When the customer is an AI agent: preparing websites and catalogs for agentic commerce before 2027

You open Google Analytics and see a strange session: direct access to a product page, reading JSON-LD, calling a price API, then exiting in 3 seconds. No scrolling, no banner clicks, no video reviews. That's not a spam bot. It's an AI agent price-checking on behalf of your customer — and it just removed you from the list because your catalog lacks the "material" attribute and the schema price mismatches the actual price.

This scenario is no longer science fiction. By 2026, approximately 73% of consumers will have used an AI assistant somewhere in their buying journey. AI platforms account for ~1.5% of US retail ecommerce sales — about $20.6 billion, nearly 4 times that of 2025. McKinsey estimates the global agentic commerce opportunity at $3–5 trillion by 2030. The customer journey has gained a new persona: a machine shopping for a human.

Infrastructure for shopping agents has taken shape

A year ago, "AI shopping" meant ChatGPT suggesting a link and the user clicking it themselves. Now it's different:

• Universal Commerce Protocol (UCP) launched by Google with over 20 partners — an open standard for agents to discover products, check inventory, and place orders without needing to scrape HTML.
• ChatGPT enables direct purchasing from Etsy right within the conversation, with Shopify following close behind. Users do not leave the chat box.
• MCP (Model Context Protocol) after one year has reached 97 million SDK downloads per month and over 10,000 production servers — becoming the default integration layer for agents to call tools, including "catalog lookup" and "ordering" tools.

The common denominator of all three: agents do not "view" your website. They read structured data, call APIs, and compare dozens of shops in seconds. Beautiful interfaces, promotional popups, parallax effects — all are invisible to this new persona.

Agents select brands based on data quality

When the customer is a human, you compete with images, price, and reviews. When the customer is an agent, you compete with three highly technical things:

1. Clean structured data

Schema.org Product with full offers, price, priceCurrency, availability, sku, gtinNot just "adding enough for SEO" but must be accurate and synchronized with reality. The Agent compares the price in JSON-LD with the checkout price; a single discrepancy loses trust, and the model will remember.

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Ghế công thái học E-Dra EEC220",
  "sku": "EEC220-BLK",
  "gtin13": "8938527211234",
  "material": "Lưới + khung thép",
  "offers": {
    "@type": "Offer",
    "price": "2990000",
    "priceCurrency": "VND",
    "availability": "https://schema.org/InStock",
    "priceValidUntil": "2026-08-31"
  }
}

2. Full attribute catalog

Buyers can zoom in on images to guess the material. Agents cannot. If a customer asks for "mesh chair for 1.8m person, load capacity over 120kg", the agent filters by attribute — but the product has not declared them. weightCapacity will be rejected outright, even if they actually meet requirements. A Catalog missing attributes is not "suboptimal," it is non-existent in the eyes of the agent.

3. Stable price–inventory API

Agent checks inventory before committing an order. Slow endpoints, 500 errors during peak times, or caching old prices for 6 hours — for a buyer, it's an annoyance; for an agent, it's an immediate reason to switch to a competitor, there is no "let's just come back tomorrow."

GEO replaces SEO at the discovery layer.

The top of the funnel is changing. AI assistants will handle about 1/4 of global searches in 2026 (Gartner forecasts more than half by 2028). The question is no longer "what is the Google rank" but "is the brand mentioned in the AI's answer" — a field called GEO (Generative Engine Optimization).

Notable point: LLM traffic is low volume but high quality. ChatGPT referrals convert at 15.9% compared to 1.76% for organic search; Ahrefs records a 24:1 value ratio. Users coming from AI answers have already completed the comparison step and visit the site almost exclusively to close the deal.

What needs to be done for GEO is not unfamiliar: content that answers questions directly, includes clear data and sources, has full structured data, and fast-loading pages that do not block AI crawlers. The difference lies in measurement — instead of tracking keyword ranking, you track the frequency of brand mentions in ChatGPT/Gemini answers for industry-specific purchasing queries. Tools already exist for this, or you can write a script to query periodically and count mentions.

Open up or block — and who handles after-sales?

There is another strategic decision that is not purely technical. Walmart chose to build its own AI experience, being cautious with external agents; Shopify chose to open up, allowing merchants to appear in ChatGPT. Both have their reasons:

For most Vietnamese businesses selling online — lacking the capacity to build a "proprietary AI" like Walmart — the practical answer is to be open, but with controlled openness: APIs with rate limits and authentication, consistent pricing declarations, and clear terms for orders originating from agents.

And don't forget the aftermath: when an order comes from ChatGPT, the buyer has never visited your site. They don't know the return policy, don't see the hotline, and have no account. The order confirmation email becomes the first touchpoint with a real human — it must carry the brand introduction, order tracking instructions, and support channels. The post-purchase experience becomes the only place where you still "meet" the customer.

Conclusion

For ten years, Vietnamese e-commerce has optimized for human eyes: larger banners, more videos, more frequent flash sales. The new persona has no eyes. It has a parser. Before 2027, the most worthwhile task is not adding another advertising campaign, but cleaning up product data — because in agentic commerce, the catalog is the storefront, and the schema is the facade.