Session recordings: what to look for in the first thirty

This is the single most useful hour a small business owner can spend on their website, and almost nobody spends it. It is unglamorous. There is no dashboard, no score, no report to forward. You sit and watch strangers struggle with something you built.

That discomfort is the value. You are watching evidence instead of arguing about opinions, which is the only way this conversation ever ends.

What is a session recording and what does it show?

It captures the visitor's screen as they use your site. Cursor movement or thumb position, scroll depth, clicks and taps, time on each element, and the last thing they did before leaving.

It does not tell you why. A recording shows a person tapping the same image four times. It does not tell you they thought it was a button. You infer that, and the inference is nearly always obvious once you have seen it happen three times.

That distinction matters, because recordings answer a different question than analytics. Analytics tells you 60 people left the contact page. Recordings tell you they all left at the same field.

One thing to set correctly before you start. Turn on input masking so form contents and any personal information are not recorded. You want to see where people hesitate, not read what they typed. Check your recording tool's privacy settings before you switch it on, and make sure your privacy policy reflects that you use it.

Why thirty and not three hundred?

Because thirty is the number where detection gets cheap, and there is arithmetic behind it rather than a round figure somebody liked.

Detecting a problem is a different job from measuring how often it happens, and they need wildly different sample sizes. If a problem affects a share of your visitors, the chance of catching it at least once in a batch of recordings is one minus the chance of missing it every single time.

Read the first row. Thirty recordings gives you a better than nine in ten chance of witnessing anything that trips up one visitor in ten, and a problem affecting one visitor in ten is exactly the class of thing worth an afternoon of your time. Below that, you are hunting edge cases, and the fix is worth less than the hunting.

Now the contrast that makes the point. Ron Kohavi, Alex Deng, Roger Longbotham and Ya Xu, writing from Microsoft and LinkedIn in a peer reviewed 2014 KDD paper, publish the sample sizes Bing needed for its own metrics. To detect a 4.4% change in revenue per user, they needed 114,000 users per variant, because that metric has a skewness coefficient of 17.9. Their published rule of thumb for normality is 355 times the square of the skewness coefficient, per variant.

One hundred and fourteen thousand users to measure a rate. Thirty recordings to find a broken button. Those are two different questions and only one of them is available to you.

Which thirty recordings should I watch?

Not thirty at random. Filter for the ones that carry information.

Start with sessions that visited a service page or a contact page and did not convert. That is where your money is being lost, and it is the smallest set that still contains the answer.

Add sessions on mobile specifically, because that is where most local traffic sits and where most breakage lives. If your tool lets you filter by device, watch mobile and desktop separately. They are two different websites.

Add any session flagged with a rage click or repeated tapping. Tools name this differently, but the pattern is the same and it points straight at something the visitor thought was interactive and was not.

Skip sessions under about ten seconds. Those are mostly bounces from mismatched search results and they will tell you about your ad targeting rather than your page.

Thirty filtered recordings will beat three hundred random ones, and you can watch thirty in under an hour at double speed.

What am I looking for exactly?

Five things. Write them at the top of a page and tally as you go.

Where scrolling stops. If most people stop at the same point, whatever sits just above that point is where you lost them. Often it is a large image, a wall of text, or a section that reads like it was written for a different audience.

Dead taps. People tapping images, headings, phone numbers that are not linked, or anything that looks clickable and is not. Every one of these is a fifteen minute fix.

Form hesitation. Long pauses on a specific field, cursor moving in and out, then leaving. That field is the problem, and you can confirm it with form abandonment tracking that costs nothing once you know which one to watch.

The Baymard Institute, which has tracked checkout flows since 2012, reached the conclusion that field count matters far more to performance than the number of steps. That is ecommerce checkout research from a company that sells UX research and benchmarking, not local lead form research, so treat it as a hint about where to point your attention rather than as a rule about your quote form. It tells you to watch the fields, not the pages.

Back and forth movement. Visitors bouncing between two pages usually means they are hunting for something you have not stated. Price and service area are the two most common. Watching this for a week will tell you what your homepage is missing better than any keyword tool.

What they scroll past to look at. When people repeatedly scroll down to your review section or your credentials before doing anything else, that is a hint that the trust elements should sit higher on the page than you put them.

Can I turn what I saw into a percentage?

No, and this is the mistake that turns a good hour into a bad report.

Say four of your thirty visitors did the same odd thing. The temptation is to divide and report that 13%derived of visitors hit this problem. Run the interval on it and that number falls apart. A Wilson score interval on four out of thirty gives a 95% range of roughly 5% to 30%derived. Your point estimate sits inside a band six times wider than itself. It is not a measurement.

What you actually have is far more useful than a bad percentage. You have four independent people doing the same specific thing on the same specific element, which is a description of a defect. Defects do not need confidence intervals. They need fixing.

The same discipline that governs experiments governs this. The Microsoft and LinkedIn authors above report that at Bing most experiments fail and those that succeed move key metrics by 0.1% to 1.0% once diluted. If an effect were that subtle, you would never see it in thirty recordings. The corollary is the good news: anything you can see in thirty recordings is far too big to be a statistical artifact. You are not detecting a whisper. You are watching someone hit a wall.

So write the finding as a sentence, not a rate. "Four of thirty mobile visitors tapped the hero image expecting it to call us." That sentence survives scrutiny. A share of visitors, quoted to two figures off a sample of thirty, does not.

What should I ignore?

Your own sessions, first of all. Filter your office IP address out or you will spend twenty minutes watching yourself.

Ignore anything that looks like a bot. Instant scrolls to the bottom, impossible movement, no pauses. Some tools filter these, most do not filter all of them.

Ignore aesthetics entirely while you watch. You will notice a font you dislike and a photo you have been meaning to replace. Neither is why anyone left. Color and typography decisions matter for reasons that are not conversion, and mixing the two conversations wastes the recordings.

And ignore anything you saw exactly once. One person doing something strange is a person, not a pattern. Two is a coincidence. Three is a defect.

Where does this approach break down?

Three places.

When the problem is invisible on screen. A form that submits successfully and then fails to deliver looks perfect in a recording. So does a phone number that dials a disconnected line. Recordings show the visitor's side, never yours. Submit your own form and place your own call separately.

When you need a rate rather than a defect. If somebody genuinely needs to know what share of visitors abandon at field three, recordings cannot give it to you and neither can thirty of anything. That is what event tracking is for, over a quarter.

When traffic is too thin to accumulate a filtered batch. Under about a hundred sessions a month you will wait a long time for thirty mobile non converting sessions on a service page. Call your last twenty inquiries instead and ask what they could not find. It is the same information, delivered faster, by people who will tell you the reason as well as the behavior.

What to do this week

Install a recording tool on the contact page and your top two service pages. Free tiers cover this volume easily for a local site. Turn on input masking before it starts collecting.

Wait a week. Then block one hour, filter for mobile sessions that reached a service page and did not convert, and watch thirty at double speed with a notepad open.

Write down every repeated behavior as a sentence with a count, never as a percentage. Rank the list by how cheap the fix is. Do the cheap ones this week, because dead taps and unlinked phone numbers take minutes.

Then look at what happens after a conversion, since the page people land on after submitting is usually doing nothing and you now have recordings that show exactly where those visitors went next.

Repeat the exercise quarterly, not weekly. Behavior does not change fast enough to justify more.

Questions about what your visitors are doing?

Email me at eric@seod.com with your website and the one page you are least confident about. I will send back a short list of exactly what to filter for and what behaviors to tally on that page, so your first hour of watching is not wasted on random sessions.

If you have already watched some and are not sure what you saw, describe the pattern in two lines and I will tell you whether it is a real problem or normal browsing noise.

Otherwise, there is more on conversion in the library.