Insights • 4 min read

Considering sampling? What you might be missing if you do.

Addison Price

Global Head of Solutions Engineering and Enablement at Fullstory

Last updated: 11/16/2026

Article summary

Sampling is a cost control, but it comes with a price. You pay it in how fast you find problems, how well you can size them, and whether the person who needs help has a record.

As pricing moves toward consumption, the question comes up in almost every planning cycle: do we need to capture every session, or would a sample do? It is a fair question, and the answer depends on what you use the data for.

A sample is good at describing the average. Much of the value in digital experience data sits somewhere other than the average.

Sampling slows two important clocks

Two numbers decide how quickly a team can act on a problem:

  • Time to identification:

    how long it takes to notice something is wrong.

  • Time to assess impact:

    how long it takes to say how many people are affected and whether it matters.

Both depend on how often the problem appears in your data. Capture 20% of sessions and every issue shows up one-fifth as often, which makes it five times slower to spot. The delay then compounds. Sizing an issue means collecting enough examples to trust the count, so work that takes about five days at full capture can stretch to about three weeks.

Try it yourself

Each dot below stands for 1,000 sessions, and 10% of them hit the same problem. Change the sampling rate, then switch to “Your data” to see what your team would actually see.

Two things stand out when you play with it. The share of problem sessions you catch changes each time, so the rate you measure can sit well above or below the real one. Teams set priorities from that number.

Second, look at the dashed red rings. Those sessions happened, and nobody can ever go back and watch them.

Trends tolerate sampling. Individuals do not.

Anything that starts from one specific person breaks down quickly. A customer writes in about a failed payment. A claim form times out. An account manager wants to know what happened before a cancellation. An AI agent is about to reply to a customer and needs to know what that customer just went through.

Each of these asks about a single session. For one person's problem, the chance that no record exists is simply one minus your sampling rate. At 20% that is four in five. Widespread problems still surface eventually. The small, urgent, one-off ones are the ones that slip through, and the readout above shows how quickly the odds improve as more people hit the same thing.

Support context for humans
Context for AI agents
Risk and personalization

A customer contacts support and the agent opens the ticket. The replay either exists or it does not.

An AI agent answering a customer can only use what was captured. Gaps in the data become gaps in the answer.

Abuse reviews and in-the-moment offers fire on one person's behavior. If the session was not captured, the trigger never fires.

4 in 5 tickets, no context

4 in 5 customers, no context

~80% of events missed

At a 20% sampling rate.

Your most valuable customers make this worse. They are usually a small share of traffic, and a random sample thins a small group out just as hard as everyone else, so the people you most want to serve well are the people you most often cannot see.

Questions to ask before you sample

  1. Which workflows start from one person's session? Support, AI agents, fraud review and personalization all need the specific session to exist. Those need every session.

  2. How long can you wait to see a new problem? Divide the time you have now by your sampling rate. That is your new time to identification, and your time to assess impact follows it.

  3. Who is in the sessions you would not keep? Check whether your highest-value customers, newest releases or smallest segments are the ones a sample hits hardest.

Sampling is a cost control, but it comes with a price. Now you know what that price is: slower detection, slower sizing, and gaps wherever a person is waiting for help. If the conversation is already underway at your company, those are the numbers to put on the table.

Google logoWant more like this in your Google results? Add Fullstory as a preferred source.
Addison Price
Global Head of Solutions Engineering and Enablement, Fullstory

Addison fell in love with Fullstory from the first time he logged-in as a customer in 2016 and has been lucky enough to live out his dream of wrangling behavioral data as a member of the Fullstory team for the last 5 years. Addison currently leads Fullstory’s Global Solutions Engineering team and has previously worked in the data privacy industry where he travelled the globe hosting speaking engagements on responsible use of data and AI.

Additional Resources

reduce shopping cart abandonment OG
Reduce cart abandonment: 8 proven solutions to recover potential sales

Struggling with cart abandonment? Learn how to improve your e-commerce experience and convert more shoppers into buyers.

Read the blog
fs-deepdive-og-web-analytics
What are website analytics? Guide to data and behavioral insights

Here we define website analytics, explain what they mean, and share tools and resources to help you track them.

Read the blog
Fullstory is shown here with other tools representing that an effective ecommerce CRO software stack contains multiple tools
Best CRO tools for ecommerce: A behavioral data view for 2026

Successful ecommerce CRO takes more than one tool. Build a stack covering behavioral analytics, monitoring, testing, and personalization.

View more
Person icon on ID badge with data tags being added from multiple sources.
What is customer data enrichment? Overview and how to use it

Customer data enrichment adds new attributes to a customer profile. See how it works and to use behavioral data without buying third-party sources.

Read the blog
Preparing your data for AI
AI-ready data: Practical tips to prepare your data for AI

A practical guide to preparing your organization for AI adoption, covering key strategies for data, infrastructure, and cultural readiness.

Read the blog
Illustration of a lit match striking against a jagged, brick-like magenta formation with circuit-style connector lines, symbolizing digital friction sparking into a bigger problem.
65% of teams have misdiagnosed digital friction, shipping the wrong fix

Shipping a fix for the wrong root cause charges a compounding tax on engineering resources and customer trust.

Read the blog