When analytics numbers are estimates, and when it matters

Analytics tools display sampled figures and unsampled figures in the same typeface, in the same box, with the same air of precision. A number like 1,847 sessions reads as a count of things that happened. Sometimes it is. Sometimes it is a projection calculated from a portion of the available data and then scaled up, and the difference is not visible in the number itself — only in a small indicator elsewhere on the screen that most people have never clicked.

This matters less often than the anxiety about it suggests, and it matters enormously in specific cases. Knowing which case you are in is the whole skill. A sampled figure is not wrong in the sense of being fabricated; it is an estimate with an error range that widens as the sample shrinks, and for broad totals it is usually close enough to act on. It becomes genuinely misleading when you slice the data thinly, because a small slice of a sample can rest on very few actual records — and a projection built from a handful of sessions can be badly off while still being displayed as a specific, confident number.

Google's own documentation for GA4 puts numbers on this, which is worth reading rather than guessing about. It describes default sampling settings of 10 million events for standard properties and 100 million for Analytics 360 properties: when a request has to process more events than that limit, Analytics uses a representative sample of the available data rather than all of it. The figure applies per request, not per month, so what triggers it is the size of the question you ask, not simply the size of your site.

Two practical consequences follow. First, prebuilt standard reports are largely served from pre-aggregated tables and rarely run into this, which is why the everyday reports most small businesses use are generally not sampled. Second, the way a small site does hit sampling is by asking an expensive question: a long date range combined with several dimensions, custom segments, or an exploration that has to reach into event-level detail. For a site with modest traffic, ordinary reporting will not approach 10 million events in a single request, and the honest summary is that sampling is a real mechanism you are unlikely to meet unless you go looking in the ad-hoc tools.

There is a second mechanism that produces missing or odd-looking numbers and is constantly confused with sampling, because both make the data seem incomplete. Data thresholding withholds data to stop anyone viewing a report from inferring the identity or sensitive information of individual users, based on demographics, interests or other signals present in the data. It is a privacy protection, not a performance measure.

The distinction is worth holding precisely because the responses differ completely. Sampling gives you an estimate of everything; thresholding gives you an accurate figure with rows removed. So a report broken down by age or gender for a low-traffic site may show a total that does not equal the sum of its visible rows — not because anything is estimated, but because small groups have been withheld to protect individuals. No amount of waiting or upgrading fixes that, and the remedy is different: broaden the segment, lengthen the period, or remove the demographic dimension that triggered it. Reading a thresholded report as a sampled one leads you to distrust numbers that are actually exact.

The interface reports this rather than hiding it, though not prominently. Near the top of a report or exploration there is a data-quality indicator, and hovering or clicking it states what the figures are based on — whether all available data was used or a sample of it, and where thresholding has been applied. Making a habit of checking it before quoting a number to anyone else costs seconds and is the entire practical skill here.

If you find you are looking at a sample and the answer matters, there are a few honest routes. Shorten the date range or simplify the query so the request processes fewer events, which often removes sampling entirely. Ask the question in a standard report rather than an exploration where one exists. Or move to the raw data: the BigQuery export gives event-level records for standard properties, and the trade-off is that it requires setup and SQL rather than being available in the interface. What you should not do is quote a sampled figure to a decision-maker without saying it is sampled, particularly a thinly-sliced one, because the number carries no visible warning once it has been copied into a document.

The useful test is what the number is for. If you are reading a broad trend — traffic direction over months, which channels bring most visitors, whether engagement is improving — a sampled figure is fine, because the sample is representative and the conclusion is about direction rather than a precise value. Broad totals from large samples are close enough that the decision would not change.

Care when the number is small, when it is thinly sliced, or when someone will act on the exact value. A conversion count for one campaign in one city over one week, drawn from a sampled report, can be materially wrong. Care also when you are comparing two figures that came from differently-sampled queries, since the comparison inherits both error ranges and can invert. And treat any figure you are about to put in a document, a lender's pack or a board update as one that needs checking, because a number stops carrying its caveats the moment it leaves the interface. Beyond that, an hour spent worrying about sampling on a small site is usually an hour that would have been better spent on whether the tracking is set up correctly at all.

Sampling is the visible, documented, quantified uncertainty in web analytics, which is exactly why it attracts disproportionate attention. The larger distortions are not labelled anywhere in the interface. Consent banners mean a share of visitors are never measured, and that share varies by region and by how the banner is worded. Ad blockers and privacy-focused browsers prevent collection outright. Cross-device journeys break one person into several. Bots inflate some figures and are only partly filtered. None of this produces a warning icon.

The honest position is that web analytics figures are a systematically incomplete record of behaviour, with a gap whose size you cannot measure from inside the tool. That is not an argument for ignoring them; it is an argument for using them as directional evidence and for being suspicious of anyone quoting them to several significant figures. Reconciling against something you can count independently — actual orders, actual payments received, actual enquiries answered — is the only reliable check available, and where the two disagree, the money is right and the analytics is an estimate.