Why Salary Sites Disagree About the Same Job

They are not measuring the same thing

Open three salary sites for the same job title and you will get three materially different numbers. The instinct is that two of them must be wrong. Usually none of them is, and the difference tells you something more useful than any single figure would.

Government wage surveys ask employers what they actually paid, drawn from payroll records across a large sample of establishments. Self-reported sites ask workers what they earn and publish the aggregate. Posting aggregators collect the ranges employers advertise in job listings. Those are three different questions with three different answers, and each has a defensible claim to being the right one for some purpose.

Once you see them as different instruments rather than competing estimates, the disagreement stops being a problem to resolve and becomes information about which one fits your question. That reframing is the whole of what this article is for. Everything below is how to tell which instrument you are holding.

Who answers a salary survey

Self-reported data has a selection problem that is worth understanding rather than dismissing. People contribute a figure when they have a reason to — usually because they are looking, benchmarking, or curious about whether they are underpaid — and that population is not a random sample of everybody doing the job. It is a sample of everybody with a reason to look.

The direction of the bias is contested and probably varies by field. People who are pleased with their pay may be more willing to state it; people who suspect they are underpaid may be more motivated to check and contribute. What is not contested is that the sample is self-selected, which means the confidence intervals these sites display describe the spread of who responded rather than the spread of who works.

There is also no verification. A figure entered on a form is a figure somebody typed, and while most people have no reason to misstate it, nothing checks. Employer surveys pull from payroll, which is a different quality of evidence entirely.

The job title problem underneath all of it

This one affects every source and is probably the largest single cause of disagreement. A Software Engineer at a fifty-person company and a Software Engineer at a global bank are doing different jobs with different scope, and both are counted under the same title in most datasets. The title itself is doing almost none of the work.

Government surveys handle this by mapping employer job titles onto standardized occupational codes, which is more consistent and coarser — a single code can cover a wide range of seniority. Self-reported sites usually take the title as given, which preserves the company’s own terminology and imports all its inconsistency along with it. Neither approach is wrong, and the two of them fail in different directions.

The practical consequence is that a wide range on any source may be telling you the title covers several jobs rather than that the pay for one job varies enormously. Before concluding that a figure is wrong, it is worth asking whether it is describing your job at all. A wide range is often two jobs rather than one.

Dates that are not the dates you think

Every figure describes a period, and the period is rarely the date on the page. Large wage surveys have a reference period well before publication, so a figure released this year commonly describes last year’s pay. Self-reported figures accumulate over time, so a page may be mixing submissions from several years without saying so.

Posting data is the freshest, which is its main advantage, and it describes what is being advertised rather than what is being paid. Those are different, particularly in a market where employers post optimistically or where the advertised range spans several levels at once. Both of those are common in current job postings.

So a comparison between three sources is frequently also a comparison across three different points in time, which accounts for more of the gap than most readers assume. Check the vintages of each figure before concluding anything at all.

What is counted as pay

The last major source of disagreement is definitional. Some figures are base salary only, while others include bonus. Some include equity at an assumed value, and some report total compensation with benefits folded in.

In fields where variable pay is a large share of earnings — sales, finance, senior technical roles — that single choice can move a number by thirty percent or more, and it is frequently not stated on the page. Two sites reporting base and total for the same role will look like they disagree wildly when they agree completely. Definition explains more disagreements here than data quality does.

The check is to look for what the figure includes before comparing it to anything. If a source does not say, that absence is itself a reason to weight it less than one that does. Silence about method is itself a finding.

How to use them together

Rather than picking a winner, use each for what it is good at. Take the government survey as your anchor for the level and the spread, because it is the largest sample and it comes from payroll rather than recollection. Use percentiles rather than the median, since the percentile range tells you the shape of the market and the median tells you one point in it.

Use posting data for freshness and for what employers are currently willing to advertise, particularly in jurisdictions where ranges are legally required — those are the most honest advertised numbers available anywhere. The legal standard behind them is good faith.

Use self-reported data for the things nothing else covers: company-specific figures, the structure of pay in a particular firm, and the qualitative detail about how compensation is composed. Treat the numbers as directional rather than precise. Used that way they are genuinely valuable.

A worked comparison

Suppose you look up a mid-level analyst role and find $72,000 on a government source, $91,000 on a self-reported site, and a posted range of $80,000 to $110,000. That looks like chaos and is entirely ordinary. Each figure is doing something different.

The government figure is a median across every employer of every size in the area, describing last year, base pay only. The self-reported figure skews toward people at larger employers who are actively benchmarking and may include bonus. The posted range covers two levels and is what employers currently advertise, which is why its floor sits above the survey median.

Read that way, the three are consistent: the market pays around $72,000 at the middle across all employers, larger employers pay meaningfully more, and current advertised roles are asking for candidates who would sit above the median. That is a considerably more useful picture than any single number would have given you. The spread between them was the information.

The questions to ask of any figure

Four questions settle almost any salary figure you meet. Who was counted — employers, workers, or postings? What period does it describe, rather than when was it published? What is included in the number — base, total cash, or total compensation? And how many observations sit behind it for your specific occupation and location?

A source answering all four is worth using whatever the number turns out to be. A source answering none is showing you a figure it has not interrogated, and the honest response is to treat it as one data point among several rather than as the answer.