Google quietly added a new category to Lighthouse called agentic browsing. If you have spent the last year adding structured data to your store, the natural assumption is that this new score rewards it. It does not. The agentic browsing category does not look at your JSON-LD at all. It measures something different, and once you understand what, it is clear why it matters for where shopping is heading.
Reading is not the same as acting
There are two separate things an AI can do with your store. It can read it, and it can act on it.
Reading is what most AI shopping work has been about so far. Structured data, llms.txt, clean descriptions, all of it helps an assistant understand what you sell so it can mention you in an answer. That is the layer schema lives in.
Acting is newer. It is an agent actually driving your page: finding the search box, typing a query, clicking a variant, adding to cart. For that, the agent does not need to understand your catalog in the abstract. It needs to operate the actual controls on the actual page. Lighthouse's agentic browsing category scores how well an agent can do that.
What the score actually measures
The category breaks down into a handful of checks, and none of them are about schema.
- Can an agent read the controls? This is the accessibility tree. Buttons and links need names. Form fields need labels. An icon-only button with no text and no aria-label is invisible to an agent the same way it is to a screen reader. If the agent cannot tell what a control does, it cannot use it.
- Does the layout hold still? This is layout stability, the same idea as Cumulative Layout Shift in Core Web Vitals. An agent that decides to click based on where something is will mis-click if the page jumps while images load. Reserved space for media keeps the target where the agent expects it.
- Is there an llms.txt? A simple, machine-readable map of your important pages. The one check on this list that overlaps with the reading work you may already have done.
- Can an agent call tools on the page? This is the newest and least settled piece, sometimes called WebMCP: pages declaring the actions an agent can take in a structured way rather than making it guess from the markup.
Put together, the score is asking a single question. If an AI agent landed on this page and tried to get something done, could it?
Why this is awkward on Shopify
Here is the honest part. Most of what this score measures lives in your theme, not in an app and not in your product data.
The names on your buttons, whether your form fields have labels, whether your hero images reserve their space before they load, these are theme template decisions. A well-built theme handles them. A heavily customized or older theme often does not. So two stores with identical, perfect structured data can score very differently on agentic browsing purely because of their theme markup.
That also means the fix for the weakest items is sometimes a theme change, which is a bigger lift than toggling a setting. It is worth knowing that going in, rather than expecting a one-click fix.
What you can actually do
You are not powerless here, and several of these are genuinely easy.
Start with image alt text, because it helps the reading layer and the acting layer at once. Every product image, your logo, your banners. We wrote a whole piece on image SEO and alt text if you want the detail.
Make sure your important controls have readable names. If your theme uses icon-only buttons for search, menu, or cart, that is the first place an agent gets stuck. A theme that adds an aria-label to those controls fixes it.
Give your main images dimensions so the layout does not shift as they load. This is the same work that improves Core Web Vitals for human visitors, so it pays twice.
And publish an llms.txt. It is the cheapest item on the list and the one most directly under your control.
The honest state of this
Agentic browsing is early. The in-page tool-calling piece in particular is sparsely supported, and the standards are still moving. At the same time, Shopify is building agent commerce into the platform from the server side, so some of the acting layer will be handled for you over time rather than through your theme. It is not yet the case that a great agentic browsing score directly translates into sales, the way a fast page translates into conversions.
But the direction is unambiguous. Agents are going to operate stores, not just read about them, and the stores that are operable will be the ones they can complete a purchase on. Treating it as a real signal now, while most stores ignore it, is the cheap version of getting ahead.
This is exactly why we added an agent-operability check to the readiness audit in AgentReady. It reads your storefront the way an agent would and flags the accessibility-tree and layout-shift issues that hurt this score, so you can see them without running Lighthouse by hand, alongside the llms.txt we already generate for you. Reading and acting are becoming two halves of the same job, and it is worth being ready for both.