Label every AI visibility observation as one of five states:
- a brand mention
- a source citation
- a product retrieval
- a recommendation
- a completed order
Each one proves only itself. A mention isn’t a citation, a product in a result list isn’t a recommendation, and a recommendation isn’t a sale.
When you see something, record the state, the platform, the date, the question and the evidence. Keep the five counts separate rather than rolling them into one visibility score. The five states are our proposed labels, not a platform metric. This page gives each one a definition, a clear example, a boundary case and the line it doesn’t cross.
The five states on one page
| State | What you observed | Counts | Doesn’t count |
|---|---|---|---|
| 1. Brand mention | An AI answer names your brand or store in its text | “For a top-fill model, [your brand] makes one rated for 500 sq ft” | A paid placement labelled Sponsored or Ad; a generic category answer that never names you |
| 2. Source citation | An AI answer links or cites one of your pages as a source | Your humidifier buying guide shown as a supporting link in Google AI Mode | A third-party review that mentions you being cited. That’s their citation and your mention |
| 3. Product retrieval | Your product appears in a set of results a system returned for a request | Your humidifier among the products in a shopping carousel, or in Shopify’s search preview | Your brand named in text with no product card or record returned |
| 4. Recommendation | The assistant presents your product to the shopper as a fit for what they asked | “For your 450 sq ft bedroom, this one fits: 26 dB on low, top fill, under $90” | A carousel slot with no reason given; a pick in a simulation; your product recommended with another store’s offer listed first |
| 5. Completed order | A shopper paid for your product through, or after, an AI channel | An order in Shopify’s agentic storefronts report for a channel | A session, a click or an add-to-cart. Also an order for a different product |
The states aren’t a strict ladder. A product can be retrieved without your brand ever appearing in the answer text, and a page can be cited without your product being shown. An order can arrive through Shop’s own agent with no mention you could ever screenshot. Record every state you can see, not just the highest.
Card 1: brand mention
Definition. The answer text names your brand or store.
How you’d see it. A screenshot or saved copy of an answer, with the question, platform, date and any context the assistant had.
Boundary cases.
- Paid placements. OpenAI says ChatGPT product results aren’t ads and that ads are separate. Microsoft says Copilot draws results from the web and from advertisers, and that sponsored links are clearly identified. A sponsored slot is a paid placement. Count it as advertising, never as an earned mention.
- Wrong facts. OpenAI says ChatGPT may generate simplified titles and descriptions, and labels like “Budget-friendly” that aren’t guarantees or verified statements. A mention that gets your product wrong is still a mention. Log it with a note that the facts were wrong, because that’s a separate problem to fix.
What it doesn’t prove. That a link was shown, that your product was retrieved, that the shopper was steered toward you, or that anything was bought.
Card 2: source citation
Definition. The answer links to, or cites, a page on your site as a source.
How you’d see it. The link itself, in the saved answer. For Google, there’s a documented route. Google says AI Overviews and AI Mode show links to supporting websites, and a supporting link must be an indexed page that’s eligible for a snippet.
Boundary cases.
- Someone else’s page about you. If an answer cites a review site that mentions your brand, the citation belongs to the review site. For you it’s a mention.
- Search Console numbers. Google says traffic from AI Overviews and AI Mode is included in the Performance report under the “Web” search type. A rise in Web impressions doesn’t tell you it came from AI features. Don’t label it as citations.
What it doesn’t prove. That your product was shown or recommended. A citation of your buying guide in an answer that then recommends a competitor’s humidifier is a real citation and a lost recommendation.
Card 3: product retrieval
Definition. Your product appears among the results a system returned for a request. That might be a catalog search, a carousel or a list of product cards.
How you’d see it. A product card in a saved assistant answer. Or a catalog record returned for a query, such as Shopify’s search preview or a catalog snapshot, with the system named.
Boundary cases.
- Catalog results aren’t assistant results. Shopify says its search preview shows the raw output of Shopify Catalog search, and that agentic storefronts often re-rank results on their own logic. Shopify calls the preview a directional signal, not an exact prediction. Our own snapshot is a good example. For “humidifier for dry winter air large room quiet top fill”, we saved the 10 listings the catalog returned. That’s retrieval in Shopify Catalog, not what ChatGPT showed anyone. Our Catalog research checklist sets the limits on claims like that.
- Occurrences aren’t products. Across our 162 searches, the catalog returned 1,620 result occurrences, not 1,620 distinct products. If you count retrievals across many questions, say “results”, not “products”, unless you’ve removed duplicates.
- The record may not say what you wrote. In that humidifier result, a decibel figure was in the description of 4 of 10 listings. Counting Shopify-generated metadata as well, it was in 6 of 10. Being retrieved with a record that lacks your key fact is still retrieval. It’s just weaker evidence for the stages after it.
What it doesn’t prove. That the shopper saw your product near the top, or that the assistant said it fit. OpenAI also notes that not all available products will necessarily be shown.
Card 4: recommendation
Definition. The assistant presents your product to the shopper as a fit for what they asked: named as a pick, put on a shortlist with a reason, or matched to a stated need.
How you’d see it. The answer text tying your product to the shopper’s request, saved alongside the full question and context. OpenAI says ChatGPT considers the query and context such as Memory or Custom instructions, so the same question from two shoppers can get different answers. Without the context, a recommendation can’t be reproduced.
Boundary cases.
- Your product, someone else’s offer. OpenAI says that once a product is picked, ChatGPT may list the merchants offering it, ranked on availability, price, quality and whether they’re the maker or primary seller. The initial price typically comes from the first merchant listed. Your product can be recommended while another store’s offer comes first. Record product recommendation and offer position separately.
- Simulated picks. In our supplied-snippet simulation, an invented humidifier with its facts as a spec line was picked in 134 of 180 Claude sessions, 20 of 60 on GPT-6.1-Sol, 19 of 59 on Gemini 3.8 Flash and 15 of 55 on Gemini 3.1 Pro. Each model picked 2 of 8 listings, so random picking is 25%. Those are controlled simulation results from snippets we supplied. They’re useful evidence about which facts models use, and they’re not live recommendations. The full write-up keeps the two apart.
- A recommendation with wrong facts. If the assistant recommends you as “26 dB” and your product is 36 dB, that’s a recommendation you don’t want. It can end in a return. Log the fact error next to it.
What it doesn’t prove. That the shopper clicked, bought from you, or kept the product.
Card 5: completed order
Definition. A shopper paid for your product through an AI channel’s checkout, or on your store after arriving from one.
How you’d see it. Shopify’s agentic storefronts reports show sales, orders, online store sessions and conversion for each AI channel. Shopify says the sales figure combines referral-based sales and direct-checkout sales. Those reports aren’t yet available for headless stores.
Boundary cases.
- Sessions and carts. A visit from an AI channel is a session. An add-to-cart is a cart. Neither is an order.
- Orders your analytics can’t see. Shopify says Google Analytics and custom pixels don’t fire in the Google, Copilot and Meta direct checkouts; only server-to-server pixels do. A direct-checkout order that’s missing from Google Analytics may still be a real order. Check the Shopify report before concluding nothing sold.
- A different product. An order from an AI referral for your filters, after a recommendation for your humidifier, is an order. It’s not an order for the recommended product.
What it doesn’t prove. That the AI channel caused the sale, that the sale was profitable after shipping and returns, or that the customer will come back. Those need a separate analysis.
Classify an observation in five questions
Work through these in order for each thing you’ve seen. Record every “yes”, not just the first one.
- Was money paid for your product, with a record you can point to? Then it’s a completed order. Note which product and which report.
- Did the assistant present your product as a fit for the shopper’s request? Then it’s a recommendation. Note the offer position and whether the facts were right.
- Did your product appear in a returned set of results? Then it’s a product retrieval. Name the system: assistant, catalog or preview.
- Did the answer link or cite one of your pages? Then it’s a source citation. Name the page.
- Did the answer name your brand or store? Then it’s a brand mention. Note whether it was sponsored.
If every answer is no, write down “not observed” with the date and context. Don’t write “not visible”. One answer on one day isn’t a rate, and absence from one answer isn’t proof of absence everywhere.
Five observations, classified
These are worked examples. Two come from our own research, and three are hypothetical observations of the kind a merchant might bring us.
| Observation | Source | State | What it doesn’t prove |
|---|---|---|---|
| A humidifier listed at position 3 for a catalog search | Our Shopify catalog snapshot, October 26, 2026 | Product retrieval, in Shopify Catalog | That ChatGPT or any assistant showed it, in any position |
| An invented humidifier picked in 134 of 180 Claude sessions | Our supplied-snippet simulation | A simulated recommendation, not an observed one | Live recommendations, rank or sales |
| Search Console Web impressions up 30% this month | Hypothetical | None of the five on its own | That AI features caused it. Google folds them into Web |
| A ChatGPT answer names your brand and links your blog post | Hypothetical screenshot | Brand mention and source citation | That your product was retrieved or recommended |
| Shopify’s report shows 3 orders from Copilot this month | Hypothetical report | Completed order | That Copilot recommended those products, or caused the sales |
Why one visibility number hides all of this
Vendors often sell a single “AI visibility score”. The problem isn’t the arithmetic. It’s that the five states above measure different things, and a single number lets the easy ones stand in for the hard ones. A pile of brand mentions can lift a score while orders stay flat.
We learned this on our own retired score. When we recomputed it, about 65% of the points it awarded came from review depth, a part designed to be only 25 of 100. The audit shows how that happened. Keep the five counts side by side and you can see which stage is actually moving.
For a repeatable way to collect these observations, see our AI product visibility monitoring plan. For how the jobs behind each state split across your team, see SEO, AEO, GEO and ACO. For what each assistant does and which controls you have, see AI shopping agents: compare capabilities and merchant controls.
Your next step
Take the last three “we’re showing up in AI” claims you’ve seen, whether from a screenshot, a vendor report or a teammate. Run each through the five questions and write the state next to it, plus what it doesn’t prove. If none of them reaches a completed order, open Shopify’s agentic storefronts report and look at orders by channel before deciding anything.