WordLift’s Andrea Volpini wants markup to work as memory for AI agents

Markup is usually added as if only a search engine would ever read it: drop in Product or Organization, check it in the testing tool, forget it.

Andrea Volpini, the founder of WordLift, offered a different view at the W3C and GS1 workshop "E-commerce for Humans and AI Agents". According to the talk, as recounted on the WordLift blog, structured data for AI agents should become a memory an agent can move through.

The full text of the talk was not available to us at the time of publication, so what follows is only what the announcement contains, plus our own explanation of it.

Markup that remembers

The workshop took place in Zurich as a hybrid event, co-organised by GS1 — the body behind barcodes and global trade item numbers (GTIN).

For e-commerce that is no accidental partner: GTIN is what ties the same product together across shops, feeds and catalogues. When a talk about AI agents runs under a W3C and GS1 banner, the conversation is plainly about product data and its standards, and it reaches beyond SEO.

The "memory" idea is simple once translated into practice. An agent buying or comparing products does not read every page from scratch.

It needs stable identifiers, relations between entities (product, brand, manufacturer, seller) and the ability to come back to those facts later. A single tag on a single page does not provide that. An entity graph, where every thing has its own address and links to its neighbours, does.

schema.org, WebMCP and the graph beneath them

The talk names three layers. The first is schema.org, the vocabulary a page uses to say what it describes. The second is WebMCP, an initiative meant to let a site show a browser agent which actions are available here. The third is the entity graph underneath both: it is what turns separate descriptions into a connected picture.

How exactly the layers fit together is set out in the talk, and we will not paraphrase that part rather than invent detail.

What can be taken from it right now without any assumptions:

• markup on every page should point to the same brand and product identifiers, because varied spellings confuse an agent;
• GTIN, manufacturer and brand should match across the markup, the Merchant Center feed and the product card;
• relations between entities are better stated explicitly than left for an agent to infer.

For a Ukrainian online shop this shows up most clearly on marketplaces: the same product has several listings there under different names, and an agent struggles to see that they are one thing.

The basics of markup and its typical mistakes are covered in our guide to schema.org and structured data, and how AI answers pick their sources is covered in our AI Overview piece.

Identifiers first, WebMCP later

WebMCP is still more of a direction than a tool you can implement over a weekend. Quick traffic results are not to be expected from it.

Order in your own identifiers, on the other hand, can be established today, and all three layers depend on it.

That is exactly where I'd start, whatever becomes of WebMCP: reducing brand and product names to one spelling across markup, feed and product card is a few days of work that pays off in any scenario.

I am curious whether agents will really use these graphs, or whether, like search engines with markup before them, they will mostly lean on the text of the page.

Related reading: SGE and AI Overview: Optimising for AI Search · SEO Website Promotion