Something I’ve noticed running my own stores and building for others: a growing share of “visits” now come from AI assistants — ChatGPT, Perplexity, Claude — checking product pages on behalf of a shopper who asked “find me X under $50.”
The thing is, they don’t see your store the way a person does. No images, no vibe, no urgency banner. They read structured data, titles, descriptions, stock status — and if they can’t parse it, your product simply doesn’t come back in the answer. The shopper never knows you existed.
A few things I’ve found actually matter: clean product titles that say what the thing IS (not just the brand poetry), stock and price in structured data, and descriptions that answer the questions a buyer would ask out loud.
Curious what others are seeing in their analytics — anyone else noticing bot-like traffic that converts later from a different device? How are you thinking about being “legible” to an AI middleman?
I’m seeing AI referrals show up, but I would not assume bot-like visits later converted on another device. That link is hard to prove unless the shopper signs in or uses the same browser.
What I check now:
In Shopify Analytics and GA4, filter referrers for ChatGPT, Perplexity, Gemini, and Copilot. Save the landing page, product, add-to-cart, and order data by source.
Test 5 product URLs in Google Rich Results Test. Fix missing price, currency, availability, SKU, brand, and variant data.
Rewrite vague titles and the first 150 words of descriptions. Include product type, material, size, compatibility, use case, and shipping limits in plain text.
Keep Merchant Center feeds current. AI answers often rely on shopping feeds and search indexes, not just live page crawls.
I’m also building ClickFrom.ai to track citations and product visibility, but the checks above are where I’d start today.
Sure! We are in the era of “agentic commerce,” where Large Language Models (LLMs) are becoming the new starting point in the customer funnel. Instead of browsing search results, shoppers now turn to ChatGPT, Perplexity, or Claude to find specific products; this means the AI itself is the entity that “visits” the store first.
Actually, these AI assistants don’t see beautifully designed user interfaces (UI) or eye-catching pop-ups. They rely entirely on machine-readable and processable information:
• Technical specifications hidden within expandable sections (accordions) are often completely overlooked
• Vague marketing copy always loses out to specific, clear technical details
• Coherent headings and accurate structured data (schema markup)
These are the deciding factors for whether you qualify to appear in result.
However, i think getting recommended by AI is the first step. Customers arriving via AI responses are usually well-informed and ready to buy. This shifts our online store’s role from product discovery to immediate validation and persuasion. So, now, onsite chat tools and automated, journey-based email systems play a crucial role.
From my point of view, AI merely guides the customer to the doorstep; the brand experience is what invites them inside to finalize the purchase.
Great points, especially on— merchants have no idea how much of their spec content is invisible to a parser. The gap I keep running into: almost nobody has a way to test what an AI actually sees when it reads their store. We check rankings and page speed religiously, but nobody’s checking “what does ChatGPT say when someone asks about my product category?” That’s becoming as basic as checking your own Google results. Agree on the post-click shift too — an AI-referred visitor arrives pre-sold, so the store’s job is confirmation, not discovery. We built a tool for this and hope it brings value to storefronts
Hi there @Robert_Dainowski
Following has been a very interesting change to observe. I’d concentrate on ensuring your product data is clear and consistent across all channels, whether the visitor is a human or an automated crawler. Having accurate titles and descriptions, consistent pricing and availability, inclusion of product structured data, and having images that are descriptive all help search engines understand your catalogue. Analytics can be used to detect abnormal referal patterns and whether such visitors turn into conversions at a later time.
You’ve described the mechanism better than most posts on this — the middleman reads structured data, titles, stock, description, and if it can’t parse a clean answer you’re simply not in the set. Two things I’d add on the “legible” question.
First, structured data is necessary but not sufficient. It gets you eligible to be read; it doesn’t get you chosen. What decides whether you’re the one cited is whether the page answers the question shape the shopper asked. “Find me X under £50” is a comparison/selection query, and pages built as comparisons (“X vs Y, when to pick each”) get cited far more cleanly than a list or a brand-poetry product page — the model can lift one self-contained paragraph without summarising the whole thing.
Second — the one that quietly costs stores the most — check the citation bots can even reach the page. The fetchers that pull a quotable chunk at query time (ClaudeBot, Google-Extended, Applebot-Extended, PerplexityBot, OAI-SearchBot) are distinct from the training crawlers (GPTBot, CCBot). If you’re on Cloudflare with managed robots.txt on, the default blocks several of the citation bots too — so you can do all the data work and still be invisible to the answer surface. Worth ruling out before anything else.
On analytics: I wouldn’t try to prove the cross-device conversion link — it’s mostly unprovable. I’d measure the input side instead: pick a fixed set of buyer questions, run them in fresh sessions a few times a week, and log whether you’re cited and how you’re described. That tells you if the legibility work is landing, more reliably than referral traffic will.
The robots.txt point deserves to be pinned — you can do all the data work and still bounce the answer-time fetchers at the door. I hadn’t seen the managed-robots default called out that plainly before.
And you’ve roughly described what I ended up building (the app from my disclosure above): a fixed set of buyer questions in fresh sessions, logging cited-or-not and how you’re described — same conclusion as yours, the conversion link was unprovable so I measure the input side. Agreed on comparison-shaped pages; the model wants one liftable paragraph, not brand poetry.
Glad it’s useful. Since you’ve already built the measurement loop, one refinement that earned its keep for me: log three outcomes, not one. Mentioned, cited-with-a-link, and picked-as-the-recommendation are different results with different causes, and collapsing them into “cited or not” hides where the work actually needs to go — plenty of brands get named in the prose but never linked, or linked but never the pick.
The other thing I’d watch is how you’re described — the model’s one-line paraphrase of what you are. That phrasing is a lever in its own right: when the description comes back wrong or generic, it usually traces to the page not stating plainly what the thing is — which loops straight back to your ten-second point.
Last one: rotate the prompt set now and then, and split it head vs long-tail. The engines route broad queries to category and forum content and specific ones to comparison pages, so a fixed set at a single specificity can quietly stop being representative of what buyers actually ask.
Adopting all three of these. The mentioned / linked / picked split is the one I’d felt but hadn’t named — a store can rack up mentions forever and never be the pick, and that difference is the whole game. The description point hits too: when the model describes you wrong, it’s usually quoting your own vagueness back at you. Prompt rotation with a head vs long-tail split goes in this week. Genuinely appreciate the operational detail — this thread has been better than most paid advice.