How we measure what AI engines tell applicants
What we ask, where we ask it, how we read the answers, and what this method cannot tell you.
1. The questions
questions in the question bank
engines: ChatGPT, Gemini and Perplexity
opinion topics each answer is read for
We ask what applicants ask: short, informal questions, for the subjects a university teaches and the countries it recruits in.
- Discovery questions name no school ("best business schools in Belgium"), so the engine chooses whom to recommend. Every university is scored on the same discovery questions, so the comparison with its competitors is fair.
- Questions about a university by name, and the same questions about each of its named competitors.
- Questions asked from an applicant's country, for the markets a university recruits in.
The question bank is versioned, and we only compare answers to the same questions over time.
2. The engines
ChatGPT, Gemini and Perplexity, each with live web search switched on, as an applicant uses them. Each engine is measured on its own, because each one reads different websites. Questions run every week.
3. The reading
For every answer we record which universities are named and in what order, every website it cites, every opinion it gives about a university (on which topic, positive, neutral or negative) and the facts it states, such as fees and deadlines.
Every figure in the Meridian Report traces back to the sentences behind it. Quotes are copied word for word; nothing paraphrased is ever put in quotation marks. University names are matched by fixed rules.
4. The statistics
Every share comes with a confidence interval, and differences from competitors are tested for significance. Week-on-week changes are shown next to the measured noise between runs, so a small wobble is never presented as a trend. Where a sample is too small, the report says so.
We report what goes together, not what causes what: a website that appears in answers naming a university is associated with those answers. We do not claim it made the engine name the university.
5. The limits
- We ask from clean sessions. A logged-in applicant with a personal history can see a different answer.
- Engines change their models and search without notice. We record the model behind every answer and compare like with like.
- Answers vary from one asking to the next. That is why we use many questions and report only what holds up across them.
- The engines decide what they recommend. We show what they read and where a university is missing; we cannot promise it will be named more often.
Questions about the method: hello@meridianstanding.com.