How AI Shopping Agents Decide Which Products to Recommend and Buy

A buyer used to open ten tabs, read reviews, and decide. Now they ask ChatGPT to compare three options, or tell Perplexity to find the best cordless drill under $150, and the agent returns a short list with reasons. The buyer never sees the pages you optimized. They see the answer the agent assembled. If your product is not in that answer, you lost the sale before the buyer knew your brand existed.

This is the same shift that hit content marketing with answer engines, arriving now in commerce. The winning move is not more ad spend or better landing-page copy. It is making your product legible to a machine that reads structured data and ignores the persuasion layer you built for humans. Here is how these agents actually decide, and what determines whether your product makes the shortlist.

What an AI shopping agent is doing when it recommends a product

An AI shopping agent is a system that takes a buyer's request in plain language, interprets the intent, retrieves candidate products, ranks them, and presents a small set with justifications. Some can also complete the purchase. The category now includes ChatGPT's shopping research, Perplexity's shopping experience, Amazon's Rufus, and Alexa+, and each reads a different pool of product data.

The retrieval step is the one that decides your fate. Before an agent can rank you, it has to have your product in its candidate set, and it builds that set from data it can read cleanly: merchant product feeds, structured markup on your pages, and third-party sources like review sites. If your product is not represented in a form the agent can parse, it is not a candidate, and no amount of on-page copy changes that. This is the commerce version of a rule OnlyAEO has documented across answer engines, explained in how AI engines read your paragraphs, not your page: the machine reads structured units of meaning, not the whole marketing experience.

The signals that decide whether your product gets surfaced

Merchants who watch agent behavior see the same inputs decide the shortlist again and again. Product feeds carry the weight. According to Shopify's guidance for optimizing a store for Perplexity shopping, the fields that matter are the ones a matching system needs: the product category, attributes like material and dimensions in structured fields, and identifiers such as GTINs and MPNs that let the agent match your item to a query and to reviews elsewhere. Miss the identifier and the agent cannot confidently connect your listing to the ratings that would have won it the slot.

Reviews are the second pillar. Shopping agents build their product cards partly from customer reviews, so the volume and quality of your ratings feed directly into whether you are recommended and how you are framed. The mechanics of turning review presence into AI recommendations are the same ones covered in how to turn G2 and review sites into fuel for AI recommendations, and they apply to consumer products as much as to software.

Live accuracy is the third. Agents deprioritize products with stale pricing or unclear availability because recommending an out-of-stock or wrongly priced item breaks the buyer's trust in the agent, not just in you. Keeping pricing, stock, and shipping times synchronized across every channel an agent might read is not hygiene, it is a ranking factor.

The pattern is consistent: the agent rewards products that describe themselves precisely in a language it can read, and it rewards them because precise self-description is what lets the agent answer the buyer's actual question.

The plumbing changed in 2026, and it favors data over ads

The infrastructure moved fast, and the direction matters for where you invest. OpenAI built the Agentic Commerce Protocol with Stripe and open-sourced it at agenticcommerce.dev, a standard that lets agents read merchant product feeds and act on them, with Target, Sephora, and The Home Depot among early integrations. Google introduced its Agent Payments Protocol, AP2, aimed squarely at completing transactions, with Etsy, PayPal, Revolut, and Mastercard involved. Stripe added a machine payments path for businesses already on its rails.

The most telling move was a retreat. OpenAI launched Instant Checkout in September 2025 and discontinued it by March 2026, saying it did not offer the flexibility they wanted, and refocused ChatGPT on product discovery rather than closing the transaction, as Exploding Topics documented in its analysis of the agentic commerce protocol. Perplexity walked a similar path, replacing its native checkout, launched in November 2024 with PayPal, with a free discovery experience for all users.

Read those two retreats together and the lesson is clear. The battle these agents are fighting is discovery and recommendation, not checkout. The value accrues to the product that gets named when the agent answers the buyer's question, which means your investment belongs in the data layer that decides the recommendation, not in a checkout integration that the platforms themselves keep reworking.

What this means for a brand that wants to be recommended

Start by auditing whether an agent can even see your product as data. Pull your feed and check that every required and recommended field is filled, that identifiers are present and correct, and that attributes a buyer would constrain on, size, material, compatibility, capacity, live in structured fields rather than buried in a paragraph. Agents rely on structured data over visual inference, so a spec hidden in an image or in prose is a spec the agent does not have.

Then treat your product pages the way you would treat any page you want an answer engine to quote. Add Product schema so price, rating, and availability carry machine-readable meaning, and add an FAQ block that answers the constraint questions buyers actually ask, structured so the engine can extract each pair. The discipline is the one OnlyAEO applies across every surface, described in how OnlyAEO works, and it is the same reason a machine-readable feed of your catalog matters: the AI Feed Engine exists to keep a clean, current representation of what you sell in front of the engines, and a free llms.txt generator gives crawlers a map to the sources you want them to read.

Do not neglect the review surface. If your ratings are thin or scattered across sites the agent does not read, the agent has no confident basis to recommend you over a competitor with a dense, current review presence. Build reviews on the sources your category's agents actually pull from, and keep them fresh, because recency counts.

B2B products get selected the same way

The instinct is to file all of this under consumer retail, and that is a mistake. B2B buyers now open their software research inside an agent, ask it to compare vendors against their requirements, and arrive with a shortlist the agent built. The mechanics are identical: the agent retrieves candidates it can read as structured data, matches them to stated constraints, and names a few. The difference is only that your feed is your category positioning, your integrations, your pricing model, and your proof, rather than a GTIN.

Getting a software product onto that shortlist is the discipline covered in how to get your SaaS into AI best-tools lists and comparisons, and the proof that structured AEO work moves the needle is in the FastTrackr AI case study, where deliberate answer-engine structure turned a category unknown into a named option. The buyer using an agent to compare vendors is running the same retrieval-and-rank loop as the buyer comparing drills. Make your product the one that describes itself precisely enough to be the confident answer.

The measurement problem, and how to get ahead of it

Agent-driven recommendations are hard to attribute because the buyer often arrives with no referrer, having decided inside the agent and typed your brand directly. That does not mean the channel is invisible, only that you have to instrument for it. Track branded search and direct traffic for lifts that follow agent recommendation, run periodic prompt tests to see whether the agents name you for your category's real buyer questions, and treat your presence in agent answers as a leading indicator the way you would treat share of voice. The point is to measure the recommendation, not just the click, because the recommendation is where the decision now happens.

The takeaway

AI shopping agents do not read your marketing. They read structured feeds, identifiers, reviews, and live data, match them to what the buyer asked, and surface a short list. The brands that win the recommendation are the ones that describe their products precisely in a machine-readable form and keep that description accurate and current. That is an AEO problem wearing a commerce hat, and the teams that treat it as one now will be the default answer when the buyer asks.