Merchants are seeing orders arrive tagged utm_source=chatgpt. How the recommendation gets made, what doesn't move it, and how to test whether changes are working.
Across hundreds of Shopify stores, sessions tagged utm_source=chatgpt grew from 115 in January to 33,548 in August of the same year — about 190x — while still roughly 0.1% of total traffic. Adobe Analytics measured AI-referred traffic to US retail up 693% year over year across holiday 2025, converting about 31% better than average visitors. Salesforce put AI-influenced US holiday retail sales near 20%.
The channel undercounts itself: assistant referrals without tags show in analytics as direct traffic.
When a shopper asks an assistant what to buy, three things happen in sequence:
1. It searches the open web. ChatGPT, Claude, and Perplexity run web searches while composing answers. Each announces itself in your server logs — ChatGPT-User, Claude-User, PerplexityBot. If your robots.txt blocks those crawlers, or your pages render nothing readable without JavaScript, you're out before ranking starts. Check this first.
2. It shortlists by the shopper's qualifiers. Shoppers ask "wide-fit trail shoes for flat feet under $150," not "best shoes." Every qualifier filters the shortlist. Specific questions are winnable now — most stores publish category pages, not answers. Broad head terms follow domain authority: a store that doesn't rank for them in Google won't rank for them in ChatGPT.
3. It ranks by proof it can quote. Between two stores on the shortlist, the assistant recommends the one whose pages make specific, checkable claims — who the product is for, what it solves, at what price, backed by what evidence. Reworded marketing copy loses to a page that reads like an answer.
Shopify's Agentic Storefronts puts your product catalog inside ChatGPT and other assistants, managed from the admin, with a "recent searches where your products appeared" report under Sales Channels > Agentic. Turn it on. Shopify reports catalog-formatted products converting about twice as well as scraped ones. Checkout still completes on your store — products appear in the chat, the purchase happens with you.
The catalog decides what the assistant can show, not what it recommends. When a shopper asks an open question — "what's a good X for Y" — the assistant builds its answer from the open web. Catalog syndication without open-web proof means your products are present and never named first.
Reported across Shopify community threads:
Rewriting descriptions "for AI" with no new information. Rewording the same copy adds nothing an assistant can quote. What works is adding checkable specifics the page didn't have — dimensions, compatibility, who it's not for, comparisons.
Schema that says what the page doesn't. Structured data and visible text must agree; markup describing content that isn't visibly there gets ignored.
An llms.txt file by itself. Serve one — some assistants read it, and on Shopify it takes a redirect or an app, since you can't place root files. But Google has said its AI features don't use llms.txt, and a pointer file aimed at thin pages points at nothing. Several paid "AI visibility" apps sell this file plus a dashboard.
Two instruments, both free:
Blind ranking tests. Open a ChatGPT temporary chat with memory off — logged-in ChatGPT recognizes you and will find your own store, which is not a measurement. Ask the questions your buyers ask, with their qualifiers. Record date, prompt, and result. Repeat monthly with the same prompts. Do the same in Perplexity from a logged-out window.
Your server logs. Watch for ChatGPT-User — it appears when ChatGPT fetches your page while composing a live answer. Rising fetches of a page mean it is entering answers. Sessions tagged utm_source=chatgpt confirm the orders.
Assistant rankings are not deterministic; no one can guarantee a placement on a schedule.
Crawlable to AI crawlers, one page per buyer question, checkable specifics the assistant can quote, structured data agreeing with visible text, monthly measurement. Which questions to build first, in what order, with what proof, is store-specific — that's the engagement. The free agent-readiness check tells you whether your store clears the crawlability layer.
Run the free agent-readiness check — no email required. Then email Tom the questions your buyers ask; he'll quote a fixed-scope build.
Email Tom →