How our AI scoring works

Why we publish this

A score prediction is only useful if you can trust it, and you can only trust a number if you can see how it was made. Most PTE tools hand you a band and leave you guessing. We think that's backwards, so this page walks through exactly how our engine scores your mock: what it measures, what's real AI versus an estimate, and where the limits are. PTE Academic is itself a computer-scored exam, so an AI-scored practice test is the right model, as long as it's calibrated and disclosed honestly. That's the whole point of this page.

What PTE Academic actually measures

The real exam reports a single overall score and four communicative skills (Speaking, Writing, Reading and Listening), all on the 10–90 Global Scale of English. Crucially, scoring is integrated: one task can feed several skills at once. Repeat Sentence feeds Listening and Speaking; Write from Dictation feeds Listening and Writing. Underneath the four skills sit the enabling skills the automated marker watches:

• Speaking: content, oral fluency and pronunciation.
• Writing: content, form, grammar, vocabulary, spelling and development, structure and coherence.
• Reading & Listening: comprehension, judged mostly by objective right/wrong items with partial credit.

Our engine is built to assess these same dimensions, not a vague “good/bad” but the specific traits Pearson's scorer rewards.

Reading & Listening: exact marking

For the objective task types (multiple choice, Fill in the Blanks, Re-order Paragraphs, Highlight Incorrect Words and the rest) there's no AI guesswork. Your answer is checked against the key with the same partial-credit and negative-marking rules as the real exam: every correct element earns a point, and on the multiple-answer questions a wrong selection costs one. This is deterministic and exact, so the same input always produces the same score.

Speaking & Writing: real AI scoring

The open-ended tasks are where the AI does its work, in two steps:

• 1. We capture what you actually produced. Your spoken answers are recorded and run through speech-to-text, so we score the words you really said (and how you said them), not a script. Your written answers are taken exactly as typed.
• 2. A state-of-the-art language model grades against a fixed rubric. It evaluates each response on the trait dimensions above (content and relevance, fluency, grammatical range and accuracy, vocabulary, spelling and discourse) and returns a structured result: an overall and per-skill score on the 10–90 scale, the six enabling skills, a per-task score with a one-line reason, and your three biggest “score leaks”.

Because the model reads the actual content, the feedback is specific. It can tell you that an essay lost marks on form (under the word count) rather than ideas, or that a Describe Image answer was fluent but thin on data: the kind of “why” most tools never give you.

Mapping to the official 10–90 scale

Every trait score is mapped onto the 10–90 band the real PTE uses, never a 0 and never a percentage. The overall is our own estimate. Pearson does not publish how it combines item scores, so we average the band you earned on every question in the test rather than averaging the four skill scores, and we label the result a directional estimate, not a Pearson calculation. Each band also maps to a CEFR level so you can see where you sit (B1, B2, C1…), and the report frames the gap to your target. For example: “you're 6 points off 79, and Writing is the skill under the line.”

What’s real AI scoring vs an honest estimate

This is the part we refuse to fudge:

• A full mock = real AI scoring of every open answer. Your speaking and writing are genuinely assessed by the model, your objective answers exactly marked. The predicted score reflects what you actually did.
• The 10-minute diagnostic uses a labelled estimate for Speaking & Writing. To stay fast, the quick diagnostic only directly tests Reading and Listening (objective items). Your Speaking and Writing are then derived from your measured Reading and Listening and shown as estimates, clearly marked as such. Because they track your real performance, skipping the test makes them fall too; there are no flattering numbers.

Wherever a number is an estimate, we say so on the report. AI scoring is a high-quality preview, not the official Pearson result.

How we keep the bands realistic

An AI score is only as good as its calibration. We tune ours against publicly available official scoring descriptors and sample responses, so a band feels like a fair preview rather than a flattering one. A practice 65 should mean roughly what a real 65 means, and that's the entire value of the exercise. If anything, we aim to be slightly conservative: it's far better to under-promise and have you walk into test day over-prepared.

Why your score isn’t shown on screen

Just like the real exam, your score doesn't flash up the second you finish. Instead we email the full breakdown: your predicted overall, per-skill detail, your three biggest leaks and a dated plan to your target. Partly that mirrors test-day conditions; mostly it's because the report is the useful bit, and a considered breakdown beats a bare number you'd stare at and learn nothing from.

The honest limits

No practice scorer is a perfect twin of Pearson's, and pretending otherwise would be the opposite of trustworthy. Three honest caveats:

• Pronunciation & fluency are assessed from a transcript, which can't fully capture acoustics (stress, rhythm, individual phonemes) the way Pearson's acoustic models do. We score these conservatively.
• It's a prediction, not the official result. Use it to find and fix weaknesses and to gauge readiness, not as a guaranteed test-day number.
• We're independent. PTE Mocks is not affiliated with, endorsed by, or licensed by Pearson; our material is modelled on the published format, not copied from the real exam.

Think a score is off? Tell us

If a score ever looks wrong, that's genuinely useful to us, because spotting our errors makes the engine better for everyone. Our full methodology, sourcing and 72-hour correction policy live in our editorial standards, and you can always reach us at hello@ptemocks.com.

2026 scoring engine update: Haiku primary, Sonnet fallback, floor-clamp defence

Two changes we shipped in July 2026 are worth pulling out, because they change what happens on your worst-case report, not your average one.

1. Two-tier model cascade with retry and backoff (15 July). The engine no longer relies on a single model call. Every open-answer task is graded by a fast primary model (Claude Haiku), and if that call fails, times out, or returns a malformed rubric (missing traits, out-of-range band, unparseable JSON) the request is retried with exponential backoff. If the primary still won't settle, the task is routed to a slower but stronger fallback (Claude Sonnet) that grades against the same rubric. In practice that means a flaky network hop or a one-off model hiccup no longer nukes a whole report; you get a graded task instead of an apologetic gap.

2. Post-parse floor-clamp on unattempted tasks (18 July). Language models are agreeable by nature and will occasionally hand a small pity score to a blank Describe Image or a two-second Read Aloud. That's the wrong signal for a mock. After parsing the model's reply, the engine now checks whether the underlying response was actually attempted (audio duration, transcript length, typed characters against a per-task minimum) and clamps the score to 10 when it wasn't. No task the user skipped can quietly contribute non-floor credit to the overall.

Together, the cascade stops the engine falling silent on you and the floor-clamp stops it flattering you. Both are the same idea from opposite ends: the report should reflect what you did, including the parts you didn't.

Quick reference: how our AI scoring works at a glance

The short version of everything above, in one scannable table.