The Addressable-Account Teardown: Census Cells to a Dated Range

This guide covers one job: working out how many companies in a vertical could actually buy, backing that number with a named list, and defending the gap between the two before it goes into a sizing model or a board deck. It is for strategy and research teams, talent-intelligence analysts, and operators sizing a market. I take one vertical, US dentistry (NAICS 621210, Offices of Dentists), through every decision point with the real intermediate counts, including the wrong turns that inflate a first draft. You can follow the same path on your own vertical.

What the procedure produces, step by step

The procedure produces four things that travel together: a dated low, central, and high addressable-account figure; a named list matched against it; a coverage ratio for each size band; and a filter log with a written reason for every step between the official count and the final number. A single number pulled from one database is none of these.

Most guides to bottom-up market sizing stop at "count the companies, multiply by ACV" (annual contract value, the yearly revenue per account). They treat the count as a lookup. In practice it is the weakest input in the model. It depends on which unit you count, which official cell you pull, how well the industry code matches the buyer, and how much of that cell your named list actually covers.

One ordering question needs settling first. The ranking pages build the list first ("use a B2B data platform to count companies"). I pull the official cell first. A list tells you what the list contains. Only the official count tells you what the market contains, and you need it before you can say what share the list covers.

The durations are my own estimates for one analyst. They are not sourced figures.

Which unit should the count use: firms, establishments, or practitioners?

Count the unit you bill. If you price per location, count establishments. If you price per contract with a whole business, count firms. If you price per seat, count practitioners. Never add two of these together.

The three official terms:

• Establishment: a fixed physical location where business is done. County Business Patterns (CBP) counts these.
• Firm: the enterprise that has operational control of one or more establishments. All of its establishments count as one firm. Statistics of US Businesses (SUSB) counts these.
• Nonemployer: a business with no paid employees. Nonemployer Statistics (NES) counts these separately from CBP and SUSB.

In the worked example, assume the product is practice software priced per location and sold to the practice owner or a group's operations lead. That makes establishments the billing unit and firms the number of buying decisions. Both belong in the model: the first drives revenue, the second drives sales capacity.

The firm-to-establishment spread shows how much multi-site concentration a vertical has. Dentistry's spread is narrower than the national average:

Rows 1 and 3 mix reference years. The 122,711 firm figure comes from a third-party page that does not name its series, so treat it as unverified until you replace it with the SUSB national row for your code. The Canadian row compares employers with non-employer or indeterminate establishments. It is not a firm-to-establishment ratio. I include it because it shows that in this industry the nonemployer row is not a rounding error.

What the spread proves: a ratio of 1.11 means most dental practices are single-site, so firm and establishment counts sit close together. When it lies: NAICS codes a dental group office by office, so a group with several offices gets 621210 on each one. Meanwhile, DSO revenue concentrates in roughly 7,800 practices. A DSO (dental support organisation) is a company that runs the business side of many affiliated practices. The narrow average hides a two-tier market: a long tail of single offices that each make their own decision, and a short head of groups where one decision covers dozens of offices.

How do you pull the official cell without inheriting its errors?

Pull three cells for your code: CBP establishments by employment size, SUSB firms by enterprise size, and NES nonemployers. Label each one with its reference year and release date before you use it. Each series has a known defect: lag, noise, or a unit trap.

CBP: establishments, about 18 months late

For 621210, secondary pages that repeat the newest edition report 135,665 establishments and 1,028,889 employees, with 95% of establishments under 20 staff. I have not checked those figures against census.gov, and you must before they go in a deck. At the time of writing, the newest edition covers reference year 2023 and was released June 26, 2025, about 18 months after the reference year. Census lists no planned release date for the 2024 edition while it evaluates alternatives to comply with a Commerce order prohibiting noise infusion. Check the CBP data page each time you refresh. Do not assume a schedule.

Noise infusion means Census perturbs cell values slightly to protect confidentiality. Noise-infused cells look like ordinary integers, and the API does not flag which values were perturbed. In practice, put a band around any county or state cell with 3 or fewer establishments rather than treating it as exact. Older editions used suppression instead, and under suppression more than half of county-industry cells were withheld. If you need historical sub-national cells, the imputed CBP files published by Eckert, Fort, Schott and Yang fill those gaps. I have not established how many cells for 621210 are noise-affected. If your model depends on state or county cells, count them from the raw file.

SUSB: firms, with a summing trap

Nationally, SUSB shows 6,395,635 employer firms for 2022. Within one geographic area, a multi-establishment firm's establishments in the same industry count as one firm. Add the state rows together and a multi-state group is counted once per state. Tabulations by enterprise industry exist only for reference years 2012 to 2015, so newer tables count a firm in every industry where it has establishments. Use the national row for your code, and never sum across states or industries. I did not retrieve the 621210 firm counts by size class. Pull them from the SUSB datasets page.

NES: nonemployers, a separate row

NES reports 29,811,495 nonemployer businesses nationally for 2022 (a secondary figure). I did not retrieve the 621210 count. Keep it as its own row, and decide in writing whether a practice with no paid staff can buy your product. If it can't, the row stays in the log as an exclusion with a reason. It does not quietly disappear.

Series | Unit | Code | Geography | Size class | Count | Reference year | Release date | Pulled on | Noise or suppression note
CBP | Establishments | 621210 | US | All sizes | 135,665 (verify) | 2023 | June 26, 2025 | <date> | National cell; band any sub-national cell of 3 or fewer
CBP | Establishments | 621210 | US | Under 20 staff | <pull> | 2023 | June 26, 2025 | <date> |
CBP | Establishments | 621210 | US | 20 staff and over | <pull> | 2023 | June 26, 2025 | <date> |
SUSB | Firms | 621210 | US national row only | By enterprise size | <pull> | <year> | <date> | <date> | Do not sum states
NES | Nonemployers | 621210 | US | n/a | <pull> | <year> | <date> | <date> | Separate row; never added to CBP

Does the industry code match the buyer?

Not fully, and the mismatch runs in both directions. Some real buyers sit under other codes or outside the business statistics altogether, and some records inside the code will never buy. You find out how big each leak is by testing, not by assuming.

The documented leaks around 621210:

• Dental laboratories are classified under 339116, not 621210. Include them only if your product serves labs, and as their own row.
• Hygienist or denturist establishments go to 621399. The same rule applies.
• Government and HMO dental clinics mostly fall outside SUSB, which excludes most government entities. In Refolk's index, the top employers in a sample of US dental practice managers were the US Air Force, Kaiser Permanente, and the US Army. None of these are 621210 private practices, but all of them employ people who run dental offices.
• DSO head offices. Each affiliated office carries 621210. If your buyer is the support organisation rather than the office, the head office is a separate account to find, and you should check its code rather than assume it.

What the code proves: the main activity at a location. When it lies: your buyer is defined by a workflow or a budget, and the activity code is only a proxy for that. The official count is too high where the code includes records that will never buy, and too low where buyers sit in other codes or outside SUSB.

The misclassification share for 621210 has not been published, so measure your own. Draw 100 records at random from whatever list you are using for the code. Classify each one as buyer, non-buyer, or wrong code, and record the shares. That sample is the evidence behind your central figure. I am not giving you a share I haven't measured.

Also watch for a classification change. The revised NAICS applies from reference year 2027, and OMB publishes the final changes in the Federal Register. My sources do not establish whether 621210 will be split, merged, or redefined. Read Part II of FR 2026-14086 for your own code, and note in the log whether your cell sits on the old or the new basis.

What do second counts measure, and how do they compare?

Second counts check the official cell, but most of them count people rather than businesses. Convert each one to your unit before comparing, or leave it out. Lined up against CBP establishments, the US dentistry counts look like this:

All ratios are my own arithmetic. The owner row multiplies a 2023 ownership share by a 2024 dentist count, so it mixes years. The DSO row is a secondary figure I have not verified against a primary source. One source puts the 5 largest DSOs at more than 5,600 supported practices.

Reading the ladder:

• Active dentists, 202,485. This proves the size of the workforce. Read as a market count, it overstates offices by about half, because several dentists can share one office.
• Owner dentists, about 147,800. This comes from 73% of dentists being practice owners. It lands close to the establishment count, which is tempting. But a practice can have more than one owner, so it does not prove a practice count.
• DSO practice totals, about 7,800 for the top ten. This is the right comparator for the multi-location tier. It is not a share to apply to the whole cell.
• Licensing rolls and association totals. I did not retrieve these. If you use them, record whether they count licensees, active practitioners, or offices.

The commercial list deserves its own check.

My hypothesis, which I have not tested, is that lists keyed on licensed individuals count dentists rather than offices, and add dormant records on top. To test it, take a sample, deduplicate by street address and NPI (the US provider identifier), and compare the survivors with the CBP cell. Until you have done that, the list is a source of names. It is not a ceiling.

How do you build the named list and measure its coverage by band?

Build the named list one size band at a time, then divide matched accounts by the official count in each band. An aggregate ratio hides what matters. Profile coverage is usable for multi-location groups and very thin for single-site practices, and the board needs to see that split rather than one blended percentage.

Two wrong turns first

Wrong turn one: the keyword "dental". Searching Refolk's index for US profiles with Owner/Partner seniority and a "dental" keyword returns 8,980 people. In a sample of 25, the leading employers included a dental software company, an M&A advisory firm, and marketing agencies, alongside real practices. A dental-software CEO counted as a practice owner is a false positive. So 8,980 against about 147,800 owner dentists, or 6.1%, is an upper bound on coverage, not a measurement.

Wrong turn two: practice managers taken at face value. Searching for US dental practice and office managers returns 3,184 profiles, about 2.3% of CBP establishments. The leading employers in the sample were the US Air Force, Kaiser Permanente, and the US Army. These may well be buyers, but they belong in the government and HMO row, outside the 621210 cell. Folding them into the practice band inflates its coverage.

Coverage is this low because 95% of establishments have under 20 staff, and owner-dentists rarely keep professional profiles. A named list built only from profiles will look like an 80 to 90% shortfall in the small bands.

A named list that covers 2% of small practices is a coverage measurement, not a market finding.

The bands to build

Match the bands to the size classes in your official pull, so every numerator has a denominator. At minimum, split under 20 staff from 20 and over, since 95% of establishments sit below 20. Then add the rows that do not come from the cell:

1. Multi-location groups and DSOs. Compare against DSO practice totals and against the establishments you can match office by office. Profile sources work well here.
2. Employer establishments, 20 staff and over. Compare against the CBP size class.
3. Employer establishments, under 20 staff. Compare against the CBP size class. Build this band mainly from address-based sources, because profile coverage is thin.
4. Nonemployers. Compare against NES, only if you decided they buy.
5. Government and HMO clinics. Keep this as a separate row with its own named list. It has no official comparator in SUSB.

The multi-location band is where Refolk removes the most work. You ask in plain English for groups by location count with a named operations lead, and get back accounts with the decision-makers attached, ready to match office by office against the establishment cell.

To match, compare each named account's offices with the official cell by address, roll offices up to firms, and assign each firm to the band its staff count puts it in. Coverage for a band is matched accounts divided by the official count for that band. Report it to one decimal place, and record the matching rule in the log.

How do you reconcile the two counts into a low, central, and high range?

The range comes from the official cell adjusted by measured filters. The named list does not set it. The list proves that accounts exist, and the coverage ratio by band explains why the list is smaller than the range.

Define the three figures in writing before you calculate them:

• High: the official cell for your billing unit, plus any separate rows you decided to include (out-of-code buyers, nonemployers if they buy), before any deductions. In the worked example, it starts from the 135,665 establishments once you have verified them. It never starts from the commercial list's 226,938.
• Central: the high figure minus the non-buyer and wrong-code shares from your 100-record sample.
• Low: the central figure, using the pessimistic end of the sample's non-buyer share, minus the noise band on small sub-national cells, and leaving out any row you included on judgement rather than evidence.

On tolerance: one ranking glossary says to reconcile any gap beyond 2 to 3x between top-down and bottom-up totals. A 20% tolerance between two unit counts also circulates, but I found no published source for it. Pick your own tolerance before you match, write it in the log, and don't move it afterwards.

In dentistry, the small employer bands usually land top-left: the code fits, the list is thin. The DSO tier can reach top-right. The government and HMO row sits bottom-right, because the buyers exist and can be named but no official business count covers them.

ADDRESSABLE-ACCOUNT LEDGER
Vertical and code: <e.g. US dentistry, NAICS 621210, current NAICS basis>
Billing unit sentence: <e.g. Priced per office location, so the billing unit is the establishment; firms give the number of buying decisions>
Date stamp: <date compiled> | Cells: <series, reference year, release date for each>
Gap tolerance (set before matching): <your figure> | Source for tolerance: <your own, stated>

FILTER LOG
1. Start | CBP establishments, national | <count> | Reason: matches billing unit | Evidence: cell register row 1
2. Add or exclude code | <code> | <+/- count> | Reason: <buyer or not> | Evidence: <code definition>
3. Misclassification | <share from 100-record sample> | <- count> | Reason: non-buyer or wrong code | Evidence: sample file
4. Nonemployers | <include or exclude> | <count> | Reason: <can a no-staff business buy?> | Evidence: NES row
5. Government and HMO row | <count of named accounts> | separate row | Reason: outside SUSB | Evidence: named list
6. Noise band | <cells of 3 or fewer> | <+/- count> | Reason: perturbed small cells | Evidence: raw file

RANGE: Low <n> | Central <n> | High <n>

COVERAGE BY BAND
Band | Official count | Named and matched | Coverage
Multi-location groups and DSOs | <n> | <n> | <%>
20 staff and over | <n> | <n> | <%>
Under 20 staff | <n> | <n> | <%>
Nonemployers | <n> | <n> | <%>
Government and HMO (no official comparator) | n/a | <n> | n/a

How does the count go wrong?

Nearly every failure falls into one of two kinds: the unit changes partway through the calculation, or a false positive gets counted as a buyer. These are the eight that show up in a vertical like dentistry, with what each one looks like and the check that catches it.

1. Adding CBP establishments to NES nonemployers and calling the total "companies". CBP counts physical locations, not distinct firms, and NES counts a different population again. The result is a large, impressive number with no unit. Check: every column you sum uses the same unit.
2. Summing SUSB firms across states. A multi-state DSO is counted once in each state where it has offices, so the total overstates unique firms. Check: use the national row only.
3. Turning the dentist-level DSO share into a practice count. 16.1% x 135,665 produces about 21,842 "DSO practices", a figure HPI's own guidance rules out. Check: read the unit stated in the source before multiplying anything by it.
4. The keyword "dental" in profile searches. It returns software vendors, marketers, and M&A advisers along with practice owners. The false positive is a dental-software CEO counted as a practice owner. Check: require the employer to be a practice, not the profile to mention dentistry. In Refolk, put the employer type in the request itself ("practice owners of independent single-location dental offices") so that vendors never enter the list.
5. Folding government and HMO clinics into the practice count. They sit outside SUSB, so adding their named accounts to a 621210 band inflates its coverage against a denominator that never included them. Check: keep them as their own row.
6. Trusting an inflated commercial list. The 226,938 figure probably mixes dentists, locations, and dormant records. Check: deduplicate a sample by address and NPI, then compare with CBP.
7. Treating noise-infused small cells as exact. A county cell of 2 establishments may not be 2, and nothing in the API tells you so. Check: band every county or state cell with 3 or fewer establishments.
8. Citing a stale reference year as current. A deck that calls a reference-year-2023 cell "current" two or three years later misses whatever DSO consolidation happened in between, and affiliation more than doubled between 2015 and 2024. Check: date-stamp every cell with its reference year and release date, on the slide as well as in the model.

The pattern behind most of these: the bigger the number, the more likely it is that a unit changed somewhere upstream. When a count comes in well above the official cell, find the unit change before you look for missing businesses.

How do you keep the range current?

Refresh the range when an input changes, not on a calendar alone, and always re-date it. Four inputs move: the official edition, the classification basis, the market's own consolidation, and your named list.

• A new CBP edition. Check the CBP data page at each refresh. At the time of writing, no date has been set for the edition after reference year 2023, so the cell you use now may be the last one on the current basis for a while.
• The NAICS revision. The revised classification applies from reference year 2027. When final changes for your code appear in the Federal Register, re-run the code test and note the basis in the ledger.
• Consolidation. DSO affiliation more than doubled over the decade measured by HPI. Recheck DSO practice totals at every refresh, because that tier moves faster than the official cell.
• The named list. Re-run the band queries, re-match, and recompute coverage. Coverage that moves without the official count moving means your list changed, not the market.

The date stamp is part of what makes the number defensible. A range with its sources and reference years printed beside it survives a skeptical CFO in a way an undated single figure does not.