How many job postings a pay benchmark needs to be meaningful

How many job postings a pay benchmark needs to be meaningful depends on the figure you read and how sure you must be. This post gives the 95% interval around a median and a lower quartile at each count of postings, from methods the US National Institute of Standards and Technology publishes, explains why statistics offices also withhold small groups, and works a rare sales role through by hand.

How many job postings does a pay benchmark need to be meaningful?

For a 95% interval around the median, a pay benchmark needs at least six job postings: with five, even the lowest and highest offers bracket the market median only 93.75% of the time. An equal-tailed interval for a quartile needs 13. At 15 postings the median's interval spans the sample's middle half; at 100, its 40th to 60th percentiles.

The reason is a count, not a property of pay. If the postings are independent draws from the market of advertised offers, each lands below or above the market median like a fair coin. With five, all land on the same side one time in 16, and then no pair of offers brackets the median. Pay level, currency and the shape of the distribution do not enter. Read the postings count first beside any median or quartile, and the companies second: postings from one employer share its pay band.

The interval for a median or a quartile

The NIST/SEMATECH e-Handbook of Statistical Methods estimates a percentile from sorted values. For the pth percentile of N values, it computes p(N+1), splits it into a whole part k and a fraction d, and reads the kth value plus d times the gap to the next one (section 7.2.6.2). A 95% interval, in its terms, brackets the true value in about 95% of repeated samples (section 7.1.4).

For the median, the handbook notes that such intervals are hard to derive in closed form and refers to the bootstrap (section 1.3.5.2). Its bootstrap plot (section 1.3.3.4) resamples with replacement, typically 500 to 1,000 times, and for a 90% interval sorts 500 medians and takes the 25th and 475th as limits.

NIST's Dataplot reference manual gives a route from the sorted values, with no simulation (median confidence limits). The number of observations below the median, W, is binomial with n trials and probability 0.5. Take the largest k for which the chance that W lies between k and n minus k still reaches the confidence level; the interval runs from the kth to the (n minus k plus 1)th sorted value. Dataplot then interpolates between neighbouring values for an approximate interval at the level itself; this post keeps the sorted values.

From here the algebra is this post's own. A quartile works the same way with 0.25 in place of 0.5: the rth and sth sorted values bracket the lower quartile with probability P(r <= W <= s - 1). The table leaves at most 2.5% on each side; where no such ranks exist, it shows the full range with its true coverage. The percentile column inverts the handbook's p(N+1) rule: the kth of N values is the 100k/(N+1)th percentile. The upper quartile mirrors the lower one.

The interval narrows slowly: from 30 postings to 100, the median's span falls from 36 percentile points to 20, roughly with the square root of the count. A quartile is harder: below 13 postings only the full range bounds it. Every coverage figure assumes independent postings.

What sample size does a pay benchmark need, from a salary survey or from job postings?

The arithmetic of a median is the same for any source: the number of independent observations sets the width of its interval. An official salary survey draws employee jobs or employers by design, while job postings record offers, several of which may come from one employer's pay band. Size a posting benchmark by distinct postings and by companies.

Two features of postings shrink the effective count.

• Employers. Ten postings from two companies say less about the market than ten from ten. Report the companies and the largest one's share.
• Weeks. A weekly group counts the postings open that week, so a posting open for a month sits in four groups. Pool weeks by distinct postings.

Official surveys publish their precision. The UK Office for National Statistics samples 1% of employee jobs from HMRC's Pay As You Earn records for its Annual Survey of Hours and Earnings, and marks each estimate's coefficient of variation by background colour, in bands up to 5%, 10% and 20% (quality and methodology information). An ONS table of 8 January 2025 notes that figures with a coefficient above 20% would usually be suppressed on quality grounds as unreliable (ASHE ad hoc release). For postings, the interval plays that part.

Why statistics offices set a minimum group size

A minimum group size protects the people and employers behind a figure as well as its precision. Three statistics offices describe the rules.

• The UK Office for National Statistics, for the Annual Survey of Hours and Earnings, suppresses every cell based on fewer than three individuals, applies a dominance rule that uses the largest employer's contribution and the standard error to judge whether that employer's pay could be deduced, and releases only weighted counts rounded to the nearest 1,000, never sample counts (quality and methodology information).
• The US Bureau of Labor Statistics may withhold an occupational estimate for reasons that include failing its quality standards or protecting respondents' confidentiality, and to protect it further gives no specific reason for any withheld estimate (OEWS questions and answers).
• The Australian Bureau of Statistics describes a frequency, or threshold, rule setting a minimum number of unweighted contributors to a cell, and a dominance rule written (n, k): the n largest units may not contribute more than k% of a cell's total, as in its (2, 75) example. It presents these as common rules: dominance applies to magnitude data such as income, a cell above the threshold can still be disclosive, and secondary suppression is sometimes also required, so that a suppressed value cannot be worked out from the rest of a table (treating aggregate data).

Advertised pay is public, so the concern in a posting benchmark is narrower: a group dominated by one company reports that company's pay band under the name of a market. This post's own adaptation of the ABS example: the two companies with the most postings in a group hold no more than 75% of them.

How to build a sales compensation benchmark for a rare role

To build a sales compensation benchmark for a rare role, read the postings count, the companies and the interval for its role and country. If the interval is too wide, widen the place, the period or the role, one at a time and by the smallest step that suffices, and label the figure with the step taken.

Widen by period first, then place, then role: each step changes more of what is measured. Official statistics widen by period and by role. The Bureau of Labor Statistics bases its May 2025 occupational estimates on six semiannual panels collected over three years, adjusting wages from earlier panels to the reference date (OEWS technical notes), and aggregates some occupations to the broad level of the Standard Occupational Classification to improve data quality (OEWS questions and answers). That classification combines 867 detailed occupations into 459 broad occupations, 98 minor groups and 23 major groups (Standard Occupational Classification), so each step up is a defined widening.

Salary benchmarking methodology for niche roles with low data availability

A salary benchmarking method for niche roles with little data has five steps: count distinct postings and companies, test whether one employer dominates, compute the interval around the median, widen one dimension if either test fails, and publish the figure with its count, its interval and the widening used.

1. Count the distinct postings that state the figure you measure, base pay or on-target earnings, and their companies.
2. Test dominance: the two largest companies hold no more than 75% of the postings.
3. Read the interval ranks from the table, against a maximum width set in advance as a share of the median.
4. If a test fails, widen by period, then place, then role, one step at a time.
5. Label the result with role, place, period, postings, companies, interval and widening.

How does compensation survey methodology handle market pricing for niche sales roles that are hard to benchmark?

Official pay surveys classify jobs into standard occupations, withhold estimates that fail quality or confidentiality tests, and merge some occupations, so a role that is hard to benchmark is priced within a broader group. Market pricing a niche sales role from postings follows suit: price the nearest role, test the interval, and widen only when it fails.

The Bureau of Labor Statistics notes that its major group totals may include detailed occupations whose separate employment estimates could not be published (OEWS questions and answers). For postings, match a niche title, such as a solutions consultant for one product line, to the nearest sales role, here sales engineer, and narrow by segment or seniority only where the count allows.

Worked example: pricing a rare sales role at Acme Robotics

This example is illustrative. Acme Robotics is an invented company, and every figure is invented to show the arithmetic; none is a Fokals figure.

Assumptions:

• Acme Robotics is hiring a sales engineer in one country and wants a reference for on-target earnings, in yearly US dollars.
• In the latest closed week, 9 open postings for that role and country state on-target earnings, from 4 companies holding 4, 3, 1 and 1.
• Over the latest four closed weeks, 20 distinct postings state on-target earnings, from 8 companies, the two largest holding 5 and 4.
• A split of the four weeks into halves showed no shift in pay.
• Acme's policy: use a median only when its 95% interval is no wider than 15% of the median and the two largest companies hold at most 75% of the postings.

Step 1: one week. The 9 values, sorted: 118,000; 126,000; 131,000; 135,000; 140,000; 146,000; 152,000; 165,000; 190,000.

1. Median: p(N+1) = 0.5 × 10 = 5, so the 5th value, 140,000.
2. Interval at 9 postings: 2nd to 8th value, 126,000 to 165,000, with 96.09% coverage. The width is 39,000, or 27.9% of the median.
3. Sample quartiles: 0.25 × 10 = 2.5 gives 126,000 + 0.5 × 5,000 = 128,500; 0.75 × 10 = 7.5 gives 152,000 + 0.5 × 13,000 = 158,500. The median's interval is wider than the middle half of the offers, 30,000.
4. Dominance: the two largest companies hold 7 of 9 postings, 77.8%.

Both tests fail, so Acme widens by period.

Step 2: four weeks, each posting once. The 20 values, sorted, in thousands: 112, 118, 122, 126, 128, 131, 133, 135, 136, 138, 140, 142, 144, 146, 148, 150, 152, 158, 165, 190.

1. Median: 0.5 × 21 = 10.5, so the 10th value plus half the gap to the 11th: 138,000 + 0.5 × 2,000 = 139,000.
2. Interval at 20 postings: 6th to 15th value, 131,000 to 148,000, with 95.86% coverage. The width is 17,000, or 12.2% of the median.
3. Lower quartile: 0.25 × 21 = 5.25 gives 128,000 + 0.25 × 3,000 = 128,750; its interval runs from the 2nd to the 10th value, 118,000 to 138,000.
4. Dominance: the two largest companies hold 9 of 20 postings, 45%.

Both tests pass. The median moved from 140,000 to 139,000, while its interval shrank from 39,000 wide to 17,000. Acme's draft offer of 135,000 sits 4,000 below the median but inside the interval, so these postings are consistent with a market median of 135,000. Acme's label: sales engineer, one country, on-target earnings, four closed weeks, 20 distinct postings from 8 companies, median 139,000, interval 131,000 to 148,000, widened by period.

Reading group sizes in Sales Pay Benchmarks

Sales Pay Benchmarks, part of our hiring dataset, give the lower quartile, median and upper quartile of advertised base pay and on-target earnings for sales roles across companies, per closed week, role and country, with all countries together as a group of its own. Roles run from SDR or BDR and account executive to sales engineer, sales leadership and partnerships or channel. Each group carries the number of open sales postings behind it that state a base or on-target figure, and a group is published once it holds at least five postings. Each week is written once after it closes and never revised.

Read the interval table against that count, as the upper limit for either pay figure, and apply your own width and dominance tests. The all-countries group of a role is a widening by place, ready to use. To widen by period or test dominance, use Job Postings, which carries each posting with its stable identifier, its company, its first-seen and close dates, and its stated pay as written and on one yearly US-dollar scale. The query pools one role and country over a window you set, each posting once, and lists stated on-target earnings in order with each company, ready to read at the table's ranks. It assumes JSON cells are loaded as jsonb.

with latest as (
  select distinct on (posting_id) *
  from job_postings
  order by posting_id, last_seen_at desc
),
pooled as (
  select posting_id,
         company_id,
         (sales_facts ->> 'ote_usd')::numeric as ote_usd
  from latest
  where sales_labels ->> 'role' = :role
    and country = :country
    and sales_facts ->> 'ote_usd' is not null
    and first_seen_at < :window_end
    and (closed_at is null or closed_at >= :window_start)
)
select
  row_number() over (order by ote_usd)    as rank,
  ote_usd,
  company_id,
  count(*) over (partition by company_id) as company_postings
from pooled
order by rank;

The data dictionary defines each field, each sales label and each pay figure. To place a plan against the quartiles once a group passes your tests, see benchmarking sales compensation from advertised pay.

Frequently asked questions

The queries and code on this page are examples to adapt. Test them in your own environment before you rely on them.

What this page says about the products it names was checked against their public documentation on 7 October 2026. Product and company names are trademarks of their owners. Fokals is not affiliated with them or endorsed by them.