How IntervYou Scores Your Interview — Our Methodology

A score is only worth what it can prove. Most interview tools hand you a number — "7.2/10," "82% confidence" — with no way to check it, no idea what moved the needle, and no answer to the only question that matters: would a real interviewer have advanced me?

IntervYou is built around that question. This page explains exactly how our scoring works: how the three-voice panel runs your interview, how it adapts to your specific role and seniority, how every score is tied back to a line you actually said, what we did to validate it against experienced human interviewers, and how we handle your audio and transcript. No black box. If you want to verify a claim here, you can — that is the whole point.

The three-voice panel: who's in the room

Real interviews are rarely one person. A loop has a recruiter screening for fit, a hiring manager probing for depth, and a future peer checking whether they'd actually want to work with you. We modeled that, because a single generic "AI interviewer" flattens those very different lenses into one bland average.

Your IntervYou panel has three distinct voices, each scoring through their own rubric:

• Layla — HR recruiter. She owns the first-impression and fit lens: your motivation for the role, your communication clarity, salary and logistics handling, and red-flag screening (badmouthing, vagueness, inconsistency). Layla is the one who decides whether you'd survive a phone screen.
• Marcus — hiring manager. He owns depth and ownership: technical or domain rigor, decision-making under ambiguity, the scope of impact you actually drove versus what your team did, and whether your seniority claims hold up. Marcus pushes back. He follows up. He is the hardest to satisfy on purpose.
• Priya — future peer. She owns collaboration and the "would I want them on my team" signal: how you handle disagreement, how you credit others, how you respond to a curveball, and whether you're someone people can be honest with. Priya catches the things that don't show up in a résumé.

Each voice scores independently against the dimensions it owns, and you see all three. When they disagree — and they often do — that disagreement is the signal. A candidate who dazzles Marcus on technical depth but worries Priya on collaboration is exactly the kind of split that derails real loops, and you find out before it costs you the offer.

How the panel adapts to your role and seniority

The same answer can be a pass at one level and a fail at another. "I shipped the feature on time" is a strong L4 answer and a weak L6 one — at L6 the panel expects you to have shaped what got shipped, navigated cross-team tradeoffs, and owned the outcome, not just the delivery.

So before your interview starts, the panel calibrates to two things:

1. The role and company. You tell us the title, the company (or company type), and the role focus — backend, product, data, design, management. The panel pulls the competencies that role is actually evaluated on. A PM interview weights product sense, prioritization, and stakeholder influence; a backend interview weights system design, tradeoff reasoning, and operational maturity. Marcus's follow-ups change accordingly.
2. The seniority bar (L1–L8). This is the part most tools skip. We map your target level to a concrete expectation of scope, autonomy, and influence. At L3 the panel wants correctness and learning velocity. At L5 it wants independent ownership of ambiguous problems. At L7 it wants you to be changing how the org operates. The questions shift, the follow-up depth shifts, and — critically — the bar each answer is graded against shifts.

This is why a generic mock interview can lie to you. Practicing senior questions at a junior bar produces inflated confidence; practicing junior questions for a senior loop leaves you blindsided by the depth of the real follow-ups. Calibrating to the real seniority bar is the single biggest reason our feedback tracks reality. If you want the bigger picture on how voice mock interviews work end to end, our complete AI mock interview guide walks through the full flow.

Every score is grounded in your transcript

Here is the rule we hold ourselves to: no score without a citation. Every number the panel gives you points to a specific line in your own transcript — the line that earned it.

When Marcus marks your system-design answer 3/5 on "tradeoff reasoning," he doesn't just assert it. He quotes you:

You said: "I'd just use a cache to make it faster." This names a solution but skips the tradeoff — what you're caching, the staleness you're accepting, and what breaks under invalidation. A stronger L5 answer states the tradeoff explicitly.

That structure — the line you said → why it scored where it did → what a stronger answer at your level sounds like — runs through the entire report. It does three things ordinary feedback can't:

• It's verifiable. You can scroll back, find the exact moment, and judge for yourself whether the critique is fair. You're never asked to trust a number on faith.
• It's actionable. "Be more specific" is useless. "At 4:12 you said we improved performance — name the metric, the baseline, and the delta" is something you can fix in the next attempt.
• It exposes patterns. When the same gap is cited across three answers — say, you keep describing team accomplishments in "we" without isolating your decision — that repetition is the real coaching, and it only surfaces because each citation is anchored to evidence.

Because the feedback is transcript-grounded rather than vibe-based, you can re-run the same interview, get a fresh transcript, and watch specific cited gaps close. That tight loop — answer, see the cited critique, fix the exact line, re-run — is what actually moves your score, and it's a big part of why people choose grounded tools over generic ones, as we cover in our roundup of the best AI mock interview tools.

How we validated the scoring against real interviewers

Transparency without validation is just a nicer-looking black box. So we tested whether the panel's verdicts match what experienced humans actually decide.

The study. We took 50 real interview transcripts spanning multiple roles and seniority levels. Each had been judged by experienced interviewers who made the call that matters most in a loop: advance or no-advance. We then had the IntervYou panel score the same transcripts cold and produce its own advance/no-advance verdict, and we measured how often the panel agreed with the humans.

The result: 78% advance/no-advance agreement. On roughly 4 out of 5 transcripts, the panel reached the same advance-or-not decision an experienced interviewer did.

We want to be honest about what that number means and doesn't mean:

• What it means. The signal is real and decision-relevant. When the panel says you'd advance, that verdict aligns with a human's most of the time — far better than a self-assessment or a friend's encouragement, both of which skew generously.
• What it doesn't mean. It is not 100%, and we don't claim it is. About 1 in 5 calls is a disagreement, often on genuine borderline candidates where two human interviewers might also split. We chose advance/no-advance as the benchmark precisely because it's the hardest, most consequential call — agreement on softer metrics would look more flattering and tell you less.
• Why we publish the method, not just the headline. 50 transcripts, real interviews, human ground truth, a binary outcome. You can interrogate the design. We'd rather show our work and report 78% than wave a bigger, unfalsifiable number.

We treat that 78% as a floor, not a trophy — it's the metric we're actively working to raise as the calibration improves.

What we do with your audio and transcript

Trust isn't only about accuracy. It's about what happens to the most personal data you can hand a product — a recording of your own voice, fumbling through hard questions. Our stance:

• Audio is deleted within 30 days. We keep your recording only long enough to generate and let you revisit your transcript and report. After 30 days the audio is removed automatically. You don't have to ask.
• No training on your data without opt-in. Your interviews are not used to train models by default. If we ever want to use anonymized data to improve the panel, that requires your explicit, separate opt-in — never buried in a checkbox you can't find. Opted out is the default and stays the default unless you choose otherwise.
• Your transcript is yours. It exists so you can study your own performance, not so anyone can resell it. We don't sell interview data.
• Try it without a credit card. Your first three interviews are free, no card required — so you can evaluate the methodology on your own answers before deciding anything.

We hold this line because we'd want it ourselves. Practicing your weakest moments only works if you trust those moments aren't going anywhere you didn't agree to.

Putting it together: what your report actually shows

When you finish an interview, here's the full picture you get — and how each piece traces back to the methodology above:

1. Three voice verdicts (Layla, Marcus, Priya), each scored against the dimensions that voice owns, calibrated to your role and target level.
2. An overall advance / no-advance call — the same binary we validated at 78% agreement with experienced interviewers.
3. Per-dimension scores, each citing the exact line that earned it, with a stronger-answer rewrite pitched at your seniority bar.
4. Cross-answer patterns — the recurring gaps that matter more than any single moment.
5. A clear next step — the one or two cited fixes most likely to flip a no-advance into an advance.

Nothing in that report is unfalsifiable. Every score points at evidence you can check. That's the standard we built toward, and it's the standard we think any tool asking you to trust its feedback should meet.

Frequently asked questions

Is AI interview feedback actually accurate?

It depends entirely on the methodology. IntervYou's panel reached 78% advance/no-advance agreement with experienced human interviewers across 50 real transcripts — the same call a real loop makes. More importantly, every score cites the specific line in your transcript that earned it, so you can verify the feedback yourself rather than trusting a number. Accuracy you can't check isn't accuracy; it's a guess with a confident font.

How does the panel know what's a "good" answer for my level?

Before your interview, you set your target seniority (L1–L8) and role. The panel grades each answer against the scope, autonomy, and influence expected at that level — so the same answer scores differently for an L4 versus an L6. The follow-up questions, their depth, and the bar each answer is measured against all shift to match the real seniority bar, which is exactly what generic mock interviews miss.

What happens to my interview recording and transcript?

Your audio is deleted automatically within 30 days, after which only you have access to your transcript and report in the meantime. We do not train models on your data without your explicit, separate opt-in — opted out is the default. Your transcript exists for your own study, and we don't sell interview data. Your first three interviews are free with no credit card.

The best way to judge a scoring methodology is to run it on your own answers and check the citations yourself. Start a free mock interview — three on us, no card — and see exactly which lines moved your score, and whether you'd advance.