Srijith Ravikumar is a Principal Engineer at Amazon building AI-powered recommendation systems at scale. Published researcher at AAAI.

getty

Amazon's Rufus has already had 250 million shoppers ask it product questions in 2026. ChatGPT shoppers can now compare laptops side by side without ever clicking through to a vendor site. A procurement agent forms a vendor recommendation by reading six pricing pages and a security questionnaire. The entity reading product pages is no longer always a human, even though most pages are still being written for one.

As a principal engineer at Amazon, I see this shift happening at infrastructure speed. Rufus delivered nearly $12 billion in incremental annualized sales, with monthly active users up 149% year over year. OpenAI launched the Agentic Commerce Protocol with Stripe in late 2025 and onboarded several large companies within six months. Google launched the Universal Commerce Protocol at NRF in January 2026, backed by Visa, Mastercard, Stripe, Amazon and more. Visa and Mastercard rebuilt the payment plumbing for agent-initiated transactions.

I believe e-commerce is the right case study because the read-recommend-buy loop closes quickly. What's true for a product page now is becoming true for every other system an agent reads on a person's behalf.

The agent is a different kind of reader. As Shopify summarizes it, "AI agents don't browse a store the way a human does. They read structured data—product titles, descriptions, images, pricing, inventory, shipping speeds—and use it to decide what to recommend." They have machine-level patience and cross-check claims against other sources. Visual hierarchy is invisible to them.​

This undermines many of the assumptions behind a typical product detail page. Specifications are tucked behind accordions because UX research suggests human shoppers rarely read them. Product descriptions are reduced to a few benefit-focused bullets and a hero image because people tend to skim. "Best Seller" badges and review-star clusters dominate valuable screen space because they help move a hesitant buyer toward a decision. Pricing is often presented in ways that obscure the true annual cost to create an opportunity for a sales conversation. An AI agent approaches the page differently. It has unlimited patience, treats badges as weak signals, compares claims against other sources and calculates the costs the page may be trying to downplay.​​

What replaces this looks less like a brochure and more like a data source. Product attributes are exposed in machine-readable formats using Product, Offer and FAQPage schema. Descriptions are detailed enough to explain not only ideal use cases but also where the product may fall short, because transparency is a credibility signal for both humans and agents. Comparison tables present specifications in a format that can be extracted directly, potentially alongside competing products. Materials, dimensions, weight, compatibility requirements and warranty terms appear in plain text rather than being buried in downloadable PDFs. Increasingly, brands are also publishing structured product feeds and integrating directly with the protocols and catalogs agents use to discover, evaluate and purchase products.​​​

Forrester analyst Emily Pfeiffer attributed part of the failure of OpenAI’s Instant Checkout experiment to a structural limitation: scraping public web pages does not reliably capture the breadth of product data required for commerce decisions.​​

I believe that observation generalizes. Anywhere an agent reads on a user's behalf, scraping the human-facing surface is a weaker signal than the brand publishing a structured feed. Brands that don't publish get represented by whatever the scraper picks up, and by whatever a competitor published cleanly. In other words, data quality depends less on page rendering and more on whether underlying product information is exposed in structured, machine-readable formats.​

​The pattern extends well beyond product pages. A B2B procurement agent evaluating vendors considers the same factors around pricing transparency, feature completeness and operational requirements. A customer support agent answering questions from a help center is more likely to rely on content that is well-structured, clearly organized and easy to retrieve. An enterprise sales agent comparing software vendors may draw from API documentation, security questionnaires and pricing pages to evaluate competing offerings. Each of these surfaces was originally designed for human readers. Increasingly, they are being consumed by software as well.​

The strategic move is to stop treating this as a marketing problem. The agent surface is not your homepage. It is your data feed and your structured documentation.

The companies best positioned for the next decade of commerce will likely be those that treat their product catalog, support content, and pricing and contract terms as machine-readable systems first, with human-readable experiences layered on top. Most product teams have an agent reader as a formal design consideration rather than a defined persona. That gap is likely to close quickly as AI-driven discovery, comparison and procurement workflows become more common across consumer and enterprise contexts.​​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?