Four separate mechanisms change what you see before you see it. Three of them publish exact numbers. The fourth, the one that silently removes entire rows from your reports, publishes none at all and cannot be turned off.

None of this is hidden. It is all documented, on pages most people never open because the report looked fine.

This page collects the thresholds in one place, says which are adjustable, and explains what each one does to a B2B account specifically.

Thresholding: the one with no number

The mechanism that removes the most data is the only one without a published figure.

What it does, verbatim. Data thresholds are applied to prevent anyone viewing a report or exploration from inferring the identity or sensitive information of individual users based on demographics, interests, or other signals present in the data.

Whether you can change it. No. The documentation states that data thresholds are system defined and that you cannot adjust them.

When it fires. When a report or exploration includes demographic data, or audiences defined by demographics, or a search term dimension with too few users.

How you know. A data quality icon appears, stating that thresholding has been applied to one or more cards and that data will only display when it meets the minimum aggregation thresholds.

The threshold itself. Not published. The documentation refers to minimum aggregation thresholds and to having enough total users, and never gives a number.

Why that matters more in B2B than anywhere else. B2B audiences are small. A campaign reaching 400 people in a narrow industry is exactly the shape of data this mechanism suppresses, so the accounts most likely to lose rows are the ones least able to spare them.

The practical workaround. Widen the date range, remove demographic dimensions from the report, and avoid demographic audiences in any report you need complete. You are not defeating the threshold; you are staying above it.

Cardinality and the (other) row

The second mechanism aggregates rather than removes, which is harder to notice.

The definition, verbatim. The number of values assigned to a dimension is its cardinality.

The guidance figure. Any dimension with more than 500 values should be considered a high-cardinality dimension, as it will significantly increase the cardinality of all the tables storing that dimension.

And the caveat the vendor attaches. The 500 values per dimension is not a limit, but a guidance.

What the (other) row is, verbatim. A row that appears in a report, exploration or API response when the number of rows in a table exceeds the table’s row limit. The product surfaces only the most common dimension values and condenses less common values under it.

Where this bites in practice. Page path on a content-heavy site. Any custom dimension carrying a session identifier, a query string, a product code or a user identifier. Campaign names, if your naming convention includes dates or unique identifiers.

The consequence people miss. The (other) row is not a small residual. On a high-cardinality dimension it can hold most of your traffic, and any percentage you calculate from the visible rows is computed against a denominator that excludes it.

The fix. Reduce cardinality at collection time. Strip query strings, group page paths, and never put a unique identifier in a dimension you intend to report on.

The four documented mechanisms altering analytics data and whether each publishes a numeric thresholdTable listing the four documented mechanisms by which the dominant web analytics product alters data before presenting it to an operator, recording for each what it does, whether a numeric threshold is published, and whether the operator can adjust it. Data thresholding removes rows entirely and is applied, in the vendor’s words, to prevent anyone viewing a report or exploration from inferring the identity or sensitive information of individual users based on demographics, interests or other signals present in the data; it fires when a report includes demographic data, audiences defined by demographics, or a search term dimension with too few users; no numeric threshold is published anywhere, the documentation referring only to minimum aggregation thresholds and to having enough total users; and the vendor states that thresholds are system defined and cannot be adjusted. Cardinality condenses values rather than removing them, causing an other row to appear when the number of rows in a table exceeds that table’s row limit, with the product surfacing only the most common dimension values; the published figure is that any dimension with more than five hundred values should be considered high cardinality, but the vendor explicitly qualifies this as guidance rather than a limit; it is addressable by reducing cardinality at collection time. Quota limits restrict event-level queries to ten million events on standard properties and up to one billion on enterprise properties with an initial default of one hundred million per query; these figures are published and are fixed by property type. Data retention limits how far back explorations can reach, offering two months or fourteen months on standard properties, affecting only explorations and funnel reports rather than standard aggregated reports, and applying a fixed two-month period to age, gender and interest data regardless of the setting chosen; it is adjustable within the published options. The pattern across the four is that the mechanism removing the most data is the only one without a published threshold and the only one that cannot be adjusted.Four mechanisms, one without a numberMechanismWhat it doesPublished thresholdAdjustableThresholdingRemoves rowsNoneNo”system defined. You can’t adjust them.” Fires on demographics and low-user search terms.CardinalityCondenses into (other)500, “a guidance”At collection”not a limit, but a guidance”. Row limits themselves vary and are not published as one figure.Quota limitsCaps query scope10M events standardBy property typeUp to 1 billion on enterprise, with an initial default of 100 million per queryData retentionLimits how far back2 or 14 monthsYesAffects explorations and funnels only. Age, gender and interest are always capped at 2 months.Why this is worse in B2B than anywhere elseB2B audiences are small. A campaign reaching 400 people in one industry is exactly the shape thissuppresses, so the accounts least able to spare rows are the ones most likely to lose them.
The one that removes rows entirely is the one with no published threshold, and it cannot be turned off. Source : Google Analytics help documentation (2026)
Effect of the other row on percentages calculated from the visible rows of a high cardinality reportWorked illustration of what happens to percentages calculated from a report in which a high cardinality dimension has caused most values to be condensed into an other row. In the illustration, a report displays several named page paths with apparently precise session counts, while an other row holds a far larger number of sessions than any individual named row. Because the product surfaces only the most common dimension values and condenses less common values under the other row, which appears when the number of rows in a table exceeds that table’s row limit, an operator calculating the share of traffic attributable to any named page has two options and both are wrong: dividing by the sum of the visible rows produces a percentage computed against a denominator that excludes the majority of sessions and therefore overstates every visible row, while dividing by the true total produces percentages that sum to far less than one hundred percent with no explanation of the remainder. The vendor states that any dimension with more than five hundred values should be considered a high cardinality dimension as it will significantly increase the cardinality of all tables storing that dimension, while explicitly noting that the five hundred figure is not a limit but guidance. The dimensions where this most commonly occurs in practice are page path on content-heavy websites, any custom dimension carrying a session identifier, query string, product code or user identifier, and campaign names where the naming convention incorporates dates or unique identifiers. Because cardinality is created at the moment of collection, the remedy is to reduce it at collection time by stripping query strings, grouping page paths and never placing a unique identifier in a dimension intended for reporting.The (other) row is not a rounding residualPage pathSessionsShare, if you divide by the visible rows/pricing1,20030.0%/contact90022.5%/about70017.5%Three more named rows1,20030.0%(other)18,400excluded from the maths entirelyDivide by visible rows: every figure is inflated.Divide by the true total: they sum to 18%.Cardinality is created at collection. Strip query strings and identifiers before they become dimensions.
The visible rows look precise. Every percentage under them is computed against a denominator missing most of the data. Source : Google Analytics high cardinality documentation (2026)

Retention: the setting that only breaks one thing

The most misunderstood setting in the product, because its scope is narrower than its name suggests.

The options. Two months or fourteen months on a standard property. Twenty-six, thirty-eight or fifty months on enterprise properties.

What it affects, verbatim. The data retention setting does not affect standard aggregated reports, including primary and secondary dimensions, even if you create comparisons in them. It only affects explorations and funnel reports.

Which explains a common experience. Standard reports go back years. The moment you build an exploration to answer a real question, the data stops fourteen months ago. Nothing is broken; that is the documented behaviour.

The clause that overrides your setting. A two-month retention period is always applied to age, gender and interest data, regardless of what you choose.

What to do about it today. Check the setting. If it is on the two-month default, change it to fourteen. Retention is not retroactive, so every month you leave it is a month you will not get back.

And the deeper implication for a B2B business. A B2B sales cycle frequently exceeds fourteen months. The product cannot show you an exploration spanning your own buying cycle, at any setting, on a standard property. That is a structural limit, not a configuration error.

Modelling: when the numbers are estimates

Two different modelling systems, two different thresholds, and they are frequently confused.

Behavioural modelling in analytics, the first condition. The property collects at least 1,000 events per day with analytics storage denied, for at least 7 days.

The second condition. The property has at least 1,000 daily users sending events with analytics storage granted, for at least 7 of the previous 28 days.

Where the estimated data appears. Only under the blended reporting identity. Most exploration templates include only data from users who consented, with two exceptions that do include estimated users.

What that means practically. Two colleagues can open the same property, look at the same period, and see different totals, because they have different reporting identities selected. This is not a bug and there is no warning.

Conversion modelling in the ads product, which is separate. It has its own threshold: a daily ad click threshold of 700 ad clicks over a 7 day period, per country and domain grouping.

And the consequence for reporting. Modelled conversions appear in the Conversions column and flow into every downstream report using that data. There is no separate column, so the reported figure is a mixture of observed and estimated events.

The reconciliation problem this creates. Your analytics property and your ads account model different things, at different thresholds, with different eligibility rules. They were never going to agree, and now you know one of the reasons why.

The two distinct modelling systems and their separate eligibility thresholdsComparison of the two distinct modelling systems operating across the analytics and advertising products, which are frequently confused with one another despite having separate thresholds and separate effects. Behavioural modelling in the analytics product requires two conditions to be met simultaneously: the property must collect at least one thousand events per day with analytics storage denied for at least seven days, and the property must have at least one thousand daily users sending events with analytics storage granted for at least seven of the previous twenty-eight days. The resulting estimated data appears only when the reporting identity is set to blended, and most exploration templates include only data from users who consented to the use of identifiers, with two named exceptions that do include estimated users. The practical consequence is that two colleagues opening the same property and examining the same period may see different totals because they have different reporting identities selected, which is documented behaviour rather than a fault and carries no warning in the interface. Conversion modelling in the advertising product is a separate system with its own threshold, namely a daily advertisement click threshold of seven hundred advertisement clicks over a seven day period, per country and domain grouping. Its consequence for reporting is that modelled conversions appear within the Conversions column itself and flow into every downstream report using that data, with no separate column distinguishing them, so the reported conversion figure is a mixture of observed and estimated events. Because the analytics property and the advertising account model different quantities at different thresholds under different eligibility rules, the two were never going to report identical figures, and these distinct modelling systems constitute one documented reason among several for that divergence.Two modelling systems, often confusedBehavioural, in analyticsNeeds BOTH conditions:• 1,000+ events/day with storage denied,  for 7 days• 1,000+ daily users with storage granted,  for 7 of the previous 28 daysVisible only under the “blended” reporting identityConversions, in the ads productA different threshold entirely:• 700 ad clicks over a 7 day period,  per country and domain groupingModelled conversions appear IN theConversions column. No separate label.And they flow into every downstream reportA consequence with no warning in the interfaceTwo colleagues, same property, same dates, different totals, because their reporting identity differs.So the two products were never going to agreeDifferent quantities, different thresholds, different eligibility rules. This is one documented reason among several.
Modelled conversions appear in the Conversions column with no separate label. The reported figure is a mixture. Source : Google Analytics and Google Ads help documentation (2026)

The attribution models that disappeared

A change worth knowing about if any of your historical reporting predates it.

What was removed, verbatim. The first click, linear, time decay, and position-based attribution models are no longer available as of November 2023.

What remains in the analytics product. Data-driven, paid and organic last click, and paid channels last click.

What happened on the ads side. Conversion actions that used the deprecated models were upgraded to use data-driven attribution, and data-driven is now the default for most conversion actions.

Why this matters retrospectively. Any comparison spanning that date compares two different attribution regimes. If a channel’s contribution appears to have changed in late 2023, that may be the model change rather than the channel.

And prospectively. Data-driven attribution is a model, not an observation. It distributes credit according to a proprietary calculation you cannot inspect or reproduce.

The one thing you can still do. Note the model in use next to any channel comparison you publish. It is a single line, and it is the difference between a defensible chart and one that will be relitigated in six months.

Attribution models withdrawn in November two thousand and twenty-three and the models remainingTable recording the attribution models withdrawn from the analytics and advertising products in November two thousand and twenty-three and the models that remain available. The withdrawn models are first click, linear, time decay and position based, all four of which ceased to be available as of November two thousand and twenty-three. The models remaining in the analytics product are data-driven attribution, paid and organic last click, and paid channels last click. On the advertising side, conversion actions that had used the deprecated models were upgraded to use data-driven attribution, and data-driven attribution is now the default model for most conversion actions. The retrospective consequence is that any performance comparison spanning that date compares two different attribution regimes, so an apparent change in a channel’s contribution during late two thousand and twenty-three may reflect the model change rather than any change in the channel’s actual performance. The prospective consequence is that data-driven attribution is a model rather than an observation, distributing credit according to a proprietary calculation that the advertiser cannot inspect or reproduce independently. The practical recommendation is to record the attribution model in use alongside any channel comparison that is published, which requires a single line of text and constitutes the difference between a chart that can be defended and one that will be disputed several months later when nobody remembers which model was active.Four models removed, November 2023WithdrawnFirst clickLinearTime decayPosition basedRemaining in analyticsData-drivenPaid and organic last clickPaid channels last clickAds: deprecated actions upgraded to data-drivenThe retrospective consequenceA channel that appears to have changed in late 2023 may be the model changing, not the channel.Data-driven attribution is a model, not an observationIt distributes credit by a proprietary calculation you cannot inspect or reproduce. Note it beside the chart.
Any comparison spanning November 2023 crosses two attribution regimes. Note the model beside the chart. Source : Google Analytics and Google Ads attribution documentation (2023)

What to check on your property this week

Six checks, in order of how much they change what you are looking at.

Open your data retention setting. If it is on the two-month default, move it to fourteen. It is not retroactive, so every week you wait is lost.

Look for the data quality icon on your key reports. If thresholding is active, note which reports it affects and stop presenting those numbers as complete.

Find your (other) row. Sort your main dimensions by row count and see whether an (other) row exists and how much traffic sits in it. If it is material, the percentages beneath it are wrong.

Check your reporting identity. Note whether it is set to blended, and note that changing it changes your totals without changing anything about your website.

Strip identifiers out of your dimensions. Query strings, session IDs, anything unique. Cardinality is created at collection time and cannot be fixed later.

Write the attribution model on any channel chart you publish. One line. It ends the argument before it starts.

Six checks to perform on an analytics property ordered by their effect on reported figuresChecklist of six actions to perform on an analytics property, ordered by how substantially each changes what the operator is looking at. The first and most urgent is to open the data retention setting and, if it remains on the two-month default, change it to fourteen months, because the setting is not retroactive so every week of delay represents data that cannot subsequently be recovered. The second is to look for the data quality icon on key reports, and where thresholding is active to note which reports are affected and cease presenting those figures as complete. The third is to locate the other row by sorting main dimensions by row count and establishing whether an other row exists and how much traffic it contains, because where the quantity is material every percentage calculated beneath it is incorrect. The fourth is to check the reporting identity setting, noting whether it is set to blended and recognising that changing it alters reported totals without anything changing on the website itself. The fifth is to strip identifiers out of dimensions, including query strings, session identifiers and anything unique, since cardinality is created at the moment of collection and cannot be corrected afterwards. The sixth is to record the attribution model in use on any channel comparison chart that is published, which takes a single line and prevents a dispute several months later. The first item is marked urgent because of its non-retroactive nature, while the remaining five are diagnostic actions that reveal what the existing reports have been concealing rather than changing what is collected going forward.Six checks, most urgent first1. Data retention, todayIf it is on the 2-month default, set it to 14. Not retroactive, so every week you wait is lost.2. The data quality iconWhere it appears, stop calling it complete.3. Find your (other) rowIf material, the percentages under it are wrong.4. Your reporting identityChanging it changes your totals. Nothing else.5. Strip identifiersCardinality is made at collection, not in reports.6. Write the attribution model on every channel chartOne line. It ends the argument before it starts.Only the first changes what you collect. The other five reveal what your reports were already hiding.
Retention is not retroactive, so that one is urgent. The rest are diagnosis. Source : Google Analytics help documentation (2026)

Where to go next

Your platform and analytics numbers disagree. Why GA4 and Meta conversions do not match.

Your bounce rate looks wrong. GA4 bounce rate.

You are deciding what to track in the first place. What not to measure.

You are building the reporting layer. B2B marketing dashboard blind spots.

You are considering server-side collection. Server-side tracking.

Your conversion rate is the number in dispute. Conversion rate and its denominator.

In short

  • Thresholding removes rows to prevent identification, is system defined, cannot be adjusted, and publishes no numeric threshold at all.
  • It fires on demographics and low-user search terms, which makes small B2B audiences the most exposed.
  • Cardinality above 500 values is high, described by the vendor as guidance rather than a limit, and the excess is condensed into an (other) row.
  • That row can hold most of your traffic, which makes every percentage computed from the visible rows wrong.
  • Data retention affects explorations and funnels only, not standard reports, and age, gender and interest are always capped at two months.
  • Behavioural modelling needs 1,000 events a day denied for 7 days, and 1,000 daily users granted for 7 of 28, and shows only under blended identity.
  • Ads conversion modelling is separate, needs 700 clicks over 7 days per country and domain, and lands inside the Conversions column unlabelled.
  • Four attribution models vanished in November 2023: first click, linear, time decay and position based.

Fix your retention setting today, then go and find your (other) row. Book a diagnostic, or see how we approach B2B websites.