The Job Interview, Tested: What the Science Actually Says

An original TeaLamp explainer based on the sources listed below. Prepared by the TeaLamp Editorial Team with the help of AI.

Almost every working adult has sat through a job interview, and almost every hiring manager has run one. Yet for most of the last century, nobody could say with much confidence whether interviews actually worked. Then researchers spent decades testing them, and the answer turned out to be both encouraging and humbling: the way you run an interview matters far more than the fact that you run one at all.

Key takeaways on job interviews

• A structured interview, where every candidate gets the same questions and the same scoring, predicts job performance far better than a free-flowing chat.
• The classic 1998 estimate put structured interviews at a validity of 0.51 against 0.38 for unstructured ones; a 2022 re-check lowered both numbers but kept structured interviews on top.
• What makes structure work is simple: the same questions, a scoring guide written before anyone walks in, and notes taken as evidence.
• Google tested its own hiring this way and dropped its famous brainteaser questions after finding they predicted nothing.
• Even the best interview is only a partial predictor, so wise employers combine it with a work sample or a skills test.

The 85-year experiment

In 1998, psychologists Frank Schmidt and John Hunter gathered 85 years of research on 19 ways of choosing people for jobs, from references and years of experience to ability tests and interviews, and asked one question of each: how well does it predict how someone will actually perform?

Their answer, published in Psychological Bulletin, was a table of numbers that became famous. A structured interview, one where every candidate is asked the same job-related questions and scored against the same standard, carried a validity of 0.51. An unstructured interview, the familiar free conversation, managed only 0.38. Validity here is a correlation between the hiring score and later job performance, on a scale from zero to one, so 0.51 is a real, meaningful relationship, not a guess, and not a guarantee.

The same table held two more surprises. A general mental ability test also scored 0.51, exactly tied with the structured interview, and combining the two reached 0.63, one of the highest figures in the study. Years of experience, the thing many job adverts demand first, was near the bottom.

A separate meta-analysis by Michael McDaniel and colleagues in 1994, built from 245 findings covering more than 86,000 people, pointed the same way.

The 2022 re-check

For more than twenty years, the 1998 numbers were quoted everywhere, including in hiring guides and textbooks. Then, in 2022, Paul Sackett and three colleagues re-ran the arithmetic and found a problem with how the old figures had been corrected.

Meta-analyses adjust for a quirk called range restriction: the people who get hired are already a selected group, so the full range of ability never shows up in the data. Sackett, Zhang, Berry and Lievens argued, in the Journal of Applied Psychology, that the corrections used across decades of research had overshot, inflating nearly every method’s score.

When they redid the corrections, most methods lost between 0.10 and 0.20 of validity. Structured interviews fell from 0.51 to 0.42. Unstructured interviews fell much further, from 0.38 to 0.19, roughly half. Ability tests fell from 0.51 to 0.31. Work samples dropped from 0.54 to 0.33.

The ranking changed in an interesting way: structured interviews did not just survive the correction, they came out on top, the single strongest predictor of any common method. But the authors also stated a warning plainly: the 80 percent credibility interval runs from 0.18 to 0.66, so in some settings the interview predicts well and in others barely at all. A good interview is a good tool, not a crystal ball.

The paperwork behind every hiring decision. Image generated by AI.

Why the same questions work better

Why does asking everyone the same questions help so much? The research points to three mechanisms.

First, the same questions remove the interviewer’s mood from the result. In an unstructured interview, one candidate is asked about their weekend and another about a crisis at their last job, so the comparison is meaningless. Structure makes it fair because everyone faces the same test.

Second, a scoring guide, written before the interviews start, keeps ratings honest. A good rubric describes what a strong answer sounds like, what an average one sounds like, and what a weak one sounds like, for each question. The interviewer judges the answer against the guide, not against a gut feeling formed in the first minute.

Third, structure makes different interviewers agree with each other. A 1995 meta-analysis by Conway, Jako and Goodman found that two interviewers rating the same candidate agreed far more often when the interview was structured: inter-rater reliability of 0.59 against 0.37 for unstructured interviews. Agreement matters because a hiring decision usually rests on more than one person’s view.

Good structured questions come in two flavours. Behavioural questions ask for a real example from the past, such as telling about a time the candidate handled a difficult customer. Situational questions ask what the candidate would do in a future scenario. Both are easier to score when the interviewer knows in advance what a good answer contains.

What Google did with the research

Google is the best-known company to have rebuilt its hiring around this evidence, and its story is a useful check on the theory. Its People Analytics team studied thousands of its own interviews and reached the same conclusion as the academics: structured interviews predicted performance, and free-form conversations did not do nearly as well.

Google’s guide, published through its re:Work project, describes four parts to its approach: vetted questions tied to the role, detailed notes on each answer so later reviewers can see the evidence, scoring rubrics shared by every interviewer, and training for the people asking the questions. The company also reports that candidates prefer it: rejected candidates who went through a structured interview rated the experience better than those who had not.

The same research habit killed one of Google’s most famous practices. For years its interviewers asked brainteasers, such as how many golf balls would fit in a jumbo jet. When the company tested whether those questions predicted anything about job performance, the answer was nothing at all. Laszlo Bock, then head of people operations, called them a complete waste of time, and Google dropped them.

Google also learned how many interviews are enough. Its analysis found that four interviews were enough to predict a hiring decision with 86 percent confidence, and extra rounds after that added almost nothing. The company now caps most hiring loops at four interviews, a rule it calls the Rule of Four.

What structure cannot fix

Structure helps, but it has limits. Even at its best, a structured interview explains only part of why one person outperforms another. The rest comes from things no interview can see: motivation, the team around the new hire, training, and plain luck.

Structure also does not remove bias by itself. A rubric can still be written around the wrong qualities, and an interview panel can still favour people who resemble them. Research on first impressions finds that interviewers often form a view within minutes and then spend the rest of the conversation confirming it. Structure fights that tendency by forcing the interviewer to collect evidence before judging, but only if the interviewer actually follows the guide.

The newest worry is automation. Many large employers now let software screen CVs or run first-round video interviews before a human is involved. Used carefully, that can be fairer than a tired recruiter skimming a pile of applications. Used carelessly, it repeats old mistakes at machine speed: in 2018, Amazon scrapped an experimental hiring tool after discovering it had learned to penalise CVs that mentioned women’s colleges and organisations, because it had been trained on ten years of hiring data from a mostly male industry. The lesson is that an interview is a measuring instrument: it works only when built carefully, used the same way every time, and checked against something else, such as a work sample, a short task like the real job.

Why it matters

Most of us will be interviewed several times, and many of us will sit on the other side of the desk too. Knowing what the research says changes both experiences. As a candidate, it explains why a good interviewer asks the same question of everyone and takes notes instead of chatting: they are trying to be fair, not cold. As a hiring manager, it is a reminder that the friendly, free-flowing conversation feels like good judgement and usually is not. The evidence has been in plain sight since 1998, and the 2022 re-check sharpened the point: interview the same way every time, score what you hear, and never let one conversation carry the whole decision.

Related reading

For how the interview itself became a fixture of modern life, see our story The History of the Interview: How Asking Questions Became News.

The same room, the same questions, for every candidate. Image generated by AI.

Frequently asked questions

Are unstructured interviews useless?

Not useless, but much weaker. The 2022 re-check put unstructured interviews at a validity of 0.19, less than half the structured figure of 0.42. They still carry some signal, and they can reveal things a scripted question misses, but they should never be the only evidence behind a hiring decision.

Do structured interviews feel cold to candidates?

Surprisingly, the opposite is often true. Google found that candidates rated structured interviews as fairer, and rejected candidates who had one rated the experience better than those who had a free-form chat. Knowing the questions are the same for everyone is reassuring, not off-putting.

Can AI interview better than humans?

Not yet, and maybe not in the way people hope. Software can score answers consistently, which is a real strength, but it learns from past hiring data, so it can also learn past biases, as Amazon discovered in 2018. The research consensus is that AI works best as a helper to a structured human process, not as a replacement for it.

Sources

• Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin, 124(2), 262–274.
• Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection. Journal of Applied Psychology, 107(11), 2040–2068.
• McDaniel, M. A., Whetzel, D. L., Schmidt, F. L., & Maurer, S. D. (1994). The validity of employment interviews. Journal of Applied Psychology, 79(4), 599–616.
• Google re:Work, “A guide to structured interviewing for better hiring practices” (updated March 2026).
• Loomstrat, “Technical Interviews That Predict Performance” (Schmidt & Hunter figures and combinations).