Ask ChatGPT for a pair of running shoes, ask Gemini to plan a nursery, or ask Rufus on Amazon what blender to buy, and something is quietly deciding what you see first. That decision used to belong to a search results page, ranked mostly on keywords and backlinks. Now it belongs to a language model reading a product feed, and enough of the rules have changed that most of what worked for search doesn't carry over cleanly.

The scale of this shift is too large to treat as a side channel now. Traffic to U.S. retail sites from generative AI platforms rose 693% year over year in November and December 2025, the largest increase of any industry Adobe tracked that season, ahead of travel, financial services and media. Digital Commerce 360, reporting on that Adobe Analytics data, put U.S. online holiday sales at $257.8 billion, with Salesforce estimating that AI agents and other generative AI tools influenced more than 20% of all online retail sales globally over the same period.

What an agent is actually weighing

None of the major platforms rank products the way Google ranks pages. There's no backlink profile to build and no meta description to optimize. OpenAI has been unusually direct about what replaces them: when multiple merchants sell the same item, ChatGPT's ranking looks at real-time availability and pricing, product quality signals, whether the seller is the original retailer, and whether that merchant supports OpenAI's checkout integration — with no ranking boost given just for having that integration turned on. It's a fair stand-in for how the rest of the field approaches the same problem.

Underneath that ranking sits a plainer requirement: the agent has to be able to read a catalog before it can recommend anything from it. Merchants feed OpenAI, Google and Perplexity a structured file, usually the same CSV or XML format built for Google Merchant Center, listing price, availability, identifiers and product attributes. A page written for a human shopper and a search crawler doesn't automatically satisfy this. A blender description that reads well but skips the wattage, capacity and material specs gives the agent nothing concrete to compare, and an agent that can't compare a product with confidence tends to leave it out rather than guess.

A handful of signals show up across every platform building this, in some form:

  • Complete, accurate, frequently refreshed product data — price, inventory, specs, images
  • Reviews and ratings, which agents lean on the way a shopper would ask around before buying
  • Price and shipping terms relative to other listings of the same product
  • Whether the retailer can be verified and trusted to actually fulfill the order
  • How precisely the content answers the specific question being asked, not just the general topic

Every major platform is building this differently

OpenAI's approach has already changed once this year, which says something about how unsettled this still is. Instant Checkout launched in September 2025 with Etsy, letting a shopper complete a single-item purchase without leaving ChatGPT, built on the Agentic Commerce Protocol it developed with Stripe. By February 2026 that had expanded to free users and a wider set of merchants. But by March 2026, OpenAI had pulled the in-chat purchase step back for its biggest retail partners. Walmart, Target, Sephora, Nordstrom and several others now appear inside ChatGPT for discovery and comparison, then hand the shopper off to finish the purchase on the retailer's own site, according to CNBC. Large retailers, it turned out, weren't eager to give up ownership of checkout.

Google has the deepest existing inventory to draw on. Its Shopping Graph, the database behind AI Mode and the Gemini app's shopping features, holds more than 60 billion product listings, with 2 billion refreshed every hour. Google's Universal Commerce Protocol, announced at NRF's Big Show in January 2026, lets an agent query that data, check real inventory and complete a purchase through Google Pay once the shopper confirms — agentic checkout is already live with retailers including Wayfair, Chewy and Quince. The advantage isn't a smarter model. It's two decades of Merchant Center feeds nobody else has access to.

Perplexity runs a separate Merchant Program, launched in November 2024, that ingests the same Google Shopping feed format most merchants already maintain. In February 2026 it opened its shopping features to free users instead of Pro subscribers only, and reported a fivefold increase in shopping-related queries the following quarter. Checkout runs through PayPal rather than a proprietary system, and ranking leans on the same core ingredients as everywhere else: how complete the feed is, how well the product matches the query, and what its reviews say.

Amazon's version has the largest built-in reach and the most disorienting name change. Rufus, the assistant Amazon rolled out through 2024, had passed 300 million users by the end of 2025, per the company's own product updates. In May 2026, Amazon renamed it Alexa for Shopping. The rebrand didn't change how it recommends: it still reads listings, reviews and Q&A content built on Amazon's product knowledge graph, and a listing has to clear Amazon's existing search ranking before Rufus, or whatever it's called by the time this is read, will consider recommending it at all.

The gap between how fast this is moving and how ready anyone is

IDC projects AI will replatform $500 billion in digital spending by 2030, with agent use among large enterprises growing tenfold by 2027, and warns that brands who aren't ready stand to lose access to a quarter of their market, according to an IDC InfoBrief commissioned by WooCommerce. Most retailers sit behind that curve, not because they're unaware AI shopping exists, but because their product content was written for humans and search crawlers, not for a system that needs machine-readable specs to make a comparison.

Merchants seem to know it. Checkout.com's 2026 survey found 72% of merchants across the UK and US agreeing that consumers will adopt agent-led shopping faster than most businesses are prepared for. A third of consumers already expect at least 10% of their purchases to be AI-driven within a year.

Trust hasn't caught up to adoption, and that's worth taking seriously

None of this reads as a settled, fully trusted system yet, and it shouldn't. The same Checkout.com survey found 27% of consumers say they trust no organization to run an AI shopping agent on their behalf, and 24% say they'll never hand a purchase decision to one at all. That skepticism sits right next to real usage: 57% of consumers said they'd let an AI agent switch them to a different brand if it found better value. People don't need to fully trust a system to let it narrow their options. They just need it to save more time than it costs them.

Adobe's own numbers back that up from the other direction. Among shoppers who used generative AI tools during the 2025 holiday season, 81% said the experience had improved their shopping, and shoppers arriving from generative AI converted 31% more often than shoppers from other channels, roughly double the gap seen in 2024. They aren't fully trusting the recommendation. They're finding it useful enough to act on anyway.

What actually moves the needle if you sell something

Most of what determines whether an agent recommends a product has nothing to do with clever prompting and everything to do with the unglamorous parts of a catalog most teams deprioritize. A clean, complete, frequently updated product feed matters more here than in almost any other part of technical SEO — if the data an agent can retrieve is thin or stale, there's nothing to rank in the first place, whatever the page looks like to a person.

Reviews carry real weight too, closer to how a person asks around before buying than how a search engine reads backlinks. A product with a hundred detailed, specific reviews gives an agent something to summarize and compare with confidence; a product with four generic five-star reviews gives it almost nothing to work with, regardless of the actual quality of what's being sold. This is the same discipline behind our GEO / AI Search work — structuring what a brand publishes so a system that's synthesizing an answer, not scanning for keywords, has a specific, extractable reason to include it.

The page copy still matters, just for a different reason than it used to. An agent answering "which one fits a small apartment" needs a real, specific answer, not marketing language that talks around the question. That's less a copywriting problem than a content strategy problem: deciding upfront which questions a buyer is actually going to ask an agent, and making sure every one of them has a direct answer sitting somewhere in the product's own data.

The platforms building this are still changing their minds about basic pieces of it, OpenAI's checkout reversal being the clearest example so far. That's not a reason to wait. Reviews, feeds and specific product data are worth having in good order regardless of which protocol ends up winning, and the brands treating this as infrastructure now are the ones an agent will still be able to find once everyone else catches up.