Your customer opens ChatGPT and types, "I need a waterproof hiking backpack for a 5-day trip in Iceland in October, with a laptop sleeve and hipbelt, under $200." Thirty-one tokens of intent. Your Shopify feed, last synced six weeks ago, gives the LLM a product title, three bullets, and a 50-character meta description that was written for Google in 2019. The LLM does the best it can. It surfaces your pack next to two competitors whose feeds happen to be richer. You lose the click to a product that is functionally indistinguishable from yours.

This is the new ecommerce front door. The old one was ten blue links. The new one is a single synthesized answer produced by a model that has read your feed, your competitors' feeds, and three Reddit threads from the same trip last year. MarTech's recent conversation with Opiversal CEO Lucas Tieleman puts a number on the shift: average Google search has been six words for a decade. The average LLM-mediated query is now around 24 words. That is not a UX preference. That is a fourfold expansion in the amount of context the buyer expects the system to absorb before recommending.

What changed in the platforms you already use

Google Shopping still rewards keyword-dense, attribute-complete feeds. That has not changed. What has changed is that AI surfaces do not consume your feed the same way Google does. An LLM does not match a query string to a title tag. It tries to satisfy the full intent of the question, and it reaches for whatever structured data is closest to a natural-language answer.

For a retailer, that means three things have to be true for any given product before an AI surface will surface it:

  • Conversational attributes are present. "Good for 4-season backpacking" beats "material: 400D nylon" when the query asks about Iceland in October.
  • Use cases are explicit. "Fits a 16-inch laptop in a dedicated sleeve" answers a question. "Padded laptop compartment" does not.
  • FAQs are wired into the feed, not buried in a tab the customer has to open. The LLM reads them. Your shopper will not.

None of this is technically hard. Every commerce platform supports the fields. The work is content work, not engineering work. That is why it has not happened at most retailers. Content work does not show up on the engineering roadmap, and the marketing team does not own the product feed schema.

Bot traffic is now your biggest referrer, and it is not buying

Tieleman flagged the part that should make every ecommerce CFO uncomfortable. LLMs are crawling retail sites at scale to answer user prompts. That bot traffic is not buying. It is researching. Your impression count looks great. Your conversion rate looks terrible. The dip does not show up in a single channel report because the LLM is not a channel. It is an upstream consumer of your catalog, and the conversion event is happening somewhere downstream that you do not have visibility into.

This is the impression-versus-sales inversion. For twenty years, marketers optimized for impressions because impressions correlated with sales. In an AI-mediated path, impressions and sales decouple. You can have the most-cited product in the category and still lose the order to whichever competitor's page the LLM decided to surface at the bottom of its answer.

What to actually do this quarter

Pick the 50 SKUs that drive 80 percent of revenue. For each one, write a conversational attribute block: who it is for, what it is good at, what it is bad at, what use cases it handles. Wire those blocks into your structured data. Add a 5-question FAQ block to every product detail page. Re-sync your feed weekly, not quarterly. And tag a single human as the owner. The owner is accountable for whether an LLM can answer the four most common 24-word queries about the product with your listing as the primary citation.

If nobody on your team can answer that question for your top SKU right now, you are already losing the AI-mediated click. The catalog page you published last year is not the catalog page the LLM is reading today. It is reading the one your competitor refreshed last Tuesday.

Sources