1. LandOffer
  2. Blog
  3. All articles
  4. Your Analytics Report Uses Sampling: Check Whether the Decision Needs Exact Counts

Data & Decisions

Your Analytics Report Uses Sampling: Check Whether the Decision Needs Exact Counts

Separate sampled event estimates, approximate distinct counts and privacy-limited rows. Compare fixed-seed samples with a complete synthetic log, then use a filled decision record to choose whether the task needs exploration, a reliable threshold or exact reconciliation.

9 min readLandOffer team

Before using a sampled analytics report, decide whether the task needs a directional estimate, a threshold decision or an exact record count. Check the report's actual data-quality message, query scope and metric definition. Sampling, approximate distinct counting, privacy thresholding and grouped rows are different limitations. “Unsampled” does not automatically mean every displayed number is an exact count of complete business records.

This LandOffer.ai guide actually compares three fixed-seed event samples with one original full synthetic log, then completes a precision-and-retrieval decision table. All events, user keys, dates and decisions are invented. No Google Analytics property, API request, export configuration or real customer log was tested or changed. The local sampling exercise does not reproduce Google's sampling, HLL++ or privacy mechanisms.

One adult examines a small tray of pebbles and shells through a magnifying glass while another reads a paper card beside a larger specimen collection.

Read the limitation before explaining the discrepancy

Google's data-sampling guidance explains that Analytics may analyze a subset when query limits are exceeded, then scale results. It also says the data-quality indicator reports the portion of data used. Record the actual message and percentage alongside the date range, dimensions, metrics, filters and retrieval time.

Open and save the limitation message; badge color alone does not explain the report. Google's separate data-quality indicator guide describes where the indicator appears in reports and explorations. That is documentation of the interface, not an observation of your account or proof that its current report is sampled.

If an exported report already contains scaled estimates, do not multiply its numbers again by the inverse sampling fraction. The multiplication later in this article applies only to unweighted rows sampled directly from our known 1,000-event fixture, where every event has the same selection probability. A displayed report may already have performed that operation or use a different estimator.

Also distinguish the available analytics population from the real-world activity you care about. An unsampled count over collected events does not prove that instrumentation captured every action, removed every duplicate or represented every eligible user. State what the records cover before calling their total exact.

Separate sampling, approximation and withheld data

Sampling uses a subset of records. Approximate distinct counting can use all available records while estimating how many unique identifiers occurred. Google's sampling page notes that even unsampled reports can use HyperLogLog++ for metrics such as users and sessions. Do not infer exact distinct counts merely from a “100% of available data” message.

Google's threshold guidance explains that some data is withheld to reduce the risk of identifying individuals or sensitive information. Those system-defined thresholds are not an analyst-adjustable sampling percentage. A hidden row does not establish a zero count. Do not infer the identity or exact value of a withheld small group from surrounding totals.

Rows grouped into (other) represent another issue: several dimension values may have been combined. Missing visibility into an individual value is different from an event sample or a missing source record. Check the relevant message and definition instead of labeling every mismatch “sampling error.”

The distinction changes the response. A larger sample may help a sampling limitation. It does not necessarily remove distinct-count approximation, restore a withheld demographic group or unpack an aggregated row. An alternate data route may also have different identity, processing and collection boundaries.

Worked asset: compare samples with the supplied full log

The fictional Cairn design-review app log contains 1,000 unique event IDs on November 4, 2026, from 00:00 through 16:39 UTC. Its 200 synthetic user keys each appear five times. There are 200 download_completed events, three download_failed events and 797 detail views. The three failure event IDs are 997, 998 and 999.

Those future timestamps are authored inputs, not already observed product activity. User keys are identifiers in the fixture, not a verified count of people. The full-log counts are exact only for these supplied records; no collection completeness is established outside the exercise.

The Python 3.12.14 script draws simple random event samples without replacement. A uses seed 20261027 and 100 events; B uses seed 0 and 100 events; C uses seed 20261027 and 200 events. These fixed illustrative draws are not a Monte Carlo study of error frequencies. Their outcomes show possible discrepancies, not how often a similar discrepancy would occur in your report.

Supplied data Raw completion events Raw failure events Completion estimate Failure estimate
Full 1,000 events 200 3 Exact supplied count: 200 Exact supplied count: 3
Sample A: 100 events, 10% 11 1 110 10
Sample B: 100 events, 10% 18 0 180 0
Sample C: 200 events, 20% 31 1 155 5

For these unweighted raw event samples, the estimate is observed event count multiplied by 1,000 divided by sample size. A's eleven completions become 11 × 10 = 110; C's thirty-one become 31 × 5 = 155. The actual full CSV, sample CSVs, selected IDs, seeds and calculation output are retained with the local article.

Sample B contains no failure events even though the full supplied log contains three. A's one sampled failure scales to ten, while C's scales to five. An absent rare event in a sample does not prove the event never occurred. Likewise, scaling one sampled failure to ten estimates does not identify ten separate incident records; only one failure row was actually observed in that sample.

This fixture's larger sample moves the completion estimate closer to 200 than A does, but does not become exact. It is also less close than B's particular draw. Sample size alone cannot label the realized error in a report whose full-population value is unknown. These three observations do not prove that increasing a sample worsens accuracy or guarantees a monotonic improvement for every realized sample.

Do not scale unique users like additive event counts

Sample A contains 77 distinct synthetic user keys. Multiplying that by ten yields 770, even though the supplied population has only 200 distinct keys. B has 84 distinct keys and C has 128; their same naive scaling would produce 840 and 640.

The failure is the estimator, not a mysterious loss of users. A user key can appear on several event rows and therefore has a different inclusion probability from an individual event. You cannot generally estimate distinct keys by applying an event-count weight to a deduplicated sample count.

We compute exact Python set counts for the fixture, not HLL++ estimates. Google's distinct-count approximation is a separate documented mechanism. No algorithmic HLL++ error, vendor confidence interval or live report discrepancy was measured here.

Likewise, “10% of events used” is not a ±10% error band. Whether an estimate is precise enough depends on the sampling design, metric distribution, segment size and decision tolerance. These fixed draws do not provide a confidence interval. If the platform does not expose uncertainty sufficient for the decision, preserve that gap rather than inventing a bound from the sampling percentage.

One adult holds a small tray of pebbles and shells and a paper card while another points into a large specimen drawer.

Worked asset: choose precision from the decision

Analyst Tove and operations owner Bram complete the following fictional decision record before reusing the example's numbers. The app's investigation rule is an illustrative trigger at five or more failure events, not a statistical test or a general incident-management standard.

Work decision Precision needed Completed choice and owner
Explore which download questions deserve closer analysis Clearly labeled provisional estimates may guide questions; no resource commitment yet Tove retains A/B/C as sensitivity examples and marks the error bound unavailable. Bram approves exploration only, not a workload forecast
Decide whether the five-failure trigger was crossed A count reliable enough to classify either side of five Use the supplied full failure records: three. A's estimate of ten and C's five would cross the threshold, but the exact supplied count does not. Tove rejects the sampled threshold decision and attaches the three supplied failure IDs for Bram's scoped review
Confirm that no failures occurred Evidence capable of finding every relevant failure in the stated scope Reject B's zero as an all-clear. The three full-log failure IDs remain in the example's investigation packet
Reconcile completed downloads to delivery receipts Exact deduplicated event/receipt identity and matching boundaries Engineer Farren must obtain an approved authoritative ledger and compare IDs. The fixture has 200 completion events but no real receipt system; operational reconciliation remains unperformed
Investigate a privacy-withheld group An approved aggregate question that respects disclosure restrictions Bram narrows the business question or uses an approved aggregate route; no identity change or attempt to reconstruct a hidden individual group is authorized

The completed conclusion is deliberately bounded: “These synthetic records contain three failure events, so this example's five-event trigger is not crossed. There were still failures; a threshold result is not an all-clear. Sampled estimates cannot settle this decision. Keep operational reconciliation and any real production conclusion pending their own source evidence.”

No staff were assigned, incident closed or account setting changed. The table demonstrates why acceptable precision depends on the action. A provisional point estimate can help formulate a question while remaining inadequate for a threshold, exact reconciliation or privacy-sensitive claim.

Choose another route without changing the question silently

Google's Data API reporting-expectations guide, last updated October 22, 2026, identifies metadata for sampling, thresholding and (other) aggregation. Where available, retain samplingMetadatas, subjectToThresholding and dataLossFromOtherRow with the request. Their presence or status must be interpreted for that response; this article made no authenticated API call.

Possible next steps include a narrower query, an eligible unsampled reporting route or an approved event-level export. Availability depends on the property, method, access and actual collected data. Changing a date range changes the population; removing a dimension can change the metric's scope. Save the modified definition rather than pretending the new report answers an identical question.

For raw-event analysis, first verify available partitions, collection boundaries, deduplication, user identity and processing cutoff. An export is not automatically an exact copy of the reporting UI. Google's threshold guidance notes that Google signals data is not exported to BigQuery. A precise query over an incomplete or differently defined export still cannot establish a complete business count.

Avoid summing daily unique users to manufacture an unsampled period total. A key present on several days must be deduplicated across the whole requested period under the chosen identity rule. Event totals and distinct counts require different reconciliation logic.

Preserve the precision boundary in the handoff

Store the question, source, query definition, retrieval time, limitation messages, metric units and required precision with the result. This lets the recipient see why a number suitable for exploration was withheld from an exact reconciliation. State whether it is a displayed estimate, a sampled raw count, an approximate distinct count or an exact count over supplied records. Include which records or identities remain unavailable.

Before a consequential decision, have the owner confirm that the evidence can distinguish the relevant alternatives. If it cannot, hold that decision and request the missing source or uncertainty information. For more practical analysis records, continue with LandOffer.ai's workplace guides; keep the meaning of “exact” attached wherever the figure travels.

Sources and Further Reading