AI Shopping Agents Are Coming – Whose Side Are They On?

The problem: shopping decisions are becoming AI‑driven

When a parent asks an LLM for advice on a gymnastics‑ring installation, the conversation quickly moves from "which product is cheapest?" to "is this hook safe for a child?" The model starts to act as a participant in the transaction, not just a source of information. It frames risk, highlights values, and can steer the buyer toward a higher‑priced option. This shift marks the birth of agentic commerce – a system where an AI assistant not only finds products but also decides where the purchase is made and how the checkout proceeds.

---

Companies building the stack

Company	Product	Core capability	Funding / traction
OpenAI	Instant Checkout (ChatGPT)	LLM‑driven product discovery, in‑chat checkout via Stripe, optional memory of user preferences	$10 B in total funding, $1 B in 2025 Series C, 150 M active ChatGPT users
Google	Universal Cart (Gemini)	Cross‑service cart that aggregates listings from the Shopping Graph (60 B SKUs), price‑history alerts, loyalty‑perk integration	$5 B in 2024‑2025 internal R&D, 1 B daily active users across Search, YouTube, Gmail
Amazon	Amazon Assistant (store‑side)	Embedded product recommendation inside Amazon, Amazon Pay integration, merchant‑controlled ranking	$12 B in AWS‑related AI spend, 300 M Prime shoppers using voice/visual assistants
Perplexity AI	Comet Browser Agent	Open‑source browsing agent that can scrape any retailer, optional user‑owned agent	$250 M Series B (2024), recent legal battle with Amazon over access rights
Anthropic	Project Deal (Claude Opus 4.5)	Negotiation‑capable agents that can bargain on behalf of users in a closed marketplace	$4 B in funding, 80 M active Claude users

These initiatives share a common architecture, but each places the decision point at a different layer of the stack.

---

How the architecture bends the recommendation

1. Intent capture – The user’s phrasing ("for my daughter", "cheapest", "quick install") determines the initial search space.
2. Memory layer – Persistent user data (budget, safety tolerance, past purchases) biases the ranking.
3. Data access – Which retailers expose structured feeds (JSON‑LD, GraphQL) and which block bots directly affect the set of candidates.
4. Ranking policy – Platforms may weight safety, margin, delivery speed, or ad fees differently.
5. Persuasion model – The language the agent uses ("safer", "recommended for children") can nudge the buyer.
6. Checkout integration – The payment processor (Stripe, Google Pay, Amazon Pay) determines who records the sale and who receives the fee.
7. Post‑purchase loop – The agent learns from returns, warranty claims, and future queries, refining future suggestions.

At each step a hidden stakeholder can influence the final answer without the user seeing it.

---

Real‑world examples

• OpenAI’s Instant Checkout – When a user asks for a "best safe gymnastics‑ring", the model pulls data from publicly available product pages, applies a safety‑risk filter, and offers an in‑chat purchase button. OpenAI claims the merchant of record stays the same and that the user must confirm each step, but the model still decides which merchants appear in the shortlist.
• Google’s Universal Cart – The cart watches for price drops across the Shopping Graph and can automatically re‑order a product when it falls below a user‑defined threshold. Because the cart lives inside Google’s ecosystem, the recommendation is automatically tied to Google Pay and Google’s ad‑ranking signals.
• Perplexity vs. Amazon – A federal judge temporarily blocked Perplexity’s agent from scraping Amazon product pages. The case illustrates the emerging legal question: must a platform grant an external agent the same data access it gives its own assistant?
• Anthropic’s Project Deal – In a closed employee marketplace, agents using the larger Claude Opus model secured better prices and more favorable terms than those using Claude Haiku. Participants with weaker agents did not notice the disadvantage, highlighting a future inequality where model quality translates directly into purchasing power.

---

What this means for consumers

1. Ask for transparency – Request a list of sources the agent consulted and any sites it could not reach. A simple "show me the URLs I’m pulling data from" can reveal whether the recommendation is based on a full market view or a limited subset.
2. Request tiered options – Instead of "best option", ask for "budget, balanced, and premium" choices with a clear trade‑off table (price, safety rating, warranty, delivery speed, return policy).
3. Separate research from checkout – Let the agent do the research, but keep the final payment step in a human‑controlled interface for high‑risk categories (children’s safety equipment, medical devices, large‑ticket items).
4. Watch for emotional framing – When the model invokes a child, health, or scarcity, pause and verify the underlying data. The model may be correct, but the framing can push you toward a higher‑margin product.

---

What this means for brands and startups

• Machine‑readable product data is now a competitive advantage. Structured feeds that include load ratings, safety certifications, warranty terms, and clear return policies will be favored by agents that cannot parse free‑form PDFs or marketing copy.
• Agent Engine Optimization (AEO) – Just as SEO once dominated search traffic, a new discipline will emerge around making your product understandable to LLMs: clean schema, consistent naming, and verified safety data.
• Risk of “prompt‑shaped” pages – Some merchants may start to craft product pages that trigger favorable LLM responses (e.g., repeating safety keywords). Platforms will need to detect and penalize manipulative content.
• Negotiation APIs – As agents gain bargaining power, retailers that expose a negotiation endpoint (e.g., price‑adjustment APIs) could attract higher‑quality agents and win the "best‑for‑buyer" slot.
• Data‑access agreements – Companies that allow third‑party agents to read their catalog via open APIs will likely capture more traffic than those that lock down their site behind bot‑deterrents.

---

The coming conflict: who gets to own the cart?

If the decision is born inside ChatGPT, OpenAI controls the recommendation flow. If it is born inside Google Search, Gemini, or Gmail, Google does. If the cart lives inside Amazon, the retailer decides the ranking. The first party to own the point of decision can shape demand, influence pricing, and collect valuable behavioral data.

The legal battles (Amazon vs. Perplexity, Apple’s App Store restrictions on third‑party payment flows) are early indicators of a broader regulatory conversation about consumer‑agent interoperability. Future legislation may require platforms to expose a standard "shopping‑agent API" that any user‑owned assistant can call, much like the current requirement for banks to share account data under Open Banking rules.

---

Practical steps for today

1. Enable memory disclosure – Ask the agent, "What past purchases am I letting you remember for this recommendation?"
2. Demand source visibility – "Show me the exact product pages you used for the safety rating."
3. Set a verification threshold – "Only recommend items where the manufacturer’s safety certification is verifiable in a PDF or official website."
4. Use a dual‑agent approach – Run the same query in two different assistants (e.g., ChatGPT and Gemini) and compare the top three results.
5. Keep a manual fallback – For any purchase above $200 or involving child safety, confirm the recommendation by visiting the retailer yourself before completing checkout.

---

Conclusion

Agentic commerce is already moving from research demos to everyday purchases. The technology stack – LLMs, memory, structured product feeds, ranking policies, and checkout integrations – gives each stakeholder a lever to influence the final recommendation. Consumers can benefit from safer, faster decisions, but only if the underlying process is made visible. Brands that invest in clean, machine‑readable data and open negotiation interfaces will thrive, while those that rely on opaque rankings may lose market share to agents that can read the whole internet.

The real question isn’t whether AI will buy for us; it’s who the AI will represent when it does.

{{IMAGE:2}}