As of 4 November 2026, Microsoft opened a new surface for product discovery: eligible Shopify and WooCommerce merchants can now have their products appear inside Copilot, Bing and MSN when a shopper asks an AI assistant for a recommendation, with the purchase completing on the merchant's own site rather than inside the assistant. The change sits inside a wider Clarity update that also added agent activity tracking and a plain-language analytics tool. The commercial detail worth noting is smaller than the headline: Microsoft's own framing of the failure mode is that an agent can find the right product and still miss a promotion.
What Microsoft actually shipped
The update was set out in two posts published on 4 November by Nabil Naghdy, director of product management at Microsoft AI, one on the Microsoft Advertising blog and a longer one on the Clarity blog. The Advertising post opens with a claim about the state of the web, that non-human traffic has overtaken human traffic, and the Clarity post maps three products to three audiences: agents get analytics, AI recommendations get Brand Agents, and people browsing get Ask Clarity. The naming is not consistent between the two posts, which is a small signal about how early the packaging still is.
The numbers behind the launch need their dates attached. Microsoft says weekly sessions from AI agents nearly doubled in three weeks to almost two million, but that comes from an analysis of 5.23 million sessions across 230,511 projects between 13 August and 9 September 2026, not from November traffic. An earlier study published on 29 October found checkout recorded in 25 percent of agent sessions against 15 percent of human sessions. Both are measurements of what agents did in a sample, not forecasts of what they will do during the holiday season.
Discovery is not inclusion
Being eligible for Copilot is not the same as being shown, and being shown is not the same as being bought. Microsoft does not publish a ranking rule, an appearance frequency or an expected traffic lift, and the announcement says only that eligible products can appear. For a brand team, that means the new surface has to be measured the way a marketplace listing is measured, with its own baseline and its own failure modes, rather than folded into an existing traffic report.
Why promotions are the weak point
The failure Microsoft describes is specific: an agent finds the product but misses a promotion, or reaches checkout and gets stuck. That is a data problem before it is a technology problem. A promotion that exists only as a banner, an image, or a rule applied at the final step of checkout is invisible to an agent reading a product feed. If the offer is not expressed as structured, machine-readable data attached to the item, the assistant will describe the product accurately and the price incorrectly.
The same logic applies to availability. An agent that recommends a variant which is out of stock has produced a worse outcome than no recommendation at all, because the shopper arrives at a product page that cannot complete the order. Retail teams that already run price monitoring have the raw material for this: the fields they check for price violations are close to the fields an agent needs in order to describe an offer correctly.
Three data sets to check before the season
| Data set | Where the agent reads it | Failure it produces | How to verify |
|---|---|---|---|
| Offer and price | Product feed and product page | Correct product, wrong price | Compare feed value with checkout total |
| Promotion terms | Structured feed field or page copy | Discount missed or misdescribed | Recompute the discount from published rules |
| Availability by variant | Feed inventory status | Recommendation for a sold-out size | Check variant status at the moment of the answer |
The third row is the one most catalogues get wrong, because inventory is usually maintained at product level while the shopper chooses a variant. A feed that marks a product in stock while three of five sizes are unavailable will keep sending agents to a page that cannot convert. The fix is not more inventory data, it is variant-level status in the same feed the assistant reads.
A checklist that fits an existing workflow
None of this requires a new system. Price and promotion monitoring already touches the same fields, and the checks below are ordered so that each one produces the input the next one needs. Running them before the peak season is cheaper than discovering the gaps through a drop in conversion that nobody can attribute.
One caveat belongs with the checklist. Microsoft has not defined eligibility, has not described ranking, and has not said what Brand Agents costs, so the only reliable measure of performance is your own before-and-after comparison on the same prompt set. Treat vendor statements about visibility as claims to be tested, not as outcomes.
The competitive context explains why this is being pushed now. Google is putting checkout inside Gemini and AI Mode, Amazon has opened its search to outside stores, and payment testing has found shopping agents failing a large share of their runs. Every one of those failures is a data failure at the merchant end as much as a model failure at the platform end, which is why the practical work sits with catalogue teams rather than with marketing.
FAQ
Does appearing in Copilot require a separate integration?
Microsoft says merchants already running Brand Agents need no extra catalog upload, and new merchants can choose discovery only during setup. Eligibility rules and covered countries are not published, so coverage should be tested rather than assumed.
Why would an agent miss a promotion that a shopper can see?
Because the promotion exists only as visual merchandising or as a rule applied at checkout. An assistant reading a feed sees structured fields, so anything not expressed as data attached to the item is effectively absent.
What should a brand measure first?
Variant-level availability and the gap between feed price and checkout total. Both are cheap to check, both produce visible failures, and both are within the brand's control rather than the platform's.
Is the two million weekly agent sessions figure current?
No. It comes from sessions recorded between 13 August and 9 September 2026. The trend is real, but the number describes that window, not November traffic.
