A prospective student opens a university website and types a question into the search bar: “What happens if I can’t pay my tuition by the deadline?” The site contains an excellent financial aid section, a scholarship database, and a page on emergency bursaries. The search returns none of them. The student gets a list of unrelated news releases, or nothing at all, and leaves.
That moment repeats thousands of times a day across higher education websites, and it explains why vector search has moved from a technical curiosity to a board-level conversation. The technology answers a problem every large institution has: websites built over decades, by dozens of departments, searched by people who phrase questions the way they speak. Understanding how it works, and what peer institutions are already doing with it, no longer requires an engineering background. It requires about fifteen minutes.
Why Keyword Search Keeps Failing on University Websites
Traditional search works like the index at the back of a textbook. It matches the literal characters a person types against the literal characters in a document. If the words line up, the page appears in results. If they don’t, it doesn’t, no matter how relevant the content might be.
This approach has real strengths. Someone searching for a specific course code or form number wants an exact match, and keyword search reliably delivers it. The trouble starts when the searcher and the institution use different vocabulary, which is most of the time the case in higher education. The institution writes “financial aid.” The student types “tuition help.” The institution publishes “registrarial services.” The student searches “where do I get my transcript.” Same intent, zero shared words, zero results.
University websites are also structurally hostile to keyword matching. They sprawl. Faculties, departments, research centres, and service units publish independently, each with its own terminology, and the result is tens of thousands of pages with inconsistent language describing overlapping things. A zero-result search is one of the highest-friction moments on any website, and for an applicant comparing five institutions, it is often the last interaction.
What Vector Search Does Differently
Vector search starts with a concept called an embedding. An embedding is a long list of numbers that represents the meaning of a piece of content: a sentence, a paragraph, a program description, a policy page. The numbers themselves are unreadable to humans, but they have a useful property. Content with similar meaning produces similar numbers.
Earlier posts on this blog cover how embeddings are generated and stored. The behaviour is what matters here. The standard analogy is a map. On a geographic map, cities that sit close together tend to share weather, culture, and language. In an embedding, concepts that mean similar things sit close together in mathematical space. “Tuition assistance,” “bursaries,” and “help paying for school” all land in the same neighbourhood, even though they share almost no words. “Parking permits” lands somewhere far away.
Measuring Meaning Instead of Matching Words
Once every page on a website has been converted into an embedding, search becomes a matter of geometry. The system converts the user’s question into an embedding too, then finds the pages whose embeddings sit closest to it. Closeness is measured mathematically, typically by the angle between the two sets of numbers, and the results come back ranked by how near in meaning they are.
This is why vector search is often called semantic search. It retrieves by what content means rather than what words it contains. The student asking about paying for school when parents can’t help gets the financial aid page, the bursary application, and the scholarship database, because those pages mean the right thing, even though none of them contain that phrasing.
The approach also crosses languages naturally. A question asked in one language can match content written in another, because meaning, not vocabulary, drives the retrieval. For institutions recruiting internationally, that property alone justifies a closer look.
The Role of the Vector Database
A vector database is the infrastructure that makes this fast. It stores millions of embeddings and finds the nearest ones in milliseconds, using algorithms designed for exactly this kind of similarity lookup. It plays the same role for meaning-based search that a conventional database plays for rows and columns. Leaders don’t need to evaluate the algorithms themselves; they need to know the technology is mature, widely deployed, and well documented by every major cloud provider.
RAG, and Why It Matters More Than the Chatbot Hype
Most conversations about AI in higher education eventually arrive at chatbots, and most concerns about chatbots arrive at the same place: what happens when the AI makes something up. A general-purpose model like ChatGPT was trained on the public internet up to a cutoff date. It knows nothing about a specific institution’s current deadlines, program requirements, or policies, and when asked, it will sometimes guess with complete confidence.
Retrieval-augmented generation, usually shortened to RAG, is the architecture that fixes this. Before the AI answers a question, the system first retrieves the relevant pages from the institution’s own content, then instructs the model to answer using only that material, with citations back to the source. Vector search is the retrieval engine inside this pipeline. It finds the right institutional content; the language model turns it into a readable answer.
A useful way to think about the distinction: semantic search is a librarian who instantly finds the right shelf. RAG is a researcher who then writes a short, cited summary from what’s on it. The grounding step is what separates an AI assistant an institution can stand behind from one that invents admission requirements.
What Peer Institutions Are Already Doing
Seneca and McMaster
Seneca Polytechnic launched a generative AI version of its virtual assistant, SAM, in March 2025, built on its Microsoft Azure AI partnership. It answers questions from students, applicants, and employees around the clock, and hands off to human staff during business hours. McMaster University runs MAC, a generative AI chatbot in the Office of the Registrar, available 24/7 with an explicit escalation path: type “human” and the conversation routes to a person. Chat is now the registrar’s primary virtual service channel.
UBC and Toronto
The University of British Columbia chose self-hosting. Its secure LLM sandbox runs open-source models on UBC’s own Canadian infrastructure, giving researchers and staff what the university describes as a secure space for AI-powered work, even with sensitive data. The University of Toronto layered multiple tools: an enterprise Copilot deployment that returns footnoted source links and AI teaching assistants built at Rotman that ground their answers in specific course materials. The pattern across both is the same. Canadian institutions are adopting AI assistants, and they are doing it on terms that keep institutional data under institutional control.
The Library World Moved First
Anyone wondering whether semantic search is proven in academia can look at library discovery, where it already shipped at scale. Clarivate added a generative AI research assistant to Primo and Summon, the discovery platforms used by a substantial share of academic libraries, and EBSCO launched natural language search and AI-generated insights across its discovery service in 2024 and 2025. The tools students use to search journal databases already understand meaning. The main university website is, at many institutions, now the least intelligent search experience on campus.
Students Changed How They Search. Most Websites Didn’t
The behavioural shift behind all of this is well documented. Google’s own data shows about 40 percent of Gen Z reaching for TikTok or Instagram search instead of a search engine. Carnegie’s research series tracked AI use in the college search climbing from 4 percent in 2023 to 10 percent in 2024 to 23 percent in 2025. EAB found 26 percent of students used an AI chatbot somewhere in their college research. A Pew survey of American teens conducted in late 2025 found a majority now use AI chatbots, with 57 percent using them to search for information. On the Canadian side, KPMG found 73 percent of Canadian students using generative AI for schoolwork in 2025, up from 59 percent a year earlier.
AI Answer Engines and the Visibility Problem
The same shift is changing how institutional content gets discovered in the first place. Google AI Overviews now answer many queries before a user ever clicks. ChatGPT referral traffic grew 206 percent year over year through early 2026, and Similarweb data shows those visitors convert at rates second only to paid search. The volumes remain small, around one percent of overall web traffic, but education stands out as a category receiving a disproportionate share of AI referrals, and the visitors who arrive carry high intent.
Two implications follow. Institutional content needs to be structured so machines can retrieve and cite it accurately, because AI engines are becoming a front door. And the on-site search experience needs to meet the conversational expectations those engines have set, because a student who just asked ChatGPT a full-sentence question will type the same kind of question into a university search bar and expect it to work.
The Privacy Question Canadian Institutions Can’t Skip
For Canadian higher education, the architecture of an AI search deployment is a compliance question before it is a technology question. PIPEDA governs federally and permits cross-border data transfer, but holds the institution accountable for safeguards wherever the data sits. Public universities in Ontario and British Columbia operate under FIPPA, and BC institutions storing personal information outside Canada must complete a privacy impact assessment first. Quebec’s Law 25 goes further, requiring assessments before personal information leaves the province.
Ellucian’s most recent sector survey, drawing on nearly 800 respondents across more than 300 North American institutions, found data security and privacy remain the single biggest barrier to AI adoption, cited by a majority of administrators. That barrier shapes the sensible architecture. Vector search and RAG systems that run on infrastructure the institution controls, with content and student queries staying in Canada, clear the compliance bar that consumer SaaS chatbots often can’t. UBC’s decision to self-host is the clearest signal of where Canadian institutional preference is heading.
Where WordPress Fits Into All of This
WordPress is the most common CMS in higher education. An eQAfy analysis of American higher education homepages found WordPress powering roughly 41 percent of them, nearly double its closest competitor. Harvard, MIT, and Boston University all run it. The platform earned that position through flexibility, a manageable editing experience for decentralized teams, and an enormous ecosystem.
Its built-in search earned nothing. Default WordPress search runs a simple database query against three fields, ranks results by title match and date rather than relevance, ignores custom fields and taxonomies entirely, and slows badly on large sites. For a university running thousands of pages of program descriptions, faculty profiles, and service content, much of it stored in exactly the custom fields the default search ignores, the search box is the weakest component of an otherwise capable platform.
Modern WordPress architectures don’t require replacing the CMS to solve that problem. Vector search can be added alongside existing publishing workflows, allowing institutional content to be embedded, stored in a vector database, and retrieved semantically based on a visitor’s intent rather than exact keyword matches. Combined with retrieval-augmented generation, this creates AI assistants and search experiences grounded in the institution’s own content while allowing universities to maintain control over their data, infrastructure, and governance.
Vector search gives a university website something it has never had: the ability to understand what people mean.
Trew Knowledge helps higher education institutions design and implement enterprise WordPress and AI solutions that improve content discovery, modernize search, and prepare websites for the next generation of AI-powered experiences. If you’re exploring semantic search, retrieval-augmented generation, or AI-ready website architecture, we’d be happy to start the conversation.
