How do you track traffic from AI search?

When somebody clicks a link inside an assistant's answer, the visit usually arrives at your store with a referrer identifying that assistant's domain. That makes it ordinary referral traffic, which your analytics already records. Nothing special is needed beyond grouping the hostnames together so the channel exists as a reportable thing.

Google's AI features work differently. AI Overviews and AI Mode sit inside Search, and Search Console reports clicks from them within the overall search performance data rather than as a separate appearance type. There is no filter that isolates them, so anyone offering a precise count of AI Overview clicks is estimating.

That leaves two different jobs. For assistants, build the channel and report it properly. For Google, read the shape of the data over time: impressions, clicks and click-through rate by page type, watching for the pattern where impressions hold or rise while clicks on informational pages fall.

Both measurements are incomplete, and it is worth saying so out loud in whatever report this ends up in, because the alternative is a stakeholder treating a partial number as the whole picture.

Create a custom channel group and add a rule that matches the assistant hostnames in the session source dimension. The ones worth including for a North American store are the ChatGPT domain, Perplexity, Microsoft Copilot, Google's Gemini, Claude, and any smaller answer engines your logs show sending visits. Use a pattern match rather than exact strings, since subdomains vary.

One detail helps a lot: some assistants append their own campaign parameters to outbound links, which means the visit can land as referral or as a tagged campaign depending on how the link was constructed. Build the rule to catch both the referring hostname and the tagged source, or the channel will undercount.

Apply the group to a report that shows sessions, engagement, add to cart rate and revenue per session, and compare it against organic search and direct. Small volumes are normal, so look at trend rather than at a single month.

Check the landing pages inside the channel too. Which pages assistants send people to tells you what they consider your site useful for, and it is often not the pages you expected.

Most of what assistants do on your site produces no visit at all. A fetcher requests a page, the answer is written from it, and the shopper never leaves the chat. Analytics cannot see any of that, but your server or CDN logs can.

Filter the logs by user agent for the assistant fetchers: the search and user-triggered agents from OpenAI, Perplexity's crawler and its user-triggered fetcher, Microsoft's crawler, and the training crawlers if you have chosen to allow them. Then look at three things: which URLs are being requested, how often, and what status codes they receive.

That report answers questions nothing else can. Whether your product pages are being fetched at all. Whether the fetchers are getting redirects, errors or a challenge page from your firewall instead of content. Whether the pages being read are the ones you would want quoted.

Cloudflare and similar edge platforms make this straightforward, and many bot-management defaults block some of these agents without anyone realising. If your logs show no assistant fetches at all, that is usually the reason rather than a lack of interest.

Three secondary measures fill in some of the gap. Branded search volume, from Search Console, tends to rise when more people are being told about a brand somewhere they cannot click. Direct traffic behaves similarly, since a shopper who reads about you in a chat often types the name later. And assisted revenue, viewed with a data-driven or position-based model rather than last click, shows sessions that touched the assistant channel earlier in the path.

Alongside those, keep a manual record. Ask the questions your shoppers actually ask, from your selling market, on a fixed schedule, and record which stores appear and how your products are described. It is a sample rather than a measurement, but it catches inaccuracies about your prices and stock that no report would surface.

What none of this supports is a promise. Nobody can commit to a number of citations or a share of assistant answers, because the systems are not disclosed, the outputs vary between identical questions, and there is no reporting from the providers.

So report it as what it is: a small, growing referral channel you can count, a Google effect you can only infer, a retrieval log you can inspect, and a set of indirect signals that move together.