Square Footage Explains 9% Of What A Building Spends On Mechanical Service

A 14x Range Is Not A Benchmark

Go looking for what a commercial building spends on mechanical service and you will be handed a dollars-per-square-foot figure. Depending on the source it will be somewhere between $0.15 and $2.15 per square foot per year. Those endpoints are third-party figures we have seen quoted, not Level measurements. They are not linked here because the publications rarely state the method, building mix or trade scope behind them. Treat them as observed examples of what gets quoted, not a measured distribution.

That is a fourteenfold spread. Nobody would accept it for labour rates or material costs. It survives here for a likely reason: few publishers hold real per-building revenue, and $/sqft is computable from a floor plan. It is the number you can produce without the data, so it is the number that got published.

The original research inventory covered 25,912 commercial buildings. Matched subsets differ by enrichment and model; that inventory count is not automatically the denominator of every test below. The published segment dataset retains the floor-area fit of r-squared 0.093. Additional historical coefficients below are explicitly limited because their exact query and eligibility populations are not reproduced on this page.

Nine Percent

Regress annual revenue per building against floor area and the r-squared is 0.093.

The fitted relationship accounts for about 9% of variation in this sample. That is association, not a causal share of spending attributable to size. The remaining variation includes unmodeled factors, measurement differences and noise; it is not evidence that 91% is inherently unknowable.

The measured variable is one mechanical contractor's annual billed revenue at a building, not the building's total mechanical spend across every vendor and in-house staff. Quoted $/sqft figures rarely state which they measure.

This is not a statement that size does not matter. It does, and we publish the gradients: in offices, the largest third of buildings bill about 2.8x the smallest third, and in multifamily the spread is wider still. Size orders buildings within a type reliably enough that it was the one enrichment field worth paying to join, and the only one that survived a paired within-contractor sign test at p=0.002.

The claim is narrower and more useful: size works as a band, and fails as a rate.

Dividing By Area Makes The Estimate Worse

This is the claim most worth testing before you rely on a per-foot rate.

If you convert revenue to a per-square-foot rate, you have not normalised the number. You have introduced a second measurement, with its own error, into the denominator of the first. The legacy methodology note reports 19% more variance after dividing by area. Its error statistic and normalization are not documented here, so this is an unverified historical research reference, not a reproducible comparison of raw dollar variance with dollar-per-foot variance.

A per-square-foot figure can be less useful when floor area is mismatched to the serviced building or the relationship is not proportional. That has to be tested with a comparable prediction-error measure, not assumed for every building.

The 19% reference cannot support a current pricing decision until the matched cohort, units and error comparison are reproducible. One mechanism is easy to see with a fictional example. Two clinics each bill $30,000 a year. One stands alone on its parcel at 30,000 square feet. The other shares a parcel with a second structure, so a parcel-level area reads 60,000 square feet. Revenue per building is identical, yet $/sqft reads $1.00 for one and $0.50 for the other. That 2x gap is created entirely by the denominator.

There is a structural reason. The relationship is not proportional. Mechanical load does not scale with floor plate, it scales with what is happening inside: a restaurant's kitchen sets its load, not its dining room, which is why restaurant spend is flat across floor area entirely. We measured that gradient and it is not even monotonic, so a size multiplier applied to restaurants would be confidently wrong on one of our largest samples.

Warehouse space, by contrast, is mostly unconditioned volume. Two buildings of identical area, one a clinic and one a distribution shed, are not remotely comparable per square foot.

Where The Other 91% Lives

The historical analysis tested several enrichment candidates. The coefficients in this section are legacy, unverified references, with source-query details and matched subsets not reproduced here. They suggest checks to run, not proof that an enrichment can never help or that its cost should be removed.

Building type is the real driver. Contractor-entered building type explains more of the variance than any physical attribute we could join, and grocery and convenience stores bill roughly 3.3x what restaurants do per building per year.

Climate looked like a 1.7x driver and was not. As a group mean it produced a clean gradient. Then we compared each contractor against its own book, and it vanished. It was telling us where contractors are located, not what buildings cost to run. The original r-squared of 0.004 was right.

Building permits looked promising and did not survive. 2.17x as a group mean, then a paired test at p=0.067. Suggestive, not established, and probably a size proxy.

A 205-code property-use taxonomy explained less than the contractor's own free text. Zero of 43 adjacent steps separated. Fragmentation, not signal.

Occupant firmographics explained nothing. Revenue and employee count at the occupying business did not predict the building's mechanical spend. In this sample, the occupant-load belief did not hold up as a predictor.

The sharpest signal we found anywhere was rooftop equipment presence, at 3.32x with standard errors between 1.5% and 2.6%. It has one problem: it lives in an asset record, which means you only know it after you already service the building.

The Honest Answer About The Remaining Variance

Unexplained variance in one fitted model does not prove that other data cannot predict it. The historical year-over-year and equipment coefficients below are retained as unverified references, not current forecast accuracy or irreversible limits on enrichment.

Within one contractor's own book, of one building type, the per-building band is still about 4.7x. Across the whole population, 59.6% of buildings move more than 2x year over year, and last year's actual invoices explain only r-squared 0.349 of next year's.

Much of it looks less like a data gap than the year-to-year randomness of what breaks, plus scope and contract changes we cannot see. No enrichment source we tested closed it, which is why any honest per-building estimate is a range and never a point.

What To Use Instead

Price off building type first, then size within type. The type sets the tier. Size moves you within it. That ordering matters because the type effect is larger than the size effect and is free to obtain.

Quote a range, and let its width mean something. A band that is tight where the data is dense and honestly wide where it is thin carries information. A single number smells like false precision and invites a customer to test whether it is exactly that.

Test it on your own book. Match each serviced structure to its own area and annual billed revenue. Within building type and comparable service scope, train building-level and area-based models on the same buildings. Compare held-out revenue prediction errors in the same dollar units. Comparing the numerical spread of dollars with dollars per square foot is not a valid test because the units differ.

Stop benchmarking yourself against a $/sqft figure. If a published rate spans 14x, then whatever you compute lands inside it, which means it can never tell you that you are wrong. A benchmark that cannot fail you is not a benchmark.

How This Was Measured

One mechanical contractor's billed revenue from one building over a year, contractor-weighted so no single firm can carry a figure, on a six-month minimum activity window annualised on the actual window. Non-operating tenants and records averaging under $300 a job are excluded.

Floor area comes from assessor records for the regression, and from FEMA/ORNL USA Structures (CC BY 4.0) where we need a per-building rather than per-parcel measurement. That distinction matters for attribution: the original joined-address analysis reported 19.4% with multiple structures and up to nine at one address. Its exact joined population is not reproduced here, so those figures are retained as unverified references. Confirm which structure the contractor actually services rather than assigning an entire parcel's area to it.

These coefficients are HVAC and mechanical only. Segment ordering is measurably trade-specific, so they should not be carried into another trade.

The full per-segment table with the size gradients, the held-out validation, and the cells we measured and chose not to publish is in the revenue per building by segment dataset, free to cite under CC BY 4.0. To get the figure for one building, the building revenue calculator takes an address.