Amazon’s AI Image Generator Turns Shopper Imagination Into Search

Amazon launched a generative AI image generator inside its Shopping app’s search bar on June 3, producing phantom product pictures in real time as shoppers type and using those images as inputs to its visual search engine to find real matching inventory. The feature covers apparel and home goods for US customers on iOS and Android; every AI-made picture appears below the autocomplete suggestions and is labeled as AI-generated.

Mihir Bhanot, Amazon’s director of search, published the generator alongside seven other visual search updates, 20 days before Amazon’s Prime Day event, which runs June 23 through June 26.

The Vocabulary Gap Amazon Is Targeting

Amazon built the feature’s rationale around a specific failure at the search bar: a shopper who can picture what they want but doesn’t know the retail term for it. The June 3 announcement described two concrete examples, a shopper who imagines a shirt with a draped collar but can’t recall the neckline style name, and one who pictures a couch with woven side panels without knowing the furniture term for the material. Typed as plain text, both searches typically return poor results from a keyword-matching engine.

A customer may want a shirt with a draped collar but can’t think of the term “cowl neck,” or a couch with woven side panels but doesn’t know the word “rattan.”

That passage is from Amazon’s own announcement. The generator’s response is to build a visual stand-in on the fly: as a shopper types a description, AI-built thumbnails appear below the autocomplete bar and update with each added word. Tapping the closest match triggers Amazon Lens technology, which runs a visual search against the real catalog and returns items with similar visual characteristics. The generated pictures don’t exist in any seller’s inventory and can’t be purchased.

Amazon described the feature as working “best where visual details matter most” and said more categories would follow after the clothing and home goods launch. Each added word progressively reshapes the generated options, narrowing the visual set before the shopper taps, so the search input converts from text to visual intent before a keyword is ever completed.

Eight Features, One Coordinated Push

The image generator was the most structurally novel of eight visual search features Bhanot published alongside it. The others extend tools already inside the Shopping app, each aimed at getting a buyer to a product listing through something other than a completed keyword query.

• Shop by Style collages: Searching a broad fashion phrase like “women’s silk shirt” now produces AI-curated outfit panels inside the results, organized under themes including “Urban luxe,” “Soft elegance,” and “Executive chic.” Tapping a collage opens a curated page where individual items can be browsed and purchased.
• Amazon Lens Live upgrades: The camera-based product scanning feature gained a real-time tracking carousel and direct integration with Alexa for Shopping, so the live camera view can now answer questions about a scanned product or suggest search terms for items it can’t immediately identify.
• iPhone lock screen Lens widget: A camera shortcut on the iOS lock screen routes a product scan to results before the phone is unlocked. Amazon described this as making shopping “ambient,” available before the user has formed an explicit shopping intent.
• Visual filter suggestions: Broad descriptive queries now trigger visual filter options as the shopper types, narrowing the visual scope before the full results page loads.

All four are part of Amazon’s broader generative AI shopping push, which positions visual and conversational tools at the front of the discovery funnel rather than behind the results page.

Alexa for Shopping, Amazon’s AI assistant that replaced the Rufus chatbot on May 13, sits across several of these features. Rajiv Mehta, Amazon’s vice president of conversational shopping, called it “a personal shopper who already knows you and remembers your preferences.” The system uses Claude Sonnet, Amazon Nova, and a proprietary model trained on Amazon’s product catalog and customer reviews, all routed through Amazon Bedrock, the company’s managed AI infrastructure service.

Three Years of Rising Visual Searches

Amazon’s investment in visual search has compounded for at least three years, and the growth figures from that period explain why a dedicated image generator is a logical continuation rather than an experiment.

The October 2024 milestone came when Amazon introduced five new visual search features, including text-enhanced Lens searches and the first visual suggestion filters. The September 2025 Lens Live launch then layered a real-time tracking carousel onto the camera search and pushed overall Lens engagement higher by another 13 percentage points, according to Amazon’s November 2025 figures.

In Amazon’s Alexa for Shopping announcement, the company tied those Rufus metrics to the decision to expand the assistant’s surface area: deeper camera integration via Lens Live, a lock screen shortcut, and the new role inside the image-generation search bar. Three hundred million customers and $12 billion in incremental annual sales are the numbers Amazon used to justify the May 13 rename and the June 3 surface expansion.

The Keyword Auction Sits One Layer Below

Amazon’s advertising business runs on keyword bids. A seller pays for a Sponsored Products placement to appear when someone types “cowl neck sweater” or “rattan side table” into the search bar. The AI image generator sits in that same bar, in the suggestion layer above where a typed query would complete and trigger the ad auction.

Where Sponsored Products Stand

PPC Land, a digital advertising publication covering search and retail media, analyzed the mechanics on June 4: the generative image feature operates in the suggestion layer below the search bar, upstream of keyword-matched Sponsored Products placements. If shoppers increasingly navigate by tapping generated images rather than completing keyword queries, the connection between a seller’s keyword bid and actual shopper behavior becomes less direct.

The scale behind that loosening matters. Amazon’s advertising services generated $17.2 billion in the first quarter of 2026, up 24% year over year, according to Amazon’s first-quarter financial results. U.S. retail media ad spending across all platforms is projected at $69.33 billion for 2026 by eMarketer, and Amazon holds the largest platform share of that market. Keyword bids are the mechanism most sellers use to compete for placement within it.

Visual search still routes a buyer to a product results page where Sponsored Products and Sponsored Brands ads appear. A shopper who taps a phantom image and lands on results for “ribbed cashmere cardigan” still arrives inside a commercial environment. The specific question is whether the typed keyword that would have triggered a seller’s bid was ever entered. For apparel and home goods sellers whose campaign structures are built around the vocabulary Amazon is now routing around, the answer shifts from certain to probabilistic.

The scale of that shift depends on adoption. A shop selling a single apparel item with unusual terminology sits in a different position than a multi-SKU brand with broad keyword coverage across dozens of style names. What changes is that the search bar now routes buyers who can’t supply the keyword Amazon’s ad auction requires, and those buyers didn’t previously surface through keyword search at all.

What Amazon Has Not Revealed

Amazon’s June 3 materials describe the image feature purely as a discovery tool and say nothing about commercial placements inside the suggestion layer itself. Whether ad formats will appear there is unanswered.

Amazon surfaces have followed a consistent pattern. Sponsored Products and Sponsored Brands prompts in other new surfaces became billable on March 25, 2026, after beta periods during which Amazon gathered behavioral data from those placements. The image suggestion layer could follow the same trajectory, introducing seller ad inventory once usage patterns are stable enough to price.

Shop by Style collages create a parallel opacity for advertisers. The outfit panels assembled under labels like “Urban luxe” and “Soft elegance” are algorithmically curated from the catalog. Which products appear in those panels, and whether Sponsored Brands or Sponsored Display placements factor into their composition, hasn’t been disclosed. Sellers running keyword campaigns have no public answer yet on how much of the new discovery surface will eventually carry paid placements.

Google Got There First

Google deployed a comparable tool inside its AI Mode roughly a year before Amazon’s launch: type a description, see AI-generated clothing or home decor suggestions, tap to find real products. The two share the visual shortcut mechanic but differ structurally in where the transaction sits.

Dimension	Amazon Shopping App	Google AI Mode
Images shown	AI-generated, labeled as such	AI-generated, labeled as such
Platform role	Transaction layer	Discovery and referral
Tap destination	Amazon product results page	Shopping suggestions with offsite links
Launch categories	Apparel and home goods	Clothing and home decor
Style or outfit feature	Shop by Style collages	AI Mode outfit suggestions

Amazon’s trust measure is the AI-generated label on every phantom image, which prevents confusing a generated render with a catalog item. The gap it doesn’t close is between the approved visual and the real products that follow it. Unlike Google, where a tapped suggestion is a referral to an outside retailer, Amazon’s version routes into a purchase funnel where advertising conversion rates average 10 to 15% for optimized listings, according to industry advertising benchmarks, making the distance between a clean AI mockup and an actual sale shorter and more consequential.

Ben Schoon, a senior editor at 9to5Google, called the Amazon version “wildly wasteful in terms of the use of AI resources” on June 3, arguing a platform with hundreds of millions of real product images generates synthetic ones to address a problem its existing catalog could solve. Amazon’s stated rationale goes to the vocabulary gap: a shopper who can’t name what they want can’t search the catalog for it.

The compute trade-off is real. Generating a set of phantom thumbnails on every partial search query, at Amazon’s search volume, costs more infrastructure than returning existing product photos. Amazon’s Prime Day 2026, running June 23 through June 26, is the first large shopping event where all eight visual search features operate together in a live selling environment.