Salary sites disagree because they measure different populations. An employer survey reports what businesses pay the people on their payroll. A self-reported dataset reports what workers say they earn. A posting scrape reports what employers advertise. All three can be honest and still differ by 20% or more, so the useful question is not which site is right but which population you belong to.
They are not measuring the same thing
Employer surveys ask organizations what they actually pay, across a defined sample, using a defined occupational classification. Slow to publish and hard to game.
Self-reported data asks workers what they earn. Fast, current, and shaped entirely by who chose to answer.
Job posting data reads advertised ranges. That measures what employers are willing to advertise, which is a claim about the market rather than an observation of it.
Three methods, three populations, three different questions. Disagreement is the expected result rather than evidence that somebody is wrong.
Who answers a salary survey
Self-reported datasets are shaped by response bias in both directions. People who are pleased with their pay and people actively researching a move are both over-represented, and neither group looks like the occupation as a whole.
Site design matters too. A platform aimed at software engineers in expensive metros will report higher figures for a job title than one aimed at everybody with it, and both can be internally honest.
The job title problem underneath all of it
Employers invent titles. One company’s Analyst II is another’s Senior Analyst and a third’s Associate. Survey data maps roles onto a standard occupational classification by duties; self-reported data groups by whatever people typed.
So two sources can report different numbers for the same title while both describe their own sample accurately. This is the single largest source of disagreement and it is almost never mentioned.
Dates that are not the dates you think
Survey data describes a reference period, not a publication date — a survey referenced to May 2025 and published in 2026 is 2025 data. Self-reported figures are usually a rolling window of submissions, so a “current” figure may include entries from two years ago.
Comparing a 2025 reference period against a rolling multi-year window is comparing two different times as well as two different methods.
What is counted as pay
Some sources count base only. Some include commission and tips. Some add bonus, some add equity, and very few explain which.
For a commission-heavy occupation, base-only and total-compensation figures for the same job can differ by a factor of two, and both are correct answers to different questions.
How to use them together
Start with the employer survey for the level and the shape of the distribution, because it has the best sample and the clearest definitions. Use posting data for what is being advertised right now, which the survey cannot tell you. Use self-reported data for texture — what people at specific companies report — while remembering who answered.
Where they agree, you can be confident. Where they disagree by a lot, the disagreement itself is telling you the occupation is heterogeneous, and that is worth knowing before an interview.
The questions to ask of any figure
Who was surveyed, when does the data describe, is it median or mean, and what counts as pay. Four questions, and most published salary figures cannot answer all four.
A source that answers them readily is usually the better source, regardless of whether its number is the one you were hoping for.
Common questions
Which salary site should I actually use?
For whether an offer is ordinary, the federal employer survey at metro level. For what one named company pays, self-reported data is the only source that attempts it, treated as directional. For whether demand is rising, posting data is the most current.
Why is the government figure usually the lowest?
It reports what is actually paid across everyone in the role, including long-tenured people whose salary has drifted, rather than what is advertised to attract new applicants.
Is self-reported salary data worthless?
No, but it is unverified and skews toward people who recently changed jobs and toward whichever audience the site attracts. It is most useful where a survey has poor coverage and least useful as a single figure.
How much disagreement is normal?
Twenty per cent between an employer survey and a posting scrape for the same occupation and area is unremarkable. Much wider than that usually means the three sources are describing different occupations under similar titles.
What if all the sites agree?
Check whether they are independent. A great deal of apparent consensus is one figure republished, which is why a source that names its own source is worth more than three that do not.
Why do salary sites disagree?
They measure different things — employer surveys, self-reported submissions and advertised postings sample different populations and define pay differently.
Which source is most reliable?
Employer survey data for the level and shape of the distribution. Posting data for what is advertised now, and self-reported for texture with its bias in mind.
What four questions should I ask of any figure?
Who was surveyed, what period it describes, whether it is median or mean, and what counts as pay. Most published figures cannot answer all four.