What Is Agentic Commerce?
Agentic commerce is online shopping where an AI assistant does the searching, comparing and sometimes the buying on a customer's behalf. The shopper states a need in plain language, and the agent returns a short list, or completes the order, without the person ever browsing a shop.
The shift matters because a shortlist is not a search results page. A search page shows ten links and lets the customer decide. An agent shows two or three options and quietly discards everything it could not read.
Traffic that arrives from AI assistants is now growing faster than any other channel in retail. Adobe reported that AI-sourced traffic to United States retail sites grew 393% year over year in the first quarter of 2026, after 693% growth over the 2025 holiday season (Adobe, 16 April 2026).
How Does an AI Shopping Agent Actually Find a Product?
An agent works in three steps: it reads structured product data, it filters against the shopper's stated conditions, then it either hands the customer to your checkout or completes payment through an agreed standard. Nothing in that chain looks at your homepage design.
Step one is the product feed. The agent needs machine-readable fields: name, price, currency, stock status, size, colour, shipping and returns. Marketing prose is not a field.
Step two is the filter. If a shopper says "under HK$800, free returns, ships within three days", the agent drops every product where those three facts are missing. Missing is treated the same as failing.
Step three is checkout. Two open standards now define this. The Agentic Commerce Protocol (ACP) was released by OpenAI and Stripe on 29 September 2025 under an Apache 2.0 licence, and its stable specification is dated 17 April 2026 (ACP repository). Google launched the Universal Commerce Protocol (UCP) on 11 January 2026 with Shopify, Etsy, Target and Walmart among the co-developers (Google, 11 January 2026).
Under both standards the shop stays the merchant of record. You keep the customer, the order data, the fulfilment and the returns.
Is Agentic Commerce Real Yet, or Still Hype?
Discovery is real and measurable today. Autonomous buying is not. The honest position in August 2026 is that AI agents are already deciding which shops a customer sees, while the number of purchases completed entirely inside a chat window remains small.
The evidence for discovery is strong. Salesforce, drawing on behaviour from 1.5 billion shoppers, found that agentic search as the first step of a shopping journey grew 200% year over year, while discovery through a brand's own website fell 7% and through traditional search fell 15% between August 2025 and May 2026 (Salesforce, 28 July 2026).
The evidence against full autonomy is equally clear, and it comes from OpenAI itself. On 24 March 2026 OpenAI wrote that the first version of Instant Checkout "did not offer the level of flexibility that we aspire to provide", and moved to letting merchants use their own checkout while OpenAI concentrates on product discovery (OpenAI, 24 March 2026).
The practical reading for a small shop: the urgent problem is being findable, not being buyable.
What Does This Mean for a Hong Kong Business Right Now?
Hong Kong customers have already adopted AI assistants for shopping faster than most owners realise, but they stop short of letting the agent pay. That gap defines exactly where a local shop should spend effort.
74% of Hong Kong consumers have used AI assistants for shopping, while 45% remain uncomfortable letting an AI complete the purchase, even after reviewing the product and price themselves (Adyen Index Hong Kong 2026, published 9 July 2026, based on a YouGov survey of 1,026 Hong Kong consumers).
Among Hong Kong Gen Z shoppers, 26% use an AI assistant for shopping daily, against 17% of Gen X and 2% of Baby Boomers.
One availability fact saves a lot of wasted effort. Google's UCP checkout is currently limited to the United States, Canada and Australia, for selected merchants only (Google Merchant Center Help). A Hong Kong shop cannot switch on agent checkout there today even if it wanted to.
So the work available to a Hong Kong owner this month is the discovery half. That half is free to fix and does not depend on any protocol.
What Makes a Product Page Readable to an AI Agent?
Readable means the facts sit in labelled fields that software can lift without guessing. Adobe measured the average United States retail product page at just 66% machine readability, meaning roughly a third of product content is invisible to a language model, while homepages scored 75%.
Shopify's own guidance for merchants names four checks, none of which requires a developer to explain (Shopify, 2 April 2026):
--- Fill in every structured product field rather than describing the product only in prose.
--- Submit the catalogue to Google Merchant Center so the data is available in a standard format.
--- Add ecommerce schema markup to product pages so price, stock and shipping are labelled.
--- Check that robots.txt actually permits AI shopping assistants to read the site.
Two specific traps are worth naming. The same shirt in five colours listed as five unrelated products confuses an agent comparing options. And product data trapped inside JavaScript or a display template may look perfect to a human and be unreadable to a crawler.
Agents also read the boring pages. Returns policy, shipping times and FAQs are inputs, not decoration. If a shopper asks for "only shops with free returns" and your policy is not written down anywhere machine-readable, you are excluded before anyone compares your price.
What Does This Look Like in a Real Hong Kong Shop?
The abstract idea becomes concrete the moment you follow one shopper's question through to the shortlist. Two ordinary Hong Kong scenarios show where an agent silently drops a business.
Scenario one: a Sham Shui Po lighting supplier. A customer asks an assistant for a ceiling lamp under HK$1,200 that ships within a week and can be returned. The supplier stocks four suitable lamps. But the prices sit inside a photograph of a price list, the delivery time is written only on a separate About page, and there is no returns policy on the site at all. The agent cannot confirm any of the three conditions, so it recommends two competitors instead. Nobody rejected this shop. It was never a candidate.
Scenario two: a Kwun Tong supplier of office chairs. Its ten chairs are listed as ten separate products with no shared model information, so an assistant asked to compare fabric options treats them as ten unrelated items and compares only one of them against rivals. The fix is not new software, it is filling in the variant fields the platform already provides.
In both cases the shop loses without ever seeing the enquiry. That is the uncomfortable part of agent-mediated discovery: there is no bounce rate for a customer who was never shown your page.
A useful habit is to test it yourself. Open a mainstream AI assistant, describe the product you sell exactly as a customer would, add two real constraints such as a budget and a delivery time, and read what comes back. Whichever shops appear have written down the facts an agent needs. Whichever do not appear, including possibly yours, have not.
Doing that test on three of your own product categories takes about fifteen minutes and produces a clearer picture than any report can.
Three Misconceptions Worth Correcting
The costliest beliefs about agentic commerce are not technical. They are assumptions about how ranking, ownership and timing work.
Misconception one: paying a fee buys better placement. OpenAI states that Instant Checkout items are "not preferred" in product results and that results are organic and unsponsored, ranked on relevance. Google says the same, that choosing its native integration "does not influence how we rank your offers".
Misconception two: you lose the customer and the data. Under both ACP and UCP the shop remains merchant of record, keeping the customer relationship, order data, fulfilment and returns.
Misconception three: a good-looking website is enough. A page can be beautiful to a person and 66% legible to a machine. These are two different jobs and only one of them has been getting attention.
Frequently Asked Questions
Do I need to join a protocol to benefit from this?
No. Clean structured product data helps agents find you regardless of protocol, and it is the only half a Hong Kong shop can act on today given UCP checkout is limited to three countries.
How much does agent checkout cost a merchant?
OpenAI says only that merchants pay "a small fee on completed purchases" and has never published a percentage. Google has published no merchant fee for UCP checkout. Treat any specific percentage you read elsewhere as unverified.
Is my Shopify or WooCommerce store automatically ready?
Partly. Shopify states that since March 2026, products in Shopify Catalog sold to United States shoppers are automatically discoverable in ChatGPT. Other channels are per-channel settings, and other platforms require the feed and schema work yourself.
Should a service business care, or is this only for retail?
The same mechanics apply to any business an assistant might shortlist. The structured facts differ, service area, price range, opening hours, languages spoken, but the failure mode is identical: unstated is treated as unavailable.
How would I know if agents are already sending me customers?
Check the referral sources in your analytics for AI assistant domains. Adobe's and Salesforce's figures are industry aggregates; your own referral data is the only number that describes your shop.
The Takeaway for Hong Kong Owners
Agentic commerce is not a future purchasing channel you must join. It is a present-day reading test your shop is already sitting, and most shops are quietly failing the parts they never knew were graded.
The action this week costs nothing. Open your three best-selling product pages and ask a simple question of each: if a machine could only read the labelled fields, would it still know the price, the stock, the shipping time and the returns policy? Where the answer is no, that is the whole job.
We understand AI. UD stands with you. Twenty-eight years of watching Hong Kong businesses adapt to new channels says the same thing every time: the ones who prepare early are not the ones with the biggest budgets, they are the ones who checked first.
Reviewed by the UD AI team.
Find Out How AI Sees Your Business
Before you change anything, it helps to know what AI assistants can currently read about your business and where they stop. We will walk you through it step by step, from the first check to the fixes that matter most.
