Shopping agents are real, and Salesforce is betting the sale still closes on your own site

Yesterday Salesforce announced what it calls its biggest Agentforce Commerce release yet. Three agents are generally available now. A Shopper Agent handles the customer conversation on a brand’s own storefront, from discovery through checkout and service. A Buyer Agent serves business buyers inside WhatsApp and text messages. A Merchant Agent does back-office work such as organizing the catalog, from plain-language requests. The release also connects to ChatGPT and Google Search, including AI Mode, with the Gemini app to follow.

The detail I would underline is where Salesforce says the sale happens. Nitin Mangtani, who runs Agentforce Commerce, said in the release that commerce will happen “predominantly on owned and operated properties.” In other words, the agent lives on your site and in your channels, and the chatbot apps send people to it.

That is a different picture from the one most people were sold nine months ago.

The first try put checkout inside the chatbot

In September 2025, OpenAI launched Instant Checkout in ChatGPT. US users could buy from Etsy sellers without leaving the chat, and OpenAI said more than a million Shopify merchants were coming. It also released the Agentic Commerce Protocol, built with Stripe, as an open standard for agents and merchants.

Walmart went in early. By November it had about 200,000 products available through Instant Checkout. In March, Walmart’s EVP of product and design, Daniel Danker, reported the result: purchases completed inside ChatGPT converted at about one-third the rate of shoppers who clicked through to Walmart’s own site. OpenAI moved toward checkout handled by the merchant, and Walmart’s plan became putting its own assistant, Sparky, inside ChatGPT, where shoppers log in and buy inside Walmart’s system.

TimelineHow agent shopping moved home

Sep 2025ChatGPT adds Instant Checkout
Nov 2025Walmart lists 200,000 items
Mar 2026Walmart: chat checkout lags
Jun 29Adobe: AI visits convert better
Jul 6Salesforce ships store agents

The traffic from AI is small and it buys

The other half of the story is what happens when the AI sends the shopper to the store. Adobe’s numbers from Prime Day week, published on June 29, show clicks from AI chat services and browsers up 235% year over year from January through May. During the four-day event, those AI-referred visits converted 40% better than visits from paid search, email and social. Last year the same traffic converted 23% worse.

Adobe is careful to say the overall volume of AI traffic is still modest. I take that at face value. This is a channel that is small, growing fast and full of people who already know what they want by the time they arrive.

Salesforce adds its own numbers to the release. It says AI influenced 20% of global online sales during the 2025 year-end holidays, or $262 billion, and that retailers running their own shopper agents grew sales 59% faster than those without. Those come from Salesforce’s own shopping data and surveys, and Salesforce sells shopper agents, so I read them as a direction rather than a forecast for your store.

Put the data sets together and the shape is clear. AI tools are good at helping people decide. The purchase works better where the merchant controls the cart, the payment, the account and the service after the sale.

Before and after

Checkout inside the chatbot

Shopper never sees your site
Walmart saw 1/3 the conversion
Harder to own the follow-up

Agent on your own site

You keep the customer record
AI visits converted 40% better
Returns and support stay with you

The research happens in the AI tool, and the sale works better at home.

What a 30-person company can copy

You do not need Salesforce to act on this. The lessons hold at any size.

Make your catalog readable by a machine. Agents, whether yours or ChatGPT’s, read product titles, descriptions, prices, stock and shipping cutoffs. If your product data is a mess of abbreviations and missing sizes, an agent will give wrong answers or skip you. Clean product data is the cheapest commerce AI project there is.

Put an agent on the channels where your buyers already talk. Salesforce put its Buyer Agent inside WhatsApp and text messages. If you sell to restaurants, salons, contractors or retailers, your repeat customers probably already text a rep to reorder. An agent that confirms the order, checks stock, confirms the delivery date and writes it into your system takes that load off the rep. Keep a person on anything involving credit terms or a price change.

Here is a hypothetical to make it concrete. Say you run a 30-person food distributor and your reps spend their mornings reading texts like “same as last week, plus two cases of the oat milk.” An agent with access to the customer’s order history and your stock can draft that order, confirm the delivery day and post it to your order system for a rep to approve with one tap. The rep stops typing orders and starts calling the accounts that have gone quiet. That is the kind of agent commerce a small company can run this year.

Keep checkout on your site. Let AI tools send you the shopper. Then make sure the page they land on loads fast, shows the right price and stock, and lets them pay in a few taps. Walmart’s numbers say that last step is where the sale is won.

Track AI traffic on its own line. Set up a view in your analytics that separates visits referred by AI assistants, using their referring domains. Watch that line monthly. It is small today, and Adobe’s growth figures say it will not stay small.

ChecklistReady for shopping agents

Product data is complete
Stock and prices are live
Checkout takes a few taps
AI traffic is tracked separately

What this means for the next six months

Salesforce’s own pitch is aimed at the 2026 holiday season. Mangtani’s point in the release is that brands with a Shopper Agent live on their own properties will be the ones that win it. Whether or not you use their product, the calendar is the useful part. If you want an agent answering product questions and taking reorders in November, the product data cleanup and the rules need to be done by September, and you want a few weeks of real conversations reviewed by a person before the volume arrives. An agent gets better as people correct it and you fold those fixes back into its rules, and that cycle takes calendar time. You cannot compress them into the week before Black Friday.

What to do this week

Ask ChatGPT and Google’s AI Mode to recommend a product like yours, and see whether your company shows up and whether the details are right.
Pick your 20 best-selling products and check that each has a complete title, description, price, stock level and shipping time.
Look at your analytics and find, or create, a view for visits referred by AI assistants.
List the three questions your reps answer most often for repeat buyers, and write the answers down as rules an agent could follow.
Put a date on the calendar in September to have an agent tested on one channel before holiday volume starts.