Structured interview guide: research, questions and scoring

The conversation flowed, you liked the candidate, and 3 months later the hire isn't working. A structured interview is the best-supported method we know to stop that happening again. It isn't a verdict on your judgment, because this happens to careful, experienced interviewers too.

An unstructured interview feels informative even when it isn't. In a foundational 2013 experiment, some interviewees secretly switched to random yes-or-no answers halfway through. Their interviewers found the interview just as useful.

In a structured interview, every candidate gets the same questions in the same order. You score every answer against written examples before anyone compares notes.

Key takeaways

• Do structured interviews really predict job performance? Yes. They ranked first of the hiring methods compared in a 2022 review of decades of studies, about twice as predictive as free-form interviews.
• Will it feel robotic to candidates? Not with a short, warm, unscored welcome. In Google's data, rejected candidates who had a structured interview were 35% happier than those who didn't.
• How much work is it to build? About 12 questions and 30 short example answers (5 per skill), written once per role and reused.
• How many interviewers do I need? Up to 4, each scoring alone before the debrief. In Google's data, each interview after the 4th added under 1% to predicting the hire decision.
• Can AI run it? It can ask the same questions and score against your criteria, but keep a person on every decision. NYC, Illinois, California and the EU add notice, audit or record duties.

1. Does a structured interview actually predict job performance better, and how solid is the evidence?

Yes, and the evidence is decades deep. In Sackett and colleagues' foundational 2022 re-analysis, structured interviews ranked first on validity: .42, against .19 for unstructured ones.

Validity measures how closely interview scores line up with how people later perform. It runs from 0 (no link) to 1 (a perfect match).

No method in the re-analysis gets near 1. So the gap is what counts: the structured version's link is more than twice as strong.

In the chart below, a job knowledge test (.40) and keyed biodata, a scored background questionnaire (.38), come next. Work samples (.33) and cognitive ability tests (.31) follow.

Hiring methods ranked by validity: the structured interview leads at .42 and the unstructured interview trails at .19. Source: Sackett and colleagues, 2022.

The estimates changed because the re-analysis revisited corrections that had inflated them. After fixing an over-correction in older studies, most methods dropped .10 to .20, while the structured interview dropped only .09.

The opening experiment comes from Dana, Dawes and Peterson's foundational 2013 study. In it, 76 students predicted classmates' grades, with each classmate's past grades in hand. For some predictions, they also held a 20-minute interview first.

The interview made things worse: predictions made after one matched real grades at .31, against .61 without one. The chat distracted them from the better evidence, the past grades. It used students, not hiring managers, so treat it as a warning, not proof.

Kahneman and colleagues' book Noise puts a number on it. If 1 candidate simply interviewed better than another, the chance they're truly the better one is only 56% to 61% (NPR). A coin flip gets you 50%.

2. Our last hire interviewed brilliantly and then flopped: what should a structured interview test instead?

Test the work, not interview skill. Ask how they last did the job's core tasks, score answers against written examples, and add a short work sample.

That kind of flop hurts, and the reason is simple. A free-form chat rewards people who interview well, and being fluent in a meeting is a different skill from doing the job.

That's why, in section 1's ranking, tests of the actual work beat the free-form chat, and only the structured interview beat them. So use both.

Say you're hiring for customer support: ask "Tell me about the last time your queue was more than you could clear. What did you answer first?" Then score what they did, not how well they told the story.

A strong answer names the order and the result: "I answered the 3 customers threatening to cancel first." A general answer, "I always stay calm and prioritize," scores lower, however well told.

For the work sample, use a small, real piece of the job, such as 3 customer emails to answer in writing. Run it live or recorded, and ask the candidate to explain 1 choice they made. A copied answer is easy to paste but much harder to defend out loud.

3. Won't a structured interview feel robotic to me and to the candidate?

It doesn't have to. Keep the welcome warm and unscored, allow planned follow-ups, and take the candidate's questions at the end.

The worry is fair. When Daniel Kahneman built a structured interview for the Israeli army in 1955, his interviewers felt "reduced to unthinking robots" (Nobel Prize autobiography). But structure fixes the questions and scoring, not the tone, so you can still sound like yourself.

The table below shows what each style holds fixed.

You don't need to ban follow-ups. Huffcutt and Arthur's foundational 1994 meta-analysis pooled 114 results from earlier studies of entry-level jobs. It sorted them into 4 levels by how fixed the questions and scoring were.

Validity climbed from .20 at level I, the loosest, to .35 at level II and .56 at level III. Level IV, the strictest, uses identical questions, no custom follow-ups and answers scored against pre-set examples, yet reached only .57. Because the categories differed in several ways, that comparison does not isolate the effect of follow-ups.

Interview validity by level of structure: it climbs from .20 at level I to .56 at level III, then flattens. Source: Huffcutt and Arthur, 1994.

OPM's guide keeps follow-ups fair: decide the range of probing in advance and use very similar probes for everyone.

The figure below weighs both sides. You gain validity of .42 against .19, answers you can compare, a written record and smaller group gaps than ability tests. The costs are stiffness, leaked questions, build time and pushback, each with a fix.

Structured interview pros and cons: 4 things you gain, and 4 costs with a fix for each.

Candidates don't seem to mind either. Google's re:Work guide reports that rejected candidates who had a structured interview were 35% happier than those who did not. It's Google's own figure, so treat it as a signal, not independent proof.

4. How do I build a structured interview for the role I'm hiring now?

Write a scorecard, pick about 6 skills, and write questions and 5 anchors for each. Then run it the same way for everyone and score alone.

It sounds like a lot, but you write it once per role and reuse it. Our illustrative example is a support hire at a small company.

Structured interview build in 6 steps: from the role scorecard to running it and scoring alone.

1. Write a scorecard. Smart and Street's 2008 book Who starts with a mission and 3 to 8 ranked outcomes. For our support hire, the top outcome is a same-day first reply to every ticket (illustrative). Every question you write should trace back to 1 of these.
2. Pick about 6 skills. Kahneman's 2011 book Thinking, Fast and Slow suggests about 6 traits, each scored on its own. That stops 1 strong impression from coloring the rest.
3. Write 1 or 2 questions per skill, with planned probes (see 4.2).
4. Write 5 anchors per skill. An anchor describes what an answer at that level shows.
5. Calibrate. Have every interviewer score 1 recorded answer, then compare. If 1 person gives a 2 and another a 4, agree why and tighten that anchor first.
6. Run it the same way, and score alone. Section 5 shows how 1 answer becomes 1 score.

4.1 How many questions, interviewers and minutes do I need?

Plan about 12 questions and no more than 4 interviews. Length matters less than you'd expect.

Thorsteinson's foundational 2018 meta-analysis found no link between interview length and how consistent or predictive the scores were. In practice, a short, well-built interview does as well as a long one.

Google studied 5 years of interview data to find when extra interviews stop helping. The scores from 4 interviews predicted the final hire-or-reject call with 86% confidence, and each extra one added less than 1%. That measures agreement with the panel's decision, not how the hire later performed.

4.2 What are 12 structured interview questions for a support hire?

Each question maps to 1 of the 6 support skills. Behavioral interview questions ask what someone did before, so use them for experienced applicants. Situational ones describe a scenario and ask what they'd do, which suits first-timers.

Both kinds work, and past-behavior questions predict slightly better: .31 against .25 before statistical corrections, per Taylor and Small's foundational 2002 meta-analysis. For first-round screens, see our pre-screening interview questions.

5. How do I get my interview panel to score a structured interview the same way?

Give them a 5-level scale written as observable behavior, have them score 1 recorded answer together, then score each skill before moving on.

Without a shared scale, 1 interviewer's 4 is another's 3, and the debrief turns into an argument about numbers. An anchored interview rating scale fixes that because it writes out what each level shows, as in this one for product learning.

Say a candidate learned a new help-desk tool before launch and kept their own notes. That's a 4, whoever runs the interview.

The figure below follows 1 answer to its score. You ask in the same words and order, using only planned probes, then note what they said and match the closest written level. You score it 1 to 5 now, before the next question, so a vivid later story can't crowd out a strong early one.

How 1 answer becomes 1 score: 6 steps from asking the question to moving on to the next one.

5.1 Is there any place for gut feel?

Yes, but last. Kahneman's rule: "Do not eliminate intuition but delay it" (Behavioral Scientist).

In his army interview, interviewers first scored 6 traits, then closed their eyes and gave 1 overall score from 1 to 5 (Ed Batista). That gut score predicted as well as the 6 ratings combined. The order made it work: by then, their gut was reacting to evidence they had just scored.

So record your gut score last, and hire the highest total even if you like another candidate more.

5.2 Why should each interviewer score alone before the debrief?

Because once anyone has spoken, say the most senior person, the others are no longer scoring independently.

The figure below, from OPM's panel guidance, compares 2 orders. Score alone first, and scores go in before the meeting, so the debrief explains the gaps and records a consensus. Debrief first, and the panel talks before anyone rates, so nobody explores the gaps.

Score alone before the debrief: the 2 orders compared, based on OPM panel guidance.

Then compare answer by answer, as Iris Bohnet recommends in HBR: all answers to question 1, then all to question 2. In her foundational 2016 lab study, people picked candidates by past task scores. Judged 1 at a time, the stronger performer got picked 62% of the time and the weaker 52%, only 10 points apart.

Side by side, it was 54% against 8%. The study used task scores, not interviews, so applying it here is our inference.

6. Is a structured interview still worth it if candidates can feed our questions into ChatGPT?

Yes, because its strength is the scoring, not secret questions. Refresh questions as Google does, ask for concrete examples and verify important claims.

It's a fair worry. In Gartner's surveys, 39% of candidates used AI in their applications, and 6% admitted to interview fraud, such as posing as someone else. So assume some candidates have seen your questions or rehearsed with a chatbot, and plan for it from the start.

A leaked question ruins a quiz, where the answer is a fact. It hurts a structured interview less, because your anchors reward specifics: what the candidate did, what happened and what changed as a result.

Say a candidate gives a polished answer about calming an angry customer. Ask what the customer did next and what the candidate changed in their own process afterwards. A script can invent that detail too, so use it to identify claims to verify, not as proof of authenticity.

Refreshing the questions helps too. Bock's 2015 book Work Rules! describes Google refreshing its questions often, so candidates can't compare notes. This won't catch every coached candidate, but it makes coaching less useful.

7. Does a structured interview protect us if a rejected candidate claims discrimination?

It helps: research links structure to smaller rating gaps between groups and better court outcomes. It removes neither risk.

Nobody enjoys thinking about claims mid-hire, but a scored record is what you'll want if one arrives.

• Group gaps: Berry and colleagues put the gap between Black and White candidates' average scores at .24 standard deviations in structured interviews. That is under a third of the .79 for cognitive ability tests.
• Court cases: Williamson and colleagues' foundational 1997 study read US federal court cases about interviews. In 84, a candidate said an employer treated them differently because of a trait such as race or sex. In 46, candidates said a neutral-looking step screened out their group. Employers won more often with more structure, and most of all with objective, job-related interviews.

Under the US Uniform Guidelines, even an unscored interview counts as a selection procedure, so the rules for hiring tests cover it too. Adverse impact means 1 group passes at a much lower rate than another.

Say 30 of 100 men pass your interview and 18 of 100 women do. Women pass at 0.60 of the men's rate, and federal agencies generally treat anything below 0.80 as evidence of adverse impact. That's when per-answer scores pay off, because they can help you see which question drives the gap.

8. Can a one-way or AI interview run our structured interview, and what do the 2026 rules require?

It can, with a person on every decision, but the evidence on AI scoring is early. A 2026 study found several AI models, averaged together, may score as well as 1 human rater, but urged caution for high-stakes use.

Stockdale, Hickman and Liu tested large language models, the technology behind ChatGPT, as scorers on 2 sets of interviews, with 954 and 144 people. The best results came from averaging several larger, newer models, each told in detail what the score measures. This means the AI, like a person, needs a clear written description of each skill.

In practice, let AI ask the same questions and draft scores against your criteria, then have a person review the recording and decide. Score what the candidate said, not emotions a tool guesses from their face or voice.

The main rules as of September 2026 cover notice and consent, bias audits, records and human oversight, plus 1 outright ban:

The NYC rule only covers a tool you rely on alone, weight above every other criterion, or let overrule people. That's another reason to keep the AI score as 1 input among several, weighted no higher than the rest. Anti-discrimination law still applies to an AI score however you weight it.

How Flowmingo helps

Flowmingo's AI Interviewer runs a structured, spoken interview that each applicant does in their own time. AI drafts the questions and criteria from your job description; you edit both and mark the must-ask ones. It asks a follow-up when an answer is thin, so follow-ups differ between candidates.

Each criterion gets a 0 to 10 score with a written justification, weighted by importance, next to the recording and transcript. You decide who gets hired.

Flowmingo sample report: importance, a 0 to 10 score (finer than 1 to 5) and justification per criterion.

In our data, top scores are rare.

Flowmingo data · 86,638 AI-interviewed candidates · 2,574 companies · Jul 2025 to Sep 2026

• 3.8 to 7.8 out of 10 is where the middle 80% of AI interview scores fell.
• 6.4% of candidates scored 8 or more out of 10 in the AI interview.

AI interview scores against each job's own criteria, not job performance, and not comparable across jobs.

9. How do I check whether our interviews are structured enough?

Check your process against the 15 components in Campion and colleagues' foundational 1997 review: 7 on what you ask, 8 on how you score.

This check matters because the same method gives very different results in different settings, and design is part of the reason. In Sackett and colleagues' figures, about 10% of hiring settings can expect validity of .18 or less, no better than a free-form chat's .19.

Give yourself 1 point for each component you do every time. A step you only do sometimes doesn't count, because inconsistency is exactly what structure removes.

Say your team asks the same questions (2), uses anchored scales (9) and has 2 interviewers (11). If they chat about candidates between interviews, number 13 is your first fix.

For a benchmark, Levashina and colleagues' foundational 2014 review looked at 104 interviews from research articles. On average they used 5.74 of the 15 components, about 6. So if you score 6, you're typical of the interviews researchers studied, not advanced.

If you're not sure where to start, write anchors. Across 19 past-behavior studies, Taylor and Small found validity of .35 with anchored scales, against .26 without them. Those studies differed in other ways too, so treat the gap as a strong hint, not a controlled test.

10. Where should my team start this week?

Start with 1 role: a scorecard, about 6 skills and 5 anchors each. Then score every answer before anyone talks.

You don't need to change your whole process at once. A realistic first week looks like this:

• Pick the role you'll hire for next, ideally one with plenty of applicants, so the work pays back quickly.
• Write the scorecard and the 6 skills with the hiring manager in 1 sitting, so you agree on what good looks like.
• Adapt the 12 questions above and write 5 anchors for each skill.
• Calibrate on 1 recorded answer with everyone who will interview, and talk through any gap of 2 points or more.
• Run the next few candidates through it, and collect every score before the debrief.

It won't feel smooth the first time. Real answers will show you where an anchor is vague. That's why 1 role comes first: you fix those gaps once, then copy the approach.

After a few hires, compare how your top scorers perform with how your old gut calls did. That comparison, not a feeling, will tell you whether the new process is working.

How Flowmingo helps

Send each applicant a Flowmingo AI interview with your must-ask questions, and get back a scored report, free for unlimited applicants. Try Flowmingo for your next structured interview.

11. Sources

Every research number in this guide comes from 1 of these sources. The data card in section 8 is Flowmingo's own, and numbers in the examples are illustrative.

• US Office of Personnel Management (2008). Structured Interviews: A Practical Guide. OPM.
• US Office of Personnel Management (n.d.). Structured Interviews. OPM.
• Huffcutt and Arthur (1994). Hunter and Hunter (1984) revisited: Interview validity for entry-level jobs. Journal of Applied Psychology.
• Sackett, Zhang, Berry and Lievens (2022). Revisiting meta-analytic estimates of validity in personnel selection. Journal of Applied Psychology.
• Dana, Dawes and Peterson (2013). Belief in the unstructured interview: The persistence of an illusion. Judgment and Decision Making.
• Kahneman, Sibony and Sunstein (2021). Noise: A Flaw in Human Judgment, as quoted by NPR Planet Money. Little, Brown Spark.
• Kahneman (2002). Daniel Kahneman: Biographical. Les Prix Nobel 2002, Nobel Foundation.
• Campion, Palmer and Campion (1997). A review of structure in the selection interview. Personnel Psychology.
• Levashina, Hartwell, Morgeson and Campion (2014). The structured employment interview: Narrative and quantitative review of the research literature. Personnel Psychology.
• Smart and Street (2008). Who: The A Method for Hiring. Ballantine Books.
• Kahneman (2011). Thinking, Fast and Slow, chapter 21, as quoted by Ed Batista. Farrar, Straus and Giroux.
• Thorsteinson (2018). A meta-analysis of interview length on reliability and validity. Journal of Occupational and Organizational Psychology.
• Shaper (2017). How many interviews does it take to hire a Googler. re:Work, Google.
• The re:Work Editorial Team (2026). A guide to structured interviewing for better hiring practices. re:Work, Google.
• Gartner (2025). Gartner survey shows just 26% of job applicants trust AI will fairly evaluate them. Gartner newsroom.
• Taylor and Small (2002). Asking applicants what they would do versus what they did do. Journal of Occupational and Organizational Psychology.
• Bock (2015). Work Rules!, excerpt in WIRED. Hachette Book Group.
• Nesterak (2021). A Conversation with Daniel Kahneman About Noise. Behavioral Scientist.
• Bohnet (2016). How to Take the Bias Out of Interviews. Harvard Business Review.
• Bohnet, van Geen and Bazerman (2016). When Performance Trumps Gender Bias: Joint vs. Separate Evaluation. Management Science.
• Berry, Lievens, Zhang and Sackett (2024). Insights from an updated personnel selection meta-analytic matrix. Journal of Applied Psychology.
• Williamson, Campion, Malos, Roehling and Campion (1997). Employment interview on trial: Linking interview structure with litigation outcomes. Journal of Applied Psychology.
• EEOC, Civil Service Commission, Department of Labor and Department of Justice (1978). Uniform Guidelines on Employee Selection Procedures, 29 CFR Part 1607. eCFR.
• Stockdale, Hickman and Liu (2026). Scoring employment interviews with large language models. Journal of Applied Psychology.
• European Parliament and Council (2024). Regulation (EU) 2024/1689, the AI Act. Official Journal of the European Union.
• European Commission (2025). Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act). European Commission.
• European Parliament and Council (2026). Regulation (EU) 2026/1744, the AI Omnibus. Official Journal of the European Union.
• Office of the New York State Comptroller (2025). Enforcement of Local Law 144: Automated Employment Decision Tools. OSC.
• NYC Department of Consumer and Worker Protection (2023). Notice of Adoption of Final Rule: Use of Automated Employment Decision Tools, 6 RCNY §5-300. DCWP.
• Illinois General Assembly (2019). Artificial Intelligence Video Interview Act, 820 ILCS 42. ILGA.
• California Civil Rights Council (2025). Final Text of Proposed Employment Regulations Regarding Automated-Decision Systems. Civil Rights Department.