Reading ecommerce customer journeys beyond the first product

It measures the relationship between the first identifiable product in a participant’s captured page events and the book that person reported selecting. We analyzed 11,460 eligible participants from a 2022 experiment conducted by Ilya Morozov and Anna Tuchman. Their online bookstore contained 100 ebooks. Participants lived in the United States and were recruited through Amazon Mechanical Turk.^[1]

This was an incentivized research task. Participants received a $3 book budget and had to select a book. After the bookstore task, they separately chose whether they would prefer the book plus any remaining budget or the full $3 in cash if they won a fulfillment lottery. The lottery gave each participant a 50% chance of fulfillment. Our outcome is the book reported in the survey, before using any information about that separate preference or fulfillment.^[2]

We connect that reported selection to product-page events captured during the matched survey interval. A mapped event is one whose label can be matched to a book in the experiment’s catalog. “First” means the first unambiguous captured product event in that interval. It does not establish the first product someone noticed, their first-ever view, or an event before their decision. The interval may include events after selection.

The original paper studied advertising content and search. We pool the experiment’s conditions to examine a different descriptive question: when the first product and the selected book differ, does the selected book appear elsewhere in the captured records? This comparison does not estimate an advertising effect or describe the conversion rate of an ordinary store.

A first-product match accounts for 72.4% of the analyzed participants, but the remaining cases describe two different situations. For 26.6%, another product appears first and the selected book is also present somewhere in the mapped records. For 1.0%, another product appears first and the selected book is absent from those mapped records.

The table counts each participant once. Its three rows contain 8,295, 3,048 and 117 people, respectively, and share the same denominator of 11,460. The first two rows both contain a record for the selected book. They differ only in whether that book is also the first identifiable product.

For a business reviewing product traffic, those distinctions change the next question. A different first product does not, by itself, mean that the eventual selection is missing from the event history. Equally, a match does not show that the first product caused the choice. The comparison tells us how two pieces of information line up, before we attach a commercial explanation.

A creator selling several products faces the same interpretation problem. Someone might encounter one item in a campaign and select another. A useful report preserves the product identities involved. It should not convert every mismatch into a failed product page, an assisted sale, or a lost customer without evidence for that particular conclusion.

No. Among the 3,165 participants whose first recorded product differed from their reported selection, 3,048 had the selected book somewhere in their mapped page events. That is 96.3% of this mismatch group. The other 117 participants account for 3.7% of the group.

This denominator is narrower than the previous table. The 96.3% figure answers a question about participants with different first and selected products; it is not a percentage of all participants, visitors or purchasers. It also does not say that 96.3% completed a purchase after viewing another product.

The selected book’s presence is useful evidence about the records, even though its timing relative to the decision is unresolved. It tells us that reducing the history to its first product discards an identifiable connection to the selection in most mismatch cases. A fuller record can distinguish that situation from one where no such connection is visible.

That is a reason to preserve a second field in an analysis: whether the selected item appears anywhere in the defined event interval. It is not a reason to assign revenue credit to earlier items. A store with reliable transaction times could ask a separate question about events before purchase, but this experiment’s matched survey intervals do not provide that validated cutoff.

The overall first-product match combines two very different record histories. Of the 7,656 participants with one distinct mapped product, 99.5% reported selecting that product. Among the 3,804 with two or more distinct mapped products, 17.8% selected the product recorded first.

The one-product group needs careful interpretation. If its only mapped product differs from the selection, the selection cannot also be present in the mapped set. That makes the middle cell exactly zero by definition. This is not evidence that people in that group never considered alternatives, and the high match rate is not a stand-alone discovery about decisive shoppers.

In the group with two or more mapped products, 80.1% have a different first product while the selected book appears elsewhere in their records. Another 2.1% have no mapped event for the selected book. These shares describe the group’s captured product identities, not a progression from browsing to buying.

For reporting, the implication is to show product breadth alongside the headline match rate. A report dominated by one-product histories is answering a different descriptive question from one with many multiple-product histories. In this study, breadth is observational and partly connected to the definitions themselves. More recorded products cannot be said to cause a person to change their choice.

The 117 participants without a mapped record for their selected book are not a single, fully explained category. For 89 of them, the interval also contains at least one page event with a blank label or no usable product match. The remaining 28 have no such retained event.

The 89 cases represent 76.1% of this selected-not-recorded group. An unidentified event could matter to the missing connection, but its label does not tell us which book it represents. It is not a verified product view, and we do not fill it in with the selected book merely because that would complete the story.

The 28 remaining cases are also worth keeping separate, with a small-count flag. Having no unidentified retained event does not prove that the captured history includes everything the person did. Selection from a listing, unobserved activity and the study’s task design remain possible explanations. The data do not establish which explanation applies.

For an agency or store owner, these categories support different questions about the available evidence. One concerns records that exist but cannot be matched to a product. The other concerns a selected item absent from otherwise identifiable records. Neither justifies telling a client that a particular product page was skipped or that tracking definitely failed.

Imagine a ceramics creator selling a mug, a bowl and a vase. The examples below are invented to explain the comparison; they are not customers or results from the bookstore study. Assume the maker has separately verified the selected product and has defined the event interval being reviewed.

In the first example, the mug is the first recorded product, and the person selects the mug. The first product and selection match. That alone leaves the reason for the choice open: product appeal, an earlier recommendation and prior familiarity would all require other evidence.

In the second example, the mug appears first, the bowl also appears in the interval, and the person selects the bowl. The first product differs, but the selected item is represented. Calling this simply a failed mug visit would omit information the records contain. Calling the mug an effective introduction to the bowl would add a causal story the comparison has not tested.

In the third example, the mug and vase appear in the mapped records, while the reported selection is the bowl. The selection is absent from the mapped set. If a further event has an unusable product label, retain that fact without guessing its identity. If there is no such event, leave the absent connection unresolved.

The practical decision is what evidence to ask for before revising a campaign or a product page. A match can be described directly. A different first product with the selection present deserves its own category. An absent selected-item record calls for checking the event and selection definitions before drawing conclusions about customer behavior. None of the three examples supplies a reason, on its own, to increase or cut a campaign budget.

A short brief can make a customer-journey discussion more precise. Start with the business question: are you examining which products appear in a history, whether a selected item is represented, or which activity happened before a verified purchase? Those questions require different evidence. Keep the observed association separate from any later experiment intended to explain it.

Outcome: name the event or answer that establishes selection or purchase. Write “reported selection” when that is what you have. For a transaction analysis, state how cancellations, refunds and duplicate transaction records are handled rather than assuming every purchase event is a fulfilled order.

Observation window: state when the history begins and ends, which clock it uses, and whether activity after the outcome is excluded. If there is no validated decision timestamp, do not describe the interval as a path leading to that decision. Retain ambiguous ties instead of inventing an order.

Product identity: specify how the same product is recognized across records and how variants are treated. Keep blank or unmatched identifiers visible in the accounting. A row that cannot be mapped to the catalog is different from a product event that was never observed.

Denominators: show the eligible group, the first-product match states, and the smaller group used for any conditional percentage. Preserve sparse counts and explain zeros forced by the definitions. This gives a reader enough information to distinguish a broad pattern from a percentage based on a small exception.

For a modern GA4 implementation, Google documents separate view_item and purchase events with items arrays. Each item requires item_id or item_name; Google recommends both. Consistent product identifiers make an item-level comparison easier to interpret. This is current implementation guidance, separate from the historical experiment’s custom page events, and is not a recommendation to change consent or attribution settings.^[5]

We used the authors’ public replication data pinned to a specific repository revision. The analysis covers the source’s two waves between May 19 and August 3, 2022, retaining the pause between waves. We fixed the linkage and coverage rules before calculating the first-product and selected-book combinations, then fixed the outcome protocol before reading those joint results.^[3]

The source contains 24,930 survey rows, including repeat attempts. After the specified exclusions for previews, comprehension, source dates and completion history, 18,373 attempts remained ready for linkage. The matching rules selected 12,813 uniquely linked attempts; 11,606 contained a usable reported book choice. Requiring an unambiguous first mapped product left 11,460 analyzed participants. These are different stages, not interchangeable sample sizes.

Linkage uses coupon codes, client identifiers and survey times. Shared or ambiguous matches are excluded, and a later attempt is not substituted because it has a more convenient outcome. We retain only the source’s exact Page View - Custom event type for the product comparison. Earliest events with conflicting product identities, or unidentified events before or tied with the first mapped product, cannot establish a first product and are excluded.^[4]

All estimates are unweighted descriptions of this analyzed cohort. We do not calculate population confidence intervals for a self-selected research-task sample. Cells below 30 participants are flagged; conditional percentages with fewer than 30 people in their denominator are suppressed in the downloads. A zero that follows from the definitions is identified separately.

The table reports both waves, a narrowly defined coupon-coded no-ad subset, and two alternative interval definitions. The first-product match is 70.6% in wave 1 and 74.6% in wave 2. The coupon subset contains 192 participants, with only one selected-not-recorded case. Its small exception cannot support a substantive claim about advertising, and the code identifies an intended condition rather than verified exposure.

Expanding each interval by 59 seconds at both ends and contracting it by 59 seconds triggers a fresh run of the whole matching process. The resulting cohorts contain 11,437 and 11,292 participants. Their first-product matches are 72.4% and 72.0%, respectively. Similar aggregate shares do not establish the timing of the selection decision; these checks assess sensitivity to the stated interval boundaries.

The downloadable tables also preserve comparisons on the same baseline participants, additions and exclusions under each interval rule, and the full joint counts. The reproduction package contains the code, fixed protocols, source manifest, aggregate outputs and README instructions. It contains no participant identifiers or raw participant records. These calculations and interpretations are by Influence IQ; they do not replace the original paper’s advertising analysis.^[6]