How to Track AI Search Traffic and Citations in GA4.

To track AI search traffic, confirm the AI Assistant channel that GA4 added in May 2026, add a custom channel with a referrer regex for the assistants Google does not yet recognize, and then measure two things GA4 cannot see: which of your pages get cited inside AI answers, and how often you are named versus competitors. That is the whole system: referral traffic, citations, share of voice. Everything below is how to set up each layer and what to report.

We built this for our own properties first. Across the 31 sites in our AI assistant referral study, ChatGPT sent 93% of all AI assistant referrals, so if you only get one source classified correctly, make it that one. The rest of the setup is for the long tail and for the two layers that never show up as a session.

The three layers: referral traffic, citations, share of voice.

• Referral traffic: sessions where someone clicked a link inside an AI answer and landed on your site. Lives in GA4. Undercounted, but tied to real leads and revenue.
• Citations: your pages appearing as sources inside answers, whether or not anyone clicks. Lives in visibility tools or a manual panel. Leading indicator for traffic.
• Share of voice: the percentage of answers to a fixed set of prompts that mention or cite you versus named competitors. Lives in tools or a panel. The number leadership understands.

Each layer answers a different question. Traffic tells you what AI search is worth today. Citations tell you which pages the models trust. Share of voice tells you whether you are winning the category or just appearing in it. Skip any one and the story has a hole in it.

GA4 setup, step by step.

Why Search Console cannot see it.

Search Console reports Google surfaces only. ChatGPT, Perplexity, Claude and Copilot are not Google, so they never appear there. Google's own AI surfaces do appear, but not separately: per Google Search Central, sites appearing in AI Overviews and AI Mode are reported in the Performance report within the Web search type, with no filter to split them out.

The practical consequence is that a drop in Web clicks with flat impressions can be AI Overviews answering the question for you, and Search Console will not label it. Use it for what it does well, query and page level data on Google, and accept that the AI layer has to be measured elsewhere. Anyone selling you an "AI Overviews report" built purely from Search Console data is inferring, not measuring.

AI visibility tools compared.

Citations and share of voice need a tool or a panel, because no analytics platform sees inside the answer. We checked each of the following on its own site. All of them run prompts on a schedule and record whether you are mentioned or cited; they differ on which engines they cover, how deep the reporting goes, and who they are built for.

Pick on two criteria. First, does it cover the engines your customers use, which for most businesses means ChatGPT and Google's AI surfaces before anything else. Second, can you see the actual answers, not just a score, because the answers are where you learn which competitor and which source beat you. Every tool is only as good as the prompt list someone typed into it on day one, so build the list before you buy.

Building a monthly prompt panel by hand.

A manual panel costs nothing but an afternoon and it is the most honest check you can run. It is also the right way to build the prompt list a tool will later automate. Here is the version we use.

• Write 20 to 40 prompts in the words customers use: problems, comparisons, "best X for Y in [city]", "is X worth it", "how much does X cost". Pull them from sales calls, support tickets and Search Console queries. Keep brand-name prompts to a handful.
• Fix the engines: ChatGPT, Google AI Mode or AI Overviews, and one or two more that match your audience. Use a clean session with no login and no history.
• For each prompt and engine, log four fields: mentioned (yes or no), cited (which URL, if any), position in the answer (first named, listed, or absent), and who was named instead.
• Run it the same week every month, same engines, same prompts. Add prompts over time but never delete the originals, so the trend stays comparable.
• Compute share of voice as the percentage of prompt and engine pairs where you were mentioned, and citation share as the percentage where one of your URLs appeared. Track both by prompt group.

After three months the panel tells you which pages the models trust, which competitors own which questions, and which third-party sources keep getting cited instead of you. That last column is your outreach list. Our guide on how to get found by AI search covers what to do with it.

What numbers to report to leadership.

• AI sessions and their share of total sessions, from the custom channel, with the month-over-month trend and a note that the real number is higher because of stripped referrers.
• Key events and leads from AI sessions, and the key event rate compared with Organic Search. On our own site, 90 days of lead data showed ChatGPT visitors converting at 5.2% against 2.4% from Google, on a small sample. Report your own version of that comparison.
• Top landing pages from AI sources. This is what proves which content earns the traffic.
• Share of voice and citation share from the panel or tool, by prompt group, against two or three named competitors.
• Gaps: the prompt groups where you are absent and the sources being cited instead. This is the list that turns the report into next month's work.

Five lines, one page, every month. Leadership does not need the regex. They need to know whether AI search is sending people who buy, which pages do it, and where you are losing.

What usually goes wrong.

• In-app browsers strip referrers. Mobile assistant apps and pasted links arrive with no referrer and land in Direct, so a chunk of your AI traffic is invisible. Watch for Direct growth on pages that only make sense as answer click-throughs.
• Bot traffic in Direct. AI crawlers that execute your tag can register as sessions with no source. Check for sessions with zero engagement and odd user agents before you credit Direct with a lift.
• A regex that is too broad. Matching google.com or openai.com at the root pulls ordinary search and documentation visits into the AI channel and inflates everything.
• Channel order left alone. A custom AI channel placed below Referral never fires, because Referral claims the session first.
• Prompt lists made of brand names. You will always be mentioned when someone asks about you by name. Track the problems and comparisons customers actually type.
• Comparing panels that changed. New engines, deleted prompts or a logged-in session make month three incomparable to month one.

Where the tools stop and the work begins.

GA4, a visibility tool and a panel will tell you what is happening. None of them change it. Closing the gaps means fixing crawler access, rewriting pages to answer first with specifics, lining up your entity across the web, and earning mentions on the sources the models already cite. That is the job our AI search optimization services do, inside the SEO program, with this exact measurement stack reporting the result. If you want the setup done and the monthly report landing in your inbox, Hire Theory Road.