How AI shopping agents work

What is an AI shopping agent?

An AI shopping agent is software that helps a person shop: it researches, compares, answers questions, and narrows the options, and on some platforms it can complete the checkout once the shopper decides. Instead of returning pages of links, it interprets the request, runs its own searches, reads product feeds and reviews, and answers with a short recommended list the shopper chooses from. It is the buying side of agentic commerce: shoppers bringing an agent along to do the research legwork, while the decision stays with them.

The category is a spectrum. A conversational assistant answers and recommends. A shopping copilot builds comparisons and tracks prices. A true agent acts: it adds to cart and completes checkout through standards like the OpenAI and Stripe Agentic Commerce Protocol (ACP) or Google's agentic checkout. Since late 2025 the lines have blurred, but transacting remains the smaller, still-settling layer; most of what agents do for shoppers is the research.

The stakes stopped being theoretical in 2025: agent-referred traffic now converts better than the traffic funnels were built around, and it is growing at triple-digit rates.

Which AI shopping agents do people actually use?

Five surfaces matter in August 2026. Compare them on where checkout happens; that decides who owns the customer relationship.

ChatGPT is the largest third-party discovery surface; how ChatGPT recommends products covers its ranking in depth. Rufus is the walled-garden giant: Amazon told investors that shoppers who engage it are 60% more likely to complete a purchase.

How does an AI shopping agent decide what to recommend?

Take one real query: waterproof trail runners under $150 for wide feet. Every major agent runs it through the same five-stage pipeline.

Step 1: Parse intent, not keywords

The agent extracts the explicit constraints (budget, waterproofing, wide fit) and infers the unstated ones: wet weather, durability, sizing accuracy. A keyword engine matches strings; the agent builds a requirements list.

Step 2: Fan out into multiple searches

It then runs several searches at once behind the scenes. Google calls this query fan-out in AI Mode: simultaneous background searches on weather resistance, terrain grip, and wide-fit sizing, not one search for the original phrase.

Step 3: Assemble candidates from feeds and indexes

Each agent draws candidates from the data it can reach: ChatGPT from merchant product feeds and its web index, Gemini from the Shopping Graph, Rufus from Amazon's catalog alone. If your products are not in the sources an agent reads, you are not a candidate, and nothing downstream fixes that.

Step 4: Evaluate against structured data and reviews

Now the agent checks candidates against hard data: price, availability, and specs from feeds and schema markup, cross-referenced with reviews, editorial roundups, and forums. OpenAI's developer documentation spells out the split: required feed fields ensure accurate price and availability display, while recommended attributes like rich media, reviews, and performance signals improve ranking, relevance, and user trust.

Step 5: Rank, answer, and sometimes buy

The agent ranks the survivors and answers with a short list of products and reasons. The weights are opaque and model-specific, and in ChatGPT the results are organic: placement cannot be bought. If the shopper says yes, checkout can happen inside the conversation. Under ACP the merchant still validates the order, charges through its own processor, and can accept or decline; the agent is the interface, not the merchant of record.

What data do the agents actually read?

Strip the branding away and every agent reads the same four layers, ranked roughly by signal strength:

• Product feeds: the highest-signal source. OpenAI's spec ingests identifiers, pricing, inventory, media, and fulfillment via daily snapshots; Google's Merchant Center feeds the Shopping Graph.
• Structured data: schema.org Product, Offer, and AggregateRating markup lets a crawling agent read price, stock, and review scores off a product page without guessing.
• Crawlable, renderable pages: agents that browse need server-rendered content. JavaScript-only product pages and blocked crawlers (GPTBot, PerplexityBot, Google-Extended) make a store invisible.
• Third-party corroboration: reviews, editorial mentions, forum threads. Agents cross-check your claims against them; a brand no outside source mentions is one agents hesitate to recommend.

Freshness is a filter: agents re-ingest constantly, and a stale price or stock status gets you excluded, not corrected.

What do merchants control, and what can they not?

Merchants control the machine-readable layer: feed quality, structured data, whether AI crawlers get in, which merchant programs and checkout protocols to join, and, under ACP, branding, fulfillment, and the right to accept or decline any order.

They do not control the ranking weights, which agent a shopper uses, the conversation before their product comes up, or how the agent summarizes them against rivals. That is the real break from the funnel you optimize today; agentic commerce vs traditional ecommerce sets the two models side by side.

The conclusion: treat the machine-readable layer as merchandising infrastructure, not an SEO afterthought. It is the only surface you own, and acting on it pays: Salesforce found retailers using its Agentforce 360 to run their own branded agents grew Cyber Week 2025 sales 32% faster than those without.

How do you get your store agent-ready?

Readiness is a sequence, not a rebuild. In order:

• Audit where AI engines send your category's buyers today, and who they recommend instead of you.
• Unblock and verify the AI crawlers (GPTBot, OAI-SearchBot, PerplexityBot, Google-Extended) in robots.txt.
• Ship complete product feeds and schema.org structured data, and keep price and stock fresh.
• Make product pages answer-shaped: specs, sizing, and policies as extractable facts, rendered server-side.
• Evaluate checkout protocols (ACP, Google's agentic checkout) as they open to your platform.

Step one is measurement, and it is free. The rest compounds: checkout is still being negotiated between the conversation and the merchant's own storefront, and the merchants agents can read today become the defaults tomorrow.