Michaels' Ask Mike Converts at More Than Double Traditional Search: Inside the AI Shopping Assistant Built in Six Weeks

The early numbers behind Ask Mike

Michaels says AI shopping assistant more than doubles conversions

Michaels said shoppers using its Ask Mike AI shopping assistant converted at more than double the rate of customers using traditional search.

www.marketingtechnews.net

Michaels, the craft and celebration retailer running over 1,300 stores across North America, formally announced Ask Mike, its first customer-facing AI shopping assistant, on July 21, 2026. The tool had actually been running quietly since May, so the announcement doubled as a disclosure of early performance.

Four figures were made public. Since launch the assistant had handled nearly 75,000 conversations, of which 27% led to a product link click or an item added to cart. More than 60% of interactions centered on product discovery, and shoppers who engaged with Ask Mike converted at more than double the rate of those using traditional search.

Availability spans desktop and mobile web plus the iOS and Android apps. Rather than typing individual product names, shoppers describe what they are trying to make. Ask Mike then asks back about project type, materials, colors, and budget before recommending products.

The project-level intent that search keeps dropping

What matters here is not that AI got smarter. It is that a conversational interface finally addressed a long-standing mismatch: the granularity of shopper intent does not match the granularity of a search box.

Heather Bennett, president and chief customer officer at Michaels, described the mechanics this way.

"

Traditional search requires customers to break projects down into isolated keywords like 'floral fabric', leaving them to sift through the results themselves. With Ask Mike, shoppers use natural dialogue, asking complete questions like 'help me plan a space-themed party,' and the tool instantly curates the exact products and ideas they need.

Source: Heather Bennett, president and chief customer officer, Michaels

Someone throwing a child's birthday party does not need the paper plates, or the balloons, or the banner. They need all of it. Keyword search, however, was designed as a device that returns one answer to one query. The shopper ends up decomposing the plan in their own head, running search after search, and checking for gaps unaided.

The purchase behavior Michaels observed backs this up. Customers arriving through Ask Mike tended to bundle core supplies and finishing touches for a party, a framing job, or a craft within a single session. Customers using the search bar more often came looking for one specific item they had already decided on.

Baymard Institute, known for its ecommerce usability research, has similarly documented that site search carries several intents beyond exact product queries, including feature-based, compatibility-based, and use-case-based searches. Ask Mike's follow-up questions are structured to settle what the customer is trying to complete before recommending anything. The value of a conversational assistant lies less in the smoothness of the dialogue than in that resolution step.

How far to trust the doubling

The doubling figure, however, does not transfer cleanly into an expectation for your own store. At least three reservations apply.

First, the disclosure is partial. According to MarketingTech News, Michaels disclosed neither the underlying conversion rates, the sizes of the comparison groups, nor average order values. Separate figures for product clicks, cart additions, and completed purchases were not provided either. Nor was any data offered showing that the observed bundling behavior actually produced larger baskets or higher order values.

Second, self-selection remains baked into the comparison. A shopper who opens a new assistant and sustains a conversation may already carry higher purchase intent than the average site-search visitor. Whether the comparison controls for that gap cannot be determined from what has been published.

Third, the Adobe data frequently cited alongside these results measures something different. Adobe Digital Insights, analyzing more than one trillion visits to US retail sites, reported that AI-referred traffic converted 54% better than non-AI traffic in May 2026, with 53% more time on site and 23% more pages per visit. Those visitors were referred from external AI services, which makes the finding not directly comparable to an assistant embedded inside a retailer's own site.

Adoption levels deserve a sober look too. A YouGov survey of 1,414 US adults in July 2025 found that only 14% had ever used an AI shopping assistant, while 56% had never used one and were not interested. The survey is a year old and the numbers have surely moved, but conversational UI is still some distance from replacing the search box. What the Michaels result shows is that it works for people who use it, not that everyone will.

The catalog, not the model, made six weeks possible

On the implementation side, the most instructive detail is what sat behind the six weeks from concept to production.

Paul Tepfenhart, global director of retail industry strategy and solutions at Google Cloud, framed the timeline as evidence of what happens when years of infrastructure investment meet Gemini Enterprise for Customer Experience. Michaels had been building on Google Cloud for years, and Ask Mike landed on top of that accumulation.

Kapil Dabi, Americas lead for retail and CPG solutions at Google Cloud, was more specific. Speaking to Modern Retail, he attributed the fast turnaround to Michaels having modernized its technology stack and, critically, having organized its product catalog in a way that is AI-ready on top of its existing cloud foundation.

This is the part worth mapping onto your own operation. Conversational assistant projects usually stall not on model selection or UI design but on product data. Attributes are missing, category structures were cut to suit merchandising rather than shoppers, and use-case or occasion information is structured nowhere. Connect the strongest model available to a catalog in that state and it still cannot assemble a coherent product set from the phrase "birthday party."

Dabi also noted that the technical foundation is only one part of the deployment. Retailers need a clear view of what kind of help customers require at specific points in a shopping journey. A customer planning a birthday party needs recommendations across several categories rather than a single-item answer. That operating knowledge does not live inside an external model.

Gemini Enterprise for CX as shared infrastructure

The platform behind Ask Mike, Gemini Enterprise for Customer Experience, is the retail and restaurant package Google Cloud announced at NRF in January 2026 to combine shopping and customer service in a single interface. Several large retailers are now building distinct implementations on the same foundation.

Lined up side by side, the disclosure levels vary sharply. Ulta Beauty shipped Ulta AI to Ulta.com and its app in April 2026, yet on the earnings call CEO Kecia Steelman offered only that initial results have been promising. By publishing figures at all, Michaels is the one leaning forward here.

The Home Depot case shows how wide the surface area has become. Its deployment targets store phone support rather than pre-purchase discovery, understanding a caller's issue within 10 seconds and reaching a resolution four times faster than traditional phone menus. That the same platform serves both discovery and service is precisely what makes the return on AI agent investment hard to capture in one KPI.

What merchants should tackle first

Distilled into an order of operations, three things follow from the Michaels disclosure.

Start by re-reading your own search logs through the lens of intent granularity. Zero-result queries and sessions that ended in abandonment after repeated searches are the residue of intent that could not be decomposed into keywords. Measuring how much demand is buried there is faster than deciding whether to deploy a conversational UI.

Next comes product data. Without structured information on use case, occasion, target age, and complementary items, a conversational assistant has nothing to assemble a recommendation from. This work is not merely a precursor to AI deployment; it pays off simultaneously in AI search visibility and agent-mediated transactions.

Then measurement design. The items Michaels left undisclosed form a ready-made list of what to define before you start. Group sizes for users and non-users, a comparison that controls for purchase intent, separated figures for clicks, cart adds, and completed purchases, and average order value. Let a "double the conversion" number circulate internally without those, and investment decisions will follow it off a cliff.

Conclusion

Michaels demonstrated two things at once: AI shopping assistants are moving past the experimental stage, and the measurement around them has not been standardized. A six-week build says the barrier to entry has shifted from the model side to the data side.

What to watch next is how and when Michaels ships the AI-driven product overviews and contextual prompts it has signaled for product detail pages. Once the assistant dissolves out of a standalone chat window and into the product page itself, how much of the site search component remains? The answer will likely become visible in implementations across the second half of 2026.