How to Measure AI Search Traffic in GA4 and Search Console

How to measure AI search traffic comes down to two copy-paste filters: a GA4 regex that isolates referral visits from AI assistants like ChatGPT and Perplexity, and a Search Console regex that surfaces the long, conversational queries AI search produces. Neither filter is perfect. Anyone who tells you AI traffic can be measured precisely is selling something. But together they turn "we think AI sends us traffic" into a number you can watch move month over month.

Both setups take about ten minutes. Here's each one, what it actually captures, and where it lies to you.

Why is AI search traffic hard to measure?

Because most of it doesn't announce itself.

AI answers touch your business in three ways, and only one of them shows up cleanly in analytics:

1. Referral clicks. Someone clicks a citation link in ChatGPT or Perplexity and lands on your site. GA4 sees the referrer. Measurable.
2. AI Overviews activity. Google folds AI Overviews impressions and clicks into regular web search totals in Search Console. There's no "AI Overviews" filter. Blended.
3. Influence. Someone asks an assistant for a recommendation, hears your name, and later visits directly or Googles your brand. GA4 logs direct or organic. The AI's role is invisible.

So real measurement means one clean number, one proxy, and one acknowledged blind spot. Let's get the clean number first.

How do you measure AI referrals in GA4?

Build an Exploration filtered to AI referral sources.

1. In GA4, go to Explore and create a blank exploration.
2. Add the dimension Session source / medium and the metrics you care about: sessions, engaged sessions, key events.
3. Add a filter on Session source / medium → matches regex → paste:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|copilot\.microsoft\.com|gemini\.google\.com|claude\.ai

4. Add Landing page as a second dimension to see which pages AI assistants send people to.

That's the whole setup. A few notes on the pattern:

• The dots are escaped (\.) so they match literal dots instead of any character.
• chat.openai.com is in there alongside chatgpt.com to catch legacy referrers.
• New assistants launch constantly. Recheck your raw referral report quarterly and extend the pattern. It's one more | per source.

Expect the totals to be small. AI referral traffic is a trickle for almost every site right now. But watch what it does, not just how big it is: check engagement rate and conversions against your organic average. Visitors arriving from an AI answer have often been pre-sold by the recommendation itself, so what these sessions lack in volume they can repay in intent. Judge the channel on that, not on raw sessions.

Save the exploration once it's built, and note the date you created it somewhere you'll see it. GA4 explorations only read data going forward from whatever range you set, and you'll want a clean "since we started counting" baseline when someone asks whether AI traffic is growing. It will be. The question worth answering is how fast, and which pages are earning it.

One trap to avoid: don't build this as a custom channel group and call it done. Channel groups are easy to misconfigure and painful to audit later. The exploration with an explicit regex is transparent, because anyone on the team can open it, read the pattern, and know exactly what's being counted. Boring and inspectable beats clever and opaque in measurement, every time.

How do you find AI-style queries in Search Console?

Search Console won't label AI activity, so you approach it sideways: filter for queries shaped like conversations rather than keywords.

1. Open Performance → Search results.
2. Click + New → Query → Custom (regex) → paste:

(\b\w+\b\s){7,}

3. Compare the last 28 days against the previous period.

What it isolates: queries containing roughly eight or more words. The pattern matches seven repetitions of "a word followed by a space," and the query's final word makes eight. Nobody types "best emergency plumber open now near downtown austin with weekend availability" into a search box out of habit. Long, specific, question-shaped queries are the signature of conversational search: AI experiences issuing detailed searches, and users who've learned to search the way they prompt.

Its limits matter too, because this filter gets waved around as more than it is:

• It's a proxy for query style, not query origin. A verbose human matches it; a terse AI-generated search doesn't. Fan-out queries in particular are often short and literal, so the long-query filter won't catch those.
• The word boundaries count hyphenated and punctuated terms in ways that occasionally over- or under-count words. Close enough for trend lines, not for forensics.
• Volumes will look tiny, and Search Console's privacy thresholds hide many rare queries entirely. Your true long-tail is bigger than the report shows.

Use it for what it is: a monthly trend line, plus a mining rig for question-shaped queries you should have pages answering. Each long query where you're getting impressions without a dedicated answer-first page is a page idea with proof of demand attached. That feeds directly into the work of showing up in Google AI Overviews.

What other signals fill in the picture?

Bing Webmaster Tools' AI Performance report comes first. It's free, and it's the most direct data any engine currently gives you: how your content performs in Copilot and Bing's AI experiences, including the grounding queries AI ran before surfacing you. One ecosystem's sample, but real query language from real AI systems. Take what it teaches you about phrasing and apply it everywhere.

Branded search volume is the blind-spot workaround. When AI assistants recommend you, some people verify by Googling your name. Rising branded queries without a matching campaign or press moment is circumstantial evidence of AI influence. Circumstantial, but worth tracking on the same monthly cadence.

The landing-page mix in your AI referral report rounds out the set. Which pages do assistants link when they cite you? Those are your citable assets. Study their shape and build more like them; we break down the pattern in how to get cited by ChatGPT.

What should your monthly dashboard track?

Keep it to five lines. Pull them the same week each month:

This is the exact measurement layer we build in the first phase of our Reforge Method. You can't reforge what you haven't assayed, and every client dashboard starts with these five lines.

One caution on reading the numbers. Small AI referral totals don't mean AI search doesn't matter; they mean clicks are the smallest part of how AI answers affect you. The recommendation itself, your name spoken into a buyer's research process, never hits your analytics at all. If the small numbers tempt you to write the channel off, read our full answer on whether SEO is dead before you do. Then build the exploration, save the GSC filter, and book twenty minutes next month to pull the five lines. Month two is when the dashboard starts earning its keep.