How to Build a Revenue Model From Drivers

Introduction

When a candidate opens a blank spreadsheet and types 8% into the revenue growth cell, they have not made a forecast. They have made a wish, and they have made it in a form nobody can argue with, check, or improve. The number says nothing about what the company sells, how many of them it sells, what it charges, or what has to happen in the real world for the projection to come true. It is the modeling equivalent of a shrug.

A driver-based revenue model does the opposite. It breaks the top line into the physical and commercial units a business actually transacts in, projects each one separately, and lets the growth rate fall out at the end. Once the build exists, every conversation about the forecast becomes concrete. Instead of arguing about whether 8% is aggressive, you argue about whether a chain can open 120 restaurants a year, whether a software company can hold $14 million of annual expansion revenue while churn runs at 12%, or whether a marketplace can defend its take rate while sellers push back on fees.

That shift matters far beyond elegance. The revenue line drives the entire projection: margins, working capital, capital expenditure, debt capacity, and terminal value all scale off it. If the top line is a guess, everything downstream inherits the guess. This post covers the driver architectures used across sectors, two full worked builds with the arithmetic shown, how to sanity-check a build you have no confidence in, and how interviewers probe whether you actually understand the business or just filled a cell.

What a growth rate actually hides

Take a specialty retailer growing revenue 7% a year. That single number could describe at least four completely different companies. It could be a chain adding 10% more stores while same-store sales fall. It could be a chain with a flat footprint pushing 7% price increases into a market that will eventually resist. It could be a business gaining traffic in a shrinking category, or losing traffic in a booming one. Each version has a different margin path, a different capital requirement, and a different value.

The growth rate compresses all of that into one digit and destroys the information. Worse, it hides the duration question, which is what valuation actually turns on. Unit growth ends when the map fills up. Price growth ends when customers substitute. Mix improvement ends when the premium tier saturates. You cannot see any of those ceilings when the driver is invisible.

Price times quantity, and finding the real unit of sale

Every revenue build reduces to the same identity, and the whole craft lies in defining the two terms correctly:

The hard part is never the multiplication. It is naming the unit. For an airline the unit is not a flight, it is a revenue passenger mile, priced at yield. For a hospital it is not a patient, it is an adjusted admission, priced at net revenue per admission after payer discounts. For a payments company it is not a customer, it is a transaction, priced at a spread on volume. Pick the wrong unit and the model produces numbers that cannot be checked against anything the company or its industry actually reports.

A useful test: if management would recognize your unit and your price on an earnings call, you have the right ones. Companies build their internal plans in the same units they use to run the business, and those units usually appear somewhere in the filings, often in the operating metrics table just before the financial statements. When you have the right unit, you also inherit the right mix question, because revenue can rise from more units, higher prices, or a shift toward expensive units, and those three stories have different durability.

Most revenue builds you will ever construct are variations on six or seven templates. Knowing them cold means you can start a model in minutes rather than staring at a blank sheet trying to invent structure, and it means you can read a company's disclosure and immediately spot which numbers you are missing.

The architectures below are the ones that come up most in interviews and in live deal work. They are not mutually exclusive, and a diversified company usually needs two or three of them stacked by segment, which is exactly why segment disclosure matters so much.

Retail and restaurants: units times average unit volume

The classic build is a store count roll-forward multiplied by average unit volume, with comparable sales applied to the mature base only. Beginning units plus openings minus closures gives ending units, and revenue uses a weighted average of units open during the period rather than the ending count, because a store opened in November contributes two months of sales rather than twelve.

The three assumptions that carry the model are the opening pace, the productivity of a new store relative to a mature one, and the comp rate. New units almost never open at mature volume, and treating them as if they do is one of the most reliable ways to overstate a retail forecast. Cannibalization then sits on top: as the footprint densifies, new locations pull traffic from existing ones, which shows up as a drag on comps rather than as a line labeled cannibalization.

Software: the ARR bridge

Subscription software is modeled as a balance that rolls forward, not as a revenue line that grows. Beginning annual recurring revenue plus new bookings plus expansion minus churn and downgrades equals ending recurring revenue, and reported revenue is roughly the average of the balance across the period rather than the ending balance.

Annual Recurring Revenue (ARR)

The annualized value of a software company's active subscription contracts at a single point in time. A customer paying $2,000 per month represents $24,000 of ARR. It is a run-rate snapshot rather than a GAAP measure, so it almost never equals reported revenue for the same period, and the gap between the two tells you how fast the balance moved during the year.

The bridge structure is what makes a software model auditable. Each line has a different owner inside the company (sales for new bookings, customer success for expansion and churn) and a different cost attached to it, which is why a company buying growth through new logos looks very different from one growing on expansion even when the headline growth rates match. This is also why software businesses that are still unprofitable can still be valued sensibly, a problem covered in more depth in the post on valuing a company with no profits.

Marketplaces: GMV times take rate

Marketplaces sell access, not goods, so the build starts with gross merchandise value and applies a take rate. GMV itself decomposes into active buyers times orders per buyer times average order value, which is where the interesting operating story lives, while the take rate captures pricing power over the sellers on the platform.

Take rate is rarely a single number. It bundles transaction fees, payments, advertising, and shipping products, each with its own attach rate, and treating it as one blended percentage means you cannot explain why it moved.

Consumer subscription: subscribers, ARPU, and churn

Consumer subscription businesses use the same roll-forward logic as software with different labels: beginning subscribers plus gross additions minus cancellations, multiplied by average revenue per user. The distinguishing feature is that churn is high, frequent, and behavioral rather than contractual, so monthly churn compounds into an annual number that surprises people who have not done the arithmetic. A 3% monthly churn rate retains only about 69% of a cohort over a year.

ARPU deserves its own build rather than a single blended figure, because pricing tiers, ad-supported plans, geographic mix, and promotional pricing all pull in different directions. A company can raise list prices in every market and still report falling ARPU if growth is concentrated in cheaper regions or cheaper tiers.

Industrial: units, price, and the backlog

Order-driven industrials add a dimension the others lack: revenue in a given year is largely already sold. The build starts with beginning backlog, adds orders, subtracts revenue recognized, and ends with closing backlog, so backlog coverage tells you how much of next year's forecast is contractual rather than speculative.

Book-to-Bill Ratio

Orders received during a period divided by revenue billed in the same period. A ratio above 1.0 means the backlog is growing and revenue should follow; below 1.0 means the company is shipping faster than it is winning new work. GE Vernova, for example, reported $59.3 billion of orders against $38.1 billion of revenue in 2025, a book-to-bill of about 1.6, and ended the year with roughly $150 billion of backlog.

Backlog conversion rates matter as much as the orders themselves. Long-cycle equipment can sit in backlog for two or three years, so an order surge shows up in revenue slowly and a collapse in orders takes just as long to appear. Modeling revenue as a percentage of opening backlog plus a share of current-year orders is usually closer to reality than growing last year's revenue.

Advertising: impressions, fill rate, and CPM

Advertising revenue is a supply-and-pricing model. Available impressions come from users times sessions times ad slots per session, fill rate determines how many of those impressions actually sell, and CPM sets the price per thousand delivered impressions. A platform with 500 million monthly available impressions, an 80% fill rate, and a $6.00 CPM generates 400 million delivered impressions and $2.4 million of monthly revenue, or $28.8 million annualized.

Splitting supply from pricing is what makes the model useful, because the two move in opposite directions all the time. Loading more ad slots into a feed raises supply and usually depresses CPM and engagement, so a build that grows both at once without acknowledging the trade-off is quietly assuming the platform found free money.

Take a chain that starts the year with 1,000 company-operated restaurants generating average unit volume of $3.00 million, for prior-year revenue of $3,000 million. Management guides to 120 openings, spread evenly through the year, and comparable sales of about 2%.

Rolling forward the unit count

Openings spread evenly means the average number of new units open during the year is half the total, so weighted average units are 1,000 plus 60, and ending units are 1,120. New restaurants do not open at mature volume; assume they run at 85% of the mature average in year one, which is a conservative but common starting point for a chain still building brand density in newer markets.

The mature base first. Comparable sales of 2% lift the existing 1,000 restaurants from $3.00 million to $3.06 million of average volume, giving $3,060 million of revenue from the legacy base. The new units come next: 60 unit-years of exposure at 85% of $3.06 million, or $2.60 million each, contributes $156 million. Total projected revenue is $3,216 million.

That is 7.2% growth, and the decomposition is the entire point: comps contributed 2.0 points ($60 million) and new units contributed 5.2 points ($156 million). A model that simply assumed 7% growth would have produced almost the same revenue number while concealing that nearly three-quarters of it comes from capital spending on new locations, which requires capex, hits free cash flow, and decelerates automatically as the base grows. Run the same 120 openings against a 1,120-unit base the following year and the new-unit contribution drops to roughly 4.9 points without anything going wrong.

Cannibalization and the honest comp

The 2% comp above is a net number, and pulling it apart is where the build earns its keep. Suppose underlying comps run at 2.5% on traffic and price, while new openings in existing trade areas transfer roughly 0.5 points of sales away from nearby restaurants. The legacy base grows $75 million gross and gives back $15 million to cannibalization, netting the $60 million above.

That distinction changes the forecast's trajectory. Cannibalization is a function of opening density, so a chain accelerating openings into mature markets should show a widening drag, while a chain expanding into new geographies should not. Chipotle's disclosure illustrates why the split matters: the company opened 334 company-operated restaurants in 2025 and grew revenue 5.4% to $11.9 billion even as comparable restaurant sales fell 1.7%, according to its fourth quarter and full year 2025 results. Every point of growth came from the unit build, and a single blended growth rate would have hidden that completely.

Revenue builds show up in every modeling test: Work through financial modeling, valuation, and accounting questions with full explanations, start practicing interview questions for free and find out which mechanics you can actually reproduce under time pressure.

Now a software company entering the year with $100 million of ARR. The bridge has four lines, and the discipline is that each one is forecast separately rather than netted:

The bridge arithmetic

Assume $24 million of new bookings from new customers, $14 million of expansion from existing customers upgrading seats and tiers, and $12 million of churn and downgrades. Ending ARR is $126 million, up 26%. Gross retention is 88% and net retention is 102%, because the $14 million of expansion more than covers the $12 million lost.

Net Revenue Retention

The share of recurring revenue a company keeps from the customers it already had twelve months earlier, including upgrades, seat growth, and price increases, but excluding anything sold to new customers. Above 100% means expansion outweighs churn and the existing base grows on its own; below 100% means the company must sell new customers just to stand still.

Reported revenue will not equal $126 million. If ARR builds evenly, recognized revenue approximates the average balance across the year, roughly $113 million. Note that averaging alone does not create a gap: if ARR compounds at a constant rate, revenue compounds at the same rate. Reported growth lags the ARR growth rate only when ARR growth is decelerating, which is the usual case, because revenue is still catching up to a balance that grew faster last year. Interviewers ask about that gap specifically, and the answer is that ARR is a point-in-time balance while revenue is a flow.

Why the retention curve outranks the growth rate

Consider two companies that both enter the year at $100 million and both exit at $126 million, a 26% growth rate that would look identical in a screening table. Company A adds $11 million of new bookings, $20 million of expansion, and loses $5 million, for net retention of 115%. Company B adds $36 million of new bookings, $8 million of expansion, and loses $18 million, for net retention of 90%.

Hold new bookings constant for each and roll forward two more years. Company A compounds its base at 115% and reaches $155.9 million, then $190.3 million. Company B compounds at 90% and reaches $149.4 million, then $170.5 million, with growth decelerating from 18.6% to 14.1%. Company B is spending on three times the new-logo volume to end up 10% smaller, and every dollar of that volume carries sales and marketing cost that Company A did not have to spend.

Cohort analysis is how you get the retention line right instead of guessing at a blended average. Group customers by the period they signed, track what each group is worth in each subsequent year, and the shape of the decay becomes visible. The shape matters enormously: a cohort losing 8% in year one and then only 4% a year retains 78% after five years, while a flat 8% annual decay retains just 66%. Same starting churn, a $120 million difference on a $1 billion book, and no growth-rate assumption anywhere in the model would ever have surfaced it.

A driver build can be internally consistent and still be impossible. The reconciliation step asks two separate questions: can the company physically do this, and does the world have room for it?

Capacity is a hard ceiling

Every revenue driver has a resource behind it. New bookings require quota-carrying sales representatives, so a plan for $36 million of new bookings at $1 million quotas and 75% attainment requires 48 fully ramped reps, not the 40 currently employed. Store openings require a real estate pipeline, construction crews, and general managers who take months to train. Manufacturing revenue requires plant hours, and services revenue requires billable headcount at achievable utilization.

The check is simple: divide the revenue driver by the productivity metric and see what the model is implicitly hiring, building, or leasing. If the implied resource never appears in the operating expense or capex lines, the model is producing revenue for free. Working this through also keeps the revenue build honest about timing, because ramp periods delay revenue while costs start immediately, which is exactly the pattern that shows up in how working capital and the cash conversion cycle behave during a growth phase.

Reconciling top-down and bottom-up

The bottom-up build gives revenue. The top-down check converts that revenue into implied market share and asks whether the share gain is credible. Suppose the addressable market is $40 billion growing 6% a year, and the chain from the first worked build does $3.2 billion today, an 8.0% share. Two years out the market is roughly $44.9 billion and the model produces about $3.7 billion, an 8.2% share. That is a 20 basis point gain, which is defensible.

Change the model to $5.0 billion and the implied share jumps to 11.1%, a gain of more than 300 basis points in two years. That is not impossible, but it now requires naming which competitors lose that volume and why, and if you cannot name them, the build is wrong somewhere. The reconciliation does not tell you the answer; it tells you which claim you are making.

Plenty of companies do not hand you the unit and the price, and private targets in a live process often disclose less than public comparables. The workaround is triangulation rather than surrender.

Segment disclosure is the first place to look, since companies must report results for reportable segments and often include operating metrics there that the consolidated statements omit. Management commentary on earnings calls is the second: executives routinely quantify drivers verbally that appear nowhere in the tables, such as attach rates, pricing actions, or the productivity of a new format. Industry proxies come third, using a listed peer's disclosed unit economics as a benchmark and adjusting for scale and positioning.

The same handful of mistakes turn up in modeling tests and in real staffing work, and each has a distinct signature that a reviewer can spot in seconds.

Every one of these is caught by the same habit: label every row with its unit and its period, and keep all assumptions in a single input block where they can be read at a glance.

Flexing revenue growth by plus or minus two points is a non-analysis. It tells the reader that revenue could be higher or lower, which they already knew, and it gives no indication of which real-world event would cause it.

Sensitivity on drivers answers a different question. Flex the opening count from 100 to 140, or net retention from 95% to 110%, or take rate by 100 basis points, and the output tells you what each specific business risk is worth. Those cases can then be defended in a meeting, because each corresponds to something a management team or a lender can be asked about directly.

Scenarios go one step further by bundling drivers into coherent stories. A downside case where comps fall and openings slow together is realistic; a downside case where openings slow while comps accelerate usually is not, because the same weak demand environment drives both.

The question is rarely asked as "build me a revenue model." It arrives disguised, and the disguise is the test.

The most common version is a company prompt: "How would you forecast revenue for a gym chain?" The wrong answer starts with a growth rate. The right answer names the unit within the first sentence (locations times members per location times monthly dues times twelve), then adds the losses (member churn, promotional pricing) and the ceiling (capacity per club, catchment population). A good follow-up is always about the ceiling, so have one ready.

The second version is a live case where the interviewer feeds you disclosure and watches which numbers you reach for. They are checking whether you can find the operating metrics table and whether you notice that a metric changed definition. The third version shows up in an Excel test, where structure carries as much weight as arithmetic and a clean input block matters as much as the formulas, a pattern covered in what interviewers look for in Excel modeling tests. Expect a margin question immediately afterward, since the natural follow-up is what the revenue mix does to profitability, which is where what a good EBITDA margin looks like by sector becomes the relevant benchmark.

A 160-page reference for the technical rounds: Our PDF walks through valuation, modeling, accounting, and deal-structure questions with worked answers, access the IB Interview Guide and use it as your revision spine in the weeks before superdays.

The difference between a revenue line that holds up in a committee meeting and one that gets torn apart is not sophistication. Plenty of driver builds are simple: two or three rows of units, one row of price, one row of leakage. What separates them is that every number in the build refers to something real, so when someone challenges the forecast, the conversation moves to whether the chain can really open 120 restaurants or whether net retention can hold at 102%, which are answerable questions.

Building the revenue line properly is also what makes the rest of the model meaningful, since margins, working capital, and capital expenditure all key off it, and the full three-statement model build inherits whatever quality you put into the top line. A well-driven revenue build makes the downstream statements arguable in the same way, which is the entire point of modeling.

Practice this on companies you can actually check. Pull the operating metrics table from a filing, build the unit and price rows yourself, and compare your output to what the company reported. You will be wrong at first, and the wrongness is diagnostic: it points at the driver you misunderstood, which is exactly the feedback a growth-rate assumption can never give you.