#The challenge
Expansion was the growth engine and nobody could say what drove it. Loss reasons were rarely populated, growth notes were empty, and the recurring-revenue field was blank on the vast majority of records in one product's opportunity file. Win rates had been falling for two years and the decline was visible in every major vertical, but there was no agreed explanation. Leadership wanted a prioritised list of accounts to work and a defensible number attached to it, and the honest first answer was that the data quality would not support one without a lot of preparatory work.
#The approach
Start from the CRM and establish what actually triggers an expansion
The first wave read roughly 73,000 CRM records across seven exports: account qualifications, three years of expansion opportunities, line items, account firmographics, an expansion timeline, product-specific opportunities and contacts. Two trigger reasons accounted for the large majority of all seat additions — an over-subscription signal fired by the product itself, and a direct customer request. Both convert at high rates. Trigger-category comparison was stark: seat-addition-triggered opportunities converted from qualification to won at a rate more than fifteen times higher than outbound-triggered ones, and won at the opportunity stage more than thirty times more often.
Measure velocity, because it reframes what the motion even is
Median time from opportunity creation to close on these expansions was zero days. Over half closed the same day, and roughly three quarters closed within a week. That single fact changes the interpretation of everything else: this is not a sales cycle, it is an order-taking motion triggered by a product event. It also became the rebuttal to a stakeholder's reasonable objection that the pre-expansion usage ramp might be a mechanical artefact of opportunities being created retroactively — with a median creation-to-close lag of about a day, there is no window for that artefact to form.
Segment the base by expansion behaviour, not by size
Roughly a fifth of accounts are multi-expanders and produce close to eighty percent of expansion volume, averaging four expansions each. Another third expand exactly once. Nearly half never expand at all. That distribution is the entire argument for account prioritisation: the question is not which accounts are large, it is which accounts have already shown they expand.
Find the motion nobody is running
Across roughly 6,700 expansions, only about forty were tier upgrades — a rate of about one percent. Almost all of the ones that did happen were bundled with a seat addition rather than sold standalone, and bundled upgrades won at a materially higher rate than standalone ones. That is a specific, actionable finding: the tier upgrade is not a campaign, it is an attachment to a motion that is already converting.
Rank products by win rate and stop pitching the ones that lose
Seat additions across all three tiers won at roughly 75 to 82 percent. Every add-on product won in the teens to twenties, and one integration add-on won at single digits across two dozen attempts. The AI product had a further timing finding: nearly all of its purchases came from existing customers rather than new ones, the median time from becoming a customer to buying it was around four years, and under four percent bought it in their first year. The recommendation followed directly — stop pitching it to new and early-tenure customers.
Connect the customer-success health score to expansion, then look inside it
Accounts in the healthiest health-score band expanded at around sixty percent with an average of two expansions each; accounts in the highest-risk band expanded at around forty percent with well under one. Crossing health score with product usage produced the single most operational cut in the study: low-risk-score accounts with high usage expand at roughly forty-five percent, against roughly twenty-three percent for high-risk accounts with low usage. That quadrant is the account list.
Test the timing hypothesis with non-parametric statistics, and say why
The distribution failed a normality test decisively, so the analysis ran on rank-based methods — Wilcoxon, Mann-Whitney, Kruskal-Wallis, Theil-Sen, Spearman — with log-transformed parametric tests only as confirmation. Every reported effect carries its p-value and its sample size, and the effects that are not significant are reported as not significant rather than quietly dropped.
Report the findings that kill your own recommendations
Three negative results made the deliverable. External corporate events — funding rounds, acquisitions, leadership changes — collected across roughly 300 events on about 150 top accounts showed no significant relationship with expansion timing, and the funding-event subgroup showed the worst usage outcomes of any. Survey signals do not predict expansion: satisfaction respondents and promoters both won at slightly lower rates than everyone else. Meeting volume and meeting completeness showed no meaningful correlation with expansion outcomes at all. And downsells ran at well under a tenth of a percent, so the explicit recommendation was to allocate no resources to downsell prevention.
Backtest the headline signal before selling a pipeline built on it
The pre-expansion usage ramp was the study's marquee finding, so it was backtested as a trigger rather than assumed. It came back at roughly five percent precision and twenty-six percent recall, meaning about twenty accounts have to be touched per conversion. Sweeping the growth threshold from ten to twenty-five percent changed precision by almost nothing. The reason is simple and fatal: about three quarters of active accounts fire that signal at some point. The sized opportunity attached to that segment was then risk-adjusted down by an order of magnitude to reflect the backtested conversion rate, and the recommendation became to stack the signal with health score and vertical rather than to run it alone.
Convert vertical performance into a forecasting discount factor
One vertical won at around thirty-eight percent against a roughly fifty percent standard, which translates to a discount factor a little over three quarters that gets applied to pipeline in that vertical. Tenure was the discriminator inside it: new customers in that vertical won at over half, tenured ones at under a quarter. The vertical was framed as an explicit fix-or-exit decision with a defined sprint on one side and sunsetting new acquisition on the other, rather than left as an observation.
Separate execution gaps from market gaps
One region won at around fifty-five percent against another region's seventy. Because the gap persisted within the same verticals, it was characterised as roughly eighty percent execution-driven rather than market-driven, and sized as incremental opportunity at parity rather than treated as a structural difference. That distinction determines whether the answer is enablement or a territory change.
Revise your own conclusion in public when the deeper cut disagrees
The third wave concluded that usage does not rise after expansion and that one expansion type showed a significant decline. The fourth wave, with a longer window, revised this: usage stabilises above the pre-expansion baseline and is still roughly ten percent above it at five to six months out. Both readings are in the record with their statistics attached. A study that never contradicts itself across waves is usually a study that stopped looking.
#Outcomes
Seven prioritised segments, sized three ways
A sized addressable expansion opportunity across roughly 5,200 qualifying accounts, split into seven actionable segments with conservative, base and optimistic cases, and with the largest segment explicitly risk-adjusted down to its backtested conversion rate rather than carried at face value.
A ranked action list, not a report
The recommendations are ordered and specific: wire the health-score-plus-usage quadrant as a play in the customer-success platform and route a defined subset to named reps; work a scored hot list of multi-expansion accounts showing the pre-expansion ramp; run an attach campaign for the highest-value add-on to a defined account set; apply the vertical discount factor to that pipeline; diagnose the regional execution gap; and stack the usage signal with health score and vertical rather than firing it alone.
A documented reason not to build several things
External-event triggers, satisfaction and promoter signals, meeting-activity metrics and downsell prevention were all tested and all came back without predictive value. Knowing what not to instrument is a deliverable.
Analysis delivered; implementation not evidenced
Every recommendation in the record is a recommendation. There is no artefact showing the health-score play wired, the usage-ramp alert built, or the hot list worked. A further enrichment phase covering hiring signals, company age and technographic markers was requested and scoped, pending the enrichment run. Treat this as an analytics engagement, not an implemented programme.