Consumer & Beverage · Commercial analytics · 6 weeks
We found a 2.6× revenue lever hiding in two years of till data.
185 products, 22 months and $2.57M of transactions — reconciled, normalised for footfall, and turned into a tap list that pays for itself.
A craft beverage producer with a high-traffic direct-to-consumer taproom, US Northeast
2.6×
Take-home sales lift when a product is on tap
185
Products analysed across 22 months
$2.57M
Revenue placed under measurement
2.9%
Of all volume was invoiced at zero, undetected
Recognise any of this? Book a workflow audit.
At a glance
Goal
Find out whether putting a product on draft actually drives take-home sales of that same product — and if so, which products, and by how much.
Result
A measured, significance-tested conversion rate per product, a ranked list of which products earn their tap line and which cost money to keep, and two source-data defects fixed that had been quietly distorting every report in the business.
Project duration
6 weeks — 2 weeks data reconciliation, 3 weeks analysis, 1 week handover.
The problem
The taproom was this company's highest-margin channel and its least-instrumented one. Leadership believed that pouring a product on draft drove take-home sales of the same product — the tasting-room-as-shopfront theory that most of the industry runs on — but nobody had ever tested it. So the tap list was set by intuition and brewer preference rather than by measured conversion.
Two years of point-of-sale data existed. It sat in spreadsheets that nobody trusted enough to make a decision with, because the totals never quite reconciled and comparing a quiet February against a peak August felt like comparing nothing at all.
What it was costing them
- Roughly a dozen tap lines allocated each month with no evidence behind the choice
- Seasonal swings of 6.5× between the slowest and busiest months made every raw comparison meaningless
- A zero-priced line item was suppressing per-product figures across every report
- The same products appeared under several spellings, splitting their sales history in two
What we did
Method
Five steps, in order. Each one exists because the step before it produced something we could not yet trust.
- 01
Normalise for footfall
Raw unit counts are meaningless when monthly visitors swing 6.5× between winter and summer. We rebuilt every metric as units per 1,000 visitors and dollars per 1,000 visitors, which made a February pour directly comparable to an August one for the first time.
- 02
Reconstruct the tap calendar
We built a monthly on-tap / off-tap flag for all 185 products. That single derived column turned two years of static sales history into a natural experiment — every product that rotated off the tap list became its own control group.
- 03
Measure the lift, then test it
For every product with enough off-tap months, we compared take-home performance on-tap against off-tap and significance-tested the gap. This is the step most analyses skip, and it is the step that separates a real effect from a product that happened to be on tap during a good summer.
- 04
Profile what converts
We regressed conversion against product and release attributes, moving the answer from 'which products convert' to 'which kinds of product convert' — so the finding applies to next year's releases, not just last year's.
- 05
Fix the source, not the report
The zero-price defect and the duplicate product records were corrected at the point of entry rather than patched in the analysis. Every downstream report got more accurate at once, permanently.
What we found
Does being on tap actually drive take-home sales?
Yes — 2.6× more units and 2.8× more revenue.
The clearest statistically significant case sold 15.1 units per 1,000 visitors while on tap against 5.7 once it rotated off — a 2.6× difference — and generated $265 per 1,000 visitors against $96, a 2.8× revenue lift. The tasting-room-as-shopfront theory holds. It just doesn't hold for everything.
Take-home performance, on tap vs off tap
Units / 1,000 visitors
Dollars / 1,000 visitors
Same product, 14 months, significance tested · 2.6× units / 2.8× revenue
So should everything go on tap?
No. Ten products convert negatively.
Ten products showed a negative effect — on-tap availability actively cannibalised their take-home sales, because customers who would have bought a bottle to take home drank a glass instead. The worst offender cost roughly $150 per 1,000 visitors in lost take-home revenue for every month it held a tap line. Those lines were being given away.
Tap line ledger — keep or cut
Δ units per 1,000 visitors · on tap minus off tap
- SKU A+25.4
- SKU B+16.5
- SKU C+13.5
- SKU D+13.1
- SKU E+11.3
- SKU F+10.4
- SKU G+10.3
- SKU H+10.1
- SKU I+9.7
- SKU J+9.4
Below this line, the tap line costs more than it returns
SKU P−2.7- SKU Q−3.5
- SKU R−3.7
- SKU S−4.7
- SKU T−4.7
- SKU U−4.8
- SKU V−5.0
- SKU W−6.2
- SKU X−6.5
- SKU Y−10.3
+129.7
10 lines earning their place
−52.1
10 lines being given away
+67%
Available from reallocation alone
Reallocating the ten negative lines is worth 52.1 units per 1,000 visitors — against the 77.6 the list nets today
Can we predict which new products will convert?
Partly — style predicts about half of it.
Adjunct-ingredient, barrel-aged and fruited profiles dominated the top converters; hop-forward and traditional Belgian profiles dominated the bottom. The attribute models explained 47% and 41% of the variance in popularity and revenue. That is enough to guide a tap list and not enough to automate one — which is exactly what we told them, rather than shipping a black box that would be wrong half the time.
Is the underlying data trustworthy?
Not yet — 2.9% of volume was invoiced at $0.00.
A single tasting-flight line item recorded 10,725 units at zero price. Because it was the highest-volume item in the entire dataset, it was silently deflating every per-product revenue metric in the business. Two products were also recorded under multiple spellings, splitting their sales history. Neither defect was visible in any existing report.
The business felt busier last year. Was it?
Volume rose 21% while revenue moved 1.6%.
Year over year on matched months, units sold rose 21.0% while revenue rose just 1.6% and footfall actually fell 0.7%. The same number of people were buying substantially more, cheaper items — a mix shift toward low-price pours that no existing report surfaced, and that explains why the floor felt busy while the bank balance did not.
Revenue against footfall, indexed to 100
One axis, both series indexed to month 1 · year two: units +21.0%, revenue +1.6%, visitors −0.7%
What changed
Tap list set by intuition and brewer preference
Tap list set by measured conversion, with a rule per product
Seasonal swings made month-to-month comparison impossible
Every metric normalised per 1,000 visitors and directly comparable
2.9% of volume invisible behind a zero-price defect
Defect fixed at source; all downstream reporting corrected
Two years of data nobody trusted enough to act on
A reconciled dataset and a repeatable monthly measurement
Projected annual impact — modelled, not realised
$180k–$240k
Applying the measured 2.6× conversion ratio to the products currently rotating off-tap during peak season implies $180,000–$240,000 in recoverable annual take-home revenue. This is modelled from the observed lift — it is a forward estimate, not a realised client result.
What this doesn't tell you
Every analysis has a boundary. We publish ours, because a finding you can't see the edges of isn't one you can safely act on.
- Under two years of history. Some products never left the tap list, so their conversion effect cannot be measured at all.
- 'Profitability' here means revenue per visitor, not gross margin — no cost or COGS data existed in the source system.
- Product attribute data was coarse, which caps how much of the variance any attribute model can explain.
- The analysis covers on-site sales only. It says nothing about wholesale or distribution behaviour.
How we handle your data
Your name never appears
Clients are described by sector, scale and region. Names, products, addresses and identifiers stay out of anything public — including this page.
Nothing raw gets published
Charts are rebuilt from the analysis with identifying labels stripped. No raw data, no notebooks, no exports leave your side of the engagement.
It can stay in your building
Where data cannot leave your environment, we build so it does not have to — models running on your hardware, no third-party API calls.
Next step
Have a version of this problem?
If any part of this looked familiar, the first conversation is free and costs you half an hour.
What happens next
- You pick the workflow. We ask what already exists, who touches it, and where it stalls.
- We reply within one business day with a rough shape and a rough timeline — usually smaller than people expect.
- If it makes sense, the next step is a working system in two to six weeks, tested by your own staff.
- You keep your existing tools. Private deployment is available where data cannot leave.
- And if it is not worth building, we say so instead of selling you one.