This page explains how the scoring works, how good it is and where it gets things wrong. We update it whenever something changes (as of October 2026). GoetheCoach is not affiliated with the Goethe-Institut; a practice score is not an exam grade.
How does GoetheCoach score a text?
A language model reads your text and reports what it sees; code turns that into points. The model does not hand out points by feel.
- Read: the model checks every Leitpunkt (covered, partly, missing) and quotes the sentence that covers it. It quotes every mistake and rates Kohärenz, Wortschatz and Strukturen in bands.
- Calculate: code turns those observations into points by fixed rules. Words are counted by code, not guessed by the model.
- Read twice: every text is read twice independently and the results are averaged. If one reading is below 60, one above, and the average below, a third reading is added.
That is the same logic as the exam: two examiners score independently, and a third rating only happens when pass or fail is at stake. The real exam rules are in how Goethe writing is scored.
Does GoetheCoach use the real Goethe criteria and points?
Yes, the four criteria are the same: Aufgabenerfüllung, Kohärenz, Wortschatz, Strukturen. For B1 and B2 we also use the point values from the official rating forms.
| Level | Point values | Source |
|---|---|---|
| B1 | Teil 1 and 2: 4 × 10 points each, Teil 3: 4 / 4 / 6 / 6 | rating form in the B1 Übungssatz |
| B2 | Teil 1: 14 / 14 / 16 / 16, Teil 2: 4 × 10 | rating form in the B2 Modellsatz |
| A1, A2, C1, C2 | four equal criteria | our assumption, not the official forms |
Each criterion only gets one of the official band values (A to E = 100, 75, 50, 25 or 0 % of its maximum). There are no values in between, just like the exam.
What this means: for B1 and B2 our calculation is as close to the exam as it can be without examiners. For the other levels it is an approximation.
When does a task get 0 points?
As in the exam, only at band E for Aufgabenerfüllung. That happens when your own text has less than half the required words, misses the topic, is blank, or is mostly copied from the task.
Missing Leitpunkte alone never give 0 points. They cost bands, but the task is still scored.
What do "safe", "borderline" and "not yet" mean?
Instead of only "pass" or "fail" we show three zones:
| Zone | Score | What it means |
|---|---|---|
| Safe | 70 % and up | You pass with a margin. |
| Borderline | 55–69 % | It could go either way. We show you what is missing. |
| Not yet | below 55 % | Something important is still missing. |
Why not simply "pass from 60"? Because between 60 and 69 % even trained raters disagree (more below). A "pass" at 62 % would be false comfort. Why 70 is the better target is explained in 60% on a practice test: is it enough?
How closely does it match real examiners?
We measured our scoring on 124 real exam texts that trained raters had placed on a level (MERLIN corpus, telc tasks A1–C1). For comparison, a plain chatbot prompt ("grade it like a Goethe examiner") ran on the same language model.
| plain chatbot prompt | GoetheCoach before (rubric 1.3) | GoetheCoach now (from 1.4) | |
|---|---|---|---|
| verdict matches the raters | about 75 % | about 75 % | about 75 % |
| below-level text called "safe" | – | 18 % of gradings | 9 % |
| at-level text called "not yet" | – | 7 % | 13 % |
| typical mistake | too strict | too lenient | on the cautious side |
Three things we take from this:
- On the bare yes/no we are not more accurate than a good prompt. Both get about three in four texts right.
- We were too lenient. With rubric 1.3, 43 % of the texts the raters placed below the level passed. We fixed that with the official point values and two readings.
- We would rather err on the cautious side. A false "not yet" costs you one practice round; a false "safe" can cost you an exam.
Limits: telc rather than Goethe tasks, CEFR levels rather than Goethe points, and we chose the cut-offs on the same texts. The numbers are an honest guide, not a seal of quality.
Where does GoetheCoach get it wrong?
Every scoring system has weak spots, ours too. The main ones:
- C1 and C2: the verdict varies most here. That is not only an AI problem; research also finds the top levels hardest to score.
- One paragraph off topic: if an otherwise fitting text has a paragraph that does not belong to the task, we often miss it.
- No examiner comparison yet: no certified Goethe examiner has rated texts for us so far. That is why we do not claim to be "as accurate as an examiner".
- Writing only: we do not score Sprechen.
- Repeats: if you resubmit the same text, you get the same result. Two different texts of similar quality can still land in different bands.
What this means: use the score to see what is missing. Do not read the exact number more precisely than ±5 points.
Does the scoring change over time?
Yes, and every change gets a version number. Before we change the language model or the rules, we measure the effect on fixed test texts.
| Version | Date | What changed |
|---|---|---|
| 1.3 | October 2026 | Missing Leitpunkte, mistakes and short texts cost more points. |
| 1.4 | October 2026 | Official point values for B1 and B2; two readings, averaged; third reading by the Goethe rule. |
| 1.5 | October 2026 | The official E-rule: 0 points only below half the words, off topic, blank or copied. |
One example of such a check: when we switched language models in October 2026, 9 of 10 test runs gave the same points. With a different setting, A2 and B1 texts would have scored 10–15 points lower, so we did not use that setting.
Frequently asked questions
Is GoetheCoach more accurate than ChatGPT?
On the bare pass/fail verdict, not automatically: in our comparison, we and a good chatbot prompt both agreed with trained raters about 75 % of the time. The difference is fixed rules, counted words, quoted Leitpunkte and a borderline zone instead of a narrow "pass".
Can I rely on my score?
As a direction, yes; as an exact number, no. Allow ±5 points. From 70 % you are safe; in the borderline zone between 55 and 69 % you should work on the gaps.
Does the scoring replace a teacher?
No. It quickly shows which Leitpunkte are missing and where the mistakes are. A teacher sees more, especially style and argument. That is why the "Prüfungsvorbereitung" plan includes a teacher check.
Does GoetheCoach score Sprechen?
No, only writing. For speaking you need a partner or a different tool.
Why don't I get "pass" at 62 %?
Because 62 % on a practice test is not a safe pass. Between 60 and 69 %, even trained raters in our comparison split about half and half. So we show "borderline" there and tell you what is missing.
Why is scoring different for A1, A2, C1 and C2?
For these levels we do not use the official rating forms but four equal criteria. The score is an approximation there, and least certain at C1 and C2.
Sources
- Goethe-Institut: Durchführungsbestimmungen B1 and B2, § 4.3 (as of 1 September 2025)
- Goethe-Institut: B1 Übungssatz and B2 Modellsatz for adults, Prüferblätter (rating forms)
- MERLIN Written Learner Corpus (Wisniewski et al., Eurac Research, CC BY-SA 4.0)
- ChatGPT vs GoetheCoach: one B2 text, graded five times
How many points would your text get?
Write a task and see the points per criterion – by the same rules this article explains.
Score my text