How AI Assistants Choose Which Stores to Recommend

The shift from links to recommendations

For two decades, being found online meant ranking in a list of blue links. That is changing. When a shopper asks ChatGPT, Claude, Perplexity, or Gemini what to buy, the assistant does not return ten links. It names a few specific stores and says why.

The behavior is already mainstream, and the channel is growing fast.

That raises a practical question for any store owner. When an assistant picks which stores to recommend, how does it choose, and can you influence it? To find out, we did not guess. We looked at the brands AI shopping programs already surface.

The question, and how we studied it

We assembled a set of brands with publicly announced AI-commerce partnerships, confirmed through engine, platform, and corporate primary sources rather than marketing blogs. These are brands the major AI shopping programs, run by OpenAI and Google, have already committed to surfacing.

Then we read their public storefronts the way an AI crawler would, and compared what each one exposes. We were not looking for who ranks highest. We were looking for the shared structure: what these brands have in common that a machine can read.

Two honest limits. We observed public signals only, not the private feeds and data deals that large retailers use behind the scenes. And a scan is a snapshot, accurate for the moment we looked, not a permanent grade.

Finding 1: stores reach AI through three different paths

The most important finding is that there is no single path to being readable by AI. The brands assistants recommend get there in three structurally different ways. A useful audit recognizes whichever path a store has taken instead of demanding all three.

For an independent store, Paths A and C are the lanes you control. Path B belongs to retailers with private data deals. That distinction matters for how you read any AI-readiness score, including ours.

• Path A, public and readable. The store writes its product and business facts directly into the page in a format machines read, sometimes alongside the agent endpoints platforms like Shopify now expose for eligible merchants. This is the most measurable path, and where most independent DTC brands live.
• Path B, private feeds and deals. Large retailers feed their catalog straight to AI partners through private pipelines. Their public pages can look thin to a crawler while the real integration happens out of view. This path is open mainly to enterprises, and no public scan can measure it.
• Path C, signals hidden in code. The same useful facts are present, but tucked inside JavaScript or rendered in a way that takes real work to read. Headless and custom storefronts often sit here. The data is public, just harder to reach, so a naive checker misses it.

Finding 2: most popular AEO advice did not hold up

The reverse-engineering also falsified a lot of popular advice. Several widely repeated claims simply did not match what the recommended brands actually do.

• Knowledge-graph links are not load-bearing. Zero of the audited AI-partner brands link to Wikipedia, Wikidata, or Crunchbase. Their identity links are social accounts. Chasing knowledge-graph anchoring is mostly wasted effort.
• You do not need to name every AI bot in robots.txt. Almost all the brands rely on a normal wildcard rule with no blanket block. Requiring an explicit per-bot allowlist is stricter than reality.
• Structured data alone is not enough. AI engines missed product facts that lived only in hidden markup when the visible page did not repeat them. Some flagship partners even ship sloppy schema, a placeholder image here, a wrong brand field there. The visible page has to carry the same facts.
• llms.txt is not moving shopping visibility today. Crawler-log studies show the major shopping bots fetch it in near-zero volume. It is a forward-looking convention, not a current driver.
• FAQ schema is oversold. No flagship partner depends on it, and assistants read plain prose answers fine. It is a nice-to-have, not a requirement.

Finding 3: what actually matters is unglamorous

So what does matter? The signals that showed up consistently are concrete and a little boring. They are about making your real product facts easy for a machine to read, on the page, not buried.

There is a content angle too. Clear, specific, well-sourced product and policy content helps an assistant quote you.

These signals are about plain public readability, so they help with any assistant that reads your pages, whether that is ChatGPT, Claude, Perplexity, or Gemini. They map to the seven dimensions our scanner measures, and the full per-check rubric is public on our methodology page.

• Product facts in machine-readable form on product pages: name, description, image, price, currency, availability.
• The same facts written into the visible page, not only in hidden markup.
• A robots file that does not block AI crawlers, and a sitemap that actually reaches your product pages.
• Clear business identity, and reachable trust pages for shipping, returns, and contact.
• Product attributes named in plain language: the material, size, and use case a shopper would ask about.

What a public scan can and cannot see

Any honest AI-readiness score has to admit what it cannot see. A public scan reads Paths A and C, the signals on your own pages. It cannot see Path B, the private feed deals large retailers use. So a low public score means low public readiness, not that AI can never find you.

That is also where the line between free and paid sits. Reading your public pages is something we give away. Confirming that an assistant actually names you, and where you land when a shopper asks, takes watching the engines over time, which is a different and heavier job.

One note on scope. We have validated scoring on select non-English storefronts, including Hebrew and Japanese stores, rather than claiming full multilingual coverage.

Why this matters more if you are not on Shopify

There is a reason this matters more for some stores than others. Starting in late 2025, Shopify built native agentic-commerce wiring into its platform, adopting the OpenAI and Stripe Agentic Commerce Protocol, so eligible Shopify stores got a head start they did not ask for. Everyone else, on WooCommerce, Wix, Squarespace, BigCommerce, Magento, PrestaShop, or a custom build, has to earn the same readability deliberately.

That is the gap we focus on. The work is rarely a rebuild. It is usually filling in product facts, making them visible, and keeping your pages reachable. The brands assistants already recommend are not doing anything magic. They are doing the readable, unglamorous things consistently.

The brands assistants recommend are not doing anything magic. They are doing the readable, unglamorous things consistently.

How to see where your store stands

The fastest way to know is to check. Our free scan reads your public site the way an AI crawler would and scores seven dimensions, with a specific list of what to fix. Nothing is installed and nothing on your store changes.

If you want the detail behind the score, the full rubric is on our methodology page, and the terms are defined in our AEO explainer.

Background reading: what AEO means, the full methodology, the WooCommerce guide, and how we compare to Commerce GPT.