What does an interview score measure, and what can it not?

Updated: 16 September 2026

In short

An interview score measures how well the text of a spoken answer fits a rubric: whether it contains something concrete, has a visible shape, and suits the job and the person who asked. It does not measure whether the story is true, how you sound, who else is applying, or what the employer will actually decide. It is a language model's judgement of an automatic transcript, which makes it a useful mirror for your answers and not a forecast for the job.

Practise the interview out loud, and see how you scored

An AI panel asks the questions and you answer by voice, the way the real call goes. Afterwards you get a pass probability and the moments that cost you.

Start a live practice interview

This page is about the limits. How the score in the SwissJobs.app practice interview is built is covered on "How is an interview answer scored?", and how to read the number on the page about a good score. Here the subject is the chain between the sentence you speak and the number you see, and the points along that chain where something stays invisible.

The short version: check the transcript first, then the reasoning, and only then the number. Compare attempts only within the same setup. And treat every score as a statement about one answer, never as a statement about you.

What gets scored is a transcript, not your voice. Whatever speech recognition mishears, the model scores as misheard.
The score is a language model's judgement: reasoned, but not exactly repeatable. A one-point wobble between runs is noise.
Plausible is not true: an invented example scores the same as a real one. Only a reference check can tell them apart.
The end-of-session verdict reads only the opening of each answer. A result that arrives after several minutes never reaches it.
Out of sight: other candidates, budget, salary band, permit, the chemistry in the room and the employer's real criteria.
Used properly, the score is a comparison tool: same question, one change, new score.

What the score actually measures

In the practice interview you answer a question out loud, and straight afterwards a language model assesses the text of your answer. It knows the position, the company size and the role of the panel member who asked, and it returns a score from zero to ten with a short justification and a note on what a strong answer would have added. The rubric behind it asks about substance, structure and fit.

That is a narrow definition of quality, but an honest one. It covers the part of an interview you can prepare: whether a question about a difficult project gets a real project in reply, whether the answer has a beginning and an end, whether it connects to the advert. Swiss recruiters look hard at exactly this. The register is sober and evidence-led, and self-promotion that plays well in a London or New York interview tends to land flat in Zurich or Lausanne.

One detail worth knowing: the overall score is not calculated from the three sub-scores. The model gives it as a judgement of its own, and the report shows that overall score with its reasoning. An answer with strong substance and weak fit can therefore land above or below the average of its parts, and the reasoning tells you which part carried the weight.

The chain before the number: voice, transcript, judgement

Two automatic steps sit between your spoken sentence and the score. First, speech recognition turns your voice into text. Then the language model reads that text. The model never hears you. If recognition drops a word, mangles a technical term or cuts a sentence short, the mangled version is what gets scored.

This does not hit everyone equally. Technical terms, company names and acronyms get lost more often than everyday words. Strong accents cost accuracy, and so does answering in a language you are still building. If you are moving to Switzerland and practising in German or French rather than English, a low substance score may simply mean your example never made it into the text, even though you said it.

The practical rule is simple: in the report, read the transcript of your answer before you read the score. If it does not say what you said, the score for that answer is worth little. Try the answer again a little slower and clearer rather than rewriting its content. It does no harm for the real interview either: a panel across the table also follows calmly spoken technical terms better.

A judgement, not a ruler

A score from a language model is a reasoned assessment, closer to an experienced reviewer working from a checklist than to a measurement. It is not exactly repeatable. The model is set to vary its wording slightly so the explanations read naturally, which means the same answer scored twice can come back a point apart.

That is not a defect so much as a property of any judgement, human ones included. But it does mean that moving from six to seven after a small rewording does not yet prove the change worked. Moving from four to eight, with reasoning that explicitly praises the new example, does. Watch for large movements and read the justification. Single points tell you little.

There is also a length limit. Scoring for each individual answer reads long transcripts only up to a ceiling of roughly eight thousand characters, which almost no real answer reaches. The verdict at the end of the session is tighter: it reads only the first two thousand characters or so of each answer, which is about two minutes of speech. Anything after that does not feed into it.

Plausible is not the same as true

The rubric rewards answers that sound checkable: a specific project, a figure, an outcome. The model can check none of it. It does not know whether the project existed, whether lead time really fell by a third, or whether you actually led that team. An invented example with a clean structure scores exactly as well as a real one.

In a real Swiss hiring process that difference has consequences. An application dossier normally includes Arbeitszeugnisse (the written references Swiss employers issue when someone leaves), and many employers call referees before making an offer, often the very manager your example mentions. A panel also follows up: how big was the team, what did your manager say, what would you do differently now? An invented story survives the first question and rarely the third.

So use the score to tell real experience better, not to invent experience. A high score is a verdict on how the story was told. Whether the content holds up gets decided later, in places no practice tool can see.

A worked example: good score, buried result

Say the operations lead on the panel of a mid-sized logistics firm asks: "How did you handle it when returns started rising?" The answer opens with background, explains the old system, the departments involved and two dead ends, and after a little over three minutes gets to the point: returns fell clearly within a quarter because the product descriptions were rewritten. The individual scoring reads the whole transcript and gives a seven, noting the answer is long but supported.

The end-of-session verdict sees something else. It reads only about the first two minutes of each answer, and those two minutes hold context but no result. The strength of this answer does not appear in the list of strengths, and the concerns may well say the answers ramble. That is less a flaw in the scoring than a fairly accurate picture of what a tired panel remembers at the end of a long afternoon.

The fix is a reordering, not new content: result first, then how you got there. "We brought the return rate down clearly within a quarter, through the product descriptions. Here is how we found that." The same story, told in about ninety seconds, keeps its seven or better in the individual score and arrives whole in the final verdict. The numbers here illustrate the mechanism; they promise no particular score.

What no score can see

The room. A real hire is a comparison against other candidates, and no practice tool knows them, nor the internal applicant, the budget freeze or the role that was informally filled weeks ago. A score, even a very high one, says nothing about whether you will be hired, and no threshold on this scale is meant as a prediction.

The hard criteria. Whether your salary expectation fits the band, whether your permit is enough, whether your notice period matches the start date and whether the workload percentage works are decided on facts that the interview at most touches on. For anyone applying from abroad, the permit question often decides more than any answer. The score rates how you talk about these topics, not whether the facts fit.

The impression. Tone, pace, pauses, eye contact and the handshake are not in a transcript. If you know you speak too fast or come across as unsure, a video recording, an honest friend or a coach will teach you more. And chemistry, whether the panel can picture you in the team, often counts for more in a small Swiss company than any single answer. A person prepares you for that part better than any rubric.

How to use the score well

Read in this order: transcript, reasoning, score. The transcript tells you whether the assessment saw your answer at all. The reasoning and the note on a strong answer tell you what to change. The score only tells you how far you still have to go.

Change one thing per attempt and keep the setup the same; the redo button carries over position, company size and panel for exactly that reason. If the score moves clearly and the reasoning names your change, it worked. If it moves by a point, that is more likely noise than progress. A missing score means "not scored", not "failed": if the assessment fails technically, the transcript is still saved.

And keep in mind what the number is good for. It shows which of your answers still lacks a checkable core and where the result comes too late. That is the part of an interview you can rehearse alone. The other part, presence, chemistry and the facts behind your answers, is measured by no score. You can try the scoring without an account, with one scored answer.

Scoring flow (transcript, per-answer score, final verdict), the per-answer reading limits and what the report displays checked in the SwissJobs.app practice interview code on 16.09.2026.

Practise an interview — live, with an AI voice and scoring

What our job index says about the Swiss market

Computed live from our own index, not quoted from a study. Shares only, as of today.

Language the advert is written in

Deutsch: 60%
English: 23%
Français: 13%
Italiano: 3%

Of adverts that state a language requirement, the share asking for

Deutsch: 70%
English: 43%
Français: 21%
Italiano: 3%

19% posted in the last 7 days · Largest markets: Zürich 18% · Bern 10% · Genève 5% · Basel 5%