Question about ai visibility for shopify stores!

AI Build Your Business
28

the technical signals piece is real. biggest gap i keep seeing on shopify stores is that product schema is half-broken depending on the theme, missing aggregaterating, brand, or availability fields even when reviews show fine on the page. quick check is to run a product url through google’s rich results test, takes 30 seconds.

the other thing that’s worked for us: comparison content. pages that explicitly say “x vs y” or “best [thing] for [use case]” get pulled into chatgpt and perplexity answers way more than generic product or category pages. a single comparison post indexed against our category has done more for llm mentions than any technical tweak.

29

To answer @gemaster’s question directly: the thing most merchants keep postponing is structured AI context for their store.

Not just schema (though that is often broken as the post above correctly flags). The specific thing that keeps getting deferred is the llms.txt file and the richer product-level context that AI models need to represent a store accurately.

Right now ChatGPT, Perplexity, and Gemini are making recommendations about products in millions of niches. When they surface a store, it is almost never because the product description was good. It is because the model found something that answered a clear question in structured language it could parse. For most Shopify stores that context either does not exist, is buried in thin product descriptions, or is scattered in a way no model can stitch together.

The thing merchants keep postponing: checking whether ChatGPT can accurately describe their store at all. Quick test in a fresh session: ask “What does [your brand name] sell?” If it does not know, or gives wrong information, that is the gap. The fix is structured brand context at yourstore.com/llms.txt plus cleaner schema on product pages.

The comparison content point in the post above is the external piece. The llms.txt plus schema is the internal piece. Both matter and most merchants have done neither.

31

I’m replying here because most of these answers seem generated by ChatGPT lol

I’ve worked on AI visibility for hundreds of Shopify stores at this point, and honestly most of the advice in this thread is stuff people repeat to each other without ever testing it.

The uncomfortable truth: when we audit a store that “doesn’t show up in ChatGPT,” the problem is almost never schema or page structure. Shopify themes already output decent product markup. Adding more of it changes nothing. We’ve tested this on dozens of stores. Markup heavy stores with zero outside presence stay invisible. Stores with messy markup but strong off site mentions get recommended constantly.

That’s because these tools don’t recommend you just based on your site (they did at one point, but not anymore). They recommend you based on what the rest of the internet says about you. When someone asks for “best linen shirts for hot weather,” the AI runs a search, pulls a handful of sources, and summarizes the consensus. If you only exist on your own domain, there is no consensus to summarize. The brand that appears in three Reddit threads, two comparison posts and a niche blog wins, even with a worse site. Or the store with a lot of TrustPilot review.

So the strategic question isn’t “how do I make my pages more readable for AI.” It’s “in how many places, that I don’t own, is my product described as the answer to a specific question.” That’s a positioning and marketing issue. Most merchants don’t want to hear that because a schema app is easier to install than earning mentions.

Second thing we see over and over: stores optimize for generic queries they can never win. You will not be the AI’s answer for “best skincare brand.” You can absolutely be the answer for “fragrance free moisturizer for rosacea that ships to anywhere in Europe in under 7 days.” AI answers are weirdly specific because the questions are. The stores getting real traffic from this carved out narrow questions and made sure both their pages and their outside mentions repeat the same specific claim.

And third, the catalog itself matters more than content now that Shopify is piping product data straight into these assistants. Incomplete attributes, vague titles, missing variants, that’s what actually breaks you in agentic shopping, not your blog.

Quick way to check where you stand: ask ChatGPT and Perplexity for products in your niche, then look at which sources they cite. Those sources are your real to do list.

If you want to understand if you store is optimized for AI, you can also use the free trial of IndexGPT to get a good picture of where you currently stand and where you have to go.

32

I think the hard part is that “AI visibility” is not one single ranking factor.

For Shopify stores, I’d separate it into three layers:

  1. Clean product data
    Titles, descriptions, variants, attributes, availability, price, reviews.

  2. Machine-readable structure
    Schema markup, product feeds, collections, internal linking, maybe llms.txt as the space evolves.

  3. External confidence
    Mentions, reviews, comparison pages, Reddit/forum discussions, niche blogs.

My guess is that AI tools will trust stores more when all three layers tell the same story.

The practical challenge for merchants is that most stores don’t know where the gaps are. They may have decent SEO, but weak product attributes, missing structured data, or unclear policies that make AI recommendations less reliable.

Curious if anyone here has actually tested prompts like:
“best [product type] store for [use case]”
and tracked which sources the AI cites?

33

Hello, @gemaster
Hope you are doing well
Really interesting observation. I have also noticed that a lot of stores focus heavily on SEO but miss key technical elements like structured data, product context, and trust signals. It make sense that Ai tools would rely on these signals more when generating recommendations.

34

Something that hasn’t been fully unpacked in this thread yet, the gap between citation volume and citation consistency.

We’ve been tracking how different Shopify brands show up across ChatGPT, Gemini, Perplexity and Claude, and the pattern that stands out most isn’t who has the most mentions. It’s who gets described the same way across independent sources.

A brand might appear on Reddit, a roundup blog, and their own product pages but each source describes them differently. Different positioning, different problem association, different category language. When that happens, the model can’t build a clear mental picture. It either gives a vague answer or skips to a brand it can describe with confidence.

The brands consistently surfaced tend to have what I’d call coherent signal density, not just volume of mentions, but the same problem-brand association repeated across sources it didn’t control.

Practically, that means the off-site work matters as much as the on-site technical work. Getting your JSON-LD right is table stakes. But if the only place that describes you clearly is your own website, you’re invisible the moment an LLM weights third-party sources more heavily, which most do.

Curious whether others are seeing this split between brands that have clean on-site data but still don’t surface, vs ones with messier technical setups that somehow do.

35

Hi @gemaster .
Good observation, well, JSON-LD schema, clear return policies, and genuine product context are commonly missing and do matter for AI visibility.
That one thing most store owners keep postponing is structured data setup.

36

This is the clearest version of this debate I have seen, so let me try to put a frame on it, because the thread keeps circling.

Two different things are getting mixed together, and they have different fixes.

One is AI search citation. ChatGPT or Perplexity answering a buying question by reading pages and citing sources. This rewards legibility. Can a crawler read your page at all, are your products clean machine readable entities, is the answer stated plainly, do your pages connect into one map.

The other is agentic shopping. Gemini, Copilot, and the Shopify agentic storefront pulling structured catalog and feed data through protocols. This rewards your product and variant data being complete, consistent, and governed, which is exactly what Gabe is describing at 8500 SKUs.

They overlap, and the foundation underneath both is the same. Most stores fail that foundation before any of the advanced work matters. In order:

First, the Front Door. Can a machine read the page at all. A surprising number of stores serve a near blank page to crawlers, because the content only loads with JavaScript, or a bot wall blocks them outright. Nothing downstream counts if this is broken.

Second, Entity Integrity. Complete Product and Organization data. Price, availability, identifiers, real social links, ratings. This is the layer AI actually cites, and it is where most stores fail hardest.

Then the rest. Readability, internal structure, and yes off site credibility, which this thread is right about. Reddit, editorial, reviews, and consistent third party descriptions all raise a model’s confidence to name you. But off site signals do not save a store the AI cannot read in the first place.

On llms.txt and ai.txt, I would not spend much time there yet. There is little evidence the major engines meaningfully read them today, and it is easy to mistake for progress. Fix the things they demonstrably use first.

The simple version. This is not a new playbook, it is the old foundations plus a new unit. Search rewarded the page. AI rewards the entity.

I will be happy to run those five checks on a specific store if anyone wants to see where they actually stand. Drop a URL.

37

Gabe, you asked how to validate this without guessing. Shopify now exposes a public UCP catalog for the store, so I ran one read-only test against Stillwater.

The good news: the compatibility work is real. The Redington Original Kit page exposes staff-validated compatibility and structured grain-window fields.

The break is between the product page and the catalog response. In the UCP product object I reviewed, those grain-window fields were not returned as structured data. The same catalog search returned test-accessories on all five products, and three of five had no specific Shopify taxonomy category or returned na. I also found the free-shipping threshold split between $49 and $50 across official surfaces.

I put the exact request, response and repair order into a short audit. I can post the relevant excerpts here if useful.

38

@Gabe_Stillwater Here is the short audit I mentioned.

Stillwater_Audit_Part_3_pages_9-11
Stillwater_Audit_Part_3_pages_9-111920×3240 476 KB

Stillwater_Audit_Part_2_pages_5-8
Stillwater_Audit_Part_2_pages_5-81920×4320 680 KB

Stillwater_Audit_Part_1_pages_1-4
Stillwater_Audit_Part_1_pages_1-41920×4320 868 KB

39

Hi Ian — your skepticism is fair. The missing part in most AI visibility advice is evidence.

I do not judge a store from one prompt or give it a generic score. I separate two things: what crawlers and shopping agents can actually retrieve, and what AI systems return across a defined prompt set and repeated runs.

I track mentions, recommendations, first choice, purchase-intent visibility, citations, answer stability and factual conflicts. Every finding points back to a captured answer, citation or reproducible request.

In Stillwater’s case, the product page carried staff-validated compatibility data, but the reviewed public UCP product object did not return those fields as structured product data. That is the kind of gap I think a useful table-stakes checklist should make visible. I posted the 11-page read-only example below. No ranking claims.

42

Answering the actual question: the thing most stores postpone isn’t a file or a schema. It’s instrumentation. Everyone ships the metafields, the JSON-LD, the llms.txt, and almost nobody sets up a way to tell whether any of it changed what the agents actually do.

@kestrel-ian’s skepticism upthread is the right instinct, and it’s exactly why I’d argue for measuring over optimising. Nobody can verify a claim like “this improves your AI visibility.” You can verify whether ClaudeBot fetched your product page on Tuesday.

@Gabe_Stillwater on the “educated guesswork” problem: part of that loop can close today. Server logs will show which agents are fetching which URLs. It won’t tell you whether you got recommended, but it tells you whether you’re being read at all, which removes half the guesswork. Worth splitting them into two groups, because they get conflated constantly. GPTBot and CCBot are training crawlers, their visits say nothing about today’s answers. ChatGPT-User, OAI-SearchBot, ClaudeBot and PerplexityBot fetch at answer time, when an actual person has asked something. Only the second group is a signal about now.

That gives a cheap fork before you spend more days building. If the inference bots never touch your store, it’s an access or discovery problem, and content work won’t help yet. If they hit your PDPs regularly and you’re still not surfacing, they’re reading you and picking someone else, which is a content problem. Two different fixes, and the logs tell you which one you have.

43

The metafield + KB work @Gabe_Stillwater describes is exactly where the leverage is — and it lines up with what I’m finding from the other side of the equation. I’m running a research study scanning how ChatGPT, Claude, Perplexity and Gemini actually answer buying questions in a merchant’s niche (who gets named, who gets skipped, which sources the engines cite). A few patterns from the scans so far:

1. Crawl access is the silent killer — stores block OAI-SearchBot / PerplexityBot / ClaudeBot in robots.txt without realizing, and no amount of content fixes an engine that can’t read you.

2. Structured product data decides who’s “quotable” — assistants recommend the store whose attributes (materials, sizing, compatibility min/max like Gabe’s modeling) they can extract confidently, not the one with the best prose.

3. Third-party footprint beats on-site content — the engines lean hard on Reddit threads, listicles and review sites. The fix is often “get cited in the 3 threads the AI already trusts for your category” rather than “write more blog posts.”

Happy to run a free scan for anyone here as part of the study — you get the full report (mention frequency across 4 engines, who shows up instead, the exact sources citing them, prioritized fix list). It’s research, not a product; all I ask is 15 min of feedback after. Reply or DM your store URL if you want in.

44

@PromptSightApp Thanks for tagging me in this and providing analysis with your observations.

Your “silent killer” is an excellent example of this. I went through my robots.txt output and found a myriad of issues that needed to be addressed. Apparently, the default build made by Shopify (which most stores are likely using) needed a bit of re-write/edit to ensure compatible/accurate/readable outcomes not just for Agentic - but also general search crawlers.

45

Solid thread. On the original question (“what’s one thing you keep postponing?”) — from the stores I’ve scanned, it’s almost always complete product schema: people have basic JSON-LD but it’s missing price, availability, or sku/gtin, so the AI can read the page but can’t trust the details enough to recommend it. Category pages with ItemList schema are the other big one people skip.

Also agree with the folks saying build for humans first — I’d just add that structured data is the cheap, boring layer that makes all that good human content machine-readable. No black-hat needed, it’s just plumbing.

I built a free scanner around exactly these checks (it scores crawlability vs. recommendation-readiness separately and shows the specific gaps, no signup) if it’s useful: solkendo.com. Happy to compare notes on what people are seeing across categories.