Survey
300 readers told us how they search for news. They wanted AI that makes search faster, not different. 69% wanted quick summaries, but 41% never use AI chatbots for news.
Replacing The Globe and Mail’s keyword search with an AI assistant that subscribers can trust, from research to a live launch.
My role
I led design end to end: research, interaction design, prototype testing and handoff, working with data science, engineering, product and research. My first push was to get research done before the vendor deadline locked the scope.
The problem
The Globe and Mail’s search was a third-party keyword matcher. It couldn’t tell what a reader meant, and there was no way to refine a query. Readers had to phrase things the way the system wanted.
Data science had built a retrieval model (RAG) that could answer plain-language questions across articles, market data and author profiles. It needed an interface, and the vendor launch date was set before design joined.

How do you design one search box for a model that gives a different kind of answer depending on what you ask?
Research
I planned a three-phase research program. We ran the first two before any interaction design; the diary study is set for phase 2. The goal wasn’t to confirm what we believed. It was to find the beliefs worth challenging.
300 readers told us how they search for news. They wanted AI that makes search faster, not different. 69% wanted quick summaries, but 41% never use AI chatbots for news.
10 Globe and Mail subscribers used an interactive prototype that tested two search modes, the tab structure and grouped results.
The launch timeline didn’t leave room for one. Phase 2 will track how readers use the assistant over several weeks, beyond a first session.
We assumed
We learned
Separate news and markets modes would help readers set context.
Picking a mode before typing slowed people down. One input that detects intent worked better.
A conversational interface would get people writing full questions.
People typed short keywords anyway. The look of a chat changed what they expected to see, not how they searched.
The hard part was making AI answers feel impressive.
The hard part was making them feel safe. Readers wanted reliable results and clear sources, not synthesis.
A ranked list is the clearest way to show results.
Grouping results under headings that match the question helped people judge coverage faster.
For market questions, a written AI answer would be most useful.
A live price, chart and key numbers felt more credible than generated text.
Key insight
Readers weren’t asking for AI. They were asking for search that understood their question.
Research prototype
Exploration
Before the prototype, I explored three empty states. Each one bet on a different way readers would get started.

A time-of-day greeting and an example query for signed-in readers.

Prompt chips by topic, with shuffle for readers who don’t have a question yet.

A news / markets toggle inside the input.
Ruled out in research (see 01 above)Interaction flows
I mapped every path the model could return for each of the three query types.



General news: a topic or event query, with branches for clear, broad and no-result answers. Open full size
Key decisions
News search and market search became a single input. The model works out the intent, so readers never have to choose before they type. A question about Shopify is about the stock and the news at the same time.
Tradeoff
One model is harder to tune at the edges than two specialist ones.

Best matches shows AI-ranked results grouped by intent. All articles shows everything retrieved, with filters. The tabs leave room for AI summaries later without a redesign.
Tradeoff
Without summaries, the first release is less conversational than the vision.

Four result types, each matched to a kind of query: article cards for news, live price and chart for a company, author cards for people, and a carousel for comparisons. I worked with data science to define the output each one needs.
Tradeoff
Four result types meant a strict contract with the backend. Design helped define what the model returns.


A BETA badge, a link back to classic search, sources and a disclaimer on market data, and a feedback link. Each one sits where the doubt happens, not in a help page.
Tradeoff
More on every screen. Each element answers a specific doubt we saw in testing.
Chat history, follow-up suggestions and AI summaries are fully designed and on hold. Summaries can get facts wrong, and in a newsroom that risk waits until the model has proven itself in real use.
Tradeoff
The first release feels less conversational, but it is reliable and ready to grow.
Final design

All articles: everything retrieved, with filters.

Mobile Best matches, grouped under Latest updates.

Mobile All articles, with filter and actions on each item.

Author bio is collapsed by default. People come for the articles.

The bio opens on demand.

Stock result: live price, chart, key metrics and a data disclaimer.

Filters apply to All articles only. Best matches stays unfiltered on purpose.
The system
The design work sat at the end of the pipeline: what the model returns, how it signals trust, and which component shows it.
Outcome
Shipped
Live for subscribers as a beta, on the vendor date, with research, design and engineering agreed on scope.
Trust
In beta testing, readers were more willing to try AI results when the trust signals were on screen.
Ready
Summaries, follow-ups and chat history can switch on without a redesign.
Reflection
The most useful trust feature was one line of text: Switch to classic search. People will try something new when they can see the way out. The hardest call was holding back summaries after we had designed and prototyped them. The model hadn’t yet earned that trust in a newsroom.
What I’d look at next