How the test finds the correct level
A fixed test spends most of its questions on things that are far too easy or far too hard for the person answering. NGPT spends every question near the candidate's own level, which is why 18 to 30 are enough.
What the candidate sees
A link in an e-mail, then one question on the screen at a time. No account, no app, no instructions to study first.
- One question on the screen
- Usually a sentence with a gap: the candidate chooses the missing word from three or four options, or types it. Typed answers are compared sensibly: stray spaces and a final full stop do not turn a right answer into a wrong one.
- “I don’t know” is an answer
- A candidate who has no idea can say so. That tells the test more than a lucky guess would, and it keeps beginners from sitting through questions they cannot read.
- It survives real life
- If the phone rings or the connection drops, the same link opens the test again at the question where it stopped.
- It looks like your school
- Your logo, your colour and a few words of welcome from you are on every page, from the invitation to the result.
If I had more time, I’d learn Italian too.
- have
- had
- would have
- am having
Powered by NGPT
After every answer, the test thinks again
It starts in the middle
Before the first answer the candidate could be anywhere from A1 to C2, so the first question comes from the middle of the scale.
It asks where the answer will matter the most
A question far below the candidate's level will be answered correctly, one far above it wrongly; neither says anything new. The test picks at random among the questions that fit its current estimate best.
It adjusts and narrows
A right answer moves the estimate up, a wrong one down, and either way the range of levels still possible gets smaller. Lucky guesses are allowed for: on a four-option question the test knows one candidate in four gets it right by chance.
It stops when one level is clear
Never before 18 scored questions, never after 30. In between, it ends as soon as one level is clearly more likely than its neighbours. Up to 3 new questions that are being tried out are mixed in and do not count.
One candidate, 22 questions. The line is the estimate after each answer; the shaded band is what is still possible. Here the test stopped once the band lay inside B2.
Difficulty is measured, not assumed
Ask ten teachers how hard a question is and you get ten answers. When we compared the level teachers had given our questions with how candidates really answered them, the label turned out to say surprisingly little: questions marked with the same level differed widely in real difficulty.
So NGPT does not trust the label. The difficulty of a question is worked out from the answers of real candidates, more than 190,000 of them. New questions start from an estimate made in review and are measured as candidates answer them. The test then knows two things a paper cannot: how hard a question really is, and how much a right or wrong answer to it says about the person in front of it.
Two candidates rarely see the same test. Questions are drawn from a bank of more than 7,900, each is picked at random among the best fits for that moment, and no question comes twice in a test.
The life of a question
The bank began with questions written by teachers at the partner schools. As of 2026, all questions are monitored and moderated by advanced AI agents, removing all subjectivity from the process and ensuring a true machine-learning approach.
Stage 1
Written
Based on CEFR descriptions of what a learner at each level can do. Never copied from an exam board.
Stage 2
Answered blind
A second, independent AI reviewer answers the question without seeing the key. If the two disagree, or another answer would also be correct, the question goes back.
Stage 3
Given a starting difficulty
The reviewers rate the level, and the rating is converted using questions whose real difficulty is already known.
Stage 4
Tried out
Where a level has enough measured questions, new ones are first mixed into real tests without counting towards the result.
Stage 5
Measured and watched
As candidates answer, measurement replaces the estimate. A question that behaves oddly is taken out.
What it measures, and what it leaves to you
A placement test has one job: to put a new student in a group where they can follow and still have something to learn. It is not an exam and gives no certificate.
Grammar and vocabulary in use
The core of the test and the basis of the level.
Measured
Reading and listening
One short text and one recording at the candidate's level, from A2 up, each reported next to the level as a score out of 5.
Measured
Writing
A short written task, rated against the CEFR criteria, is being added.
Coming
Speaking
Not tested. A few minutes of conversation on the first day confirms the group.
Yours
How sure is the result?
We test the test. Thousands of simulated candidates with the spread of ability of our real ones took it on the real question bank, and we counted how often the result was the candidate’s true level or the one next to it.
No short test puts everyone on the right side of every line: plenty of real students sit on the boundary between two levels. That is why NGPT never gives a bare label. The score and its margin show how close to the boundary a candidate is, so that you can decide with your eyes open.
| Scored questions | On the correct level, or the one next to it |
|---|---|
| 15 | 96% |
| 20 | 98% |
| 25 | 99% |
| 30 | 99% |
Simulation of 3,000 candidates on the English question bank, eight levels from A1 to C2. A simulation assumes the measured difficulties are right. The other four languages have far fewer answers behind them, so their figures will be worked out as tests are taken.
The quickest way to judge a placement test is to give it to people whose level you already know.
Read on
Place your next intake with NGPT
Tell us about your school and we add 10 free tests to your account. Use them on real candidates and see the results for yourself.