AI citation tracking: the GA4 setup, the tools, the blind spots

AI citation tracking splits into three layers, and only the first is free. Some AI referral traffic is trackable in GA4 with a custom channel you set up in about ten minutes. Some is only visible through a paid tracker. And a large share is invisible by design, because a model can quote you to someone who never clicks. If you bought into answer engine optimization and now have to prove it worked, this is the honest map: what you can measure for free, what you have to pay for, and what no tool will ever show you.

The measurement gap AEO created

The gap is simple: brands show up in AI answers faster than they can measure it. A Goodfirms survey of 100 marketers (April 2026) found that 89% of brands already appear in AI search results, while only 14% track their AI citations, even though 43% call AI search optimization a strategic priority. Most teams are, in the report's words, visible but blind.

That is an awkward place to be if you are the person who approved the AEO spend. You can feel the shift happening, your competitors get name-dropped in ChatGPT answers, but the dashboard you report on does not have a row for it. The rest of this guide is the stack that fills in that row, with its blind spots labeled instead of hidden. If you are still choosing tooling, the companion piece is our breakdown of the best answer engine optimization platforms; this post is the measurement method, not the buying guide.

Step 1: a free GA4 channel for AI traffic

The fastest way to start tracking AI referral traffic in GA4 is a custom channel group that catches the assistants by referrer. It is free, takes about ten minutes, and it is the floor everything else builds on.

What GA4 already does, and what it misses

On May 13, 2026, Google added a native "AI Assistant" channel to GA4, with broad rollout around June 7. It auto-classifies traffic from recognized assistants with no setup. The catch is coverage: the native AI Assistant channel currently recognizes ChatGPT, Gemini, Deepseek, Copilot, and Grok, and notably leaves out Perplexity, whose visits still fall into Referral. Claude is absent from the current published definition too. So the native channel is a head start, not a complete picture, and you still want a custom channel to catch what it drops.

The custom channel regex

Create a custom channel group (Admin, then Channel groups), add a channel called "AI Traffic," and give it one condition: Session source/referrer matches regex. Paste this pattern, which covers the platforms behind over 99% of measurable AI referral traffic as of mid-2026:

chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|bard\.google\.com|copilot\.microsoft\.com|bing\.com/chat|deepseek\.com|grok\.com|meta\.ai|you\.com

It includes Perplexity and the OpenAI and Google variants the native channel misses or splits, which is the whole reason to keep a custom channel alongside the built-in one. ChatGPT is the source that matters most here: it drives about 87% of measurable AI referral traffic (Conductor, 2026), so if your regex gets the ChatGPT domains right, it gets most of the volume right. We go engine by engine on the content side in how to rank in ChatGPT.

Put the rule above Referral

Drag the AI Traffic channel above Referral. GA4 assigns each session to the first channel whose definition it matches, evaluated top to bottom, which Google's own channel-group docs state plainly. If AI Traffic sits below Referral, every AI visit matches the generic Referral catch-all first and your new channel stays empty. The configuration looks fine and reports nothing. This ordering step is the single most common setup mistake, and it takes about two minutes to fix.

The catch: most AI sessions never carry a referrer

Here is the part the tutorials skip. An estimated 60-70% of real AI sessions arrive with no referrer at all and land in Direct, not your new channel. Mobile in-app browsers, copy-paste between apps, native AI apps, and the ChatGPT Atlas browser strip the referrer header before the visit reaches GA4. So your AI Traffic channel is a floor, not a count. It undercounts by design, and you should report it that way, with a note that the real number is higher than the channel shows.

Step 2: manual prompt testing, the baseline you can run this week

The cheapest way to measure AI visibility is to ask the assistants yourself, on a schedule, and write down what they say. It measures citation directly, which is the one thing GA4 cannot see.

Build a list of the 15 to 30 questions your buyers actually ask: the comparisons, the "best tool for X" queries, the problem-shaped questions that lead to your category. Then run them weekly across ChatGPT, Perplexity, Gemini, Claude, and Copilot. For each prompt, log three things: whether you were cited, which competitor was quoted instead, and where you landed in the answer. A spreadsheet is enough to start.

This is the honest floor because it measures citation, not clicks, and citation is the metric that actually moves first. Most AI visibility never becomes a visit. As of April 2026, 64.82% of Google searches ended without a click, Google's AI Mode runs around 88% zero-click, and ChatGPT Search about 82%. A click-based metric misses the majority of the value, so you watch the citation directly while the traffic, where it exists, trickles in behind it. The same zero-click pressure is reshaping classic search, which we cover in how to show up in Google AI Overviews.

Step 3: the paid trackers, compared honestly

A paid visibility tracker automates that manual log. It runs your prompts across the engines on a schedule, scores your share of voice, and shows which competitors get quoted instead of you. What it cannot do is make you citable. It grades the page; it does not change it. Buy one to measure, not to improve, and never trust one that promises citations.

Three tools cover most of the market in mid-2026. Prices and plan details below are current as of this writing and change often, so confirm them before you buy.

The traps live in the fine print. Otterly's entry plans include four engines, with Gemini and Google AI Mode sold as paid add-ons, so the headline $29 Lite is real but the all-surfaces price is higher than it first looks. Annual billing lowers Otterly's effective prices to roughly $25, $160, and $422. Scrunch has run since 2023, raised $19M, and serves 500-plus brands, with SOC 2 Type II reserved for its enterprise tier. Profound is the enterprise option: self-serve starts around $499 a month, and the company has raised roughly $155M at a reported $1B valuation, with about 10% of the Fortune 500 on the platform. Across all three, the accuracy caveat is the same: every tool samples a fixed number of prompts on a schedule, so a small prompt allowance means you are tracking a narrow slice of how people actually ask. More prompts cost more money.

Do you even need one yet?

Probably not, if AI referral is still a rounding error in your analytics and you publish a handful of posts a month. A subscription is premature when a free manual log answers the same question. Buy a tracker when AI search drives enough traffic or leads that you need to watch it weekly, or when you are managing several brands and competitors and the spreadsheet has become its own job. For a one-time "where do I stand" read, a single audit beats a recurring bill. The same skepticism that protects you from bad data protects you from bad spend, which is the discipline we bring to fact-checking every published claim too.

Step 4: the invisible majority, and how to proxy it

The largest layer of AI visibility produces no analytics event at all, so you measure it indirectly or not at all. A model can describe your product, quote your docs, and recommend you by name to a buyer who never clicks a link. That mention is real and it shapes the decision, but it never reaches GA4. This is the no-click reality behind the 89%-appear, 14%-track gap: visibility is outrunning measurement, and a lot of it will never produce a trackable visit.

You cannot count what leaves no trace, but you can proxy it. Four signals are worth watching together:

• Branded-search lift. When someone hears about you in an AI answer and later searches your name, that shows up as branded query volume in Search Console. A rising branded trend after AEO work is one of the cleaner proxies for no-click exposure.
• Direct-traffic trend after AEO work. Those stripped-referrer sessions pile into Direct. A Direct trend that rises in step with your AEO pushes is a noisy but useful signal that AI mentions are sending people who type your URL straight in.
• Self-reported attribution. A "how did you hear about us?" field on signup catches what analytics structurally cannot. When "ChatGPT recommended you" starts appearing, that is data no channel report will give you.
• Assisted conversions. AI often touches the journey early, not at the last click. Look at assisted and multi-touch paths, not just last-click attribution, or you will under-credit the whole channel.

None of these is exact, and you should never present them as if they were. The point is to triangulate a trend, not to fake a precise number.

Why a small channel is worth proving

Because it converts far above everything else. AI referral is still only about 1.08% of all website traffic (Conductor, 2026), but the quality is the story. AI search referral converts at 4.4x to 23x the rate of organic search, on measured Q4 2025 and Q1 2026 data from Semrush, Seer Interactive, and Ahrefs. Ahrefs found that 0.5% of visitors from AI search drove 12.1% of total signups, a 23x multiplier. Per platform, Seer Interactive measured ChatGPT at a 15.9% conversion rate, Perplexity at 10.5%, Claude at 5.0%, and Gemini at 3.0%. A channel that small with conversion that high is exactly the kind of thing worth proving to the person holding the budget, which is the whole reason to measure it carefully rather than wave it away as noise.

A measurement stack honest about its blind spots

Put the four steps together and you have a stack that is cheap to start and clear about what it cannot see:

1. Set up the free GA4 AI channel, ordered above Referral, and label it a floor.
2. Run a weekly manual prompt log across the major assistants to measure citation directly.
3. Add a paid tracker only when volume justifies watching it weekly.
4. Track proxy signals for the no-click layer, and never claim that layer is fully measured.

The discipline is in the labeling. The GA4 number is a floor, the manual log is a sample, the tracker is a sample you pay for, and the no-click layer is an estimate. Report each one as what it is, and you will keep the trust that a single overconfident chart would burn. Measurement that admits its limits is the kind a CFO actually believes.

Where Lyra fits

Lyra is a content tool, not a tracker, and the distinction is the point of this whole guide. Measurement tells you where you stand. Content is what moves it. She will not show you a dashboard of your citations, and she does not pretend to. What she does is write the posts answer engines cite: the answer in the first line, headings shaped like real questions, verified facts with dates, and clean structure, then she opens each one as a pull request you review and merge. Nothing auto-publishes. We built her after running this exact loop by hand, a story we tell in why we built Lyra.

So pair her with a tracker you trust. The tracker tells you the score; Lyra is one half of what changes it. If you want to see how she would write for your blog, you can request early access and tell us about it.

Measurement shows you the gap; content is what closes it. Lyra writes posts built for AI citation and ships each as a PR you merge, so the thing you are measuring actually has a reason to move.

Talk to the founder → · Join the waitlist