Agentforce Commerce just went GA: what it means for your product pages

What the GA actually changed

Before July 2026, Agentforce Commerce was available in limited release. The GA opens it to all Salesforce Commerce Cloud and Einstein licences and, more significantly, wires it natively to three of the highest-traffic AI surfaces: ChatGPT Shopping, Google AI Mode and Gemini. A shopper asking "which running shoe suits wide feet for trail running under £100" on any of those surfaces can now receive a Salesforce-powered recommendation that resolves to a specific product at a specific retailer.

The agent does not browse your site. It reads the structured data you have already published: product schema, review schema, llms.txt, any feed you syndicate. It assembles a recommendation from that, then routes the buyer. If your data is thin, the agent routes elsewhere.

What the agent actually reads on your PDP

Every AI agent reading a product page is looking for the same thing a late-stage buyer looks for: evidence that the product does what it claims for a person like them. The hierarchy of evidence that agents cite, based on how retrieval systems weight trust signals:

• Verified buyer reviews with structured attribution (author, date, product variant, verified purchase flag in JSON-LD)
• Video and photo reviews that can be associated with a specific SKU
• Q&A threads where the question maps to a real objection and the answer is attributed to a real person
• Brand-written descriptions and spec sheets (lowest trust; agents deprioritise self-reported claims when third-party evidence is available)

The pattern holds across all three surfaces now integrated at Agentforce GA. A brand that has published rich, structured UGC content (video reviews tagged to SKUs, Q&A with attributed answers, a properly formatted llms.txt) is feeding the agent the exact signal hierarchy it weights most. A brand that has only a product description and a star rating is competing on the weakest evidence tier.

The PDP gap this creates

The practical problem is not that agents exist. The practical problem is that most PDPs are not set up for agent retrieval. The evidence that agents weight highest (structured UGC, attributed Q&A, variant-level review data) is also the evidence that is hardest to publish correctly without a platform that generates it in the right format.

How UGC is the answer, and why format matters

UGC is the right answer to the Agentforce retrieval problem for a specific reason: it is the only type of content on a PDP that reads as independently verified. An agent trained to surface trustworthy evidence over promotional copy will consistently prefer a real customer's tagged video over a brand-written feature bullet, because the video is third-party by origin.

The format caveat matters as much as the content itself. A UGC gallery of embedded Instagram clips that only loads client-side is invisible to the agent. Review content that lives in a JavaScript framework without a structured fallback gets missed. The content has to be published in a form the agent can read: JSON-LD for reviews and Q&A, a properly formatted llms.txt for the broader content map, an MCP endpoint for agents that request structured data directly.

Three things to check before the next agent read

Structured review JSON-LD on every PDP. The Product schema block needs to include a review array with real attributed reviews (author, reviewBody, datePublished, verified flag). A star rating aggregate with no individual reviews attached gives agents nothing to quote.

UGC that is server-side readable. If your gallery only renders after JavaScript executes, agents miss it. Either serve a structured feed (llms.txt, MCP) or confirm that your UGC platform renders structured HTML that crawlers reach without JS.

A Q&A block with attributed answers. Q&A is the format that maps most directly to how a shopper prompts an agent ("does this fit wide feet", "is this suitable for outdoor use"). A structured Q&A with a real answer and a real person attached to it is the format agents quote. An FAQ section that is just a brand-written paragraph does not carry the same attribution weight.

The twelve JSON-LD shapes agents quote covers the specific markup needed to pass each evidence type through to AI retrieval. The AEO 2026 playbook has the full setup sequence if you are starting from a thin PDP.