Shopify + OpenAI on AI discovery
Notes from a recent webinar.
57% of European consumers are already using gen AI for shopping recommendations and decision support - that's a Bain study that Anna Jenkins, Solutions Engineer at OpenAI, cited on a recent webinar (and 14% higher AOV from AI-referred orders)
57% is past early-adopter territory - it's mainstream now.
This issue covers what Anna and Peyman Naeini (Field CTO, Shopify) walked through on the webinar - three technical mechanisms that fit together into one picture of how AI retrieval works, plus the practices they recommended.
How a shopping prompt runs
When a shopper types a query into ChatGPT, the prompt does not run as one query - it fans out into several parallel queries running across search engines at the same time.
Anna's example from the deck:
Prompt: "What are good quality sheets that don't cost too much?"
Fan out queries:
→ best affordable sheet sets
→ high quality budget bed sheets
→ cotton sheet sets under $50
Three parallel queries from one prompt, each hunting a different evidence angle - Anna's words on the mechanism: "AI actually breaks that prompt into multiple fan out queries that run across search engines and other sources at the same time."
The three retrieval layers underneath
Those fan-out queries hit three retrieval layers.
Trained knowledge inside the live model on ChatGPT - stable, but bounded by the training cutoff.
Public web search index, and for ChatGPT's live-search side that index is Bing.
An offline OpenAI search index - a separate retrieval surface available to security-conscious customers, distinct from the Bing-backed public search.
Each fan-out query hits each layer independently, so you can sit in trained knowledge from before the cutoff but be invisible to live search, and you can be on the public web but missing entirely from the offline enterprise index. The layers don't share state.
The implication most brands I audit are missing: Bing matters more than the last decade of Google-only SEO has treated it. If Bing is the live-search backbone for public ChatGPT shopping queries and your category competitors are visible on Bing while you aren't, you're missing from that retrieval layer entirely. Bing Webmaster Tools is free, the index is smaller, and submission is faster than Google - the cheapest fix in this list.
The three signals that decide what gets cited
After the fan-out queries return findings, the model uses a three-signal framework to decide what makes it into the answer with citations.
quality data (pricing, inventory, dimensions, attributes)
citation-worthy content (reviews, guides, proof points online)
user sentiment (what real customers say about the brand)
The deck framed these as a cycle, with directional arrows running around a triangle, and each signal feeds the next. User sentiment generates the third-party content that becomes a citation, citations point at structured quality data, and quality data drives the buying experience that produces more user sentiment.
Anna's framing on how all three work together: "structured data helps you get into the conversation, while the strong citations and positive sentiment really help build enough confidence to actually like win the customer over."
The brands I see winning in AI citations have all three working in concert, but most have one or two strong signals and a third that lags - and the visibility gap usually traces back to whichever signal is missing.
The 30/70 split between on-site and off-site citations
About 30% of AI citations come from a brand's own site, and about 70% come from off-site sources - Reddit threads, YouTube reviews, trade publications, comparison guides, and community forums where customers describe the product in their own words.
Most content budgets at Shopify brands I audit are weighted close to 100% on the 30%: PDP copy, blog posts, and the editorial calendar. The 70% gets minimal attention because it's not directly authorable.
What Anna recommended for the 70%:
real engagement in forums and communities where the category gets discussed
PR placements in trade publications that produce independent coverage
making it easy for customers to write reviews with specifics, not just star ratings
A review that names who used the product, what they used it for, and why it worked carries much more weight than five stars with no text - the model is reading the review body for substance, and a rating on its own carries no weight.
What Shopify has shipped
Peyman walked through three pieces of Shopify infrastructure for AI discovery.
Shopify Catalog - His words: "What Shopify has done is for every brand on Shopify, out of the box, we make your data available from a product perspective to LLMs. We made it all open to OpenAI." Announced approximately a month or two before, and currently rolling out in the US. Europe is on the roadmap.
Knowledge Base app - The native interface for non-product data: FAQ, brand voice, content guidance - feeding into the same propagation layer as Catalog.
ACP, the Agentic Commerce Protocol - co-developed by OpenAI. Anna confirmed it's live in the US and Canada, with Europe in active development - the delay is tied to regional data privacy requirements.
The audit that any merchant can run
Watch the full webinar here
Three actions any merchant can run without platform dependency or a new budget.
First, a brand audit through ChatGPT itself. Search your category and your brand, compare what surfaces against what you expect, and note which competitors appear in places you should be and which queries return no mention of you at all.
Second, a product data audit inside an LLM. Paste a PDP into ChatGPT and ask it what the product is, who it's for, what it costs, what the dimensions are, and what it's made of - the gaps in the model's answer are the gaps in your structured data.
Third, systematic measurement across the funnel:
top-of-funnel: do you make the consideration set on a broad category query
mid-funnel: when the shopper narrows by attribute, do you survive that filter
bottom-of-funnel: in a head-to-head "X or Y" query, what does ChatGPT say about you
Peyman's framing: "It just needs love and attention." No ERP-style rebuild required. He compared it to social commerce - early movers stayed years ahead while competitors were catching up, and AI discovery is shaping up the same way.
Best practices (from the webinar)
Keep SEO foundations tight, especially on Bing - since OpenAI uses Bing as the live-search index, Bing visibility matters as well.
Set the metadata in Shopify admin that controls how products surface in ChatGPT. Peyman noted there are native admin fields for this.
Use the Knowledge Base app for non-product data - FAQ, brand voice, content guidance. Do it once, let it propagate.
Build answer-ready content alongside brand pages. Anna said: "AI also needs those like answer ready content. Comparisons, buying guides, fit advice."
Reviews with specifics beat generic ratings - a review that says who used it, what they used it for, and why it worked weighs more than a 5-star with no text.
Engage in third-party communities - Reddit, YouTube, forums where the category gets discussed. These are the source of 70% of AI citations.
Feel free to connect with me on LinkedIn.
Ankit
Atomz AI
If this was useful, the next one will be too.
Weekly. Free.
Unsubscribe with one click.
