For fifteen years, ecommerce visibility had two steps. Fall '26 put a third one in front of everyone at once, and it behaves nothing like the first two.
Getting the coordinate system right matters here, because the instinct is to file new mechanics under old categories — and this one does not fit.
Key takeaways
- Step one, retrievable: get indexed, ranked, clicked. Referee: the crawler. Failure decays over weeks.
- Step two, citable: get mentioned and cited in AI answers. Referee: the model. Failure is gradual.
- Step three, callable: an agent queries you, parses you and closes the order without leaving the chat. Referee: one tool call. Failure is binary.
- The three stack rather than replace each other, but their failure modes have nothing in common.
- Step three is the only one you can monitor with a boolean instead of a trend line.
Step one: retrievable
The objective is an index entry, a ranking and a click. The assets are pages, links and Core Web Vitals. The referee is a crawler.
Its defining property is gradualness. Rankings slip, traffic erodes, and you have weeks or months to respond. Everything about SEO practice — monthly reporting, incremental testing, patience — is shaped by that gradualness.
Step two: citable
The objective is to be mentioned and cited inside an AI answer. The assets are fact density, structured expression and source distribution. The referee is a model.
This is the layer most GEO work currently addresses, and it is still gradual. You are absent from an answer, you improve the sources and the content, and weeks later you appear. Metrics look like share of voice and mention rate — see share of model for how we measure it.
Step three: callable
The objective is no longer appearing in an answer. It is that an agent can find you in a tool call, read a structured product object, build a cart and complete payment without the buyer leaving the conversation.
The assets are catalog fields, protocol capabilities and settlement eligibility. The referee is a single tool call, and the outcome is binary: it returns your product or it returns nothing.
Step | Success looks like | Failure looks like | Repair time |
|---|---|---|---|
Retrievable | Keywords rank, clicks arrive | Rankings slide, traffic decays | Weeks to months, incremental |
Citable | Mentioned and cited as a source | Absent from the answer, or described wrongly | Weeks, depends on sources and content |
Callable | The tool returns your product and payment can complete in-chat | The tool returns nothing, or only a redirect. No middle state | Hours to fix, but zero until you do |
Why binary changes the operating model
An SEO practitioner says "we are in position seven". A GEO practitioner says "our mention rate is 32%". Someone working on step three can only say one of two things: an agent can buy from us, or it cannot.
Three consequences follow.
Monitoring becomes a boolean. A trend line of a value that is only ever 0 or 1 is a waste of a chart. These belong in an uptime dashboard with an alert, not in a monthly deck.
The work becomes engineering. The fix for a failed tool call is a configuration change, a field that was never filled, or a missing capability — not better content. Ownership moves accordingly, which is why the role split is part of the work rather than an afterthought.
There is no partial credit. You cannot be 60% callable. A store with excellent product data and an empty payment_handlers is in the same position as a store with no data at all, on this particular path.
The three steps stack
This is not a replacement narrative, and anyone selling it as one is overreaching.
An agent generally has to know your brand before it searches your name — that is step two doing its job. The product URL an agent returns is eventually opened by a human, and that page still has to rank, load and convert — step one. Step three decides whether the transaction can close inside the conversation.
What changes is not the importance of the earlier steps but the order of repair. If step three is broken, improving steps one and two makes an unreachable interface nicer. Fixing step three is usually an afternoon of configuration. Start there.
How to find out where you stand
Three commands tell you whether step three works at all: are /agents.md and /llms.txt returning 200, is payment_handlers populated, and does a broad category term in a tool call recall your own products. The full set is in the ten-minute readiness audit.
Most brands we probe land in one of two squares: good data behind a broken protocol layer, or a reachable store with product data an agent cannot use. The first is a configuration problem. The second is a product data problem. They need different teams and different weeks.
FAQ
Does step three apply outside Shopify? The pattern does; the implementation differs. Shopify ships protocol support as a platform default, which is why the gap between a Liquid store and a badly-audited headless build is so visible. On a self-built stack, every layer is your engineering budget.
Is step three where the revenue is today? Not yet, for most brands. Agent-sourced volume remains small next to organic search and paid. The argument for acting now is cost asymmetry: the fix is cheap and one-off, while absence during a channel's formative period is not recoverable later.
Does this make classic SEO less important? No. It adds a layer with a different failure mode. GEO vs SEO covers how the first two relate; this article adds the third.