11 November 2026 · 9 min read

GA4 sessions: why the count does not match your traffic

Three numbers sit in front of most site owners at the end of a month. Traffic acquisition says 12,400 sessions. The pages report, added up row by row, says 19,000. A spreadsheet pulled from BigQuery says 12,700. Nobody is lying and nothing is broken.

Sessions in GA4 follow a small set of rules, and each one bends a number in a direction the report does not label. The rules are documented, the settings are visible, and the gap between two honest figures usually has a one-sentence explanation.

What starts a session, and what ends it

GA4 starts a session when a visitor views a page or a screen and no session is active. On an app, opening it in the foreground does the same job. From that point Google collects a session_start event and attaches two values to every following event in the session: ga_session_id, which is the start timestamp in seconds, and ga_session_number, which counts how many sessions that user has had on your property.

The end is simpler. A session times out after 30 minutes of inactivity by default, and Google states that there is no limit to how long a session can last. An app session begins timing out when the app moves to the background, and Google offers an extend_session parameter with a value of 1 for apps that genuinely keep working there, such as navigation or music players.

Two consequences follow. A visitor who reads for an hour without a break holds one session, so the metric measures visits rather than attention. And a tab left open while someone makes lunch will start a new session when they come back, because the inactivity clock does not care that the tab stayed in front of them.

The two clocks in the data stream settings

Both clocks live in the same place: Admin, Data Streams, the web stream, Configure tag settings, Show all, Adjust session timeout.

The first is the inactivity timeout, set in hours and minutes, with a maximum of 7 hours and 55 minutes. The second is the engaged session timer, set in seconds. A session counts as engaged if it lasts longer than 10 seconds, contains a key event, or includes at least two page views or screen views. Ten seconds is the default.

The second clock decides engagement rate and bounce rate, which are the same figure read from opposite ends. It also decides who lands in an engaged-session audience. Two properties serving identical traffic with different timers will publish different engagement rates, and neither is wrong.

This is where comparisons quietly break. If a client set the engaged timer to 60 seconds last year and your benchmark uses 10, their engagement rate is lower for reasons that have nothing to do with content. Check the setting before comparing two properties, and note the date whenever you change it, because the chart will show the change and a reader in six months will want to know why.

Why the page report adds up to more sessions than you have

The clearest source of a mismatch is the report most people use to build a session total. Google's developer documentation puts the mechanic in a table, and the table is worth reproducing.

One session: the visitor loads /home, goes to /products, comes back to /home, leaves.

Page pathSessionsViews
/home12
/products11
Total, deduplicated by GA413

Add the rows and you get two sessions. The property total is one. Both are correct, and the reason is the definition. When a page dimension is paired with Sessions, the metric counts the unique sessions in which that page was loaded at least once. A session that touches five pages appears in five rows. Views behaves differently, counting loads, so the return to /home adds a second view inside the same session.

Three habits keep this under control:

The same trap catches entry reports built on page path. A homepage entry figure taken that way counts every session that touched the homepage, including sessions that started somewhere else. Landing traffic needs the landing page dimension, not the page path dimension.

Estimates in reports, exact counts in BigQuery

GA4 does not count sessions one by one in the reporting interface. Since October 2021 it estimates the number of unique session IDs with an algorithm called HyperLogLog++, chosen because it stays accurate on large cardinalities at a fraction of the processing cost. Google's wording is that it estimates distinct counts for the most used metrics, sessions and active users among them.

BigQuery is a different machine. It has the time and resources to count exactly, so it does not use the approximation. Google's own example is worth keeping. Last week reads 1,463 sessions in reports against 1,501 in BigQuery. This week reads 1,828 against 1,876. Both pairs show a 25 percent increase. The absolute numbers differ, the direction does not.

Two habits follow. Compare like with like, so a report figure against a report figure. And when the question is "did traffic grow", the report figure answers it. When the question is "how many sessions exactly", BigQuery wins, and the unique key there is user_pseudo_id joined with the session ID, because session IDs are not unique across users on their own.

One edge is worth knowing before someone reports a data loss. Session metrics carry no data for date ranges that include days before October 2021. If a long-range comparison comes back with a gap at the start, the gap is a metric definition rather than a tracking failure. Historical session activity is still reachable in the events report by filtering on the event name session_start.

Session attribution is decided once

The session_start event carries the information that determines where the session is credited: the gclid, the UTM parameters, the referrer. GA4 keeps the non-direct last click model for sessions, and the lookback window for that model comes from the All other conversion events setting, which is 90 days by default.

Google's example makes the mechanic clear. A visitor arrives through a Google organic link on day 1, so the session is credited to google / organic. On day 68 the same visitor returns directly, and that session is also credited to google / organic.

Read that against the source and medium reports most people build. They describe the start of a session, and the credit survives later direct visits inside the lookback window. They do not describe where this person came from today. If that is the question, the user-scoped dimensions answer it, for example First user default channel group.

The safe pairing rule is short. User-scoped dimensions go with user metrics, session-scoped dimensions go with session metrics. Google's developer documentation flags mixed scopes as a common reporting mistake, and the symptom is a figure that describes neither the user nor the session.

Events that arrive without a session

Not every event carries a session. Google states that session ID and session number are attached automatically by gtag.js and the Firebase SDK, and are not included automatically in events from Measurement Protocol or Data Import.

That matters in three places. Offline conversions imported for a CRM loop belong to no session, so they never raise your session count. A pipeline sending events through Measurement Protocol produces events outside the session structure unless you supply the identifiers yourself. And a subproperty can filter out session_start, in which case GA4 still generates a session ID, so a session exists with no start event to explain it.

When someone compares a count of session_start events against the Sessions metric and finds a gap, this list is where it lives. Imports are the usual suspect, server-side sends come close behind, and a filter is the quiet one.

The keep-alive ping that inflates everything

There is an old trick, and Google has now written it down as bad practice. Send an event every few minutes so the session never times out. The documentation is direct: do not implement custom code to ping GA4 to prevent inactivity timeouts.

The reasons are measurable. A visitor who walks away is genuinely inactive, and holding the session open creates what Google calls zombie sessions, inflating average session duration so the moment interest was lost disappears from the data. It also pushes sessions past the engagement threshold, so engagement rate rises with no change in behaviour. And every ping is an event, which adds noise and, on an export, cost.

Some cases are real. A 45-minute embedded video, a long technical read, a dashboard someone watches without clicking. Google's advice there is meaningful progress events rather than blank pings: a video milestone, a section read, a filter applied. Pair them with the browser Page Visibility API so nothing fires while the tab is hidden.

const INTERVAL = 5 * 60 * 1000;
setInterval(() => {
  if (document.hidden) return;
  gtag('event', 'passive_engagement', { engagement_time_msec: INTERVAL });
}, INTERVAL);

Enhanced measurement already covers part of this ground, and knowing which part saves work. YouTube embeds send video_start, video_complete and progress milestones at 10, 25, 50 and 75 percent. Scroll sends a single event at 90 percent, not a series. Non-YouTube players and incremental reading progress need custom work.

Why the ads platform always reports more

Google Ads counts clicks. GA4 counts sessions. The click is recorded by Google before the browser reaches your site. The session is recorded by your tag after a redirect, after the page loads, and only if the tag runs and consent allows storage.

Between those two points sit a slow page, a consent banner that denies analytics storage, a script blocker, a container that failed to load. Each one removes a session and keeps the click. A gap is normal, and a gap that doubles overnight is not, because it usually traces back to a release, a consent change or a container edit.

What to check, in order

  1. Open the page report and confirm you did not sum its rows. If the question is a total, use the property-level figure.
  2. Check the session timeout and the engaged session timer for the stream, and check both properties when you compare two.
  3. Pull the same date range from reports and from BigQuery. Accept a small gap, and compare direction rather than digits.
  4. Count session_start events for the period against the Sessions metric. A wide gap points at imports, Measurement Protocol sends or filters.
  5. Search the container and the site bundle for timed events that exist only to hold sessions open, and remove them.
  6. Rebuild any entry report that used page path on top of the landing page dimension.

None of this is a defect. A session is a derived number, assembled from rules that Google publishes and settings that sit in plain sight, and the same property can produce three correct totals from three tools. What separates a useful dashboard from a confusing one is knowing which total each panel is showing, and saying so on the panel.

Session counts that stopped making sense are usually a setting rather than a bug. Book a free tracking audit and we will check the timers, the report pairings and the export against each other, then tell you which number to trust.