Your Next Customer Might Not Be a Person

Quick question for anyone running a store: what if the next customer who finds you isn't human?

Not in a science-fiction sense. In a boring, already-happening sense. Someone opens ChatGPT and asks for "a good linen shirt under £120 that ships to the UK." Perplexity gets asked to compare three moisturisers for sensitive skin. Claude gets asked which running shoe suits a flat-footed marathoner.

In each case an assistant reads a slice of the web, synthesises an answer, and names two or three products. The buyer never sees a results page. There is no page two to be on.

Most Shopify stores are effectively invisible in that moment. Not because the products are wrong — because nothing on the site is readable by a machine that is trying to recommend it.

What actually changes

Traditional SEO is a competition for position on a page of links. You optimise, you rank, a human scans ten results and picks one.

AI search collapses that. The assistant does the scanning and returns a conclusion. You are cited in it or you are absent from it, and being 11th has exactly the same value as being 400th.

That changes what the work is:

• Structure beats keywords. An assistant is trying to extract facts it can stand behind. Ambiguous marketing copy gives it nothing to quote.
• Product data depth matters more than it used to. Materials, dimensions, care instructions, use cases. The stuff most stores treat as an afterthought is precisely what gets synthesised.
• Entity clarity is the whole game. The system needs to resolve "who is this brand, what do they sell, are they credible" before it will put you in an answer. If that resolution is fuzzy, you lose by default.

The llms.txt question

A lot of people have latched onto llms.txt — a clean Markdown file at your site root that describes what your site is and where the important things live. Think of it as a plain-language guide for a machine that has landed on your domain with no context.

I implement it. I also want to be straight about what it does.

Adoption is uneven. No major assistant has committed to it as a ranking input. Anyone telling you that dropping one file at your root fixes AI visibility is selling you something.

What it is good for: it costs almost nothing, it does no harm, and writing it forces a genuinely useful exercise. Stating clearly what your site is, in a few hundred words, without marketing language, is harder than it sounds — and most teams discover their positioning is muddier than they thought.

The heavier lifting lives elsewhere. Specifications and product data written to be parsed rather than skimmed. Entity clarity — who you are and what you sell, resolvable without guesswork. Render paths that don't hide your content behind JavaScript an agent won't execute. Schema has its place — wired properly it earns you rich results — but piling on more markup doesn't buy citations, and the evidence increasingly says so.

How you'd know it's working

Here is where I'd push back on most of what's being marketed right now: there is no rank tracker for this. Anyone selling you a clean dashboard of "AI rankings" is overstating what currently exists.

What you can actually watch:

• Whether your brand appears in assistant answers for the prompts your buyers really use — checked manually, repeatedly, across assistants
• How often you appear relative to the competitors showing up in the same responses
• Referral sessions arriving from AI surfaces, instrumented properly so they don't get misfiled as direct traffic

That's a real signal. It's also slower and messier than a rankings chart, and it moves unevenly across assistants. Expect to be reading something directional at around three months, not in week two.

Why bother now

Because it compounds, and because it's early.

The structural work — entity clarity, product data, clean render paths — is the same work that makes you legible to conventional search and to humans. It doesn't become worthless if AI search plateaus. And the brands establishing clear entity signals now are the ones that get cited as these systems mature.

The uncomfortable version: your competitors' catalogue might already be easier for a machine to recommend than yours. Not because they outspent you. Because someone wrote better structured data.

We build this into storefronts as part of our AI Search / GEO program, and it's one of the things we look at in a GTM + technical audit sprint.