Productivity

Over half of US shoppers now use AI to buy — and small-business sites that AI can't read are the ones losing the sale

NIQ says 51% of US consumers used an AI tool to shop last month. AudioEye's same-day study shows agents complete just 31% of tasks on inaccessible sites, versus 96% on accessible ones.

By Esi Boateng· Productivity tools·2026-09-27·4 min read

Photo: Unsplash / rupixen — Person shopping online on a laptop with a credit card in hand

Two studies dropped on September 24, 2026, and read together they describe a structural break in how Americans buy things. NIQ’s Agentic Commerce Tracker found that 51% of US consumers used an AI-powered tool to help them shop in the past month, up from 42% in the May 5 release using the same Quick Question method. A nine-point rise in roughly four months, at a ±4.4-point margin on a ~500-consumer monthly sample. Liz Buchanan, President of North America at NIQ, is now presiding over a tracker that’ll refresh monthly, alongside a joint product with Similarweb announced September 2 and due in Q4 2026 to measure how brands surface inside AI tools.

Product recommendations lead adoption at 20%; personal shopping assistants sit at 16%. The story isn’t any one feature. It’s that discovery itself is being intermediated.

The second study, from AudioEye (Nasdaq: AEYE), quantifies who wins that intermediation and who quietly disappears. Running 1,560 independent agents across six commercial AI models against six websites, five retail, travel, and DTC properties plus W3C’s own accessible/inaccessible demonstration site, the company measured a 68% collapse in task completion between clean and broken markup. Agents completed 96% of 13 real-world tasks on the accessible version of a site, and 31% on the inaccessible one, with each combination repeated ten times. Ohio State’s WebJudge evaluator agreed with the scoring 95% of the time.

There’s also a compute tax. Agents burned 43% more tokens on average against inaccessible sites, a median of 128,000 versus 90,000, spiking to 6x at the extreme. On one task requiring numbers rendered inside an unlabeled image, all 60 attempts failed. Websites that AI can’t read are websites AI can’t sell from — that’s the practical translation, and it lands harder against Kelly Georgevich’s data than any accessibility argument of the past decade.

The parallel here isn’t mobile-first indexing in 2015. It’s earlier. It’s the mid-1990s realization that a business without a searchable web presence had become illegible to the people trying to spend money with it. That transition took a decade. This one is compressing into quarters, and the September 24 data is the receipt.

Sources