In the late 19th century, Francis Galton noticed something odd while studying inherited height. Very tall parents tended to have children who were tall, but usually not as extreme as the parents themselves. Very short parents had children who were short, but again, closer to average than the parents were. Neither group was becoming more average over generations, height wasn't disappearing. Each individual measurement was just partly the underlying trait and partly noise, and the noise doesn't carry forward the same way twice.
Galton called it regression to the mean, and it applies to any measurement that mixes a genuine underlying level with a random component, which describes most B2B performance metrics far more than teams usually assume.
Why an extreme result is rarely as informative as it feels
An outlier result, in either direction, feels like it's telling you something important. A rep who closes 180% of quota one quarter feels like proof they've levelled up. A campaign converting at three times the account average feels like proof that channel is simply better. Both readings treat the entire result as signal, when a meaningful part of it is very likely to be noise, a good stretch of timing, a couple of large deals landing in the same window, a sample too small to be stable.
The tell is what happens next. If the extreme result reflected a genuine, durable change, the following period should look similar. If it was substantially noise, the following period will look more average, not because anything got worse, but because the lucky component doesn't repeat on demand.
A return to average after an outlier result isn't a decline. It's what you'd expect from a stable underlying performance level once the unusual timing that inflated the outlier doesn't repeat.
The two places this bites hardest in B2B: reps and campaigns
Two decisions get made on the back of outlier results more than almost any others in a B2B revenue org, and both are vulnerable to the same statistical illusion.
Why "the intervention worked" is often just this in disguise
This is where regression to the mean becomes genuinely misleading rather than just a curiosity. A team introduces coaching after a rep's worst quarter. The rep's next quarter improves. The natural conclusion is that the coaching worked, and it might have. But an improvement was statistically likely regardless, simply because an unusually bad quarter contains more noise than a typical one, and that noise was always likely to partially reverse on its own.
Coaching after a bad quarter, a new campaign brief after a weak month, a process change after a slow week, all tend to be followed by improvement even when the intervention did nothing, simply because the starting point was an outlier likely to regress upward.
Praise, a bonus, or extra budget following an exceptional result is often followed by a more average result. It's tempting to read this as complacency or the extra resource being wasted, when it's frequently just the same statistical reversion working in the other direction.
The fewer data points behind a result, a single rep's quarter, a single month of campaign data, the larger the noise component relative to the genuine signal, and the more dramatic the reversion toward average tends to look.
How to tell genuine improvement from statistical noise
None of this means ignore outlier results. It means treat a single one as a hypothesis worth checking, not a conclusion worth acting on immediately.
- Look at trend across several periods, not one: A rep or campaign that's genuinely improved will show it consistently across multiple quarters or months, not in a single spike surrounded by more typical results either side.
- Check the sample size behind the result: A campaign's conversion rate from 15 conversions is far noisier than one from 500. The smaller the sample, the more scepticism an extreme result deserves before it drives a budget decision.
- Separate the intervention from the starting point: If a change followed an unusually bad result, ask whether the improvement seen afterward is bigger than what regression to the mean alone would predict, not just whether things got better at all.
- Resist adjusting targets after a single outlier in either direction: Raising a quota after one exceptional quarter, or cutting one after a weak one, bakes a noisy result into a structural decision. Wait for a second or third data point before treating it as the new normal.
- An extreme result mixes genuine performance with noise, and the noise rarely repeats identically the next time
- Raising a rep's quota or scaling a campaign's budget off one outlier period assumes the lucky component repeats too
- An intervention that follows a bad result will often look like it worked even if it did nothing, because reversion toward average was already statistically likely
- Check trend across several periods and sample size behind any single result before treating it as a structural signal
Our attribution modelling work looks at performance across enough data to tell a genuine trend from a single outlier period, so budget and quota decisions aren't built on noise that was always going to revert.
Explore Attribution Modelling →