Session Duration and Depth: Reading the Metrics Your Analytics Actually Records

Average session duration looks like the most approachable metric in any analytics dashboard, which is precisely why it does the most damage. It is a mean of a highly skewed distribution, calculated from timestamps that your analytics platform may never receive, on sessions that get chopped by arbitrary timeouts. Before you optimize a landing page because “time on site” dipped eleven percent, understand what the number actually represents.

The mechanical problem is that duration is derived, not measured. Classic Universal Analytics computed session length as the timestamp of the last hit minus the timestamp of the first hit. That means the final pageview in every session contributes zero seconds, no matter how long someone actually read it. A user who lands on a long-form guide, reads it thoroughly, and leaves without clicking anything else registers a session duration of zero and a bounce. GA4 sidesteps the worst of this with engaged sessions, which fire after ten seconds, a conversion event, or two or more pageviews, but the underlying timestamp arithmetic hasn’t changed. A session with one pageview still tells you nothing about attention.

Then there is the sampling and sessionization layer. Session timeouts default to thirty minutes of inactivity, but a tab left open in a background window can keep firing heartbeat events and inflate a session into something that never happened. Conversely, someone who reads the same article for twenty minutes, opens a new tab, and returns will often be recorded as two sessions with zero depth. Mobile users switching between apps hit the same fragmentation. If you’re making decisions from a thirty-day lookback report with sampling enabled, you’re partly optimizing noise.

Scroll depth and engagement events fill the gap, but only if they’re implemented with intent. A scroll-triggered event at fifty percent is a proxy, not a truth, and on a page that loads above a cookie banner, real scroll depth gets suppressed. The better pattern is to pair a few scroll thresholds with a visibility-based timer. The Page Visibility API lets you pause the clock when the tab loses focus, which brings duration closer to actual attention rather than wall-clock time. That single change usually shrinks reported session duration and simultaneously makes it more honest.

Depth deserves its own definition. Pages per session is the crude version, and it’s badly skewed by paginated archives, faceted navigation, and infinite scroll implementations that fire virtual pageviews. If your site uses client-side routing, you need to be explicit about which transitions count as new pageviews. Otherwise, depth metrics reflect your front-end architecture more than user interest. The more diagnostic signal lives in distribution, not averages: how many sessions hit exactly one page, how many reach five, and what the exit page is for each cohort. Segment those curves by landing page template and acquisition channel, and the story sharpens fast. Sessions from organic search to a comparison article rarely behave like sessions from paid social to a product page.

Percentiles do more work here than the mean ever will. Pull the median and the ninetieth percentile for the same cohort and you’ll see whether a duration shift came from the middle of the distribution or from a handful of outlier sessions that never closed. Benchmarks from third-party tools are close to useless for this, because they aggregate across industries and traffic mixes that resemble your site only on paper. Build your own baseline, hold it for a rolling ninety days, and watch the shape of the curve rather than a single number.

Beware of the blending trap. Sitewide session duration averages across radically different user intents, and the mix shifts whenever your traffic sources do. A spike in top-of-funnel blog traffic will mechanically lower average duration while your product pages perform identically. Look at duration within channel and page type before you attribute a change to content quality or design.

Finally, treat duration as a hypothesis generator rather than a KPI. A rising number can mean better engagement or a broken checkout that traps people on one screen. Pair it with conversion rate, scroll completion, and event frequency before drawing conclusions. Export raw event data to BigQuery when you can, build the duration distribution yourself, and exclude internal traffic and known bot signatures at the source. The metric hasn’t gotten worse; it was always this ambiguous. Your job is to stop treating a derived, skewed average as a verdict.