Adela Mincea
[ The Marketing Economist ] 26 April 2026 · 10 min readAIMarketing

What is agentic commerce and how AI assistants pick products

Shoppers are asking ChatGPT what to buy. Here is how AI assistants pick which products to recommend, and what e-commerce stores need to do to be selectable.

A customer opens ChatGPT and types: "what's the best trail running shoe under $150." Three named products come back, with reasoning. The customer asks one follow-up question, gets a single recommendation, clicks the link, and buys.

The retailer that won the sale never appeared in a Google search result for that query. They were not optimised for the shopper's keyword, did not bid on it, did not show up in any SERP the customer ever saw. They were selected by the assistant. Selected for reasons few e-commerce teams I speak with have looked into yet.

This is agentic commerce. My judgement is that the shift could prove as large as mobile or social commerce were for retail. Nobody has measured that yet, so treat it as a view. Most store owners I speak with are aware that AI is changing how shoppers find products. Few have changed anything in their operations, because the playbook is not yet written down.

This article writes part of it. What agentic commerce is, which assistants are doing it now, what signals they use to pick products, and the three actions every store owner can take in the next 90 days to be selectable when a competitor is not.

What is agentic commerce?

Agentic commerce is purchase activity where an AI assistant evaluates products for the shopper, recommends a short list, and increasingly transacts on their behalf.

The shopper delegates the comparison work to an assistant. The assistant queries multiple data sources at once: product feeds, structured data on retailer sites, aggregated reviews, editorial content, third-party comparison platforms. It ranks the options against the shopper's stated criteria, returns two to four named products, and frequently links directly to checkout. In a growing share of cases, the assistant initiates the transaction itself through integrated checkout APIs.

Three things change for the retailer in this model. The algorithm doing the picking is different from the search engine algorithms most stores have spent a decade optimising for. The data sources the assistant pulls from are different and broader than what any single platform indexes. And the surface area for influencing the outcome shrank from "rank, click, convert" to "be in the data the assistant trusts." Optimisation moved upstream. The retailer who understands the new inputs has a structural advantage that compounds.

How does agentic commerce differ from traditional ecommerce?

The traditional ecommerce flow is well understood. A shopper has a need. They search Google or Amazon. A SERP returns ten results. The shopper evaluates them, clicks through to two or three sites, compares offers, and buys. Three of those steps belonged to the retailer: ranking, the click, the conversion. The comparison work belonged to the shopper.

The agentic flow rearranges almost all of it. The shopper still has a need. They ask an assistant in natural language. The assistant evaluates products across multiple data sources without the shopper ever seeing a SERP. It returns two to four recommendations with reasoning. The shopper trusts the answer or asks a follow-up. Increasingly, the assistant transacts directly.

The shopper's role shrank. The assistant's role grew. The retailer's surface area to influence the decision shrank with it. You used to compete with whoever ranked above you in the SERP. You now compete with whoever the assistant cites. A small store with clean structured data, complete product feeds, and consistent third-party validation can be cited over a large store with weaker data, even if the large store dominates traditional SEO.

The economic implication is significant. Traditional SEO and SEM rewarded scale and budget. Agentic commerce rewards data quality and credibility across the web. Those are different inputs, with different cost structures and different competitive dynamics. A category leader on traditional search can be a category laggard in agentic commerce, and vice versa, on a much shorter timeline than most teams expect.

[ Keep reading ] Free, by email

Read the rest of this piece

Subscribe to The Marketing Economist: marketing and paid media through an economics lens. The full article opens right here, along with every other one on this site. New issues arrive about once a month.

About one email a month. Unsubscribe any time. Privacy

Which AI assistants pick products today?

Five matter right now.

ChatGPT Shopping. ChatGPT added web search in October 2024 and shopping results in April 2025 (Zeff, 2025). In September 2025 OpenAI added Instant Checkout, which lets a user buy inside the chat (OpenAI, 2025). It pulls product data from a combination of merchant feeds, public web sources, and partner integrations. When a user asks for a product recommendation, ChatGPT returns named products with descriptions, prices, and links. The potential audience is large: in February 2026 OpenAI cited more than 900 million weekly active users for ChatGPT (Search Engine Land, 2026).

Perplexity Shopping launched in November 2024 (Perplexity, 2024) and has positioned product search as a core use case rather than a feature add. It pulls heavily from structured data, third-party reviews, and editorial content. Smaller user base than ChatGPT, but high purchase intent and strong influence in categories where shoppers research carefully before buying.

Google AI Mode and AI Overviews began surfacing product information inside AI-generated answers in 2025. The data sources are the Merchant Center catalog, organic search results, and structured data on retailer sites. This shift is the most consequential of the five because Google still controls the largest discovery surface, and AI Overviews are intercepting queries that previously went to standard SERP results. When an AI summary appears, Google users click a result link in 8% of visits, against 15% when none appears (Pew Research Center, 2025). Similarweb data puts US organic search referrals at 1.8 billion visits in June 2025, down from over 2.3 billion in July 2024 (Guaglione, 2025). Neither figure is specific to retailers, but the direction is the same.

Gemini Shopping is Google's standalone assistant integration with Merchant Center and the Shopping graph. Less mature than ChatGPT or Perplexity for discovery today, but its reach inside Android and Google Workspace makes it strategically important on a two-year horizon.

Amazon Rufus is a closed system that only recommends products from Amazon's own catalog. If you sell on Amazon, Rufus is its own discipline. If you do not, it does not apply directly, but it is training shoppers to expect AI-mediated discovery everywhere else, which changes their behaviour on your site.

How do AI assistants pick which products to recommend?

None of these assistants publishes how it ranks products, so what follows is an inference from the outside. From the answers I have watched them give, six factors appear to shape whether your product gets cited.

1. Structured data quality. AI assistants read structured data first because it is unambiguous. Schema.org Product markup, accurate GTINs, brand identifiers, current price, and live availability give the assistant a reliable record to work with. Missing or stale structured data does not only hurt rankings, it makes the product invisible. Assistants prefer products they can describe accurately. If your data is incomplete, they describe a competitor instead.

2. Product feed completeness. For Google AI Mode and Gemini in particular, the Merchant Center feed is the source of truth. Missing attributes like color, material, size, certifications, or compliance information reduce the chances of being matched to a specific shopper query. Feed quality is the highest-leverage technical fix available to most stores. A complete feed gets considered. An incomplete one matches fewer queries, so it gets considered less often.

3. Third-party authority. The assistant cross-references products against external signals. Aggregated review scores from independent sites, mentions in publications, ratings on comparison platforms. A product with consistent third-party validation gets cited more often than a product with only first-party marketing claims. The question the assistant is effectively asking: "does the rest of the web confirm this is a credible product?"

4. Editorial mentions in publications AI trusts. Assistants weight content from sources they treat as authoritative. Wirecutter, Consumer Reports, vertical specialists like Outdoor Gear Lab in their categories, established review sites. A product mentioned in those sources accrues citation weight that the brand cannot generate alone. This is the AI version of digital PR, and it compounds. Once a product appears in several trusted reviews, it tends to be recommended more consistently across assistants.

5. Product page content depth. The assistant reads your product pages. Specifications, materials, dimensions, use cases, FAQs, comparisons. Pages with depth give the assistant more to match against the shopper's specific query. Pages that are eighty words of marketing copy and a buy button give it nothing. The product pages that win in agentic commerce read like a structured knowledge base, not a brochure.

6. Brand recognition signal. This is the slowest-moving factor and the hardest to influence in the short term. Assistants weight brands they have seen mentioned consistently across many sources. A brand that appears in editorial content, structured data, third-party reviews, and merchant feeds together accumulates a recognition signal that appears to raise its odds of being cited. New brands can compete, but only by being conspicuously present in every other layer.

The pattern across all six is the same. The assistant trusts signals that are consistent and verifiable. It distrusts marketing claims that exist only on the brand's own site. The economic consequence is that the work of being recommended by AI is mostly the work of being credible across the web, not the work of optimising one channel.

How I rank the six, strongest first. This is my reading of how the assistants behave. No weights were measured and no assistant publishes its ranking method, so the order is a judgement.

  • Structured data quality (strong)
  • Product feed completeness (strong)
  • Third-party authority (strong)
  • Editorial mentions (moderate)
  • Product page depth (moderate)
  • Brand recognition (weaker in the short term)

What should store owners do now?

Three actions, in order of leverage.

1. Audit your structured data. In the stores I have reviewed, Schema.org markup is usually partial and breaks under inspection. Test every product page with Google's Rich Results test. Confirm GTINs are present, prices update in real time, availability reflects actual stock, brand and aggregateRating fields are populated where applicable. Fix gaps in the catalog template, not page by page. This is a one-time engineering job that pays compounding returns. The pattern across the e-commerce stores I have reviewed is consistent: structured data is partial, often broken on a meaningful share of SKUs, and rarely audited end to end.

2. Get products mentioned in third-party content. Identify the publications, review sites, and comparison platforms in your category. Most have either an editorial pitching process or an affiliate program. Both can work for AI citation purposes. From the outside I see no sign that assistants treat affiliate coverage differently from editorial coverage when they cite it, though nobody outside the labs can confirm how they weigh either. The goal is to be present in a handful of trusted external sources per product line. This is a quarterly content and PR effort, not a one-off campaign, and the citation weight builds slowly, with consistency, over several quarters.

3. Make product feeds AI-readable. Most Merchant Center feeds I see have missing attributes. Material, color, size variants, certifications, target audience, age groups, gender, condition. Audit the feed against the full attribute list for your category. Fix the gaps systematically. A complete feed is no longer a nice-to-have. It is the entry ticket to Google AI Mode and Gemini Shopping, and increasingly the data ChatGPT and Perplexity cross-reference when they verify product claims. The work is unglamorous and high leverage, which describes most of the operational improvements that actually move revenue.

These three actions are not optional. They are the minimum required to be in the consideration set when an assistant evaluates your category. Skip them and your product simply is not visible, regardless of how good the product is or how much you spend on traditional acquisition channels.

Why act now?

The window is real, and it is short. Most e-commerce stores I see have not addressed agentic commerce in any structured way. Most have not audited their structured data in years. Their product feeds have known gaps that have been ignored for budget reasons. They have no editorial presence outside their own marketing channels and no plan to build one.

This is the same condition that existed in early Google SEO in 2003, in the first years of Facebook ads after their 2007 launch, and in the first wave of influencer marketing in 2015. A small number of operators that did the unglamorous foundation work captured a disproportionate share of the channel before it became competitive. The cost of doing it later, after the channel matures, is always higher than the cost of doing it now, when the bar to be visible is still low.

Organic search is already sending fewer clicks as AI answers intercept queries that previously triggered ten blue links, as the Pew and Similarweb figures above show, and retailers have no reason to expect an exemption. The question is no longer whether AI-mediated commerce becomes the dominant discovery surface. It is which retailers are visible inside it when it does.

This article is based on patterns I have observed across the e-commerce accounts I have reviewed. The implementation effort scales with catalog size, but the diagnostic does not.

References

If something here was useful, would love to know!