The search box became a chatbot, and it has never heard of you

Field notes on staying findable while the search box learns to answer back.

Ask ChatGPT to recommend the kind of product you sell. It names three brands, none of them yours.

The store ranks on Google. The reviews are real. No injustice in it, either: the model answers from what it can read, and it cannot read you.

The reflex is to write your way in

The first instinct is editorial, always. More blog posts, an FAQ page tuned for AI, an app with "visibility" in the name. Content got stores ranked for twenty years; surely content gets them cited now.

We assumed the same (we went looking for the new copywriting rules, notebook open).

Meanwhile the old front door keeps narrowing. Ahrefs ran the numbers on 300,000 keywords: when an AI Overview sits on top of the results, the top-ranking page loses 58 of every 100 clicks it used to earn. When they first measured, in April 2025, the cut was 34.5. By December it was 58. Our bet is that the line does not bend back.

So the visible channel shrinks while a new one opens. The new one has an entry ticket, and it is not written in prose.

The new front door runs on a schema

Shopify's move is Agentic Storefronts, announced at the Winter '26 Edition: products sellable inside the AI conversations themselves. By March 2026 that meant four channels: ChatGPT, Microsoft Copilot, AI Mode in Google Search, and the Gemini app. Underneath sits Shopify Catalog, a layer that, in Shopify's words, "infers categories, extracts attributes, consolidates variants, and clusters identical items" before handing products to the agents.

Note the first verb. Infers. That is a machine doing forensics on your catalog, reconstructing what the data never said outright.

Then we read the spec

We opened OpenAI's product feed spec for agentic commerce expecting a marketing document. It is a schema. Title, capped at 150 characters. Description, plain text, capped at 5,000. Availability is an enum with five permitted values. And near the bottom sit two fields named is_eligible_search and is_eligible_checkout.

Whether the new buyer can see you is a boolean.

The machine interviews your store

The mechanism, kept short: when an AI shopping surface assembles an answer, it pulls from feeds and structured product data, not from your brand voice at its finest. An attribute in a field is usable. An adjective buried in a paragraph is a guess, and Catalog's inference will guess on your behalf. That cuts both ways: inference is a favor to clean data and a rumor mill for messy data.

Shopify's own launch notes are plain about it: these surfaces favor structured data, clean attributes, and accuracy in real time.

There is even a mirror to look into. Shopify ships a free Knowledge Base app, in the App Store since May 2025, that shows you the questions AI agents ask about your store and lets you correct the answers. The first analytics tool where the visitor asks the questions.

Small channel, one-way math

The honest part: Shopify's own guidance puts AI engines at 2 to 3 percent of search usage today. That is small, and anyone selling you panic is just selling. The same guidance cites a Similarweb estimate of about two billion site visits a month driven by generative tools, and the click math on the classic results page has moved in one direction so far.

Which leaves the useful realization. The work that makes products legible to answer engines already has a name. Attributes in metafields instead of adjectives in prose. Variants consolidated instead of cloned. Categories mapped to the Standard Product Taxonomy instead of a type string only your theme understands. Across the 642,000+ SKUs we have processed, that was simply the job description; Google's feed and the agents' feeds eat from the same table.

The industry keeps coining acronyms for this. AEO is not a new discipline for product data. It is catalog ops wearing a new lanyard.

What we keep telling merchants

• If an attribute matters, put it in a field. Prose is invisible to the new buyer.
• Fix the source catalog once; every surface feeds from it. Per-channel content is a treadmill.
• Install the Knowledge Base app and read what the agents ask. It is free reconnaissance.
• Ignore the acronyms until your variants are consistent. A feed exposes the catalog you actually have, not the one on your homepage.

In a few months, ask the chatbot about your products again and see who it names. The machine answers with whatever it can read.

Give it better data.

If you want to know what the answer engines can and cannot read from your store today, OKART's Infrastructure Stress-Test maps that surface, read-only, in a few days.

Sources

1. AI Overviews Reduce Clicks by 58%: Ahrefs study of 300,000 keywords; February 2026 update measuring December 2025 data against the April 2025 figure of 34.5%.
2. Introducing Shopify Agentic Storefronts: Winter '26 Edition announcement; Shopify Catalog's category inference, attribute extraction, and variant consolidation.
3. Millions of merchants can sell in AI chats: March 2026 update; the four live AI channels and the structured-data requirement.
4. OpenAI Agentic Commerce product feed spec: required fields, character caps, availability enum, and the search/checkout eligibility flags.
5. Shopify Knowledge Base app: free first-party app; FAQs used by AI agents; launched May 2025.
6. What Is Answer Engine Optimization (AEO)? (Shopify blog): Shopify's own AEO/GEO guidance; AI engines at 2 to 3% of search usage; the two billion monthly visits figure is a Similarweb estimate cited there.