereas previously it might have been attributed to organic (search). Candidates often interpret the decline in organic share as a negative signal, while it is a measurement artifact. To diagnose, it is necessary to apply Marketing Mix Modeling (MMM) at the level of geographic segments (where rollout occurred at different times) or analyze blended CAC and overall LTV/CAC ratio, not channel decomposition. It is also useful to measure time-to-purchase—a reduction confirms the mechanism of channel substitution, not a loss of organic demand.