Data & Decisions
Session Counts and User Counts Tell Different Stories: Choose the Right Unit
Choose between visits, recognized accounts, device keys and completed tasks before calculating a rate. Follow a complete synthetic mapping and corrected reporting memo, with explicit shared-device and anonymous-identity limits.
Use sessions to count defined visits, user identifiers to count recognized identities, and task records to count completed work. Name the decision first, then choose the unit and calculate its rate. One account can generate several sessions across devices, while one shared browser can contain activity from several accounts. Neither a session count nor a device count is automatically a count of real people.
This LandOffer.ai guide maps one original synthetic event log into several units, then corrects a completed reporting memo. The offline calculation was actually run; the product, account relationships and business decisions are fictional. No analytics property, cookies, real identities or private customer data were inspected or changed. The example is an authored counting exercise, not a Google Analytics implementation test.

Start with the question the unit can answer
“How often do people return to work?” “How many recognized accounts used the tool?” and “How many requests were saved?” require different counts. A visit-level report can describe repeated opportunities or interactions without establishing audience growth. An account-level report can describe recognized reach without showing the amount of work each account did.
Write the outcome as well. A page view, button click and persisted request are different events. If the decision concerns completed requests, use an authoritative saved-request record and a request identifier. A person can save two requests during one session; a single request might also require several sessions before it is saved.
Google's Analytics session guidance describes a session as a period of interaction, with a default timeout after 30 minutes of inactivity. That default does not mean every session lasts thirty minutes, and it is not a universal rule for every analytics system. The guide also explains session IDs and numbers and recommends combining a user identifier with a session ID when analyzing sessions outside Analytics.
Our fixture already contains assigned session tokens. We do not reconstruct sessions from inactivity, run the Google SDK or claim that the tokens are actual ga_session_id values. That boundary matters: a correct distinct-count query can still count the wrong sessions if the underlying sessionization is inappropriate or incomplete.
Define which kind of user you mean
A user label in a dashboard is a metric definition, not proof of a verified person. A browser identifier can change, and the same browser may be shared. A signed-in account can span several browsers. Shared accounts can combine people, while duplicate accounts can split one person into several records.
Google's user-metrics guidance distinguishes total, active, new and returning users, with different qualifying events or conditions. Therefore, do not compare an export's distinct identifiers with an active-user headline without checking both definitions. Our calculations below deliberately use custom names such as “recognized reporting keys,” not a claim to reproduce a GA4 user metric.
Record the identity rule actually available. Is the report device-based, based on authenticated account IDs, or using some other permitted reporting identity? Preserve anonymous traffic as an explicit category. Removing it may be useful for an authenticated-account question, but changes the population and must remain visible.
You do not need personal names to perform these checks. Use approved pseudonymous identifiers and aggregate reporting. Do not attempt to identify anonymous visitors, merge accounts or change tracking configuration just to make two totals match. Those actions require their own purpose, permission and implementation review.
Worked asset: map events to sessions and identities
The fictional Willow equipment-request tool has this one-day log. The analysis window is November 3, 2026, from 00:00 inclusive to November 4 at 00:00 exclusive, in UTC. All future dates and records are invented. Event IDs are unique; the only completed-work event is request_saved, with its own request ID.
Here a device key is a stipulated browser or installation label. It is not a verified physical-device inventory, so six such keys do not establish six pieces of hardware. K1 is stipulated to be the same signed-in account on D1, D2 and D3. D5 is a shared workstation browser used by K3 and K4 in separate sessions. D6 has no observed account ID; its relationship to the signed-in accounts is unknown.
| Visit start, UTC | Device key | Account on every event | Session token | Events | Saved request |
|---|---|---|---|---|---|
| 09:00 | D1 | K1 | 900 | 2 | R1 |
| 13:00 | D1 | K1 | 901 | 2 | None |
| 14:00 | D2 | K1 | 902 | 1 | None |
| 16:00 | D3 | K1 | 903 | 3 | R2 |
| 09:00 | D4 | K2 | 900 | 2 | R3 |
| 10:00 | D5 | K3 | 904 | 1 | None |
| 15:00 | D5 | K4 | 905 | 2 | R4 |
| 11:00 | D6 | None | 906 | 1 | None |
For this custom exercise, a session key is the pair device key plus session token. Token 900 occurs on two different device keys, so counting distinct tokens alone would collapse two visits. The eight pairs remain distinct. The session tokens are illustrative strings, not Google-generated timestamp values. The script verifies eight composite pairs and deliberately reports the incorrect token-only count of seven for comparison.
The reporting-key rule uses the account ID when present and otherwise a namespaced anonymous device key. That yields K1, K2, K3, K4 and anonymous:D6: five recognized keys. It does not imply that five different people visited. No historical identity stitching or inference about D6 is performed.
The actual offline Python calculation expands the table into fourteen event rows and counts them under these rules:
| Unit counted once within the window | Result | What the number establishes here |
|---|---|---|
| Unique event IDs | 14 | Recorded interactions in this synthetic log |
| Device-plus-session pairs | 8 | Visits under the assigned session definition |
| Session tokens alone | 7 | An intentionally incomplete key; it merges the two 900 visits |
| Device keys | 6 | Distinct browser/installation labels |
| Nonempty signed-in account IDs | 4 | Distinct recognized authenticated accounts |
| Account-or-anonymous reporting keys | 5 | Four accounts plus one anonymous-device surrogate |
| Distinct saved request IDs | 4 | R1, R2, R3 and R4, not four distinct savers |
K1 contributes four visits, three device keys and two saved requests. That is consistent with one account returning across devices; it is not four newly acquired users. D5 contributes one device key but two account keys. The anonymous D6 row has no known link to an account. Together with the unverified account-to-person assumption, it leaves an exact real-person count unknown.
Keep the numerator and denominator at the same unit
Four of the eight session pairs contain a saved-request event: 50% of sessions. Four of the six device keys contain one: 66.7% of device keys. Three of the five reporting keys contain one: 60% of reporting keys. Among authenticated accounts only, three of four contain one: 75%.
Those are four different descriptive questions. None is an ordered funnel conversion rate, because this exercise asks only whether the unit contains at least one saved event during the window. It does not impose a required preceding action, order or completion duration. Keep such constraints in a separately defined funnel.
Dividing four saved requests by five reporting keys gives 0.8 requests per reporting key, not an 80% user success rate. K1 saved two requests, so request totals and distinct savers differ. A rate labeled “share of accounts that saved” must deduplicate the numerator by account, just as its denominator does.
The authenticated-only 75% also excludes the anonymous visit. It cannot replace the all-key 60% as if the difference were improvement. Keeping both population labels prevents a smaller authenticated denominator from appearing to improve the product. Neither fraction demonstrates that signing in causes more successful work.
Treat cross-device identity as a reporting assumption
Google's reporting-identity documentation describes User-ID, device identifiers and modeling, and distinguishes device-based, observed and blended options. A property can therefore present recognized activity differently depending on the identity spaces available. That documentation is not evidence that a particular organization collected valid IDs or configured its property correctly.
In the synthetic fixture, collapsing D1, D2 and D3 into K1 reduces recognized units, while separating the two accounts on shared D5 increases them relative to device counting. The resulting six-versus-five difference is a change of unit over identical activity, not evidence that one visitor disappeared or that tracking lost an event.
Do not repair an unexplained discrepancy by forcing every anonymous row into the nearest signed-in account. Validate the available identifier contract, consent constraints and scope first. This article chooses a counting unit; a production identity-merge rollout or historical reassignment needs separate design and verification.

Worked asset: correct the reporting memo and choose an action
Analyst Lucan receives this fictional draft: “We acquired eight users. Four completed requests, so user conversion is 50%. Device tracking shows six users, but the account view shows five; two users must be missing.” The figures come from different units, and the draft adds unsupported acquisition and missing-person conclusions.
Lucan's completed replacement reads: “The synthetic one-day log contains eight assigned session pairs across six device keys. It contains four authenticated accounts plus one anonymous-device reporting key. Four sessions include a saved request; three authenticated accounts saved four distinct requests. These observations establish neither new-user acquisition nor an exact real-person count.”
He completes the decision record with product owner Mirela and engineer Pascal:
| Decision needed | Selected measure and result | Action and remaining boundary |
|---|---|---|
| How much request work occurred? | Four distinct saved request IDs | Use the saved-request ledger for workload reporting; Pascal must verify real persistence and duplicate-event handling before production use |
| How broadly did authenticated accounts participate? | Three of four accounts saved at least one request | Report the 75% authenticated-account fraction with its four-account denominator; exclude no anonymous record silently |
| How often were visits associated with saving? | Four of eight defined sessions contain a save | Retain 50% as a visit-level diagnostic; do not call it people reached or a causal success probability |
| Are we acquiring more people? | No first-visit history, account-creation history or comparable period is supplied | Mark acquisition unknown; Mirela defers an acquisition claim |
| Must identities be merged now? | K1's stipulated mapping explains the cross-device example; D6 remains unresolved | Preserve the anonymous bucket. Pascal reviews identifier validity separately, and Mirela keeps acquisition claims on hold; no merge or settings change is approved |
The completed fictional outcome is a corrected internal memo and a reporting hold on real acquisition claims. Lucan does not rewrite past dashboards or alter identifiers. Pascal's production checks remain pending; the arithmetic exercise does not turn them into verified instrumentation.
Freeze the definition before comparing periods
Store the date range, timezone, included events, exclusions, session key, identity rule and numerator-deduplication rule beside the report. Distinct counts are not generally additive: the same recognized key appearing on two days is still one key across the combined period. Summing daily users answers a different question from counting unique users over the full period.
Before interpreting a change, check whether session timeouts, event collection, login coverage, reporting identity or eligibility changed. A different count may come from a new measurement definition. An unexplained difference should remain an investigation, not a claim of growth, churn or missing people.
Choose the next action the records support: reconcile the keys, verify the task ledger or collect the missing comparison history. For more practical measurement records, continue with LandOffer.ai's workplace guides; keep each number's unit attached wherever it is reused.
Sources and Further Reading
- Google Analytics: About Analytics sessions — session initiation, inactivity timeout, session identifiers and external-analysis key considerations.
- Google Analytics: Understand user metrics — total, active, new and returning definitions; not a guarantee that a custom distinct count reproduces those metrics.
- Google Analytics: Reporting identity — User-ID, device identifiers, modeling and reporting options. All current primary bodies checked October 27, 2026; exact publication dates are not displayed. No authenticated property was tested or changed.


