The Personal Pay Baseline, From Public Data to a Defensible Target Range

Setting a pay target for yourself is a different job from building a company pay structure, and almost every guide online does the second one. This is the candidate-side procedure: how to produce a floor, midpoint, and stretch number for your exact role, matched on job content, aged to a common date, and blended across at least three independent public sources. It is for anyone who is about to apply, about to name a number, or about to be asked "what are you looking for?" and wants an answer they can defend when a recruiter pushes back.

The hard part is not finding salary data. There is more of it than ever. The hard part is that the sources disagree, they are different ages, they measure different things, and most of them are matched to the wrong role. This guide carries the arithmetic that compensation teams use - content-matching, aging, and weighted blending - end to end into a single number you own.

Why a few averages is not a target

A target range is three numbers you can defend, not one average you found on a free site. A good target has a floor near the 25th percentile of blended data, a midpoint at the median, and a stretch near the 75th, all matched to your real duties and aged to one date.

The reason "check a few averages" fails is that the four kinds of public data each lie in a different direction. Government statistics are free and defensible but use broad occupation codes with no scope detail. Employer-reported surveys verify the job match but cost money and lag because collection is annual. Crowdsourced sites are timely and granular but self-reported and self-selected. Job postings are current but post ranges that are wide by design. Average them raw and you blend four different errors into one confident-looking wrong number.

The whole method exists to turn that pile of conflicting data into one band you can read aloud. Everything below is in service of that: fix what you are measuring, keep only the data that matches, make every figure the same age, then blend.

The four source categories and what each hides

Each source type trades freshness against reliability, and you want at least one from each side of that trade so no single weakness dominates your result. Pull one government source, one or two crowdsourced aggregators, and posting ranges.

Government data means the BLS Occupational Employment and Wage Statistics program (OEWS), which reports five percentiles across more than 800 occupations and draws on roughly 1.1 million establishment responses per cycle. Its weakness is lag: OEWS pools three years of panels, and the May 2025 estimates drew on panels running from November 2022 to May 2025. It publishes about six months after its reference period, so the effective vintage is older still. That slowness is also its strength. OEWS is defensible precisely because it pools, which is why it belongs as the stable anchor in your blend rather than the current number.

Crowdsourced sites such as Levels.fyi, Glassdoor, and Payscale are the opposite: near real-time and granular, but self-reported and self-selected, so the people who enter data are not a random sample. Posting ranges reflect genuine hiring intent and are the most current input you have, but employers write them wide on purpose, so treat a posted band as a boundary, not a point estimate.

Fix these attributes before you pull a single number

Decide what you are measuring before you look at any data, because the same job title spans wildly different pay depending on scope, and mixing pay elements is the fastest way to a fake gap. The market range comes from comparable role, level, function, industry, geography, company size, and remote status.

The one attribute people forget is the pay element. Decide up front whether you are measuring base salary only, base plus target bonus, or total direct compensation including equity, and then compare like with like. Mixing base-only survey data with a total-cash internal number is the fastest way to conclude you are underpaid by 15 percent when you are not. Label every figure you record with its element.

Geography and remote status are separate cuts, not one. A role priced to a specific metro pays differently from a remote role priced nationally, and conflating them produces a number that matches neither. Write your decision as a single line you will not change mid-analysis. If you revise the spec halfway through, every figure you already pulled is now matched to a different role and the blend is corrupt.

Why market thickness changes how much to trust a percentile

A percentile is only as trustworthy as the pool behind it, and pool depth varies enormously by market and by how you cut it. The thinner the pool, the more sources you need to compensate.

Look at the same title across two countries in Refolk's index.

In Refolk's index of professional profiles, the US has about 8 times more people matching "Software Engineer" than the UK, 353,689 against 44,106. A UK percentile therefore rests on a far thinner pool. If you are pricing a UK role, widen your source count and lean harder on live postings to offset the thinness.

The cut you choose matters as much as the market. Compare a title cut against a skill cut in the same country.

A skill tag returns about 8 times more people than a specific title. That is the trap: an unscoped "data" search pulls a population roughly eight times broader than your actual role, so its median is meaningless for you. This is why fixing attributes comes first. A thick, well-scoped pool gives you a percentile you can trust; a thin or over-broad one gives you a confident number built on the wrong people.

A percentile built on the wrong population is not conservative, it is just precise about the wrong thing.

Run the procedure end to end

This is the full sequence from a blank sheet to three defensible numbers. Budget roughly four to six hours the first time; it drops to under two once you have a template. The order matters, especially the aging step, which sources disagree on - age before you blend to stay on the safe side.

The 80 percent content test

The 80 percent rule is a discard rule, not a judgment of degree. For a match to be defensible, at least 80 percent of the duties and responsibilities must align with the source description; below that threshold, drop the data point or adjust it with an explicit premium or discount. The reason this is binary is that below 80 percent you are not slightly off, you are priced against a different population entirely, and a weak match cannot be rescued by aging it to the right date.

Before you trust any single cut's percentiles, check the pool is thick enough. A common validity threshold is 5 to 7 reporting organizations and 10 or more incumbents per cut. A confident P50 built on three people is noise wearing a number.

Refolk can do the scope-matching work for you: it reads your own history, so when you tell it the role you are targeting it already knows your real duties and level rather than just your title, which is exactly the input the 80 percent test needs.

Aging every source to one date

Aging means pro-rating a figure forward from its effective date to today using the annual market movement of salaries. The simple documented formula is:

Simple aging factor = (months to age / 12) x annual market movement rate
Compound aging factor = (1 + annual rate) ^ (months / 12)

Aged figure = original figure x aging factor

Example, 18-month-old survey at 3.1% annual rate:
  Simple:   (18 / 12) x 0.031 = 0.0465, so x 1.0465
  Compound: (1 + 0.031) ^ (18/12) = 1.0468, so x 1.0468
  A $120,000 figure ages to about $125,600

The standard reference rate for a general population is the BLS Employment Cost Index. For private-industry workers, wages and salaries rose 3.1 percent for the 12 months ending June 2026, with that release dated July 31, 2026. The total-compensation figure held flat at 3.4 percent for three straight quarters through that period. Before May 2021, OEWS itself used the ECI to update older wages, so this is the method the government uses on its own data.

Weighting and blending

Blending means combining your aged, matched sources into one composite percentile. Assign each source a weight reflecting its sample size, recency, and relevance to your market, multiply each source's value by its weight, and sum the results. The output is your composite value for that percentile. Do this separately for P25, P50, and P75.

Guidance converges on three or four sources, with two or three the floor and most compensation teams using three to five per job family. More than that adds noise; fewer leaves you exposed to one source's sampling quirk. Weight OEWS for its sample size, weight a well-scoped crowdsourced source for recency, and weight postings low because their ranges are wide.

Turn composites into floor, midpoint, and stretch

Map your composite percentiles to a band, then shift the band up for negotiation. Floor sits near composite P25, midpoint at P50, and stretch near P75; for a candidate target specifically, shift the whole band toward P60 to P80.

This is the one place candidate advice diverges sharply from employer practice. Employers anchor structural midpoint at P50, set the range minimum near P25 for developing performers, and the maximum near P75 for fully experienced high performers. A typical professional range spreads 30 to 50 percent from minimum to maximum. But once you have your P50 figure, published candidate advice says to add 10 to 15 percent to account for negotiation compression, putting your low end at P60 and your high end at P75 to P80. The same percentile data, read from opposite sides of the table.

Chanelle Howell, an ex-Goldman recruiter, frames this as the difference between a $100k candidate and a $150k candidate who are otherwise identical: the one who has done this work names a number with structure behind it. That structure is your defense file.

How this goes wrong

Most bad pay targets fail in one of seven predictable ways, and each has a specific check. These are the failure modes to audit before you call a number defensible.

The two that cost the most money are skipping aging and title-matching, and they are related. Title-matching puts you against the wrong population, and skipping aging leaves you against last year's version of it. A number can look current and specific while being wrong on both counts. The 80 percent test and the common-date check together catch most of the damage, which is why they are non-negotiable steps rather than optional polish.

Thin-cut errors are the quietest. A percentile looks authoritative whether it was built on 500 incumbents or 5, so the burden is on you to confirm depth. When a market is thin, as the UK software pool is at roughly one-eighth the US depth, you compensate by widening your source count rather than trusting a single thin cut.

Before you name a number

Run this checklist before you quote a range to anyone. If every item is true, you have a target you can defend line by line.

Keeping the number current

A pay target has a shelf life. The inputs move, and so should your range. The two things that drift are the aging rate and the underlying data, so re-check both on a schedule rather than when a recruiter surprises you.

The ECI news release updates quarterly. Note the release date, the reference period, and the private-industry wages figure each quarter, and re-age your composite from the last common date to the new one. OEWS publishes annually in the spring with a reference period from the prior year, so when a new vintage lands, re-pull your government anchor and re-blend. Watch for disruptions to the data pipeline too: a federal shutdown delayed one recent OEWS panel, though response rates stayed in normal range, so a late release is not automatically a corrupt one.

When you are ready to validate your level and scope against real people rather than a title, search for the exact transitions you are targeting.

Keep the defense file as a living sheet. When you re-age or re-pull, update it in place and date the change. The first time a recruiter pushes back and you walk them through your sources, matches, aging assumptions, and weights without hesitating, the hours you spent building it pay for themselves.