Article · Traffic quality

Traffic quality: investigate before you exclude

Preserve the signal and check the explanation before acting.

G2P Media · · 6 min read

Hands reviewing phone ad-placement crops beside an inspection checklist at a desk.

Start with the signal

Decide whether a quality concern requires investigation, temporary containment, or exclusion. Start by writing the observation without assigning a cause. A placement may show unusually concentrated click timing, weak downstream activity, a high uninstall rate, or a mismatch between reports. Each observation can justify a check. None should become a fraud finding solely because the performance is disappointing.

Describe the signal at the smallest useful scope: source, campaign, publisher or placement, market, platform, and period. Name the metric and denominator. A low purchase rate among acquired users is a different observation from a low purchase count on limited delivery. Preserve that distinction before asking a source to explain the traffic.

Separate inefficiency from evidence of invalid activity. A placement can bring real users who do not find the app useful. It can also deliver an audience outside the intended market or message. Those findings may justify changing buying conditions without establishing deception. A fraud conclusion needs evidence of the relevant mechanism and a record of how that evidence was verified.

Signal review table

The explanations below are investigation hypotheses, not diagnoses or universal detection rules.

SignalPossible explanationsEvidence to collectNext check
Concentrated click-to-install timingAttribution manipulation, timestamp handling, or a delivery patternClick and install timestamps, report definitions, and placement IDsCompare timing on a consistent basis and request the provider's assessment
Weak trial or purchase activityAudience mismatch, a product problem, or an event collection gapMature cohort events, app version, and transaction validationTrace the event and inspect the product journey
High reported uninstall ratePoor message fit, product friction, or uninstall measurement differencesAcquisition cohort, uninstall basis, and destination creativeReview the expectation set by the ad and the app experience
Source and measurement totals differDifferent windows, filters, freshness, or invalid-traffic handlingBoth exports, settings, and rejection reasonsReconcile definitions before excluding traffic

Use the quality signal framework to connect each observation to a specific check. Write what evidence would weaken the suspected explanation as well as what would support it. This helps prevent a review from collecting only the rows that make an initial suspicion look convincing.

Preserve the source and affected period

Save the evidence before editing the campaign or exclusion list. Export the relevant reporting rows with their filters, time zone, extraction time, and field definitions. Retain stable source and placement IDs, because display names may be insufficient to identify the affected inventory later. Keep original exports separate from calculations and review notes.

Record when the concern began and when it was first detected. Compare the affected period with a relevant earlier period under a consistent definition. Check whether the app release, audience, bid, creative, destination, or event mapping changed at the same time. The investigation needs that context before it can attribute the change to a publisher.

Limit collection to evidence the investigation requires and use the approved access controls for any user-level records. A working log can usually reference an export and its location rather than copying identifiers into a broadly shared document. Assign an owner to preserve the detailed evidence and answer follow-up questions.

Quality log fields

FieldWhat to preserve
Source and placementStable source, campaign, publisher, and placement IDs where available
PeriodAffected dates, comparison dates, time zone, and export cutoff
SignalMetric, numerator, denominator, and report used
EvidenceExport reference, relevant filters, and validation or rejection reasons
ConfidenceSuspected, supported, unresolved, or ruled out, with rationale
OwnerPerson responsible for the next evidence check
ActionContainment or exclusion scope, effective time, and review condition

Give each investigation an identifier and retain its revision history. If a source cannot supply a requested dimension, record that limitation. Missing placement detail should constrain the conclusion rather than be filled with a guess about which inventory was responsible.

Check measurement and product explanations

Trace the weak downstream event before judging the source on it. Confirm that the event fired, reached the measurement provider, and mapped to the intended report. Compare the event with the product or transaction record. The measurement data path can structure this review and identify who owns each stage.

Check whether cohorts have had equal time to produce the event. Separate a newly acquired audience from one with a longer observation period. Inspect the affected app version and market for a broken payment flow, unavailable feature, login failure, or localization issue. Treat these as conditions to verify from the app's evidence.

For uninstall signals, check what the timestamp represents. AppsFlyer's freshness documentation says its uninstall event time represents detection, rather than the actual moment of removal. Use the documented measurement basis before describing an exact interval between an install and an uninstall.

Review filters and rejection reasons as well. AppsFlyer's validation-rule documentation covers both targeting conditions and fraud-related conditions. It also explains that blocking attribution affects reporting, not whether the end user can use the app. A rejected attribution therefore needs its specific rule and interpretation attached; a rejection count alone does not explain every underlying event.

Ask the source and measurement owner to assess the same evidence package. A useful request identifies the affected inventory and period, describes the anomaly, and asks a concrete question about the discrepancy or detection mechanism. Keep each reply tied to its scope. A general assurance about the source does not resolve a placement-level finding.

Contain, verify and document

Choose a temporary action proportionate to the observed issue and its business consequence. A broken destination may require an immediate pause. An unresolved placement concern may justify limiting exposure while evidence is reviewed. Record why the action was selected, what inventory it affects, and when someone will reassess it.

Use the investigation workflow to keep the observation, evidence, response, and action together. Separate actions taken to control spending from conclusions about validity. Excluding a placement because it does not meet the app's quality criteria is a buying decision; it should be documented on that basis.

Our team's social chat case describes weekly Mintegral publisher pruning using high uninstall rates as a quality signal. The published Android campaign ran from September through December 2024. It does not establish that fraud caused those uninstall rates and supplies no aggregate uninstall reduction. Use the case as an example of a documented quality control, with those limits intact.

Before applying a broad rule, check its scope and possible false positives. Review whether a threshold would also capture acceptable traffic in a different market or app version. Where the selected tool provides a review or observation mode, verify its behavior in that tool's documentation before relying on it. Record the exact rule version and effective time if an exclusion is applied.

Set the conditions for the next decision

Define what would justify retaining, narrowing, or reversing the action. A return to normal reporting may resolve an event-path defect without resolving the original audience-quality question. Conversely, evidence of poor fit may justify continued exclusion even after suspected fraud is ruled out. Keep both findings visible rather than closing the investigation under a single vague status.

Investigation checklist

  • State the observed signal without asserting an unverified cause.
  • Preserve source, placement, period, filters, and original exports.
  • Record the metric denominator and relevant comparison population.
  • Check event collection, report freshness, and cohort maturity.
  • Review product, payment, creative, and destination changes.
  • Obtain the relevant validation reasons and provider assessment.
  • Distinguish inefficient traffic from evidence of invalid activity.
  • Document containment scope, rule version, and effective time.
  • Assign an owner, next review, and conditions for changing the action.

Close the log with the supported finding and remaining uncertainty. If evidence is insufficient, use an unresolved status and state which commercial action remains in place. Preserve the evidence needed for any source review, while treating credits or billing adjustments as a separate matter governed by the applicable agreement and provider process.

Plan the next test around your app.

Discuss your app