Building a certified English assessment with Cambridge
Two companies with completely different ideas of what "designing a test" means, one shared assessment, and a homemade feedback engine built out of a Google Sheet and a grudge against VLOOKUP.
The situation
Babbel's English learners had been asking for a way to prove their level for years. When Cambridge English Assessment approached Babbel about building an official digital English test together, the timing was right — our biggest competitor had recently shipped its own assessment, but without formal certification behind it. I co-led the project with a PM from New Business Initiatives and built the MVP from scratch.
Two companies, two very different instincts
At our first workshop, Babbel showed up with sleek slides and Post-its for ideation. Cambridge brought binders of photocopied exam papers and directories of rigorous fairness guidelines. We'd pictured the test running like a standard Babbel lesson; they'd pictured a downloadable, scannable PDF. Rather than push straight into design, we spent the early days learning from each other — Cambridge taught us how much precision goes into writing something as deceptively simple as a fill-in-the-gap question; we shared what we knew about how our users actually behaved on the platform.
A follow-up workshop in Cambridge combined both approaches into something simple, accurate, and technically buildable: test-takers start with A2-level questions and are routed to harder or easier questions based on performance, with no visible gaps in the experience.
The UX questions nobody expects from an exam
A few things about assessments turned out to be genuine UX problems in disguise:
- No two exams alike. A large tagged question bank — by skill and by topic — meant near-zero chance of two users seeing the same test in the same order, while still guaranteeing fair topic and skill coverage.
- Moderation without a room full of invigilators. We disabled screenshots, restricted tab-switching, and used webcam stills at intervals to flag when someone left their seat or had help — the digital equivalent of moderators walking a school exam hall.
- The certificate is the product. Research showed the actual test mattered less to users than the proof at the end. We iterated the certificate design until it mirrored a school qualification — something people told us they'd feel "proud" to share on social media or use for work and study.
- Once wasn't enough. Around 20% of test-takers returned within two months to retake it after working on their weak areas — evidence the test had built real intrinsic motivation on top of its original extrinsic purpose.
Writing content that has to measure someone
The thing I underestimated going in was how different exam content design is from product content design. Both are about precision, but they optimise for opposite things — and Cambridge's binders of fairness guidelines existed because every one of these traps has caught someone before.
Product copy tries to help the reader succeed. Exam copy has to give every reader exactly the same chance of succeeding, and then get out of the way. That flips several instincts:
The question must test one thing, and only that thing
A gap-fill designed to test the present perfect fails if the sentence around it contains vocabulary above the target level — because now you're measuring vocabulary and grammar at once, and a wrong answer tells you nothing about which. Every word surrounding the gap has to sit comfortably below the level being assessed. It's the deceptively simple part Cambridge spent an afternoon teaching us, and it's why writing a good fill-in-the-gap question takes far longer than writing the sentence.
Wrong answers have to be wrong for a reason
In a multiple-choice item, the incorrect options — distractors — aren't filler. Each should represent a plausible misunderstanding a learner at that level genuinely has, so the pattern of wrong answers is itself diagnostic. Obviously-silly options inflate scores by elimination and teach you nothing about the gap. This is what made per-skill, per-topic feedback possible at the end: the wrong answers carried information.
Assume nothing about who's reading
Test content can't rely on cultural knowledge, local idiom, or context a reader might not share — not because it's unkind, but because it makes the test measure something other than English. A question that assumes familiarity with, say, a British supermarket or an American school system disadvantages a fluent speaker who's simply never lived there. Fairness guidelines exist to stop a language test quietly becoming a culture test.
Instructions are part of the assessment, whether you want them to be or not
If a test-taker gets an item wrong because they misread the task, you've measured their instruction comprehension instead of their English. So rubrics have to be shorter and simpler than anything being tested, phrased identically every time, and ideally understood before the reader hits the content. This is the point where product content design and exam design agree completely — consistency beats elegance.
The level has to be a real level
CEFR bands aren't vibes. A2, B1, B2 and so on come with defined descriptors of what a learner can actually do, and every item has to be anchored to one of them. That anchoring is what let the adaptive routing work — starting everyone at A2 and moving up or down only means something if each question is reliably pegged. It's also what made the certificate credible, and the certificate was the product.
The MVP that ran on a spreadsheet
Positive test feedback created a new problem: support couldn't share which specific answers users got wrong, and engineering didn't have capacity to build proper feedback tooling yet. Manually reviewing each test-taker's answers took 2–3 hours per person — not sustainable.
Every question was already tagged by skill and topic in a Google Sheet, so I built a VLOOKUP-driven system that took an uploaded answer key and generated a full breakdown — "strong" (90%+), "fair" (70–89%), or "improve" (under 70%) — per skill and topic, each level paired with written feedback and a matched Babbel course recommendation.
The result: a personalised feedback profile with tailored course recommendations, generated in about two minutes instead of hours. It also opened a natural upsell — a discounted second attempt for users who'd followed their recommendations.