What actually counts as an outlier
An outlier is just a data point that sits far away from the rest of your data. That’s the whole definition. It doesn’t tell you whether the point is wrong, whether it’s a fluke, or whether it’s the most important row in your spreadsheet. It just tells you the number looks unusual compared to its neighbors.
If you run a one-person business and you look at your own revenue, churn, or conversion numbers every week, you’ve already met outliers. The customer who paid you ten times what everyone else pays. The day your signup page got hit by a bot and conversion rate looked like it tripled. The month one client cancelled and dragged your churn rate up on its own. None of these are automatically “bad data.” Some of them are the most real thing in your dataset. The job is figuring out which is which.
Why outliers show up in small business data
Small datasets make outliers louder. If you have 400 customers, one customer paying five times the average moves your average noticeably. If you had 400,000 customers, that same customer would barely register. This is worth sitting with, because a lot of outlier advice online is written for people with huge datasets where a handful of odd points wash out. When you’re a solo operator, your “n” is small, so a single weird row can swing a metric you’re using to make decisions.
In practice, outliers in small business data tend to come from a short list of sources:
- Data entry or tracking errors. A test transaction that never got deleted. A tracking pixel that fired twice. A refund logged as a new charge.
- Genuine one-off events. A customer who churned because their company shut down, not because of anything you did. A traffic spike because someone posted your product on a forum.
- Real but rare customers. The “whale” who buys the enterprise tier, or the person who never uses the free trial features and cancels day two.
- Seasonality or promotions. A launch discount that pulled in a batch of low-value signups you wouldn’t normally see.
Each of these needs a different response, which is why “just remove the outliers” is bad advice on its own.
The first question: error, or real?
Before you touch any statistical method, ask the boring question first: is this number wrong, or is it just unusual? An error should usually be fixed or removed. A genuine extreme value should usually stay, because deleting it means you’re editing reality to make your dashboard prettier.
Say you’re looking at average order value and one row shows a $9,400 order from a store that normally sells $60 items. Go look at it. Maybe it’s a duplicate charge from a webhook retry (an error, fix it). Maybe it’s a real bulk order from a corporate buyer (real, keep it, but maybe note that it’s not representative of your typical customer). You can’t tell the difference by staring at the number. You have to go check the record.
This is the part that’s easy to skip when you’re busy, and it’s the part that actually matters. A method for spotting outliers is only useful as a prompt to go investigate. It’s not a verdict.
How I actually spot them, without fancy statistics
You don’t need a statistics degree to find outliers in a small business dataset. A few plain methods cover most of what you need:
Sort and look. Sort your customer list by revenue, or your order list by amount, and scroll to the top and bottom. This sounds too simple to mention, but it’s the fastest way to catch a $9,400 duplicate charge before it feeds into any calculation.
The middle-90% trick. Look at where most of your values fall, then check anything outside that range. A common version of this uses the “interquartile range”: find the point where 25% of your values are below (the first quartile) and the point where 75% are below (the third quartile), then flag anything that falls much further than that gap below the first quartile or above the third. You don’t need to compute this by hand. Spreadsheet functions like QUARTILE do it in one line, and most analytics tools will show you percentiles directly.
Standard deviations from the average. If your data is roughly bell-shaped, a value more than two or three standard deviations from the mean is unusual enough to look at. This method breaks down fast on skewed data, which is most business data (revenue and order sizes are rarely symmetric), so treat it as a rough flag, not a rule.
Just plot it. A simple time series chart or scatter plot will show you a spike or a dip before any formula does. If you’re tracking daily signups and one day jumps to five times the usual number, you’ll see it before you calculate it.
None of these methods tell you what to do next. They only tell you where to look.
What outliers do to the metrics you already track
MRR. One large customer can make your monthly recurring revenue look stable when it isn’t. If losing that one account would knock out 15% of your MRR, that’s worth knowing on its own, separate from whether the account counts as a statistical outlier.
Churn. A single enterprise cancellation can spike your churn rate for a month even if your day-to-day retention is unchanged. Looking at churn by customer count and churn by revenue separately helps here, because they can tell different stories when one big account leaves.
CLV. Customer lifetime value calculations that use an average are especially sensitive to a few very high spenders. If your CLV number is being pulled up by a handful of unusually large accounts, that number describes those accounts more than it describes your typical customer.
Conversion rate and funnels. Bot traffic, a broken tracking event, or a single referral source dumping a lot of low-intent visitors can distort a conversion rate for a day or a week. Segmenting by traffic source before you calculate the overall number often exposes this.
A/B tests. A test can be won or lost by one large order landing in one variant by chance. This is one of the reasons a “significant” result on a small sample deserves a second look rather than an immediate rollout, especially if you can trace the win to one or two unusual transactions.
What to actually do about them
Once you’ve separated errors from real extreme values, your options are usually:
- Fix or remove confirmed errors. Duplicate charges, test data, broken tracking events. These don’t belong in the calculation at all.
- Keep real extreme values, but report more than one number. Show both the average and the median CLV, for example. The median won’t be pulled around by a few big spenders, so seeing both gives you a fuller picture than either alone.
- Segment instead of deleting. If a handful of enterprise customers behave completely differently from your typical customer, consider looking at them as their own group rather than blending them into one overall average.
- Cap or “winsorize” for reporting purposes only. Some people cap extreme values at a chosen percentile so one row doesn’t dominate a chart, while keeping the real number in the underlying data for anything that actually needs it, like invoicing. Be transparent with yourself about when you’re doing this and why.
- Rerun with and without the outlier. If a decision hinges on one number, check whether the conclusion changes when you exclude the extreme value. If it does, that’s a sign you’re making a call based on one data point, which is worth knowing before you act on it.
What outliers can’t tell you
No method here tells you whether an unusual number is good news, bad news, or noise. A method flags a value as far from the pack. Whether that value matters to your business is a judgment call, and it’s one only you can make with the context you have about your customers, your product, and what actually happened that week. Treat outlier detection as a way to know where to look closer, not as a substitute for looking.
If you want more of this kind of plain-English walkthrough on the metrics and methods behind running a business on your own numbers, you can find the rest of what we’ve covered on the home page.
Get new guides and videos first — join the Telegram channel.