One Applicant Profile, Seven Days: How SMU’s Finance Master’s Overtook NTU in AI Answers

The question this study was built to answer

Can content written for one specific buyer — one persona in one situation — change what AI models recommend to that buyer, and only that buyer? If it can, AI visibility is something a brand can aim. If it cannot, it is weather.

We ran the test on a live programme in a contested category: Singapore Management University’s Master of Applied Finance, against the four programmes applicants actually weigh it against.

Step 1: five personas from SMU’s own material

Brand name, core keywords, programme description and competitors went into JefurryAxis, which decomposed them into tags and clustered five applicant personas. Not demographic bands — decision profiles, each with a motive, a constraint and a timeline.

Step 2: persona × situation → brand-free questions

Each persona was crossed with two situations — E1, deadlines closing with the recruiting window next; E2, scholarship and return-on-investment pressure. Behaviour follows from both: B = f(P, E). From the target cell, B₃₁ = f(P3, E1), came questions like these, none of which names SMU:

• “Deadlines are close. Rank the finance master’s programmes in Singapore, Australia and Hong Kong by front-office banking placement, alumni strength and post-study work rights — which should I apply to first?”
• “Compare internship conversion, alumni access and graduate work rights for someone targeting the next investment-banking cycle.”
• “Rank them by scholarship odds, total cost and investment-banking ROI, and flag the fit for a GMAT 720 profile.”

Step 3: the baseline, 28 January 2026

Each question went separately to Doubao, DeepSeek, Qwen and Yuanbao, and each was asked more than once — a model does not answer the same question the same way twice, so one screenshot is not a measurement. The baseline Visibility Score was 69.1:

• Mention rate 76.7% — named in about three answers in four.
• Occurrence frequency 2.4 — named, rarely discussed.
• Average rank 4.5 — mid-list.
• Brand content ratio 0.152 — about 15% of each answer.

On the target persona, NTU was ahead of SMU.

Step 4: ten articles, written for one person

We did not touch SMU’s website. The baseline run showed which third-party platforms these models actually cite, so that is where the content went: Toutiao, Douyin, Sohu, Bilibili and Zhihu, one article on each on 29 January and another on 30 January. Every piece answered the target persona’s decision in that situation — which programmes to prioritise, how to choose if you can only apply to two, what actually decides a front-office outcome — using programme facts that could be sourced.

What changed

Same questions, same models, same scoring on every date. The only variable introduced was content, placed where the baseline showed the models read.

The finding that matters most: precision

Our articles were cited as sources only under the Shanghai IB Aspirant’s questions. Under the other four personas’ questions, none of them was cited.

That is the point of the study. If the content had been cited everywhere, the honest conclusion would be that we had made SMU generally more findable — useful, but not proof of anything. Instead, the models used it for the person it was written for and nobody else. The targeting held, which is what the persona-driven model predicts. It also means the other four personas are separate work: there is no spillover to wait for.

The research behind it

Limits of what this shows

• Visibility, not enrolment. This measures what four AI models said in response to a fixed question set. Applications, enquiries and enrolments are tracked in the university’s own systems.
• Seven days. Long enough to see a clean effect; too short to call a trend. Citations decay when content is not maintained.
• Chinese models only. ChatGPT, Claude and Gemini read different sources and need their own run.
• Answers vary. The same model words the same answer differently each time. Repeated sampling keeps that variation inside a known margin; it does not remove it.

Questions about this case

How a run is built and scored · The RedNote case: a closed platform, from absent to cited