AI Gift Finder Quiz: Turn Burnaby Browsers Into Buyers

It's 10 p.m. and someone in Burnaby is on your site looking for a gift for a father in law who "has everything." They open three product tabs, compare two that look the same, and close the browser. Nobody said no to your products. They just couldn't decide.

A gift finder quiz fixes that moment. It asks four or five quick questions and hands back a short list of things you actually stock. AI makes the quiz smarter, because it can read an answer like "she hikes every weekend but hates clutter" and match it to your catalog. It also adds a failure mode you have to design around. Here's how it works, what the evidence says, and when it isn't worth building.

Why gift shoppers stall

Buying for yourself is easy. You know your size, your taste and your budget. Buying for someone else means guessing at all three, and every extra option makes the guess feel riskier. A good salesperson solves this by asking questions. Most small online stores don't have that person, especially after closing time.

Online is still a minority of Canadian retail, but it's growing. Statistics Canada reported that retail e-commerce sales increased 9.9% to $5.7 billion in June 2026, accounting for 7.7% of total retail trade. So most of your gift sales probably still happen at the counter. That's fine. The same quiz can run on a tablet by the till during a busy December Saturday, so staff can help the shopper who knows exactly what they want while the quiz helps the one who doesn't.

What an AI gift finder actually does

A classic quiz is a decision tree: answer A, get product set A. The AI version keeps the structure but handles messy, human answers. A solid setup looks like this:

• Ask four to six questions. Who it's for, the occasion, the budget, a few interests, and the deadline (pickup in store versus shipping across Metro Vancouver).
• Filter with plain rules first. In stock, in budget, arrives in time. This part should be ordinary code, not the model.
• Let the model rank the shortlist. It reads the free text answers and picks the best three from the products that survived the filter.
• Lock the output to real product IDs. More on this below. It's the most important step.
• Show each pick with a one line reason and an add to cart button, plus an "ask our staff" link for anyone who wants a human.
• Log the answers, with consent, so you learn what people are shopping for.

The value isn't the questions. It's turning vague intent into three concrete choices fast, with a reason the shopper can repeat to themselves ("it's light, it packs small, she'll actually use it").

The catch: models invent products

Language models are good at sounding like a helpful shop clerk. They're less reliable at sticking to your inventory. Researchers studying AI recommendation systems put it bluntly: without explicit grounding in the item catalog, models frequently generate out of catalog or even non existent items in a zero shot manner, which the authors say severely limits their practical use. That paper (Zhu and colleagues, 2025) tested on research datasets, not a gift shop, but the lesson carries over. If you just ask a chatbot "what should I buy my sister," it may happily recommend a candle you've never carried.

The fix is grounding. You give the model only the filtered shortlist and force it to answer with IDs from that list. OpenAI's API documentation, checked in October 2026, says its Structured Outputs feature ensures the model will always generate responses that adhere to your supplied JSON Schema. You can put your shortlist's product IDs in an enum field, so the only legal answers are products you sell. The same page lists limits worth knowing: as of October 2026, a schema may have up to 1,000 enum values across all enum properties, and when a single property has more than 250 values their combined length can't exceed 15,000 characters. That's another reason to filter first and send a shortlist, not your whole catalog.

Two honest caveats from the same docs. The model can still refuse a request, and a response can be cut off if it hits the max token limit, so your code needs a fallback (show the top rule based matches). And a schema only guarantees the format. A valid product ID can still be a lousy gift. You'll need to test the picks yourself.

Privacy and email rules in BC

Quiz answers can get personal fast. "Gift for my mom, she just finished chemo" is health information about a third party, typed into your website. Treat it with care.

The federal privacy commissioner's Guidelines for obtaining meaningful consent, issued jointly with the Alberta and BC commissioners, ask businesses to emphasize four things: what information is collected, who it's shared with, why, and any risk of harm. They also call for a clear yes or no choice and favour just in time notices over a wall of policy text. For a quiz, that means one plain sentence above the free text box telling people their answers go to an AI service to generate suggestions, and a note about whether you keep them. The same guidance says that for children under 13, consent should come from a parent or guardian, so don't aim the quiz at kids.

If the quiz ends with "email me these picks," that's also a marketing list. Canada's anti spam law requires that businesses obtain consent before sending commercial electronic messages. Keep the newsletter checkbox separate and unticked. Sending the results the shopper asked for is one thing. Adding them to weekly promos is another.

How to tell if it's working

Run it as a test, not a leap of faith. Track how many people start the quiz, how many finish, how many click a pick, and how many buy. Compare against visitors who didn't see the quiz in the same period.

Here's an illustrative example, not a real result. Say 2,000 visitors see the quiz in November and 600 finish it. If 5% of finishers buy, that's 30 orders. Whether that's good depends entirely on your baseline. If your normal conversion is already 5%, the quiz added nothing but cost. Also watch returns on quiz picks. A quiz that sells more but drives exchanges isn't helping.

Where this doesn't apply

AI gift finders are not for every shop, and you're not behind if you skip one.

Small catalogs. With 25 products, a hand written "gifts under $50" page from your staff will likely do the job.

Messy product data. The quiz is only as good as your tags, descriptions and stock counts. If inventory isn't synced, you'll recommend things that sold out last week.

Low traffic. If only a few hundred people a month would see it, you won't get enough data to know whether it helps.

When a plain quiz is enough. If your answers fit tidy categories (budget, age, occasion), a rules quiz with no AI is cheaper, faster and can't hallucinate. AI earns its place when the free text answers matter.

The tooling is still young for small shops. Statistics Canada found that 12.2% of Canadian businesses used AI to produce goods or deliver services in the 12 months before its Q2 2025 survey, up from 6.1% a year earlier. Among those users, 24.8% used virtual agents or chatbots and 23.1% used AI for marketing automation. Adoption is climbing, but most businesses aren't there yet, and off the shelf quiz apps vary a lot in how well they stay grounded in your catalog.

A simple way to start before December

Pick your 30 to 50 best gift products. Clean up their tags and descriptions. Build a rules based quiz first and watch how people answer. Once you see the free text answers people would type if they could, add the AI ranking step with locked product IDs and a human fallback. You'll have something live before the holiday rush in Burnaby, New Westminster and the rest of the Lower Mainland.

Sources

• Statistics Canada. "Retail trade, June 2026." The Daily, August 21, 2026. https://www150.statcan.gc.ca/n1/daily-quotidien/260821/dq260821a-eng.htm
• Statistics Canada. "Analysis on artificial intelligence use by businesses in Canada, second quarter of 2025." Analysis in Brief, June 2025. https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2025008-eng.htm
• Zhu, Y., Steck, H., Liang, D., He, Y., Li, J., Kallus, N. "Rank-GRPO: Training LLM-based Conversational Recommender Systems with Reinforcement Learning." arXiv:2510.20150, 2025. https://arxiv.org/html/2510.20150v1
• OpenAI. "Structured Outputs." OpenAI API documentation, accessed October 2026. https://developers.openai.com/api/docs/guides/structured-outputs
• Office of the Privacy Commissioner of Canada, with the Information and Privacy Commissioners of Alberta and British Columbia. "Guidelines for obtaining meaningful consent." 2018, last modified 2025. https://www.priv.gc.ca/en/privacy-topics/collecting-personal-information/consent/gl_omc_201805/
• Innovation, Science and Economic Development Canada. "Canada's anti-spam legislation." Accessed October 2026. https://ised-isde.canada.ca/site/canada-anti-spam-legislation/en

Want a gift finder that only recommends what's actually on your shelves? Book a free call with Autana and we'll look at your catalog together and tell you honestly whether a quiz, a plain rules page or nothing at all is the right move this season.