Your Next Customer Might Be an AI Agent

On October 21, 2026, Walmart announced that its customers will soon give each AI agent that shops for them a profile of its own inside their Walmart account. The customer can require approval for every purchase, or let the agent buy within a limit such as $50, $100 or $200, and can delete the profile at any time. Walmart was blunt about why. Today agents shop its site "through a web browser, following steps designed for people," and "this can be slow and error-prone, and agents may get stuck along the way." [1]

One of the country's biggest retailers is designing accounts for customers who are software. That is the clearest signal yet that the buyer on the other side of a business's website is changing, and most businesses are still optimizing for the old one.

Three jobs a shopper used to do alone

A purchase has three stages: finding options, judging them, and paying. For twenty years, search engines helped with the first and the business owned the rest. In the last thirteen months, each stage has gained an agent rail.

Finding. Assistants now answer the question a search box used to route. In January 2026, Google announced the Universal Commerce Protocol, "a new open standard for agentic commerce that works across the entire shopping journey," co-developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by Visa, Mastercard, Stripe and others. [2] OpenAI publishes a product feed specification for merchants who want their catalog to appear in ChatGPT, with separate flags for whether an item may appear in search and whether it may be bought in the conversation, and an instruction any retailer will recognize: "Submit the current price; update the feed when a sale starts or ends." [3]

Judging. Agents now ask the questions a shopper used to ask. In November 2025, Google began letting shoppers ask it to call nearby stores about stock, price and promotions: "Now, you can save time by asking Google to call stores on your behalf." [4] Amazon reported that its shopping assistant Rufus "was used by 300 million+ customers" in 2025, and estimated it helped deliver "nearly $12 billion in incremental annualized sales." [5] That figure is Amazon's own estimate.

Paying. Then the checkout moved. In September 2025, Stripe and OpenAI launched purchases inside ChatGPT, starting with Etsy sellers, and released the Agentic Commerce Protocol. [6] Google followed with agentic checkout in Search, which buys an item when it reaches the shopper's price, after the shopper confirms. [4] On October 15, 2026, Google began rolling out Universal Cart across Search and the Gemini app in the U.S., with checkout "powered by the Universal Commerce Protocol (UCP)" and merchants remaining "the seller of record." [7]

The payment networks built the plumbing underneath. Google's Agent Payments Protocol uses "Mandates," signed records that "serve as verifiable proof of a user's instructions." [8] Visa's Trusted Agent Protocol helps a merchant recognize an agent with real commerce intent and distinguish it "from malicious automation and rogue bots." [9]

None of these rails is settled. Protocols launched in 2025 were revised within months; Google added carts, live catalog lookups and loyalty pricing to its protocol in March 2026. [10] The direction is settled, though. Google, OpenAI, Amazon, Walmart and the card networks are all building for a customer who delegates.

How big, honestly

The numbers are growing fast from a small base. Adobe measured traffic from generative AI tools to U.S. retail sites up 693.4 percent in the 2025 holiday season over the year before, and noted that the base of users "remains modest." [11] Pew found that about half of U.S. adults ages 18 to 49 use chatbots to search for information. [12] Pew did not ask about shopping, so the figure describes the habit, not the purchase.

Read it this way: most of your customers still arrive the old way, and a growing share now arrive having already asked a machine. For some categories and some ages, that share is already the one that matters.

What traditional search thinking misses

Search engine optimization trained businesses to win a ranked list of links. An agent does something different. It reads, compares, asks and sometimes acts, then returns one answer or one purchase. Four things follow.

The record beats the page

An agent reads your facts from wherever they are most structured and most current: a product feed, a merchant listing, a business profile, a page's own text. Google says so plainly. It requires no new machine-readable files or special markup for its AI features, and points businesses to the records they already keep: Merchant Center and Business Profiles "can help your products and services to be visible in both AI responses and other Google Search results." [13] A beautiful homepage with a stale price in the feed loses to a plain page with the right one.

Wrong facts now travel without you

When a shopper misread your site, you lost one sale. When an assistant restates your facts wrongly, it repeats the mistake to everyone who asks. A Minnesota solar installer, Wolf River Electric, sued Google in 2025, alleging that an AI Overview falsely said the company faced a lawsuit from the state attorney general, and listing customer cancellations it attributes to the answer. [14] Those are allegations, and the case had no ruling on the merits when we checked. The exposure is real either way. Lehvel's own September study checked 135 claims in 25 AI answers about four businesses; 64 could not be checked against the businesses' own records at all. [15]

The form is a front door

If an agent can buy, book or request a quote, your forms are part of your storefront, and agents use them differently from people. We tested this on our own site in October. A simple scripted agent that matches form labels failed to fill our booking form three times out of three, because two fields were both labeled "Email." After we gave the form real structure and distinct labels, the same script succeeded three times out of three. Two model-driven agents, from two vendors, completed the same booking (stopping before the final confirm, as instructed) on one run each. Both found our email address in the booking terms, not on the contact page. [16] Small samples; clear direction. The details a person forgives, an agent trips on.

Some callers are software

Agents already phone businesses. One of our Labs entries records an automated caller asking a building-products supplier about a window's size, lead time and price on behalf of someone it would not name. [17] The supplier's staff quickly moved from "can we answer this?" to "what are we willing to let it do?" That question now belongs in every business's policies.

What to prepare

Treat the agent as a new kind of customer with its own needs.

1. Keep one current record of your facts. Prices, availability, hours, policies, service areas. Know where each fact lives, who confirms it and when it last changed, then feed every channel from that record: the site, the feed, the listings.
2. State what an agent may do. May it get a quote? Place an order? Book? Pay? Say it in plain language on the site and enforce it where the transaction happens, at checkout, on the phone and in the booking system. Walmart's design shows the shape: identity per agent, limits per agent, approval by default. [1]
3. Make your transactions machine-legible. Real forms, distinct labels, visible contact details, clear errors. Then run an agent through your own checkout or booking flow and watch where it stalls.
4. Recognize agents before you block them. Blocking all automated traffic will, increasingly, block customers. The payment networks are building ways to tell a trusted agent from a scraper. [9] Ask your platform which ones it supports.
5. Ask the assistants about yourself. Put the questions your customers ask to the major assistants, monthly, and check each answer against your own record. Fix the source, not the answer.

The customer who never visits

Search taught businesses to compete for a visit. Agents compress the visit into a request, and sometimes skip it. The business that wins that request is the one whose facts are current, whose terms are explicit and whose transactions a machine can complete. That favors the operator who keeps good records over the one who buys the best ads, which is good news for a lot of small businesses.

Your next customer may never see your homepage. Make sure what it reads instead is true.