AI Shopping Assistants for Ecommerce: Guided Discovery vs. Quizzes and Grids

An AI shopping assistant for ecommerce is a chat agent, grounded in a store's product catalog and site content, that talks a shopper through choosing a product instead of showing them a grid or a form. The shopper who lands on a category page not sure what they need, "I have dry, sensitive skin and don't know where to start" or "I need a gift for someone who camps but I don't camp" is the audience for this. A recommendation widget can't take that input. A search bar can't either. Conversation can, because the shopper can say the actual constraint in their own words and the assistant can ask a follow-up.

That's a narrower claim than the category's marketing usually makes, so it's worth being precise about what guided discovery is and isn't before comparing tools.

The shopper this actually serves

Split ecommerce shoppers into two rough groups: people who know what they want and are checking details (does this fit, does it ship in time, is it in stock), and people who don't yet know what they want and need to be walked there. The first group is what a sales agent handles, pre-purchase questions that stall a checkout that's already in motion. The second group is discovery, and it's a different job: turning "I don't know where to start" into a specific product, before there's a cart to abandon.

Category and search pages are built for the first group. They assume the shopper can already describe what they're filtering for, in the store's own taxonomy. Baymard Institute's usability research on product list and filtering design found abandonment rates of 67 to 90 percent on sites with mediocre product-finding tools, against 17 to 33 percent on sites with even a modestly better toolset, for shoppers attempting the identical task of finding a product. Most stores hand an undecided shopper a filter sidebar and expect them to already know the vocabulary, regardless of traffic quality or price point.

Guided discovery tools exist to close that specific gap: give the shopper a way to describe their situation in normal language and get routed to the right product, without first learning the store's category structure.

Two different mechanisms, often confused

"AI shopping assistant" gets used for two structurally different tools, and which one a store needs depends on how the product catalog maps to decisions.

Quiz and guided-selling apps ask a fixed sequence of questions, skin type, budget, use case, then map the answers to product tags through rules a merchant configures in advance. RevenueHunt is a common example: a free plan for up to 100 quiz responses a month, then paid tiers starting around $39/month for 500 responses, $99 for 1,000, up to an Enterprise plan above $299/month for unlimited responses, as of publication. Octane AI runs a similar model on a credit system rather than a flat response count, starting around $50/month. These tools work well when the decision tree is genuinely finite, skincare routines, gift categories, supplement stacks, and the merchant is willing to build and maintain the tagging logic that routes quiz answers to products.

Conversational assistants, the kind grounded in an LLM and the live catalog, skip the pre-built decision tree. The shopper types whatever they'd say to a person, "camping gift for someone who doesn't already have gear," and the assistant searches the actual catalog for a match instead of walking a script. Nothing needs to be pre-mapped, which means it handles phrasing nobody anticipated, but it only works if the assistant is grounded in the store's real catalog and content. An assistant answering from general model knowledge instead of the store's actual products and policies will invent details fluently, sizes that don't exist, materials the product doesn't use, which is the most common way this category fails in practice.

Neither replaces the other. A quiz is faster to build correctly for a narrow, well-understood decision. Open conversation handles the long tail of phrasing and constraints a fixed question set won't anticipate, at the cost of needing real grounding to stay accurate.

What the results look like when it works

McKinsey's retail research includes a case study on Karaca, a Turkish homeware brand, whose generative-AI shopping assistant AIDA reportedly doubled conversion versus on-site search and converted roughly five times higher than unassisted sessions. Separately, McKinsey cites a global lifestyle retailer whose gen-AI shopping assistant lifted conversion by as much as 20 percent. Both are single-company case studies published by the vendor's own consulting partner, not an industry average, and neither discloses baseline traffic quality or how "assisted" sessions were defined against unassisted ones. Treat the direction as credible, since it matches the plain mechanism (an answered question converts better than an unanswered one), and treat the specific multiples as one company's result, not a number to plug into your own forecast.

The more useful comparison is the one you can run yourself: pick your ten most common "I don't know what to get" questions and see how many a candidate assistant answers correctly against your actual catalog, before trusting any vendor's published lift.

Honest limits of the category

Every version of this tool, quiz or chat, has real edges worth knowing before buying:

Grounding is not optional. A conversational assistant is only as good as its connection to current catalog data and real site content. Stale inventory or unwritten product details produce confident wrong answers, not a shrug.
It doesn't replace merchandising. Cross-sell grids, cart bumps, and "frequently bought together" logic are a separate discipline from guided discovery, usually handled by dedicated apps like Rebuy, and covered in more depth in our comparison of conversational recommendations against static widgets.
Quiz logic requires upkeep. A rule-based quiz is only as current as the last time someone updated the tagging when the catalog changed. Nobody maintains this indefinitely without a process for it.
It's reactive by default. Most of these tools, chat or quiz, wait for the shopper to start the interaction. A few, like Rep AI, are built around proactively opening the chat based on browsing behavior instead, which is a different product decision with its own tradeoffs.
Complex, high-stakes decisions still want a human. Custom orders, wholesale inquiries, and anything with real ambiguity should escalate with the full conversation attached rather than getting forced through either mechanism.

Where Ernest fits

Ernest is one AI agent that plays three roles for a Shopify or WooCommerce store, and guided discovery is its product-agent side. It's grounded in the store's live catalog and everything ingested from the store's own site, descriptions, specs, FAQs, size charts, so when a shopper describes a need in chat instead of picking filters, it names an actual product from real catalog data rather than a generic suggestion.

That's a conversational tool, not a quiz builder: there's no rule tree to configure by hand, but it also means the answer quality depends entirely on what the store's own content actually says, the same grounding requirement any assistant in this category has to meet.

Honest limits, since the previous section applies here too:

No merchandising layer. Ernest doesn't build cart bumps, bundles, or cross-sell grids. It answers a shopper's question in chat; it isn't a widget for browsing pages.
Reactive, not proactive. It responds when a shopper opens the chat rather than popping open based on browsing behavior.
Channels are the storefront widget and email, with escalation to a human by email when a question needs a person. No SMS, voice, or social channels.
No outbound. Ernest doesn't send follow-up or cart-recovery messages tied to a discovery conversation; it answers in the moment.

The same agent handles the two jobs that pair naturally with product discovery: pre-purchase questions on the sales-agent side, and order status, cancellations, and returns after the sale on the support side, so a shopper who starts with "help me choose" and ends with "where's my order" stays in one conversation instead of two tools.

Pricing is conversation-based rather than tied to quiz responses or visitor traffic: a free plan with 100 conversations that never expire, a real free plan rather than a trial, then $49/month for 500 conversations, $149 for 2,000, and $299 for 10,000. AI is included at every tier, with no separate per-seat or per-resolution charge.

Deciding which mechanism your store needs

Start with your actual product pages, not the vendor pitch. If the decision is narrow and well-understood, three or four questions reliably sort shoppers into the right product family, a quiz app is the faster build and the cheaper one to maintain, because the logic only has to be written once. If your product pages generate open-ended questions that don't fit a fixed question tree, "would this work for," "what's the difference between," "I need something that," that's conversational territory, and a rule-based quiz will keep missing the actual question being asked.

Either way, test before committing. Write down your ten hardest real "help me choose" conversations from support history and run them against a candidate tool on your actual catalog. A quiz that can't route an edge case isn't a bug, it just wasn't built for that path. A conversational assistant that gets your catalog facts wrong is a different problem, and it's disqualifying regardless of what the pricing page promises. Then track the number that matters: how many discovery conversations turned into an order, against your baseline conversion rate before you installed anything.