What actually drives sales performance?

Start with the obvious one. If a rep has more meetings with a customer, are they more likely to win the deal? And by how much? If you double the meetings, do your chances double too? Probably not. But how much do they actually move?

The most obvious way to check turns out to point the wrong way. If you count every email and call ever logged on a deal, more activity goes with losing. That's not because emails are bad luck. Deals that are going nowhere tend to drag on, and a deal that drags on collects a lot of follow-up emails.

So I only counted activity from early in each deal: the first 14 days for small-business (SMB) deals, which usually wrap up in about three weeks, and the first 30 days for the bigger Mid-Market and Enterprise deals. Deals that closed before that window was over are left out.

Then I tried to make the comparison fairer. A rep who books ten meetings on a deal isn't necessarily working the same kind of opportunity as someone who books three. So the model compares like with like and asks a simpler question: when two opportunities are otherwise similar, does the one with more early meetings win more often?

A quick guide to the charts. They show how the odds of winning change. Odds aren't quite the same as chances, but they move in the same direction: 1× means no change, and anything to the right means better odds. To translate: if a kind of deal normally wins 1 time in 4, 1.4× the odds works out to roughly 1 time in 3. The line around each dot shows the range the real value plausibly falls in. If it crosses 1×, I can't tell it apart from no effect.

So the answer changes by customer. Early meetings go with winning in all three segments, by different amounts. Emails, calls and demos barely register anywhere.

There's an old sales idea that whoever calls back first wins. It makes sense. Someone fills in a form on your website, and there's a decent chance they filled in a competitor's form too. So I looked at how long each rep took to make first contact after an opportunity was created, using the same like-for-like, same-rep comparison as above.

In SMB, each time the wait doubles, the odds of winning drop by about 9%. Mid-Market is similar, about 12%. Both ranges stay below 1×, so it isn't just noise.

Enterprise goes slightly the other way (1.15; range 0.99–1.34), but the range crosses 1×, so I wouldn't read anything into it. Enterprise deals take months. A day's delay at the start probably isn't what decides them.

So speed seems to matter more for smaller deals than for big ones, though how much, and how sure I am, differs by segment. One confession: I only built a speed effect into SMB. The Mid-Market result is something the analysis found on its own, possibly because reps who respond quickly also differ in other ways. It's a good reminder that a clear result isn't automatically a real one.

Some deals are easier from day one. A customer who was referred by a friend is in a different place from one who got a cold email. An existing customer buying more is different from a stranger. If you're judging reps or forecasting revenue, it helps to know how much those starting points differ. So I put the main characteristics of each deal into one model together (where it came from, whether it was new or existing business, whether a competitor was involved, and how big the deal and the customer were) and looked at each one while holding the others steady.

• Referral: about 1.9× the odds of winning compared with an outbound (cold) opportunity.
• Inbound: about 1.5× the odds of an outbound opportunity.
• Partner: 1.13, but the range crosses 1×, so I can't tell it apart from outbound.
• Expansion (an existing customer buying more): about 1.8× the odds of a brand-new customer.
• A competitor in the deal: about 42% lower odds.
• Deal size: each time the deal doubles in size, the odds drop by about 10%.
• Company size: basically nothing (1.00) once deal size is accounted for.

What this doesn't say: that referral deals win because someone referred them. Referred customers might simply be closer to buying in the first place. The model tells you which kinds of deals win more often, not why.

One wrinkle. In Enterprise alone, company size comes out at 0.87 while deal size shows almost nothing. Bigger companies sign bigger deals, and with about 1,200 Enterprise deals the model can't cleanly tell the two apart. I'd call that uncertain rather than a real company-size effect.

This one's a little tricky. When I built the fake company, I gave region no effect at all. A deal in Asia-Pacific (APAC) was supposed to be exactly as winnable as one in North America. But a straightforward comparison disagreed.

So the real question is whether this is actually a regional effect, or whether it's really about the people working those regions. APAC has only 11 reps across all three segments, and a few of them happen to win less often than average. There aren't many others to balance them out. Look at how much the reps differ from each other, compared with how little the regions differ:

The 2,146 APAC deals looked like a lot of evidence. Really, they're 11 people's worth of evidence. Treat every deal as independent and you can end up very confident about something that isn't there. (Statisticians call this confounding: deals sit inside reps, and reps sit inside regions.)

Back to my starting guess: do the best reps just do more? Partly, yes. I compared the top quarter of reps, based on how much of their sales target they hit, with everyone else. In 2025, the top reps logged 47–92% more activity, depending on the segment. But you can split a rep's revenue into three pieces:

Revenue = Deals closed × Win rate × Average deal size

That's how many deals they get to a yes-or-no decision, how often the answer is yes, and how much each yes is worth. Then you can ask which piece accounts for most of the gap between the top reps and everyone else.

In every segment, win rate accounts for the biggest share. Top reps do more, and that helps. But most of their advantage comes from winning a bigger share of their deals.

This splits up what top reps achieve. It doesn't explain why they convert better. That could be how they run a deal, which deals they choose to chase, or things this data simply can't see.

I don't think the data supports one activity target for everyone. The relationship changes depending on the customer.

• Early meetings are associated with higher odds of winning in all three segments, although I'm more cautious about the size of the SMB result.
• Speed of first contact seems to matter more for SMB and Mid-Market than for Enterprise.
• Where a deal comes from matters. Referrals and inbound leads win more often than cold outbound.
• Existing customers buying more win more often than brand-new ones.
• Competition goes with lower odds of winning.
• Bigger deals are a bit harder to win.
• Top reps do more, but most of their edge comes from converting more of their deals.
• And a variable can look important just because of who happens to be working it.

Which brings me back to the napkin math. If I'm going to tell a rep that X meetings should produce Y dollars, I probably need to know what kind of customers they're selling to first.

The data

A simulated software company: 60 reps, 5,000 accounts, 13,107 opportunities and 402,628 logged activities, running from 2023 to 2025. I analyzed 2024–2025. The first six months exist so there's already a full pipeline of deals when the analysis starts.

The tools

I used Claude Code to create the simulated dataset, and that involved Python to write the data, DuckDB to store it and SQL to turn the raw activity logs into something I could actually analyze. I pulled the resulting data and used a combination of SQL and Excel to do the analysis. Excel was basically my first database, and I still find myself using it when I'm working through an analysis because there's something about being able to visually look at the data that helps me engage with it. It also gives me another way to double check what I'm getting out of SQL.

The method

Everything is split by customer segment, because small and large deals behave differently. The main models are logistic regressions, which estimate how each factor relates to winning or losing while holding the others steady. The uncertainty is worked out per rep rather than per deal (“clustered standard errors”), because one rep's deals aren't independent of each other. The activity and speed results compare deals within the same rep. The deal-characteristic results come from one model across all segments.

Activity windows

“Early” means the first 14 days for SMB and the first 30 days for Mid-Market and Enterprise. I picked those windows knowing how long the simulated deals take, which a real analyst wouldn't know as precisely.

What I built in, and what the analysis found

Built in

When I created the dataset, I built in a few relationships that I wanted to see if the analysis could recover. Meetings should matter, especially in Enterprise. Faster first contact should matter in SMB. Referral, inbound and expansion deals should be easier to win. Competitors and bigger deals should make things harder. Region and industry should have no effect.

Found

The analysis found most of those things. It also found things I didn't put there. Speed showed up in Mid-Market even though I didn't build a Mid-Market speed effect into the simulation. The simple regional comparison also found an APAC effect that wasn't supposed to exist.

That's actually one of the reasons I like this project. The analysis didn't just give me the answer I expected. Sometimes it missed something. Sometimes it found something I didn't put there.

Fair to say

In this simulated company, activity, speed and deal characteristics relate to winning differently depending on the customer, and a simple analysis can find things that aren't there.

Still uncertain

Most things about Enterprise, which has the fewest deals. Why top reps convert better. And whether any of this would look the same in real data.

Caveats

• This is simulated data. No real company, customers or salespeople are involved.
• These are associations, not causes. The analysis shows what goes together, not what makes what happen.
• I ran a lot of comparisons. Run enough and a few will look meaningful by chance, so don't hang much on any single number.
• There are only 16–24 reps per segment, which isn't many.
• Enterprise has the fewest deals, so its numbers are the least certain.
• Company size and deal size move together, which makes them hard to separate.
• The first quarter of 2024 comes right after the simulation's warm-up period, so it isn't a normal quarter.