For close to a decade I negotiated salaries for a living. I never set the price. My job was to build the case for going above the band when I wanted someone badly enough, and to talk candidates down when they asked for more than the band would reach. There was a complaint I heard constantly in that work, and repeated myself often enough.

The complaint was that levels.fyi was inaccurate. Payscale was inaccurate. Glassdoor was worse. Candidates would arrive at the offer conversation holding a screenshot, and the screenshot would be wrong, and part of the job was explaining why.

It took me years to notice something about that complaint. It only ever ran in one direction. I never once sat in a room where a recruiting team was troubled that an aggregator had underquoted a role. Nobody flagged a data quality problem when a candidate came in low. The inaccuracy was only ever an inaccuracy when the number was too high.

The complaint was about a countervailing price signal, wearing the vocabulary of accuracy because that was the vocabulary available.

The number was justified internally like this, and I suspect it will be familiar to anyone who has worked in or near compensation.

The band was set at some percentage above average, benchmarked against market data. The market data came from a survey I did not have access to. The percentage was a policy decision made somewhere above me. When a candidate pushed, the answer was that we pay above market, and that answer was delivered with real confidence, because as far as I knew it was true.

So the person delivering the offer was reciting a figure derived from a source they had never seen, while explaining to the candidate that the candidate's source was unreliable. Authority borrowed all the way down the chain, and nobody at any link holding something they could actually verify.

This is the point where someone objects that I was never senior enough to set compensation, so what would I know. Correct, and it is the reason I noticed. Set the band and you have seen the survey, and having seen it you tend to believe in it. Negotiate against the band all day, arguing for exceptions to a number you cannot audit and against a number the candidate cannot audit either, and you eventually start wondering what either of them is.

None of this is fraud, and that is what makes it worth looking at.

The expensive benchmark data is real. It comes from firms like Radford, Mercer, and Willis Towers Watson. The mechanism is a mutual survey: participating companies submit their own compensation data, and in return they receive anonymized aggregates for their peer set.

Consider what that means. To submit, a company has to map its own internal roles into the survey provider's predefined catalogue. Radford's technology survey runs to roughly nine hundred distinct job titles. Every participating firm hand-matches its own idiosyncratic titles and levels into that schema, and there is no practical way for the provider to validate those judgments. The results are then published on an annual cycle, which means the authoritative number describes a market that existed some months ago.

So the premium source is self-reported by employers, filtered through a taxonomy each employer interprets on its own, and stale by construction.

The free source is self-reported by workers, with obvious self-selection, since people volunteer their compensation when they feel cheated or when they want to boast, and stay quiet in the middle.

Both are crowdsourced estimates. One costs six figures and requires membership.

The asymmetry between the recruiter and the candidate was never an asymmetry of accuracy. It was an asymmetry of access, and what sat behind the gate was a number of roughly comparable epistemic quality, dressed in a methodology.

There is a mechanism inside the premium number worth naming plainly. A firm selects its peer set, then targets a percentile within it. Choose a peer group of well-funded competitors and target the seventy-fifth percentile and you can say, truthfully, that you pay above market. Choose differently and the same salary is above a different market. The lever is the peer set, and the firm holds it. "Above market" is not a fact about the economy. It is a claim about a market the claimant defined.

There is a conversational version of the same move, and I heard it constantly. It's a down market. That's why the number is lower than it would have been two years ago. The statement is not false, and that is what makes it effective. Markets do move, and that one did.

But look at the shape of it. The figure is never produced. The comparison set is never named. And it only ever gets said in one direction. In 2021 nobody sat a candidate down to explain that market conditions obliged the company to exceed its own band. Upward pressure was handled quietly, as a competitive problem to be solved. Downward pressure was announced, as weather. The market appears in these conversations as a narrator rather than as a number, and it narrates on behalf of whoever brought it up.

I went looking, recently, for what a technical recruiter is worth in 2026. Not as a candidate. As someone who did the job and wanted to see what the public record says about it.

ZipRecruiter reports an average of about $69,600. Payscale puts early-career technical recruiters near $69,000. Salary.com gives a state average of roughly $67,500, with a typical band running from about $57,900 to $69,800. Glassdoor, drawing on nearly twelve thousand contributions, offers an entry-level range of $34,700 to $98,800. A 2026 industry salary guide puts the median base near $90,000, with most in-house offers between $70,000 and $130,000 and big tech recruiters clearing $200,000 in total compensation.

Every link in that paragraph goes to a live page. Follow them and watch them disagree.

Every one of those is presented as the answer to the question. The top of one source's entry-level range sits above the ninetieth percentile of another. One entry-level band alone spans a factor of nearly three.

No procedure available to an individual resolves this. You choose which incompatible answer to believe, and then discover that the person across the table chose a different one and has an institution behind them.

I have written before that job titles are not descriptions of work. They are a lever firms control, pulled for pricing, for status, for retention, and for legibility, often several at once and sometimes against each other.

That argument comes due here.

Every benchmarking system in existence, premium or free, requires a shared taxonomy of jobs. But the taxonomy is populated by the party on the other side of the transaction. Firms name the roles, firms assign the levels, and firms map both into the survey schema by hand.

Look at my own function for the clearest version. An agency recruiter, a corporate talent acquisition partner, and a technical recruiter at a large platform company all print roughly the same two words on a business card, and they are paid through three systems that barely resemble one another: commission-driven, salary-shaped, and equity-heavy. When a benchmarking tool reports a median for "recruiter," it is reporting the center of three unrelated distributions stacked on top of each other. The spread is the honest output of asking a question that has no single referent.

You cannot fix that with more data, because more data means more submissions into the same taxonomy, and the taxonomy is the distortion.

While this was going on, prices moved a great deal.

US technology job listings sit roughly thirty-six percent below their February 2020 baseline, according to Indeed Hiring Lab, with general software engineering postings down about half, even as machine learning roles have grown sharply. Stanford's 2026 report found entry-level software developer employment down roughly twenty percent from its 2024 peak. On pay, engineers who cleared $220,000 in base at the peak now report offers in the $160,000 to $180,000 range, with smaller equity grants and longer cliffs, leaving base compensation something like fifteen to twenty-five percent below 2022 marks. Average technology salaries grew under one percent year over year, which is a decline in real terms.

I lived through my own version of that curve, and I will not pretend I saw it coming. I had done this work for years, inside companies with the best compensation data money can buy, and I could not have told you in advance what my own price was about to do, or defended a figure for it afterward with any source I trusted.

That is the situation I actually want to describe. Not that workers are underpaid, though many are. That the mechanism a market is supposed to use to tell people what things are worth does not function here, and the people whose job it is to operate that mechanism cannot operate it on themselves.

The advice I have heard most often, and the advice I was tempted to give, is that you should work out the actual dollar value your work contributes to your employer.

It is pointing at something true, and as stated it fails. The reason it fails is worth working through.

Wages are not set by contribution. They are set by replacement cost. What you produce and what you are paid are outputs of two different mechanisms, and the gap between them is not a market failure. It is the reason employment exists as an arrangement. If the two numbers were equal, nobody would have any reason to employ anyone. So the worker who calculates that they generated a million dollars in avoided agency fees and walks in with that figure will be told, accurately, that the going rate for the job is a fraction of it. Both statements are true and the calculation does not convert.

There is a second problem underneath. Attribution inside a firm is exactly as unresolvable as pricing outside it. Your employer does not know what you contribute either. That is what a cost center actually is: a category for work whose value nobody can attribute. The accounting system was never built to produce the number you would be trying to compute.

The advice was reaching for something adjacent, and that survives the objection.

Knowing what your work is worth in dollars is nearly useless for getting paid, and it is essential for leaving. Contribution value is how you price a business. Replacement cost is how you price a job. They answer different questions, and the first one only becomes usable at the moment you stop asking an employer to validate it. A person who can say what their work is worth to a buyer can sell it to a buyer. A person who can only say what their work is worth to their employer is asking someone with a direct interest in the answer to grade the paper.

There is a narrower use for the same exercise, and it is diagnostic rather than encouraging. If you cannot articulate the dollar value of your work, you are in a cost center. That will not get you paid more. It tells you what kind of exposure you are carrying, and it tends to become relevant on short notice.

I have spent this entire piece arguing that no reliable number exists, so it would be dishonest to close by suggesting that a better data source is coming. It is not. Anything built from firm-submitted taxonomies will inherit the taxonomy problem, and anything built from worker self-reports will inherit self-selection. More data does not resolve a definitional problem.

The most consequential change happening in compensation right now does not try to produce a correct number at all.

The European Union's pay transparency directive came into force on June 7th, 2026. Set aside the compliance industry that has grown around it and look at the architecture, because the architecture is the argument.

It does not attempt to tell anyone what a job is worth. It does four other things. It bans employers from asking candidates about their pay history, which severs the mechanism by which one underpayment propagates forward through a career. It voids pay secrecy clauses, so the information workers do have can circulate. It requires that job categories be built on objective, gender-neutral evaluation criteria, which means the taxonomy itself becomes something an employer has to defend rather than something it simply asserts. And where reporting surfaces a gap of five percent or more within a category that the employer cannot justify and does not fix within six months, it triggers a joint pay assessment conducted with worker representatives rather than an internal review.

Underneath all of it, the burden of proof shifts. In a pay discrimination claim, the employer has to demonstrate that the difference was objectively justified. The party that defined the categories, set the bands, and holds the data is now the party that has to explain them.

That is the correct shape of the answer, and it took me a while to see why. Every remedy aimed at giving workers better data leaves the definitions untouched, and the definitions are where the distortion lives. This one goes at them.

Only four of twenty-seven member states met the transposition deadline, the first substantial reports are not due until 2027, the whole regime is scoped to gender rather than to the general question of what anyone is worth, and there is no serious federal analogue in the United States. It is also, in a sense that should be uncomfortable, a rule that treats the pay gap as a matter for lawyers and inspectorates rather than for the people being paid.

But it is the first thing I have seen in this domain that is not an attempt to improve someone's judgment. It does not require a competent allocator, or a better survey, or a smarter matching algorithm. It requires disclosure and it assigns the burden of explanation to whoever holds the information. Those are the only kinds of intervention that work in a system with nobody driving it. Almost everything else in this field has failed. This might not.

So I would not tell you to know your worth. Nobody knows your worth, including the people quoting figures at you with great confidence, including the surveys behind them, and including me.

Two things follow instead, and they run in opposite directions. Both are worth holding.

The first is that the structural repair, if it comes, will not arrive as better information for you. It will arrive as an obligation placed on someone else to justify what they already decided. That is unglamorous and slow and it is the only version of this that has ever worked. Support it where you can vote on it, and do not wait for it.

The second is that the number nobody will certify is the only one that belongs to you. Its uselessness inside an employment relationship and its usefulness outside of one are the same fact, seen from either side of the arrangement. You cannot make an employer accept your accounting of what you are worth, because their accounting is a different calculation performed for a different purpose. You can find someone willing to pay you directly for the thing you do, and in that transaction your accounting is the only one on the table.

Both of these are true at once. The system may eventually be made to explain itself. You do not have to be there when it does.

None of what follows will give you the correct number. It does not exist. What these do is widen the set of incompatible answers you are working from, which is a genuine improvement over holding one.

Look up the ranges, plural. Levels.fyi is the most useful for technology leveling and total compensation structure, and it now publishes mappings between its own leveling scheme and Radford's and Mercer's, which is the closest a private person gets to the paid taxonomy. Comprehensive.io offers free technology salary data without a survey submission requirement. The Bureau of Labor Statistics Occupational Employment and Wage Statistics program is free, government-run, broader than technology, and slower than all of them. Read three, note the spread, and treat the spread as the real answer.

Find out what your jurisdiction already entitles you to. Fifteen or so US states plus DC now require salary ranges in job postings, several require disclosure on request, and many ban questions about salary history. Jackson Lewis maintains a current state-by-state breakdown. It is written for employers, which is exactly why it is useful. Separately, the National Labor Relations Act protects most private-sector employees who discuss their wages with coworkers, and contractual pay secrecy clauses are unenforceable against that right in most cases.

If you work for or with a European employer, the ground is moving under you this year. Pinsent Masons tracks implementation state by state, and Ravio's guide covers what each national law actually requires and when.

Worth reading if you want the evidence rather than the summary. The Minneapolis Fed on the trade-offs in pay transparency policy design, which is the least ideological treatment I have found. Bharat Chandar's assessment of what is and is not known about AI and labor markets, which is unusually honest about uncertainty for the genre. Brookings on how early the research still is. And on how posted ranges behave in practice, the Cornell finding that wide ranges deter women from applying and dampen how assertively they negotiate, which is its own edition and which I will come back to.