How AI shopping agents are changing ecommerce product discovery

Product discovery used to start with a search box, a category page, and a list of results ranked by relevance. That model is changing as AI enters both sides of the shopping journey: external AI platforms are helping shoppers discover products before they reach a retailer’s site, while brands are deploying AI agents to guide discovery on their own storefronts.

External AI shopping experiences influence whether a shopper discovers your product. On-site AI agents influence what happens once they arrive. These two experiences are often grouped under “AI shopping agents,” but they solve different problems, and a brand needs a plan for each. This piece looks at how discovery itself is changing, what the current data shows, and what it takes – technically and strategically – to stay competitive as more of the shopping journey moves into conversation.

What counts as an AI shopping agent

An AI shopping agent is a system that acts on a shopper’s behalf to find and evaluate products, rather than simply answering a question when asked. That distinction matters: a scripted chatbot responds to a fixed set of inputs, while an agent interprets intent, reasons about the request, and takes the next useful action without waiting for step-by-step instructions.

External AI shopping experiences, offered by platforms such as ChatGPT, Gemini, and Perplexity, can search, retrieve, and synthesize product information across many retailers on a shopper’s behalf, with the specific mix of capabilities varying by platform. On-site AI agents, deployed directly on a brand’s storefront, interpret what a visitor is looking for and guide them through that brand’s specific catalog. A shopper describing “something warm for cold winter evenings” isn’t naming a product category, and both kinds of agents are built to bridge that gap between a stated need and a specific match.

How discovery is changing at the top of the funnel

Traditional search asked shoppers to translate a need into keywords. AI agents reduce that translation step – shoppers can describe a need in plain language, while the agent interprets relevant constraints such as material, size, compatibility, and budget. According to Riskified’s Q1 2026 Agentic Commerce Pulse, a survey of consumers in the US and UK, 61.5% have already used AI tools for product discovery and recommendations, a clear sign that this shift is no longer early-stage.

That changes what “showing up well” actually means. Keyword-focused copy alone is not enough: AI systems also need structured, machine-readable product information to determine whether an item fits the shopper’s requirements.

What the current data shows

A few more figures give a sense of how fast this is moving:

• AI-referred traffic to US retail sites grew 62% year over year in July 2026, and converted 60% better than non-AI traffic – the 11th consecutive month AI traffic has outperformed on conversion – according to Adobe Analytics, based on more than one trillion visits to US retail sites.
• That conversion advantage has been widening steadily: in March 2025, AI-referred traffic converted 38% worse than non-AI traffic; by March 2026, it converted 42% better, according to Adobe’s Q2 2026 AI Traffic Report; by July 2026, the gap had grown to 60% better.
• As of eMarketer’s December 2025 forecast, AI platforms were projected to drive $20.9 billion in US ecommerce sales in 2026 – about 1.5% of total US retail ecommerce sales, and nearly four times 2025’s figure.

Put together, the pattern is consistent: adoption is broad, and the conversion gap between AI-referred and traditional traffic keeps moving in AI’s favor rather than leveling off.

Why brands need to plan for both kinds of agents

Being findable by external agents and building an on-site agent solve different problems, and most brands will eventually need both. External agents influence discovery before a shopper reaches a retailer’s site; on-site agents shape what happens once they arrive. In different ways, both depend on the same underlying requirement: product data that’s structured and reliable enough for a system to reason over, whether that system belongs to a third-party platform or the brand itself.

What that looks like in practice differs on each side. Showing up well in external agents is largely a question of exposing accurate, well-structured data to systems a brand doesn’t control. Building an on-site agent is a question of designing something the brand does control, end to end – and it’s worth breaking down what that actually involves before getting into the technical foundation both sides share.

What on-site agents do, and why they’re worth building

Keeping some discovery on a brand’s own site, instead of outsourcing it entirely to a third-party platform, is the less obvious half of this shift. When a shopper’s research and comparison happens inside someone else’s chat window, the brand loses the relationship along with the sale. An on-site agent that can offer a comparable conversational discovery experience – interpreting intent, comparing options, narrowing to a recommendation – keeps that discovery moment, and the first-party behavioral and intent data it generates, with the brand instead. Building that kind of agent requires more than a scripted FAQ bot: it involves intent interpretation, product retrieval, recommendation logic, and reliable grounding, which is what AI agent development work actually involves.

Not every on-site agent plays the same role, and the distinction matters when deciding what to build:

• Agents that guide help shoppers narrow down options by answering questions and surfacing recommendations – their job is reducing friction during discovery, not closing the sale.
• Agents that sell read behavioral signals like hesitation or repeated comparisons and engage proactively to move a shopper toward checkout.
• Agents that do both handle the full arc from discovery through purchase, which is what most brands end up needing once an agent proves its value in one part of the funnel.

Regardless of type, an agent works through the same basic loop: it perceives inputs (what the shopper is asking, plus live catalog, inventory, and pricing data), reasons about what the shopper actually needs, and takes an action – a recommendation, a comparison, an answer. That loop is only as good as the data feeding it, which is exactly why the technical foundation below isn’t an afterthought – it determines whether the agent’s output is useful or just plausible-sounding.

Done well, the payoff for store owners shows up in a few places: higher conversion from traffic a brand is already paying to acquire, since an agent can qualify intent and guide high-intent visitors toward a relevant next step; more relevant upsell and cross-sell, since recommendations that reflect the actual conversation tend to land better than static “customers also bought” blocks; and less manual load on support for routine, repetitive questions that an agent can resolve directly from catalog and policy data. The Adobe data cited above – on both engagement and conversion – reflects part of this pattern already playing out at the traffic level, before an on-site agent even enters the picture.

The technical foundation: data, integrations, grounding

Both external visibility and a working on-site agent come back to the same question: what has to be in place for either one to work reliably?

For external visibility, retailers need product information that AI systems can reliably retrieve and interpret, whether it comes from structured page data, product feeds, APIs, or other connected sources. Standards like the Model Context Protocol are one mechanism among several: they can give AI agents a standardized way to access tools and data that a retailer chooses to expose, although accuracy still depends on how the integration is implemented. Building that connection correctly is often a substantial technical project, particularly for stores with large catalogs or multiple connected systems – which is the kind of work an AI integration build is meant to handle.

For on-site agents, the same reliability question shows up as grounding. An agent recommending from a stale catalog or an outdated product description creates a bad outcome for the shopper and a lost sale for the retailer. Retrieval-augmented generation is one approach to addressing this: a retrieval-augmented generation setup can ground an agent in current product information retrieved from connected data sources – details, availability, pricing, and policies – which reduces, though doesn’t entirely eliminate, the risk of outdated or inaccurate recommendations as the catalog changes.

Getting ready for agent-driven discovery

None of this requires rebuilding a store overnight, but it does call for a deliberate response on both fronts:

1. Audit product pages for structured data completeness, including product attributes, availability, pricing, and return terms, so AI systems can reliably interpret the information needed to evaluate a product.
2. Confirm that feeds, pricing, availability, and inventory data stay up to date, because stale product information can lead to inaccurate recommendations.
3. Measure how much product discovery is happening through third-party AI platforms, and assess whether an on-site agent could keep more of that discovery experience and first-party data with the brand.
4. Treat the technical foundation as one connected capability rather than three separate initiatives.

For the tactical checklist on structuring product data itself, our guide on preparing your store for AI shopping agents covers the schema and content side in more depth.

The shift is not simply from search to chat. It is a shift from optimizing pages for human navigation to making product information usable by AI systems – both third-party platforms and the agents brands build themselves. Brands that treat structured product data, real-time integrations, and conversational discovery as one connected capability will be better positioned as more purchasing decisions move into AI interfaces.

Deveit’s AI services team works with ecommerce brands on exactly that shift – from wiring up the integrations external agents rely on to building on-site assistants that stay accurate as a catalog changes – so the technical groundwork is in place as AI-mediated product discovery becomes a larger part of ecommerce.