As Wikipedia will tell you, sampling is the practice of “using a subset of individuals from within a statistical population to estimate characteristics of the whole population.”

In layman’s terms, sampling is when instead of gathering every single data point in a particular data set, you choose a smaller portion of data and use that sample to estimate what the whole set of data would look like based upon the smaller portion. Polling data is the most commonly known form of sampling.  Pollsters survey a percentage of the population, then extrapolate the results to apply to the population as a whole.  

Sampling does offer benefits for data analysis since it speeds up and simplifies the process as seen in the example above.  

However, sampling also has many drawbacks, not the least of which is the fact that a sample is NEVER a complete data set so you may be missing the long tail and outliers that actually contain meaningful data. And the smaller the sample size is when compared to the full data set the less accurate the conclusions drawn from the sample.

While this might not be an issue when you’re looking for a representation of demographics or similar factors it’s a big deal when doing any kind of comparative analysis. If you choose to use the sampled data you’re making decisions or providing advice to your team based upon inaccurate data and this can mean loss of your reputation at best and financial loss at worst.

In general, Google Analytics Standard Reports are unsampled reports. This is true even for their free accounts. Google Analytics captures the raw sessions data (up to the actual data limit of 10 million hits) and prepares that data into aggregated tables so that any user can access unsampled data quickly.

However, beyond the predetermined default reports, any segments, views, or dimensions being applied to ad-hoc reports run by a user can trigger sampling.

This is because Google Analytics is collecting and processing billions of hits daily, so in order to increase efficiency and save computing power, they have set limits upon how many sessions can be included in a single ad-hoc query before they will employ sampling.

According to Google Analytics Help documents, users are informed that Google Analytics will employ sampling when a  report exceeds 500k sessions at the property level for the date range you are using. In other words, Google Analytics allows you to pull up to 500K sessions per tracking ID, not per account and not per View, for a single date range without sampling the data.  Once your query exceeds that threshold, sampling will be employed.

That's it! Canvas has returned an unsampled dataset which you can now process further using the tools available in the Block Library (filtering out bad data, creating your own custom channel groupings, cleaning page paths, etc.), or you can export the data to your preferred databases, file types, or our Google Data Studio community connector. 

What's more is that this can all be done without a developer, so the business user can make as many queries as they want, updating them as their analysis unfolds, and staying in the flow to understand the data and provide meaningful analysis. 

While Analytics Canvas provides powerful automated solutions to resolve sampling issues, there are circumstances in which sampling cannot be avoided for Google Analytics free accounts.   

When running a report, Analytics Canvas will automatically partition the results, finding the range at which unsampled data can be accessed. Canvas will go so far as to partition the report down to a single day in order to eliminate sampling. However, it is not possible to request a smaller unit than a single day, which means if your property is receiving more than 500,000 sessions a day, sampling cannot be avoided without upgrading to Google Analytics 360 where the threshold for sampling is much higher.

The good news is, Analytics Canvas will reduce sampling as much as possible by partitioning down to the day and collecting the full data set on days when the threshold has not been exceeded.  As well, Canvas will tell you both in the interface and in the dataset whether the data has been sampled for each row in the dataset.  

There are a few other limitations that can be encountered when using partitioning to eliminate sampling, specifically when it comes to Calculated and Unique metrics.

Unique metrics are those which are filtered by Google and only counted once per report no matter how many times that metric may have occurred.

For example, "users" is a unique metric. If you were looking at data for the months of March, April, and May, and John Doe visited your site every month, then the date parameters for your report will significantly affect the totals returned by Google Analytics.

If you pulled all three months at once then John Doe would only show on the report once, as “1” user. But if you pulled the three months separately then summed them together, John Doe would show up 3 times (once for each month) as 3 unique users. This result is incorrect and will lead to bad conclusions. 

For this reason, you cannot use unique metrics in a partitioned query.  

Similarly, calculated metrics are those where Google Analytics has pre-calculated a metric for you.  Any metric that includes the term "rate" or the term "avg", such as Bounce Rate and Average Time On Page, are examples of calculated metrics.  Google has used standard metrics to create these calculated metrics for you.  The problem is, when partitioning, you cannot simply aggregate these calculated metrics - you will return the wrong result.

The solution here is much simpler: include the metrics required for the calculation, such as bounces and sessions to calculate bounce rate, and create your calculated metrics at the right level of granularity either in Canvas or in your reporting tool.  If you include the date dimension this does not apply as the partitions will not be summarized- the partitions will simply be combined together to create the full data set.