Sizing a Private Company From Public Headcount Signals

Estimating how many people work at a private company that publishes no official figure is a triangulation problem, not a lookup. This guide is for strategy and research teams, talent-intelligence analysts, and operators sizing a market, and it gives you an end-to-end procedure that combines several public signals, corrects each for its documented bias, and forces a confidence range instead of a false-precision single number.

The failure most analysts commit is trusting one source. A public profile count, a revenue-per-employee back-out, and a layoff filing each carry a different error in a different direction. Read alone, any one of them will mislead you. Read together, with the right corrections, they converge on a number you can defend in a partner meeting or a market model.

Why one public number is never enough

Every public headcount signal misfires, and the direction of the error changes by industry, so no single source or single correction factor produces a trustworthy count on its own.

Start with the platform employee count, the number most people reach for first. It has a well-documented accuracy profile. For large companies over 1,000 employees with strong platform adoption, the count usually lands within 10 to 20 percent of actual. For smaller companies or low-penetration industries it can be off by 30 to 50 percent. The direction of that miss is not random. For manufacturing, construction, and hospitality, expect the count to underrepresent true size by 20 to 40 percent, because most of the workforce never builds a profile. Meanwhile employee ranges are self-reported, and companies are far more likely to inflate their count than underreport it, so a polished software firm's stated range almost always skews high.

That is the whole problem in one sentence: the same tool over-counts a SaaS company and under-counts a warehouse operator. If you apply one blanket haircut, you make one of those estimates worse.

There is a second, deeper reason a lone number lies: only 10 to 20 percent of profiles truly reflect real-world, up-to-date employment. Recent hires have not updated. People who left still show the old employer. In hot sectors, inflated or duplicate profiles are common. A raw count is a noisy snapshot of a moving population, and you cannot fix noise with a coefficient. You fix it with a second and third independent measurement.

The public signals worth pulling

Four public sources carry usable workforce data, and each answers a different question: what people say, what revenue implies, what the law forces companies to disclose, and what the company says about itself.

The first is the self-reported profile count and its department-by-function split, covered above. The second is revenue per employee. The third is the primary legal record, which is the only category that gives you a hard number rather than an estimate. In the US, the WARN Act requires employers with 100 or more employees to give at least 60 days advance written notice of a plant closing or mass layoff affecting 50 or more workers at a single site. Some states set lower thresholds, for example New Hampshire triggers at 75 or more employees. Aggregated WARN databases hold more than 82,000 layoff notices across 49 states covering over 8.87 million affected workers, searchable by state, company, or year. A single filing is a rare gift: it proves a site crossed those thresholds, an absolute floor no profile scrape can give you.

For UK entities the primary record is even richer. All limited companies, including small, inactive, and subsidiary companies, must file annual financial statements with Companies House, and those are public. Critically, a company that chooses to withhold its profit and loss account still has to include an employee-numbers note in the accounts it files. That note is a disclosed, audited figure. The public register holds structured financials for nearly 3 million companies.

The fourth source is the company's own careers page and "About Us" count. Most companies disclose a headcount somewhere on their own site, and for smaller businesses it is often more accurate than any third-party estimate, though it demands manual capture and can be stale.

The revenue-per-employee back-out, and where it breaks

Revenue per employee (RPE) is net revenue divided by average full-time-equivalent headcount, so you can rearrange it to back out headcount as estimated revenue divided by an industry RPE benchmark. It is a powerful independent estimator, but only where the benchmark band is tight.

The spread is enormous. NYU's Stern School tracked RPE across more than 90 US industries and found annual figures ranging from under $10,000 to well over $1 million per employee. Only three of those 90-plus sectors average more than $1 million: entertainment software leads at $1.76M, real estate development at $1.42M, and brokerage and investment banking at $1.3M. Technology companies often average around $300,000 per employee, retail closer to $150,000, and labor-intensive assembly operations may fall below $150,000. Capital-intensive sectors routinely exceed $500,000.

The lesson is that RPE is a strong estimator inside a well-defined segment and near-useless across a mixed one. For private SaaS you have an unusually tight anchor: a 2025 survey of more than 1,000 firms found a median of $129,724 per FTE, up from $125,000 the prior year, and it is one of the few sector figures with a disclosed methodology and sample size. A simpler gut-check for SaaS is the ARR-to-headcount ratio of roughly $100k plus or minus $50k per employee, meaning a $1M ARR business tends to have between 7 and 20 people.

The following benchmarks are the ones I reach for. Treat the column-two figures as anchors within a segment, not as comparable across rows, because the underlying methodologies differ.

Two cautions before you divide. First, decide whether your benchmark uses point-in-time or average headcount for the period; average smooths out spikes from acquisitions or hiring waves. Second, segment before you back out. A firm with one division at $400,000 RPE and another at $150,000 produces a blended figure that neither team should own, and dividing by the blend gives you a headcount that is wrong for both.

The procedure, start to finish

Run these eight steps in order. The whole pass takes a focused analyst roughly four to eight hours depending on how much primary-record digging the target requires. There is one real disagreement in the field: vendor write-ups anchor first on RPE and treat the profile count as a cross-check, while practitioner pieces anchor on the profile count and treat RPE as the cross-check. I put the profile count first because it also gives you the department mix you need for step seven; RPE alone gives you a total and nothing else.

The capture date in step two matters more than analysts expect. Because profiles go stale, a count is only meaningful with a timestamp, and step seven's growth math is impossible without two dated snapshots you can compare.

Turning three estimates into one range

The reconciliation step is where the discipline pays off: you take the corrected profile midpoint, the primary-record anchor, and the RPE back-out, and you convert their spread into an explicit low/high band rather than averaging them into a fake point estimate.

The rule is simple. When the three methods cluster inside 10 or 15 percent of each other, your band is tight and your confidence is high. When they scatter, you widen the band to cover the spread and drop your confidence label. A primary record, when you have one, is not just another vote; a WARN floor or a Companies House note overrides the estimates below it, because it is a forced disclosure rather than a guess.

Here is a reconciliation you can copy and fill for any target.

Target: [company], [industry], [HQ country]
Capture date: [YYYY-MM-DD]

Method 1 - Corrected profile count
  Raw platform count: [n]
  Industry correction: [+/- %]  (e.g. +30% for manufacturing undercount)
  Corrected midpoint: [n]

Method 2 - Primary record (if any)
  Source: [WARN / Companies House note / careers page]
  Figure or floor: [n]

Method 3 - RPE back-out
  Revenue anchor: [$]
  Industry RPE benchmark: [$/employee]
  Implied headcount: [n]

Reconciliation
  Low bound: [n]   High bound: [n]   Best estimate: [n]
  Confidence: [high / medium / low]
  One-line rationale: [why the band is this wide]

This is the exact place where getting the underlying people data is the bottleneck, because a corrected profile count is only as good as your ability to see who actually works there by role and location. Asking in plain English and getting the right people back removes the manual scrape and the guesswork about coverage.

Growth rate and department mix

Compute growth over two dated snapshots, and always read function-level growth alongside the headline number, because the mix reveals a company's strategy while the total conceals it.

The math has two forms. For total-period growth use (Ending minus Beginning) divided by Beginning times 100. For an annualized rate across several years use (Ending divided by Beginning) raised to the power of (1 divided by years), minus 1. Then place the result against reference bands: mature, stable organizations in manufacturing, utilities, and financial services typically grow 1 to 5 percent a year, while growth-stage companies in technology and healthcare services may sustain 15 to 30 percent or more during expansion.

Department mix is where a total hides the plot. Notion's engineering headcount grew 42 percent year over year to 55 percent of the workforce while the company overall grew 28 percent, a clear signal of engineering-led investment you would never see in the top-line number. Track the function split across your two snapshots and you get strategy, not just size.

Refolk's index shows why role coverage matters when you build these splits from role searches. Profile depth varies about 30x by role even within a single country, which means an estimate assembled from role-level searches is far more complete for clinical and engineering functions than for niche go-to-market roles.

Coverage also varies sharply by country, which matters the moment you size a company with offices abroad. In Refolk's index the same title, Software Engineer, returns 348,312 profiles in the US against 21,663 in Germany, a 16.1x ratio. That gap is mostly platform adoption, not a true talent-pool difference, so a raw profile count will understate a German engineering office far more than a US one.

One interpretation guardrail keeps growth reads honest: headcount should generally trail revenue growth. A widening gap between revenue and headcount growth often signals rising efficiency, while a closing gap can be a red flag for over-hiring. Rising RPE increasingly reflects deliberate non-backfilling of attrition and automation of workflows, so a flat profile count during rising revenue may be real efficiency rather than stale data.

A total tells you how big. The department mix tells you what the company decided to become.

How this goes wrong

The estimate fails in predictable ways, and every failure has a tell and a fix. Work this list before you publish a number.

The failures split into two families: those that inflate your estimate and those that deflate it. Knowing which family you are in tells you which direction to correct.

Two of these deserve extra weight because they are the ones that pass silent.

Subsidiary rollup is the quietest inflator. Platforms typically fold all subsidiary employees into the parent company's listing, which simultaneously inflates the parent and empties the subsidiary. If you are sizing either entity without first confirming the legal structure, both numbers are wrong. Generic-name aggregation is the second silent one: a company with a common name can have profiles from entirely unrelated firms aggregated into a single listing. The only defense is to open a sample of profiles and confirm they belong to the target.

The net-versus-gross error corrupts growth math specifically. A company that hired 40 people and lost 30 shows only 10 net, which looks nearly flat while hiding significant activity in both directions. If you only have net snapshots, say so, and do not describe the company as "not hiring."

Before you publish the number

Run this checklist as the final gate. If any item fails, the estimate is not ready to leave your desk.

Keeping the estimate current

A headcount estimate decays the moment you publish it, so treat it as a dated snapshot and set a cadence to refresh the signals that move fastest. Profile counts drift as hires update and leavers age out; RPE benchmarks shift year to year; and a new WARN filing or a freshly filed Companies House account can reset your floor overnight.

Re-pull the profile count and department split quarterly for active targets, and note the delta rather than overwriting, so you always have two dated snapshots for growth math. Re-check RPE benchmarks annually, because the disclosed private-SaaS median moved from $125,000 to $129,724 in a single year and cross-industry figures shift with margins. Set an alert on primary records: a WARN notice or a new annual filing is the strongest evidence you will ever get, and it arrives on the government's schedule, not yours.

Above all, keep the confidence label honest as the sources age. An estimate that was high-confidence at capture becomes medium once its primary record is a year old and its profile snapshot is stale. The discipline that makes this method defensible is not the arithmetic. It is refusing to state a number more precisely than the evidence allows, and saying out loud where the evidence is thin.