Microsoft Now Lets You Watch an AI Agent Try to Shop Your Store

On 4 November Microsoft added three things to Clarity, its website analytics tool, under a title that is the whole story in six words: agents act, LLMs recommend, people browse. Two of the three matter if you own a product catalog. One runs an AI agent against your own store and reports whether it finished and where it stalled. The other publishes a Shopify or WooCommerce catalog into Microsoft Copilot, Bing and MSN, where Microsoft's assistant recommends products from it.

A quarter of agent sessions go to checkout

Both tools land on the back of a study Clarity published on 29 October. It covers more than 5.2 million sessions Clarity classified as agentic, across 230,511 of its projects, between 13 August and 9 September, with agents detected from the OpenAI, Google and Anthropic ecosystems. The 4 November post rounds the trend into one line: weekly sessions from AI agents nearly doubled in three weeks to almost 2 million.

What the agents did once they arrived is the part to read twice. Checkout was the most common action in agent sessions, at 25.21 percent. In human sessions it was 14.65 percent, and the top human action was an outbound click, at 26.99 percent. Clarity's own reading is that agents do the same things as people, weighted toward the task they were sent to finish.

Two caveats, both Clarity's. It cannot yet split the agent's part of a session from the human's, so these are sessions involving agents, not proof of autonomous purchases, and events are counted independently rather than as funnels, so a checkout event is not a completed order. It still says something the holiday forecasts could not: the agents already on your site are not browsing.

A rehearsal before the real agent arrives

The first new tool is a dashboard Clarity calls Agent Activity, generally available since 29 October for any project that receives agent traffic. It shows how that traffic changes over time, which agents are visiting, and what they do, from searches and logins to appointment bookings and checkouts.

The second is the one to go and try. Simulation assessments send an agent through a key task on your site and report whether it completed the task, how long it took, how many interactions it needed, and what might be improved. Microsoft says the first scenarios are commerce and travel, that the supported agents come from Google, OpenAI and Anthropic, and that the capability is available globally to all Clarity users. If an agent needs three times the interactions a person does to reach your cart, something between the product page and the cart is in its way: a variant picker it cannot read, a popup it cannot dismiss, a checkout field it cannot fill. Now you can find that out on a Tuesday afternoon instead of when a shopper's agent gives up.

WooCommerce gets a seat at Copilot's table

The third addition extends Brand Agents, which began in January as an on-site assistant for Shopify stores and, per Microsoft Advertising's April post, was already expanding to WooCommerce. As of this week, eligible WooCommerce and Shopify merchants can have their products appear in Copilot, Bing and MSN when a shopper asks Microsoft's AI for a recommendation. The purchase completes on the merchant's own site. Microsoft says the agent is grounded in the live product catalog, the shipping and return policies, and the brand's voice, and that a merchant can run it on the storefront or only inside the AI experiences.

The documentation explains the mechanism. Accepting the Brand Agents terms publishes a copy of the catalog to Microsoft Merchant Center, which is what makes it visible to Bing, Copilot and what the docs call other Bing-grounded AI assistants. Brand Agents is still in beta, works for websites and not mobile apps, and asks merchants to add a specific disclosure to their privacy policy. On WooCommerce the setup is the Clarity WordPress plugin, an approval screen, and a publish button with a traffic percentage on it. Microsoft's post does not define eligible, so expect to find out by trying.

What the simulation cannot tell you

A simulation measures the journey. It does not read the record. An agent can sail through your variant picker, your cart and your checkout and report a clean completion, and none of that tells you whether the product it carried had anything worth recommending written on it. The catalog that goes to Merchant Center is your titles, descriptions, images and policies. Copilot recommends from that copy. A candle described as “smells amazing” passes the simulation and loses the recommendation to the candle whose description says soy wax, 45 hours, cedar and orange peel, made for a small room.

1. Run one simulation on your best-selling product before anything else, and read the interaction count against what a person needs.
2. Open the Agent Activity dashboard and note which agents are already visiting.
3. Before accepting the Brand Agents terms, read ten of your product descriptions with the photos hidden. That is the copy Merchant Center receives.
4. Decide whether the agent runs on your storefront, only in Copilot and Bing, or both. The privacy disclosure applies either way.

Scoring the half the simulation skips

This is where CortexCommerce fits, as the other half of the test. The free scan gives every product page, and the store as a whole, an AI-visibility score from 0 to 100, measured by crawling your live storefront the way an assistant's crawler does, never a simulated number. The score is built from six categories: whether GPTBot, ClaudeBot, PerplexityBot, Google-Extended and Bingbot can reach the page at all, Product structured data, the title and meta description, canonical and Open Graph tags, image alt text, and whether the description says enough to be matched to a shopper's question. Bingbot earns its place on that list this week, because a Bing-grounded assistant is exactly what Copilot is.

Clarity's simulation tells you whether an agent can finish on your store. The score tells you whether it found anything worth recommending on the way. Enriched products are re-scanned, so the score shows the before and after for every product you fix against the rest of the catalog, and you start with the lowest-scoring pages instead of the ones you happen to remember. Reading 3,000 product pages the way a crawler does is not a job for a person. A scan does it for every page and ranks them.

We cannot tell you whether Copilot will recommend your products, and neither can Microsoft. What we can say is that when an agent gets through your checkout in record time, it will have had something to say about what it was buying. To see how your product pages score today, run a free scan at cortexcommerce.ai/scan and start with the pages at the bottom of the list.

Sources

1. Microsoft Clarity blog: Agents act, LLMs recommend, People browse, Clarity connects all three (4 November 2026)
2. Microsoft Clarity blog: What AI agents do on websites, insights from 5.2 million sessions (29 October 2026)
3. Microsoft Learn: Brand agents overview (catalog discovery and Microsoft Merchant Center)
4. Microsoft Learn: Install Brand agents from WooCommerce
5. Microsoft Advertising blog: Win across all three eras of the web (21 April 2026)