To see what changed this week, write down the normal range of each key number from your last eight settled weeks. A number inside its range is usually noise, so leave it alone. A number outside the range is a signal, so ask what else moved that week. Five minutes, three questions, one decision per number: act, watch or ignore.
Subject: “Orders down 8%. Thoughts?”
That is the whole email. It came from your ad agency at 8:02 this morning.
By 8:30 you have a theory. By 9:00 you have paused two ad sets.
By Friday, orders are back to normal, and nobody can say why.
A weekly number moves. Most of the time, that move is noise, and the right reply is to do nothing. The skill is knowing the few times it is not.
This post gives you a method that takes five minutes, uses only your own history, and ends in one decision per number. All brands and numbers here are made up.
Why a dip is not yet news
Every business number swings from week to week. Weather, paydays, a slow carrier day and plain luck all push it around a little.
A quality-control pioneer named Walter Shewhart found that people make two mistakes when they judge swing. The Deming Institute lists them (Deming Institute):
| Mistake | In plain words | What it looks like in a brand |
|---|---|---|
| Mistake 1 | Reacting to normal swing as if it had a special cause | Pausing ads after one soft week |
| Mistake 2 | Treating a real change as normal swing | Calling a refund spike “just a bad week” |
The Institute calls the first one tampering. It quotes Deming on the cost: acting on a stable process in response to its normal ups and downs only increases future mistakes and costs.
Both mistakes cost money, and the first one feels like work. That is why it is so common.
The five-minute method
You need three things. A short list of numbers, eight weeks of each, and a pen.
Minute 1: pick your numbers and make sure they are settled
Pick three to five numbers that matter. Good starters are orders, refund rate, ad cost per order, new customers and days of stock cover.
Then check the week is finished. Many tools keep updating a week after it ends. Google says Analytics data can take 24 to 48 hours to process, and data in reports may change in that time. Ad tools can take longer, as our post on Amazon ad numbers that are not final shows.
Compare weeks that are the same age. A week read one day after it ended will look worse than the same week read two weeks later.
Minute 2: write each number’s normal band
Take the last eight settled weeks of one number. Work out two things:
- The average.
- The typical swing around that average. A spreadsheet calls this the standard deviation.
The normal band is the average, plus and minus three times the typical swing. Quality engineers call these 3-sigma limits. The US government’s measurement lab says control limits are chosen so that almost all points fall inside them while the process stays in control. It adds that a point outside the limits is a reason to search for a cause.
Here are eight made-up weeks of orders for a made-up brand selling bike lights.
| Week | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|
| Orders | 412 | 398 | 431 | 405 | 420 | 389 | 415 | 408 |
The average is about 410. The typical swing is about 13. Three times that is about 39. So the normal band runs from 371 to 449 orders.
Minute 3: mark each number as noise or signal
This week the brand had 376 orders. That is down about 8% from last week’s 408.
An 8% drop that still sits inside the band. Illustrative numbers.
376 is above 371. It is inside the band. That dip is within the range of weeks the brand has already had. The email at 8:02 was about noise.
Now run the same check on three numbers.
| Number | Usual | Normal band | This week | Verdict |
|---|---|---|---|---|
| Orders | 410 | 371 to 449 | 376 | Noise |
| Ad cost per order | $11.20 | $9.40 to $13.00 | $12.60 | Noise |
| Refund rate | 3.1% | 1.9% to 4.3% | 6.4% | Signal |
Ad cost per order is up 12.5%, which sounds alarming. It still sits inside its band. The refund rate is more than double its usual level and far outside its band. The loud number was not the important one.
Minute 4: for each signal, ask what else moved
A signal says look here. It does not say why.
For the refund rate, list everything that changed that week:
- a price or a pack change
- a new supplier batch
- a deal or a coupon
- a stock gap, a late delivery or a new carrier
- a change in who the ads reached
- a listing edit
Part of the cause usually sits in a different system from the symptom. Refunds show in the store. A bad batch shows up in the support inbox and the returns notes. No single tool shows both, and that is how a real problem hides. Our post on telling whether two numbers are really linked covers the next step, which is testing a link before you trust it.
Minute 5: decide, and write one line
Each number gets one of three labels.
- Act. It is a signal and you know the likely cause.
- Watch. It is a signal and the cause is unknown, or it is noise but drifting.
- Ignore. It is inside the band, with no drift.
Then write a single line with the date. For our example: “Refund rate 6.4% against a band of 1.9% to 4.3%. Check the latest supplier batch. Review next Monday.”
That line is the whole output. It is also what you will thank yourself for in a month.
What it cost, in the example
Illustrative numbers. At 376 orders and a $28 average order, the week brought in about $10,528. A refund rate of 6.4% instead of 3.1% means an extra 3.3% of that went back out, about $347.
A quiet signal costs about $2,800 in two months. Illustrative numbers.
Pausing the ad sets over the 8% dip would have cost more in lost sales than the whole refund leak. The owner did not miss anything. The orders chart was loud and the refund line was quiet.
Run this test: does your own review pass?
You can try the method on last quarter’s numbers this afternoon. It takes about 20 minutes.
- Pick one number you reacted to in the past three months. Write down what you did.
- Pull the eight settled weeks before it. Work out the average and the typical swing.
- Draw the band (average plus and minus three swings).
- Check where the number you reacted to landed.
What a pass looks like: the number was outside its band, and your reaction matched the cause you found.
What a fail looks like: the number was inside its band. You reacted to noise, and the reaction cost you something, such as paused ads, a price cut or a rushed reorder.
Run it on three past reactions. If two of the three were noise, the weekly review is your biggest cheap fix.
Where this method goes wrong
Honest limits matter here. The band is a rough tool.
- Slow drift hides inside the band. A number that falls a little every week for six weeks can stay inside its band the whole time. Look for six weeks moving the same way, not just one week.
- Eight weeks is thin. A short history gives a loose band. A launch, a sale or an out-of-stock month inside those weeks will stretch the band and hide real changes.
- Seasons move the middle. A band built over a calm spring does not fit the holiday run. Rebuild it when the season turns.
- Not every number is bell-shaped. The NIST page notes that 3-sigma limits are common practice even when the data is not normally distributed, and that skewed data can change the risk of a false alarm. Treat the band as a prompt to look, not a proof.
- Some numbers are too young. If last week’s data is still being filled in, wait.
- A signal can still be good news. A number far above its band deserves the same question: what moved?
The short version
Every week, ask three questions per number. Is it settled? Is it outside its normal band? What else moved that week?
Do nothing about noise. Look hard at signals. Write one line for each, and read last month’s lines before you read this week’s.
Synthesis is built for the “what else moved” step. It reads your stores, ads, email and books as one brain, checks every number against its source, and sends a Monday report to the AI you already use. You can read more in what a brand brain is, or follow our weekly business review checklist.
If you have fewer than eight settled weeks, or one simple store, a spreadsheet and the five minutes above will do the job well.
Questions people ask
How do I know if a weekly change in my sales is real?
Compare it with the normal range of your last eight weeks. If this week sits inside that range, it is most likely ordinary swing. If it sits outside, treat it as a signal and look for what else changed that week.
What is a normal range for a weekly business number?
It is the band your number usually stays in. A simple version is the average of the last eight weeks, plus and minus three times the typical swing around that average. Quality engineers use the same idea in control charts.
Why do my numbers change after I read them?
Many tools keep processing data after the day ends. Google says its Analytics data can take 24 to 48 hours to process, and reports may change in that time. Compare weeks only after the data has settled.
Should I react to every dip in my weekly report?
No. Reacting to normal swing is a known mistake that can make results worse, not better. React when a number leaves its normal range, or when it drifts the same way for many weeks in a row.
How many weeks of history do I need?
Eight is a workable minimum for a rough band. More is better. If you have fewer than eight settled weeks, or a launch or big sale has changed the business, treat any band as a guess and lean on other checks.
Can I ask an AI to do this review?
Yes, if it works from settled numbers with clear definitions and shows its band. Ask it for the normal range, this week’s value and what else moved. If it gives a verdict with no band, ask again.