The Lemon Market

Picture 500 people whose résumés look equally good. Same titles, same keywords, same confident bullet points. On paper you couldn't tell them apart.

Each of them has a truth ratio: what they actually deliver, divided by what their résumé promises.

Below 1, they oversell. Above 1, they're better than they look. On average, people deliver exactly what they advertise. But most fall a little short and a few are far better, so 59% of them oversell to some degree.

We're modeling professional, complex white-collar work, where real output varies a lot around what a résumé predicts: the spread is about 48% of the average. The switch above the chart shows unskilled jobs, where output varies less from person to person, so a résumé has less room to mislead and matters less. It also shows the source essay's worst-case assumption.

Now the market runs, with one asymmetry.

Getting hired depends on the résumé. Because the résumés are identical, everyone gets offers at the same pace.

Staying hired depends on the work. Oversellers get found out and leave sooner, in proportion to how much they oversold. Someone at 0.5 is back on the market twice as fast as someone at 1.0.

Everyone starts out looking.

A few years in, the market has settled. Look at who is still looking.

75% of them oversell.

We started at 59%. Nobody started lying more. The market kept sending the oversellers back.

The average truth ratio among those looking is 0.82: they deliver about 82 cents on each dollar their résumé promises.

34% of them come from the bottom fifth of true ability. The top fifth, the people you most want, are just 10%. Each group is 20% of the population.

The employed, meanwhile, are fine. On average they deliver what they advertise, or a little more.

But only about 5% of people are out of work at any moment, and they are the people a job posting reaches most easily.

People who search while still employed are counted with the employed here. Act III looks at them.

Take one person from the looking pool. Before you talk to them, all you know is where they came from.

In that pool the typical truth ratio is 0.74, and 80% of people fall between 0.41 and 1.32. That's the range this person could be in.

You interview them the way most companies do: an unstructured conversation. It goes well.

How much did you learn? An unstructured interview predicts job performance with a validity of about r = .18. That narrows the range of what this person might be worth by 1.1%. The curve barely moves.

"Goes well" here means a better interview than about five in six candidates give.

Switch to a structured interview: the same questions for everyone, scored against a rubric. It's the best-validated interview format, at about r = .44.

Same good result. The range narrows by 8.2%. Most of your uncertainty is still there.

Validities from Sackett, Zhang, Berry & Lievens (2022). Narrowing is the drop in the spread of what the person might be worth, averaged over every possible interview result in this pool.

Now suppose you had a near-perfect read, r = .9, better than any hiring method ever measured.

Even that narrows the range by only 52%. 48% of your uncertainty about this person survives.

Why so little? A validity of .44 means your read accounts for r² = 19% of the differences between people. Uncertainty is a spread, and a spread shrinks with the square root of what's left: √(1 − .19) ≈ 0.90, or 10% narrower on a textbook bell curve. In this skewed pool it's a little less, 8.2%.

So uncertainty falls slowly at first and drops fast only near r = 1, which no interview comes close to.

Each row is a way of hiring. Each time, you draw five candidates, read them as well as that channel allows, and hire whoever reads best. The bar shows where your hires really land, by fifth of true quality across everyone.

Start on the open market with no real assessment, hiring on charm alone. 34% of hires come from the bottom fifth. The average hire sits at the 38th percentile.

Better interviews help. An unstructured interview gets you to the 43rd percentile, and a structured one to the 50th.

But that's the ceiling for the open market. The best interview format is still ranking five people from a pool that's full of oversellers.

Now change the pool instead of the interview. Hire at random from people who already have jobs, with no interview at all.

The hire lands at the 51st percentile, against the 50th for the best interview on the open market. In professional jobs, the employed pool starts about where the open market's best effort ends.

Put the structured interview back, and poaching reaches the 65th percentile. Same interview, better pool.

Poaching has costs this model leaves out: you usually have to pay more, and the current employer may counter. Act IV covers counter-offers.

Referrals are the third route. You only get one name, so the choosing happens in the referrer's head: they think through the people they could send and pick the one they rate best. We assume they weigh three people, which is generous. Most referrers probably consider between one and three.

It matters who's being referred. A friend who is currently job hunting still comes from the looking pool. Even with a well-informed judge, it lands at the 52nd percentile, about where a structured interview on the open market gets you.

Someone who has worked with a person reads them with a validity of about r = .49, the figure for peer ratings (Schmidt & Hunter, 1998).

Now ask someone who has worked with a person who is employed elsewhere. The judge knows the work, and the pool isn't full of oversellers.

The hire lands at the 67th percentile.

Best of all, ask your strongest people. Good people tend to know good people. Say the referrer is at the 90th percentile, and their contacts' quality correlates ρ = .3 with theirs.

The hire lands at the 77th percentile, or the 82nd at ρ = .5. 67% of those hires are from the top fifth.

ρ is assumed, not estimated. Nobody has measured how closely a strong performer's network tracks their own quality.

Those values of ρ may be cautious. People sort strongly by background, education and ability into friend groups and marriages, so a strong performer's network could track their quality far more closely.

At ρ = .8, the hire lands at the 88th percentile, and 80% of those hires are from the top fifth. At that point the network does most of the work. Even if the referrer sends whoever comes to mind first, the hire lands around the 80th percentile.

You've found someone good who already has a job, and made them an offer. Now their employer gets a say.

This chart is everyone with a job and an outside offer. On average they deliver about what their résumé says.

Most employers let people go. About 12% of people with an outside offer get some kind of counter-offer.

The employer has watched this person work for years, a far better read than any interview: about r = .7, against .44 for the best interview. Say it counters when its read of them is in its top 12%. Those people average 1.75. Everyone else averages 0.91.

The 12% is from Faberman, Mueller, Şahin & Topa (2022). For the employer's read: two supervisors rating the same person agree at about .52 (Viswesvaran, Ones & Schmidt, 1996), so one supervisor tracks true performance at about √.52 ≈ .7. Several people weighing in on a counter would push it higher. Counters made for other reasons, like replacement cost, would pull it lower.

So a counter-offer is the best-informed party in the market telling you this person is good.

It's common to treat a counter as a reason to walk away. In this model, it's the strongest evidence you'll get that you picked well.

There's a catch, the winner's curse. The employer will match your offer up to what it thinks the person is worth, so you only win where your offer beats its valuation. That means the countered people it valued least.

Say you win half of these fights, the half the employer valued least. The people you win still deliver 1.34× what their résumé suggests. Someone nobody fought for delivers 0.94×. That's 43% more output from the same résumé.

What's that worth? Not salary: you only open a role if the hire creates noticeably more value than they cost. Say an as-advertised hire here earns $150,000 and creates $300,000 a year.

The gap between 1.47× and 0.91× is 0.57 of an as-advertised hire's output, or about $120,000 a year of extra value. That's your break-even: you could pay a countered candidate up to that much more than someone nobody fought for and still come out even. Anything less is profit.

The calculator below lets you change the counter rate, the employer's read, how many fights you win, the value of the role and the expected tenure.

One more thing a counter can't do: improve your pipeline. It can't add anyone, only take away someone you already wanted. And it goes after your best people, because the employer counters the people it values most.

So the best channels pay the most. Your first choice from an A-player referral (ρ = .8) gets countered 40% of the time, against 12% for a random poach. Walk away from every counter and those referrals slip from the 86th percentile to the 80th. Win the half the employer valued least and you recover to the 82nd.

The last decision is when to stop. Candidates arrive one at a time, and you usually have to decide on each before long.

Each dot is a candidate from the looking pool. For now, pretend you can see exactly how good they are: the higher the dot, the better.

The textbook answer is the secretary problem. Look at the first 37% without hiring anyone, then hire the first person who beats all of them.

That finds the single best candidate about 37% of the time. If all you can judge is who beats whom, and you can't go back to people you've passed on, no rule does better.

The problem reached a wide audience through Martin Gardner's 1960 column. Lindley (1961) and Dynkin (1963) worked out the 1/e rule, and Ferguson (1989) tells the history. The figures on this page come from our own simulations, which reproduce it.

But you don't need the single best. You need someone great. Say a top-5% hire is good enough, and, as the textbook does, that you can interview 100 candidates.

Look at only the first 20%, and you land a top-5% hire 67% of the time, against 61% at 37%. You also stop after about 52 candidates instead of 74.

The textbook also assumes you can never go back. Employers can: the candidate you passed on last week may still be available. Few employers ever do it.

In this idealized setting, recall looks powerful: the same 20% rule lands a top-5% hire 85% of the time. It's insurance against the case where nobody beats your early benchmark. Without recall you'd be stuck with whoever comes last. With it, you only need someone good to still be available.

Other employers are interviewing the same people, so the ones who interview best go first: someone in the top 10% of interviews gets hired away about twice as fast as the median. The whole search lasts about one average job search. Recall with backward solicitation is studied in Yang (1974).

Now the catch from Act II. You don't see how good people are. You see how they interview.

These are the same candidates, now placed by a structured interview's read. The colours still show the truth. The rule is the same; it just ranks by the wrong thing.

With a real interview, looking at 20% first lands a top-5% hire only 24% of the time, or 30% with recall. At the textbook 37% it's 21%.

And nobody interviews 100 people for a professional role. With a more realistic ten, looking at the first two lands a top-10% hire 18% of the time, 19% with recall. Even a perfect read would manage only 30%, because ten people rarely include a star.

The stopping rule and recall move the result by a few points. With ten interviews and a noisy read, what decides your hire is who's in the ten. That's Act III.