Pin down what the claim actually asserts — which occupation, which area, which period, which statistic — then follow the citation chain until it reaches an instrument rather than another article. Read the methodology rather than the summary, check the internal arithmetic, and where the chain dead-ends without naming a source, record the claim as unsettled. A source that will not name its own source is usually not withholding it for a good reason.
Four questions that settle most claims
Most salary claims fall apart or hold up on four questions, and asking them takes a couple of minutes. Most claims fail on the first or the fourth.
Which occupation is it describing, precisely and as defined by whom? “Software engineers earn X” covers an enormous range of roles and levels, and a standardized occupational code covers a different set of people again. A claim that cannot name the occupation it means is not making a checkable statement.
Which geographic area does the figure actually cover? National figures conceal metro variation of around 2.3 times in the median occupation, so a national number tells you very little about a specific place. Which period does it describe, as opposed to when it was published? And which statistic is it — a median, a mean, a percentile, an average of advertised ranges?
That last one matters more than people expect. Mean sits around six percent above median at the midpoint across large occupations, and for some it is far worse: personal financial advisors show a median of $105,070 against a mean of $156,670, a gap of nearly fifty percent. Quote the mean there and you describe almost nobody.
Follow the citation until it stops
The mechanical part is simple and it is where most claims fail. Click through the citation and see where it actually leads. If it leads to another article, click through that one. Keep going until you reach either an instrument — a survey table, a dataset, a published methodology — or a page that names no source at all.
The chain usually resolves in two or three steps, and the failure mode is recognizable. An article cites a second article, which cites a third, which cites the first, or which simply states the figure with no attribution. At that point the number has no provenance and should be treated as unsourced regardless of how many places repeat it.
When the chain does reach an instrument, you have not finished. Check that the figure quoted matches the figure in the table, because transcription errors and quiet rounding are common, and check that the population matches what the article claimed — a figure for full-time workers presented as a figure for everybody is a real change of meaning. Populations are the quiet way a figure becomes wrong.
Read the methodology, not the summary
Statistical agencies publish technical notes and almost nobody reads them, which is where the interesting caveats live. They tell you what was sampled, over what period, with what exclusions and what known limitations apply. Statistical agencies are usually candid about every one of those points.
Three things in particular are worth checking every single time. What is excluded — many wage surveys exclude the self-employed, which materially affects some occupations. How annual figures were derived, since annualizing hourly wages usually assumes a fixed number of hours and produces full-time equivalents rather than take-home for part-time-heavy occupations.
And what the reference period is, which is nearly always earlier than the publication date. A figure released this year commonly describes last year’s pay, and that gap widens every time the number is copied onto a page with a fresh timestamp. Nothing in the copying records the original period.
Sanity checks that catch most errors
A few arithmetic checks catch a surprising share of mistakes without any expertise. Simple arithmetic is much the cheapest check available to anybody.
Check that the hourly and annual figures actually agree with each other. Most annualization assumes 2,080 hours, so an hourly figure times 2,080 should land close to the annual one. A large discrepancy means either a different assumption or an error somewhere in the chain.
Check that percentiles are ordered and plausibly spaced. The tenth percentile below the median below the ninetieth, with a ratio between the tenth and ninetieth that is usually somewhere around 2.2 times across occupations. A spread of ten times in an ordinary occupation is a signal that something has gone wrong, frequently a thin sample.
And check the sample behind a specific cut. A metro-level figure for an occupation with a hundred people in that metro can be moved by a single employer, which is how absurd comparisons get published — one occupation showed nearly 26 times variation across metros until the thin cells were filtered out, after which it was 2.9. Applying that one filter changed the finding entirely and correctly.
The claims that deserve the most skepticism
Certain shapes of claim are wrong often enough to warrant checking before believing them. Pattern recognition saves a great deal of time here.
Any figure attached to something being sold — a certification, a course, a bootcamp — where the organization publishing the salary survey is also selling the training. The sample in those is usually their own customers, which is a population rather than a market. The finding may be true and it is not about you.
Any comparison that produces an implausibly large ratio, which usually means a thin sample or an apples-to-oranges comparison rather than a genuine finding. Any figure that would need a data source nobody publishes: there is no US published breakdown of wages by years of experience, none by sex or race in the main occupational survey, and no metro-level projections either. A figure requiring one of those was constructed somewhere.
And any claim quoted without its population, period or statistic. That is not necessarily wrong; it is unverifiable, which for practical purposes is the same thing. You cannot act on what you cannot trace.
What to do when the chain dead-ends
Record it as unsettled rather than as false. A claim with no traceable source is not thereby untrue, and treating it as disproved is its own error. Unverified and false are two genuinely different states of a claim.
What you can do is look for an independent instrument that speaks to the same question. If a government survey covers the occupation, that is a stronger reading than any secondary claim, and if it disagrees with the circulating figure, the disagreement is itself the useful finding. It usually points at a difference in definition rather than an error.
Where two credible sources genuinely disagree, resist the urge to average them. They are usually measuring different populations or different periods, and working out which is a faster route to understanding than splitting the difference between two things that were never the same quantity to begin with. An average of two definitions measures nothing.
Citing a figure properly
If you are going to use a number — in a negotiation, a report, or an argument — carry its provenance with it. Name the source, the occupation as the source defines it, the geography, the period it describes, and the statistic used. Five short clauses do the whole job.
That sounds pedantic and it is what makes a figure persuasive rather than merely asserted. An employer can check a properly cited number in a minute, and a number they can check is one they can act on. A number they cannot trace is one they can dismiss without engaging with it at all.
It also protects you from the most common failure in a pay conversation, which is quoting a figure confidently and being asked where it came from. Having the answer ready is worth more than having a slightly larger number.
Common questions
What four questions settle most salary claims?
Which occupation precisely, which area, which period it describes rather than when it was published, and which statistic — median, mean, percentile or advertised average.
Why does median versus mean matter?
Mean sits around six percent above median at the midpoint, and far more in skewed occupations — personal financial advisors show $105,070 median against $156,670 mean.
How do I follow a citation?
Click through until you reach an instrument — a survey table, dataset or methodology — or a page naming no source. Chains that loop between articles mean the number has no provenance.
What should I check in the methodology?
What is excluded, such as the self-employed; how annual figures were derived, since annualizing usually assumes fixed hours; and the reference period, which is nearly always earlier than publication.
What arithmetic checks help?
Hourly times 2,080 should approximate the annual figure. Percentiles should be ordered with a tenth-to-ninetieth ratio around 2.2 times. And check the sample size behind any specific cut.
Why do thin samples matter?
A metro figure for an occupation with a hundred local workers can be moved by one employer. One occupation showed nearly 26 times metro variation until thin cells were filtered, then 2.9.
Which claims deserve most skepticism?
Figures attached to something being sold, implausibly large ratios, and any claim needing data nobody publishes — there is no US wage breakdown by years of experience, for instance.
What if the chain dead-ends?
Record it as unsettled rather than false, and look for an independent instrument on the same question. Where two credible sources disagree, do not average them — work out what each measured.