The Pay You Can't Explain
Ask a compensation leader to publish the pay data.
Not the ranges — those are fine, most companies will show you those. The actual numbers. What every person in the same job is paid, and why they differ.
Watch what happens. You will not get a policy objection, at least not first. First you get a flinch. Then, a beat later, the reasons arrive: it's more complicated than it looks, people would take it out of context, we'd spend six months in one-on-ones.
Every one of those is true. None of them is the reason.
The reason is that they already know what's in the file. They know there are two people doing the same work a level apart in pay because one was hired in a hot quarter and one wasn't. They know somebody has been in the seat eleven years and a decade of flat percentages compounded into a number nobody would set today. They know the person who threatened to leave got twelve percent and the person who didn't got three, and that the difference had nothing to do with the work. They're not afraid of the numbers. They're afraid of the question — and specifically of the moment they have to answer it out loud and hear how it sounds.
That flinch is the most useful diagnostic in compensation, and almost nobody reads it as one. It isn't squeamishness. It's an organization telling you that its pay cannot currently be explained.
The standard is explicability, not sameness
Let me say clearly what I'm not arguing, because the field's default assumption will fill the gap otherwise.
The goal is not that everyone is paid the same. It never was. Large differences in pay are appropriate, and in most organizations they are not large enough — a pay distribution flatter than the distribution of value being produced is itself an error, and a common one. Google's Work Rules! makes this case as "pay unfairly," and the argument holds: if your best people are producing several times what the median produces and being paid a few percent more, you have a problem that looks like fairness and behaves like a slow leak of your best people.
The goal is that you can explain the pay, and the differences in it, in a way that is credible, reasonable, and related to profits or to the success criteria of the job.
That standard cuts in both directions, which is what makes it a standard and not a slogan. A twofold gap between two people in the same job is fine if you can say why in terms both of them can check. And an identical rate for two people producing very different value is also an error — a more expensive one, because the person producing more can see it perfectly well and will draw their own conclusion about what this place rewards.
So the question is never "how equal is our pay." It's: for each difference, can we state the reason, and does the reason hold up?
Notice that "related to" is doing deliberate work there. Not caused by — I've argued elsewhere that we don't have the outcome variable that would let anyone make causal claims about individual contribution in dollars, and I'm not going to pretend otherwise here. Related to. Connected to something the business would defend in public. That's a much lower bar than causation, and most pay systems still can't clear it.
How pay becomes unexplainable
Nobody designs an indefensible pay structure. It accumulates, and it accumulates for a structural reason worth naming precisely: the absence of disclosure removes the feedback loop.
That's all opacity does, mechanically — it severs the signal that would tell you a pay decision was wrong. And a system with no correcting signal doesn't merely tolerate error. It compounds it, and then it exports it.
Watch a single mistake travel. Someone is slotted a level low at hire, for a reason that made sense that week — a soft market, a candidate who didn't negotiate, a hiring manager working from a stale range. Nothing corrects it, because nothing can: the one party positioned to notice can't see the comparison. The error then persists through review cycles as a percentage of itself, so every merit increase compounds it rather than closing it. It gets baked into the range when the range is next rebuilt from actual paid rates. And then it leaves the building entirely — because pay surveys are assembled from what employers report paying, and employers set what they pay by reading pay surveys. The market data is partly a mirror. An error entered anywhere in that loop circulates back as evidence, wearing the clothes of an external benchmark, and is used to justify the next one.
That's what I mean by propagation. Not a metaphor. A defect with no error-correcting channel, inside a system that feeds on its own output.
Which is why the flinch is so common. It isn't that these organizations were careless. It's that every one of them has been running an accumulator with the correction mechanism switched off, and the file now contains a decade of decisions nobody would make again and nobody has any way to unwind.
"We pay market" is the canonical non-explanation
The most common answer to why is this person paid this is "we pay market," and it's worth being precise about how little that sentence contains.
I've made the basic case elsewhere — that a market is a list rather than a point, that "we pay market" usually means the median and therefore names a deliberate decision to be uncompetitive to the better-paid half of your own talent pool, and that narrowing by country, industry, city, or actual competitor set produces a different number each time. All still true.
Here's the part I haven't written down before, and it's the one that ends the argument.
Take the sentence literally and price it. Take your job architecture, put every job at the median of its market, and total it. Then compare that to what you actually spend. It will not match, and in most companies it won't be close — because paying the median for every job simultaneously is not affordable, and wasn't ever the plan.
Which means you are already paying above market for some jobs and below for others. You have always been differentiating across job families. The only question is whether anyone chose where, on what basis, and whether they could defend it.
And there's an arithmetic joke underneath all of this that I've never seen anyone say out loud: if everyone paid market, there would be no market. The median exists only because half of employers are below it. Somebody is below. "We pay market" is a claim that, universalized, dissolves the thing it refers to.
Then the second dimension, which "market" doesn't touch at all: does it mean the same number for every person in the job, regardless of performance, tenure in role, or how far into the learning curve they are? Obviously not — and you can tell it doesn't, because you built a range with a spread precisely so that it wouldn't. Nobody designs a range and then pays everyone the midpoint. The range exists because differentiating within a job is the intention. But the intention is usually where it stops; the rule for moving through the range is the piece almost nobody writes down, which is exactly why eleven years of flat percentages can happen to someone without anyone deciding it.
So "we pay market" names, ambiguously, one of at least five decisions a real pay system has to make — the level structure, the market definition per segment, the target position and why, the spread of the range, and the rule for progression through it. Four are unstated. That isn't a strategy. It's an abstention, and an abstention still produces a distribution. Just not one anybody chose.
The number worth computing
All of this is measurable, which is the part that frustrates me most.
The conversation about pay is conducted almost entirely in terms that were never defined well enough to test. Is it fair. Are we competitive. Do people feel valued. Those aren't bad questions, but in the form they're usually asked they cannot be wrong, which means they can't be settled — so the argument runs indefinitely, on conviction, while the underlying quantity sits there waiting for anyone willing to define it.
The testable version isn't about the spread at all. Decompose the variance in what you pay into three parts:
- Variance explained by criteria the business would defend in public — level, scope, measured performance, scarcity of the skill, geography if you've decided geography counts.
- Variance explained by criteria it would not defend — hire date, negotiating posture, whether someone had a competing offer, who their manager was in 2019.
- Variance explained by nothing at all.
The first is your pay strategy, whether or not you wrote it down. The second is the accumulated residue of a thousand individual decisions. The third is the number worth knowing, and almost nobody computes it.
Everything else follows from that decomposition. How much correction costs. Which direction each error points — and note that they point both ways, which is why fixing this is far cheaper than it sounds. Some of the money to pay the underpaid is currently sitting on people who are overpaid relative to any criterion anyone would state.
Here I'll be explicit that I'm handing you an observation from practice rather than a finding, because I don't have the evidence in hand and won't dress it as though I do: in twenty-odd years of this work, the money required to hit essentially all of a company's stated compensation objectives is usually already in the budget. It is simply not distributed well. I have never run a clean test of that claim. It is testable — which is the point of saying it out loud — and I'd rather offer it as a hypothesis someone can knock down than smuggle it in as a premise.
The move that actually corrects it
Once you can price content, an obvious commercial play falls out, and my first instinct was to refuse it on principle. I've since decided that instinct was wrong, and the reasoning is worth showing because it's where a lot of people's ethics on this go soft in the wrong direction.
The play: price the content of every job against what the market pays for that content. Find the people whose skills command more elsewhere than they're paid here. Hire them — at, say, twenty percent above what they make now, and materially below what the credentialed external candidate would have cost.
My reflex was that the margin there is the employee's ignorance of their own value. That's wrong, and it's wrong in an important way. The margin is the pedigree premium you decline to pay — you're arbitraging the wrapper, the title and the résumé and the brand-name employer, which is exactly the thing this whole method says is mispriced. The employee's realistic alternative was not the credentialed hire's salary. It was staying where they were, underpaid, indefinitely.
So run the moral arithmetic honestly. What's the alternative on offer? Leaving someone to languish in a job paying less than their skills command, because it would be unseemly to profit from moving them? That isn't the ethical position. They are already being paid less than they should be — right now, today, in the status quo. A twenty percent raise and a bigger job is a correction of that, not an exploitation of it. Both sides gain. That is not a loophole in a functioning market; it is the definition of one.
And here's the part that makes it more than a good deal for two parties: it's self-correcting. Pull enough people out of an underpaid role and demand for that role rises, which should push its wages up — the correcting signal the internal system couldn't generate on its own. And if wages there don't move, that's not a failure of the theory. It's a finding: those skills were easier to replace than anyone believed, and the premium people assumed was there wasn't. Either way the market returns information that the opaque system was structurally incapable of producing.
Which reframes the whole thing. Consistent talent arbitrage isn't the flinch productized. It's one of the few forces that actually corrects mispriced pay, from the outside, when the inside has no mechanism to.
The conditions are load-bearing, though, and they're where this can genuinely go wrong:
Consistency. Do this as a declared staffing strategy applied across the board, not case by case on whoever happens to be visible. Selective application is precisely where the pay-equity exposure lives — a disparate-impact analysis is designed to find exactly the pattern of "some people got the correction and some didn't," and if who got it correlates with a protected class you have manufactured a finding against yourself. The strategy is defensible; the exceptions are not.
A real raise, not a technical one. Twenty percent is a different proposition from three. The number has to be large enough that the person would describe it as a good deal in their own words, not one that requires a spreadsheet to look like a win.
A path after the transition. If they clear the ramp, the next move has to be substantively larger again, relatively. Otherwise you've hired someone into a ceiling — the trade felt fair exactly once, and the second year is when they work out that the discount was permanent. That's not a bolted-on retention program; it's the thing that makes the original bargain true rather than merely accurate.
Notice what all three conditions have in common. None of them is about the size of the differential. All three are about whether you could state the arrangement plainly to the person it applies to and have them agree it's a fair one. Which is the same standard as everything else in this essay. The defect was never the difference. It's the inability to explain it.
Then transparency stops being the hard part
Which brings me back to where this started, and to the thing I had backwards.
I used to frame disclosure as the forcing function — the reckoning that arrives and reprices everything. It isn't. Disclosure isn't arriving on its own; it's blocked, and what blocks it is precisely that the explanations won't hold water. Employers aren't withholding pay data because they're ideologically committed to secrecy. They're withholding it because they've read the file.
So transparency is not the lever. It's the exam you can pass once the system is right. And the order matters enormously, because the two get attempted in the wrong sequence constantly — disclose first, and you've published a decade of unexplainable differences and bought yourself six months of meetings you cannot win.
The evidence on what happens when disclosure lands on an unprepared system is not encouraging. When California required cities to post municipal salaries, Alexandre Mas found city-manager pay fell about 7 percent — and quit rates rose about 75 percent.1 Read those together. A modest pay change produced a three-quarters increase in departures among senior executives, and the departures, not the repricing, are where the cost is. Nobody left over seven percent. They left because an arrangement they'd been operating under got restated in public in a form nobody had ever had to defend, and it didn't survive the restating. Mas reads the effect as public aversion to high pay rather than accountability working, which is the same point from the other side: the pay was defensible on the fundamentals and indefensible out loud.
Run it the other way and the picture inverts. Get the system explainable first, and disclosure becomes the thing that makes it pay. We build an instrument that measures how much current pay an employee will trade for credible future opportunity — a real, elicited number — and it has a property that looks like a bug until you see what it's telling you: when credibility fails, the accepted discount goes to zero.2 Not "shrinks." Zero. A promise nobody believes is worth nothing, and a ladder you can't see is a promise nobody believes.
That gate is the whole economics of this in one line. If your arrangement is defensible, its value is a function of how visible it is, and every disclosure makes it worth more. If it isn't, disclosure sets it to zero. Same dollars, opposite exposure to the one event you don't control.
I should be honest that the aggregate research on transparency doesn't hand me a clean win here, and I'd be doing the thing I complain about if I skipped it. Within a firm, transparency tends to lower wages — about 2 percent — with the decline concentrated on college-educated and non-unionized workers and essentially absent for everyone else, and the resulting compression comes from the top down rather than the bottom up.3 Job-posting transparency runs the other way, raising wages 1.3 to 3.6 percent by intensifying competition between employers.4
Both, at once. Which is my point rather than an inconvenience: asking whether transparency raises or lowers the wage bill sums two opposing movements and returns something small. The quantity that matters is allocation, not level — whether the money moved toward the work that earns it — and none of those study designs can see that, nor do they claim to.
The exam and the curriculum
So: what were we actually trying to do.
Not equalize pay. Not disclose pay. Explain pay — every difference in it, credibly, reasonably, against criteria connected to whether the business succeeds and whether the job is being done well. That's the whole objective, and it's the thing the pricing method exists to serve: a defensible account of why this job is paid this and that one that, in terms drawn from the work rather than from who negotiated hardest in a good quarter.
Everything else is downstream. Get it right and pay equity is mostly a report rather than a project, because the differences were already explained. Get it right and the retention conversations change, because you can tell someone precisely where they stand and what moves them. Get it right and transparency stops being the terrifying part, because there's nothing in the file you'd be embarrassed to read aloud.
You cannot disclose your way to a defensible system. You can only defend your way to a disclosable one.
And you'll know you've done it on the day someone asks to see the pay data and nobody in the room flinches.
This is a companion in the Content Pricing program — the method for producing exactly the explanations this essay says are missing. An Occupation Is Not a Job fixes the unit any such explanation has to run on; Moneyball Never Made It to Work argues why the explanation has to be built against the market's own revealed prices rather than against an outcome nobody has agreed on. What Are You Paying For? is the fuller treatment of the three kinds of value a pay number reconciles, and of why "we pay market" says so little. Show Your Work takes up the procedural-justice case for legibility in individual decisions.
One claim here is deliberately unsourced and labelled where it appears — that the money required to meet a company's stated compensation objectives is usually already in the budget and simply misallocated. That is my observation from practice, offered as a testable hypothesis; nothing else in the argument depends on it.
Footnotes
-
Alexandre Mas, "Does Transparency Lead to Pay Compression?" Journal of Political Economy 125, no. 6 (2017): 1683–1721. California's 2010 municipal salary-disclosure mandate, comparing cities where the mandate was the first disclosure against the 63% already disclosing; 482 cities, payroll and CalPERS data 1999–2012. City-manager compensation fell ~7 percent, largely in nominal terms, with quit rates rising ~75 percent. No relative decline at the 50th, 75th, or 90th percentiles of the city wage distribution — compression at the top rather than a citywide budget effect. Mas reads the pattern as more consistent with public aversion to high compensation than with accountability effects. ↩
-
The managed-progression program and its perceived-opportunity instrument are internal (people-analytics-toolbox,
anycomp): a conjoint/MaxDiff design eliciting the current-pay discount employees will accept for credible future opportunity, with movement modeled against per-level exit-risk hazard. The credibility gate is literal — measured credibility failure sets the accepted discount to zero rather than merely reducing it. Empirical calibration of the discount ceiling is open work, and nothing above depends on a particular value for it. ↩ -
Zoë B. Cullen & Bobak Pakzad-Hurson, "Equilibrium Effects of Pay Transparency," Econometrica 91, no. 3 (2023): 765–802. Event-study difference-in-differences over ACS 2000–2016, 13 states enacting laws protecting workers' right to discuss pay. Wages −2.2% at t+1, −2.7% at t+3, −1.8% mean across the post period. Heterogeneity: college-educated −3.2%, no-college statistically null; low-unionization occupations −3.4% at t+3 versus a statistically insignificant −1.5% in high-unionization occupations. Six of nine studies in their meta-regression find men's wages falling more than women's — compression achieved by lowering the top. The mechanism is that observability lets employers credibly refuse individual exceptions, reducing individual bargaining power. ↩
-
David Arnold, Simon Quach & Bledi Taska, "The Impact of Pay Transparency in Job Postings on the Labor Market," NBER Working Paper 34480 (November 2025). State laws requiring salary ranges in postings raised the share of postings carrying salary information by ~30 percentage points and wages by 1.3–3.6 percent across three datasets, with no detected effect on pay dispersion, employment, or posting volume. Working paper, not peer-reviewed. The authors explicitly distinguish cross-firm from within-firm transparency and cite Cullen & Pakzad-Hurson for the latter. ↩