How AI shoppers buy, and why your analytics call them direct

Why it gets misattributed

Assistants often strip or never set a referrer, and they do not always add campaign parameters. With no referrer and no UTM tag, the default bucket is direct: the same bucket as someone typing your address.

Some do leave a trace. ChatGPT, for example, can add utm_source=chatgpt.com to the links it shows. Detecting the channel properly means checking referrer domains and UTM sources together, and being honest about the share you cannot see rather than asserting a number you cannot support.

How the traffic behaves differently

Assistant-referred shoppers usually land deep, straight on a product rather than the home page, because the assistant has already done the browsing.

They arrive with more intent and fewer questions, which tends to make them convert well, and it also makes friction more expensive: a problem that hits this group costs more per session than the same problem elsewhere.

How assistants read your store

There are two routes. Crawlers such as GPTBot, ClaudeBot and PerplexityBot read your pages, subject to your robots.txt. And when a shopper asks an assistant for a product, ChatGPT, Perplexity and Copilot can ask your store directly, through the shopping endpoint Shopify runs for every store, rather than browsing your theme.

Either way, the assistant recommends what it can find and describe, and answers a shopper’s questions from what your store tells it. A product with no description, or a returns policy no page links to, is a product it cannot recommend with confidence.

What to do about it

Make your product data legible. Assistants and search engines lean on structured data for price, availability and variants. If yours is missing or invalid, you are relying on a model parsing your page correctly.

Put plain answers within one click of the home page: shipping, returns and, for apparel, sizing. Check that your robots.txt does not block the AI crawlers you want, and consider an llms.txt that lists your collections, shipping and returns pages.

Check the links assistants already cite. A product that was renamed or removed without a redirect sends a high-intent shopper to your home page or an error. Then segment the traffic: once you can see the group, you can see whether your problems hit it disproportionately, which is the argument for fixing them first.

What Liftable checks and measures

The free scan’s AI checks look at whether product structured data is present and valid, whether shipping and returns answers are within one click of the home page, whether robots.txt blocks AI agents or llms.txt is missing, and whether AI shopping agents asking your store’s own endpoint can find and describe your products. Two more are planned: whether URLs that assistants cite redirect or return a 404, and the share of your sessions referred by AI.

That share stays unknown until the pixel is on, because a crawl cannot see where visits come from. With the pixel on, Liftable classifies each session’s source by referrer domain and utm_source (ChatGPT, Perplexity, Copilot, Gemini and Claude) and flags headless browsers and known AI crawlers, so the channel stops hiding in direct.

Questions