A while back we wrote about catching our own valuation engine giving two different answers to the same CIM, and the fix — separating judgment from arithmetic so the same facts always produce the same number. That fix addressed consistency. It didn't address a separate question we'd been sitting with since: consistent with itself is not the same thing as correct.
Our valuation engine cites industry-standard multiple ranges — SDE multiples, EBITDA multiples, RMR multiples for recurring-revenue businesses — pulled from the model's general training. That's reasonable, but it's also the one place in the whole system where "defensible" was softer than everywhere else. So we went and checked it against something we couldn't just reason our way into being right about: real, sourced transaction data.
What we actually compiled
Not just alarm companies, which is the industry we'd been testing with most. We pulled data across twelve Main Street and lower-middle-market verticals — pest control, home care, HVAC and skilled trades, fire and life safety, IT services, accounting practices, veterinary, property management, auto repair, distribution, landscaping, and general services.
The important part wasn't the breadth, it was the sourcing discipline. For every figure, we drew a hard line between two categories: transactions a public company actually disclosed in an SEC filing or press release, versus multiples reported by trade press citing unnamed "people familiar with the matter." Both categories can be useful, but they are not the same kind of evidence, and treating them as interchangeable is exactly the sort of quiet imprecision we try not to build into this product.
Some examples of the first category: Rollins' 8-K filings on its pest control acquisitions state actual purchase prices and, for some deals, the target's prior-year revenue — so a real multiple can be computed, not estimated. Addus HomeCare's filings go a step further; one 2018 acquisition disclosure states the company's own computed EBITDA multiple outright. That's about as verifiable as small-business M&A data gets.
Some examples of the second category: a $2.5B private equity acquisition of a home-services platform, reported at "approximately 18.5x EBITDA" by trade press citing sources — a real number, probably a reasonably accurate one, but not something the acquirer itself confirmed. Useful as a data point. Not the same tier of evidence as a filed 8-K.
What held up, and what didn't
In aggregate, our engine's general ranges weren't far off. The industry-survey benchmarks we cross-checked against — IBBA Market Pulse, the Pepperdine Private Capital Markets Report, BizBuySell's closed-transaction data — landed close to what the model was already citing for SDE and EBITDA multiples by business size.
One structural detail was worth actually encoding as a fact rather than leaving to inference: the earnings basis itself shifts around the $2M valuation mark. Below roughly $2M, Main Street transactions are conventionally priced on a multiple of SDE. Above it, the convention shifts to EBITDA — a different earnings measure entirely, not just a bigger number on the same scale. A model reasoning about this from general training might blur that line; a documented industry convention shouldn't be left to chance.
The widest real disagreement showed up in the recurring-revenue verticals — which is exactly where you'd expect it, since that's also where the most money rides on getting the multiple right. Home care in particular: five separate Addus HomeCare acquisitions, same acquirer, same general industry, and the computed revenue multiples ranged from 0.39x to 2.4x depending on whether the underlying business was standard personal care or higher-margin hospice care. "Home care" isn't one multiple. Neither, it turns out, is almost anything else once you look at real transactions instead of a category label.
What we're doing with it
The full reference set is now something we check the model's judgment against, vertical by vertical, the same way we check its arithmetic against a fixed formula. It doesn't replace the model's reasoning — picking exactly where in a range a specific business belongs is still a real judgment call, and we still think that should stay a judgment call, disclosed and bounded, not force-fit into false precision. What it does is give us, and eventually you, something to point at besides "the model said so."
We'll keep this updated as the underlying surveys refresh and as we find more disclosed transactions worth adding. If you're curious how a specific number in your own deal compares to real, sourced comps rather than a general estimate, that's exactly what this exercise was for.