Every session on a DTC storefront generates a stream of behavioral events. Mouse position, scroll depth, hover duration, click sequence, filter interactions, time between page loads. Most analytics platforms capture some subset of these. Very few actually rank them by predictive value for the thing that matters most: purchase within the session.
We have been building Cartlyzer's inference model on live session data since we started instrumenting storefronts. Across the brands using the product, we have accumulated a reasonably large body of labeled sessions: sessions where we know the behavioral event stream that preceded a purchase, and sessions where we know the behavioral stream that preceded exit without purchase. That labeled data lets us ask which signals, measured in real time, actually move the needle on conversion probability estimation.
Here is what we have found. Not a theoretical framework: what the data says.
Tier 1: High-Confidence Purchase Predictors
Add-to-Cart (even without checkout completion)
This sounds obvious, but it is worth stating precisely. The act of adding a product to cart, even if the session does not immediately proceed to checkout, is the strongest single in-session predictor of eventual purchase we measure. Sessions that contain at least one add-to-cart event convert at dramatically higher rates than those that do not, even when controlling for session length and page views.
The practical implication: when you see an add-to-cart event early in a session, the inference priority shifts from "which product should be ranked first" to "how do I keep this session on the path to checkout." The merchandising response should change. Cross-sell and complementary product surfacing becomes more valuable than continued top-of-grid rank optimization.
Hover Duration Over a Specific Product Card (Desktop)
On desktop sessions, cursor hover time over individual product cards is a strong purchase predictor when it crosses the 2.5 to 3 second threshold. Below that threshold, hover is largely noise: the cursor passes over cards as the user scrolls. Above that threshold, the hover represents deliberate attention. The user stopped, looked at the card, considered it.
Hover over a product followed by a PDP visit converts at a substantially higher rate than PDP visits without prior hover attention. The hover predicts the visit. The visit predicts the purchase. Hover duration is therefore an early-stage leading indicator that updates the per-product conversion probability before the user has even clicked anything.
Tier 2: Useful Signals with Context Dependence
Category Filter Usage
Using a filter is a strong intent signal, but its predictive value depends heavily on which filter and what point in the session. A color filter applied within the first 30 seconds of a session suggests high-specificity product intent and correlates well with purchase. A size filter applied after a PDP visit correlates very strongly with intent to purchase that specific product. A price filter, however, is more ambiguous: it may indicate price sensitivity and potential abandonment rather than purchase proximity.
We track filter type, filter sequence, and filter timing as separate signals rather than treating all filter usage as equivalent. The combination "category filter, then PDP visit within 2 clicks" is a much stronger purchase predictor than generic filter usage.
Return to Collection Page After PDP
When a user visits a product detail page and then navigates back to the collection page, the resulting collection page session is behaviorally different from a fresh collection landing. The user rejected or deferred the first product and is continuing to search. The collection page ranking model should respond to this by promoting alternatives in the same price range and category, not re-showing the same first product the user just walked away from.
This signal is medium-strength because it could indicate either continued high intent (the user is shopping seriously and narrowing) or deteriorating intent (the user looked at the product and found something off that now extends to the category). Pairing it with time-on-PDP helps disambiguate: long PDP visit followed by return to collection is a positive signal. Sub-15-second PDP visit followed by return is negative.
Scroll Depth Within a PDP
Scroll depth on a product detail page correlates with purchase, but less strongly than most DTC brands assume. Users who scroll deep into a PDP are engaging with the page, but deep scroll is also correlated with hesitation and information-seeking. A user who scrolls to the reviews section, spends time there, and then returns to the product images is in consideration mode, not commitment mode. That is different from a user who scrolls only halfway down, looks at the size chart, and immediately adds to cart.
PDP scroll depth is a useful signal but should not be treated as a strong purchase predictor in isolation. It is most valuable when combined with interaction events on the PDP itself: swatch selection, size chart engagement, review filter use.
Tier 3: Commonly Overrated Signals
Time on Site
Total time-on-site is one of the most commonly cited engagement metrics in e-commerce analytics and one of the least predictive of same-session purchase. Long sessions can mean high engagement, or they can mean confusion, comparison-shopping across tabs, or distraction. In our labeled session data, time-on-site alone is a weak predictor of conversion, significantly weaker than add-to-cart or hover duration.
We are not saying time-on-site is useless as a metric. For brand engagement analysis and content effectiveness, it is relevant. But as a real-time purchase predictor that should influence collection page ranking in a live session, it does not carry enough signal. Brands that treat long sessions as high-intent sessions and short sessions as low-intent are reading the data backwards in many cases.
Scroll Depth on the Collection Page Itself
Scrolling deep into a collection page means the user has not found what they want in the top rows. That is not a positive signal for the products being shown. Deep scroll on a collection page is weakly correlated with exit without purchase in our data, particularly when it is not accompanied by any hover attention or filter usage. It suggests the page is not surfacing the right products early enough for that visitor.
The common mistake is to interpret deep scroll as high engagement. In a navigation context, deep scroll without interaction often means low relevance of visible products. The merchandising response should be to ensure the products most likely to match that visitor's intent appear higher in the grid, not to interpret their scrolling as a sign they are going to buy.
Building a Signal Hierarchy for Your Specific Context
The ranking above is based on aggregate patterns across multiple storefronts and categories. Your specific brand may have category-specific signal behaviors that differ. A skincare brand where users routinely spend time in the ingredients section of PDPs may find that ingredient-tab engagement is a stronger purchase predictor than our baseline suggests. A high-consideration furniture brand may find that saved-to-wishlist is far more predictive than it is for impulse-buy categories, because the purchase cycle is inherently longer.
The methodology for building your own signal hierarchy is straightforward: take a 90-day labeled session dataset, split by purchased vs. did-not-purchase, and compute the lift ratio of each signal type. Which events appear significantly more often in converting sessions, after controlling for session length? Rank by lift ratio, weight by measurability and latency, and you have a brand-specific signal hierarchy.
Cartlyzer does this continuously as new session data comes in, updating signal weights at the per-brand level rather than applying universal weights. What predicts purchase for an activewear brand's traffic is not identical to what predicts it for a home goods brand, even if the broad hierarchy is similar.