AI Search and Recommendations: How Shoppers Actually Find Products Now
A shopper used to type a few keywords into Google, scroll through blue links, and click through to a store. Increasingly, they’re asking an AI assistant a full question instead, or typing something into a store’s own search bar that sounds more like a sentence than a search term. Both changes are reshaping how a product gets found, and they’re related but not the same problem. This post covers AI search from both directions: the AI answer engines and shopping agents operating outside your store, and the AI-powered search and recommendation tools running inside it.
What Changed in Search
Google’s AI Overviews now appear above traditional results for a large share of searches, and when they do, click-through rates to the actual websites listed below them often drop by close to 60%. Google’s AI Mode, a fuller conversational search experience, passed a billion monthly users during 2026. Some analysts expect traditional keyword search volume to keep shrinking as more of that traffic moves into AI assistants instead. None of this means search traffic is disappearing. It means a growing share of it never reaches a results page at all, and a store that only optimizes for the old kind of search is optimizing for a shrinking part of the picture.
AI Answer Engines and Shopping Agents
Tools like ChatGPT, Google’s AI Mode, and Perplexity don’t just answer questions anymore. They research products, compare options, and in some cases complete a purchase without the shopper visiting a store directly. ChatGPT Shopping now recommends products to US users and can route a purchase through participating retailers, including Etsy and over a million Shopify merchants. Perplexity goes a step further for its Pro subscribers, letting them complete a purchase inside the chat itself through a PayPal-powered checkout. Google built its own checkout path into AI Mode through a protocol called UCP, with early partners including Wayfair, Chewy, and Etsy.
This space is still sorting itself out. OpenAI scaled back ChatGPT’s original direct-checkout feature in March 2026 after only a small number of merchants had signed on, largely because it hadn’t solved real-time inventory accuracy or sales tax collection. Some retailers have also pushed back directly: Amazon blocks the crawlers that ChatGPT and similar tools use to read product listings, which keeps Amazon’s own catalog out of those AI shopping results entirely. Research from Bain found that shoppers trust a retailer’s own on-site AI tools roughly three times more than a third-party AI agent like ChatGPT or Perplexity, which suggests the AI assistant built into your own store may earn more trust than the one shopping on a customer’s behalf from outside it.
Getting Found by AI Search
AI answer engines decide what to cite and recommend differently than a traditional search engine ranks a page. Research on AI-generated answers has found that longer, well-structured pages get cited more often, and that content with specific numbers, comparisons, and evidence carries more weight than vague marketing language. Sites that block AI crawlers outright tend to disappear from AI Overviews and similar results entirely, which is worth checking directly in your site’s crawler settings rather than assuming. None of this replaces the product feed data covered in the platform side of AI shopping, things like Schema.org markup, accurate stock status, and GTIN codes, but it does mean the actual writing on a product or category page matters for AI visibility in a way that pure keyword stuffing never captured.
AI-Powered Search on Your Own Store
Search inside your own store has changed just as much as search outside it. Traditional site search matches keywords literally: a shopper searching “something cozy for winter evenings” gets zero results if no product title contains those exact words, even if the store has exactly what they’re looking for. AI-powered, semantic search tools interpret the intent behind a query instead of just the words in it, matching “notebooks” to a listing titled “laptop” or understanding that “cocktail dress for wedding guest” points to color, formality, and fit rather than a literal product name. Retailers running this kind of search commonly report a sharp drop in zero-result searches, and search users on a site tend to convert at a noticeably higher rate than visitors who only browse, since a search is a much stronger signal of buying intent than casual scrolling.
Tools like Algolia, Klevu, Bloomreach, and Netcore Unbxd all offer this kind of search as an add-on to platforms like Shopify, BigCommerce, or WooCommerce, layering natural language understanding and behavioral ranking (which products actually convert for a given query, not just which ones technically match it) on top of a store’s existing catalog. The right choice depends on catalog size and budget more than any single feature, since most of these tools now offer broadly similar core capabilities.
Recommendation Engines
Search and recommendations solve different problems. Search answers a question the shopper already asked. A recommendation engine surfaces something the shopper didn’t search for but is likely to want, on a product page, in the cart, or in a follow-up email. These show up as “customers also bought,” “complete the look,” or a personalized homepage grid, and they range from simple rule-based logic (same category, similar price) to machine-learning models that adjust based on an individual shopper’s actual browsing and purchase history.
The revenue impact is well documented, even though the exact number varies by source and industry. Recommendations commonly account for a quarter to a third of ecommerce revenue in sessions where a shopper actually clicks one, and shoppers who engage with a recommendation tend to add more to their cart than those who don’t. Moving from a fixed rule-based engine to a machine-learning one typically improves both click-through and downstream conversion, though the size of that improvement depends heavily on how much clean data the engine has to learn from.
The Data Behind Both
Neither AI search nor recommendations work well on top of messy product data. A recommendation engine trained on inconsistent categorization or inaccurate stock levels will recommend the wrong things confidently. A semantic search tool can only match intent to inventory that’s actually described and tagged correctly in the first place. This is less an AI problem than a groundwork problem: the stores getting real value from these tools usually cleaned up their product data and inventory accuracy first. Our guides on stock control management and small business stock management systemscover what that foundation looks like in practice.
Practical Steps for Merchants
Start by checking whether AI crawlers are blocked anywhere in your site’s settings, since that alone can remove a store from AI search results entirely. Review your product pages for the kind of specific, evidence-based writing that AI answer engines tend to cite, rather than generic marketing copy. On your own site, run a search analytics report if your search tool offers one, and look specifically for high-volume searches returning zero results, since each one represents real demand your store isn’t currently catching. Before investing in a new AI search or recommendation tool, confirm your product data (categories, attributes, stock status) is actually accurate, since that’s what determines whether the new tool performs better than the old one or just makes the same mistakes faster.
Takeaway
AI hasn’t replaced search and recommendations, it’s split them into more paths than before: an AI assistant researching products outside your store, and an AI-powered search bar or recommendation widget inside it. Both depend on the same underlying thing, accurate and well-structured product data, more than they depend on which specific AI tool a merchant picks. Getting that foundation right is what determines whether AI search becomes a real source of sales or just a faster way to surface the same gaps a store already had.