We’ve argued that your site is the source AI assistants pull from when someone asks about your category, and covered how to make it citeable. This post answers the question that comes next, and it has the least satisfying answer of the three: how do you tell whether any of it is working?
The starting point is that you mostly can’t, at least not from Google Analytics alone. When an assistant recommends you, or an AI Overview summarizes your page, the visit that follows often arrives with no referrer and lands in the “direct” bucket, next to the people who typed your URL from memory. Frequently there is no visit at all, because the answer was complete on its own. The influence is real, the measurement lags behind it, and the gap quietly changes how teams make decisions.
Here is the failure mode. Your measurable organic traffic looks flat, so it reads as SEO stalling. Meanwhile the traffic you can’t measure is the part that’s growing: the average site’s AI-assistant traffic grew close to tenfold in the year to mid-2025, even as classic search traffic slipped. Read the flat line at face value and you cut the exact content the assistants are recommending. Measuring AI search exists to stop you from defunding your best channel because it doesn’t show up on a chart.
Why AI search traffic is invisible in Google Analytics
Analytics attributes a visit using the referrer the browser sends. A click from Google search carries one, so it lands in your reports as organic. AI referrals behave less predictably, for three reasons.
- The platforms disagree on referrers. Perplexity and Gemini usually pass a referrer, so those visits can be attributed. ChatGPT’s website tags its citations (you can filter for
utm_source=chatgpt.com), but the mobile and desktop apps, in-app browsers, and copy-pasted links that a lot of people use arrive with no source at all. That last group is where most of the attribution loss happens. - Many answers end in zero clicks. When an assistant answers completely, or an AI Overview sits at the top of the results, the reader has what they need and never clicks. There is no session to attribute because there is no session.
- What remains blends in. The visits that do come through land in
director genericorganic, indistinguishable from the rest of your traffic.
This is not a niche problem. Pew Research found that when an AI summary is present, users click a traditional result 8% of the time, versus 15% without one, and they click a link inside the summary just 1% of the time. In the same study, most users encountered at least one AI summary in a single month of searching. A large and growing share of the buying journey now happens in a surface your analytics was never built to see.
What Google Analytics can now see, and what it still misses
The tooling is catching up. In May 2026, Google Analytics added a native AI Assistant channel that groups recognized assistant traffic (ChatGPT, Gemini, Copilot, Grok, and others) under its own line with the medium ai-assistant, instead of scattering it across direct and referral. For the first time you can see AI traffic as its own channel in standard reports.
Two limits are worth knowing. First, the channel only classifies a session that arrives with a referrer it recognizes, so it captures the attributable slice and misses the no-referrer majority. Second, it deliberately excludes Google’s own AI Overviews and AI Mode, which still count as organic search, and at least for now it doesn’t list Perplexity, which passes a referrer and lands in referral. It is a genuine improvement, and it is partial. For fuller coverage, add a custom channel group with a regex on the AI domains that do pass a referrer (chatgpt.com, perplexity.ai, gemini.google.com, claude.ai, and the like) plus ChatGPT’s utm_source=chatgpt.com tag, which also picks up platforms before Google classifies them natively. Neither approach recovers the traffic that shows up with no source, which is why the next section matters more than the GA4 setup.
How to track AI referral traffic without a referrer
No single number stands in for the missing data. A few signals together get you a defensible read.
Server logs: who is actually reading you
Your access logs record every crawler that fetches a page, by user-agent. Filtering for AI user-agents tells you whether the models are reading you at all, which separates “not cited” from “never seen.” It helps to know the three tiers:
- Training crawlers (
GPTBot,ClaudeBot) gather content to train future models. - Search and index bots (
OAI-SearchBot,PerplexityBot) build the index an assistant cites from. - User-triggered fetchers (
ChatGPT-User,Perplexity-User) fire in real time when a person asks something and the assistant pulls your page to answer.
A one-line filter over your logs shows which are hitting you and how often:
grep -E "GPTBot|ClaudeBot|OAI-SearchBot|PerplexityBot|ChatGPT-User|Perplexity-User" access.logThat last tier is the closest thing to a live citation signal: it means a person asked a question and a model reached for your page to answer it. A rising trend there is the strongest log-based evidence you are being used in answers, not just harvested for training.
Branded search: are more people looking you up
Open Search Console and watch branded queries, the searches that include your name. A steady climb means something upstream is introducing people to you before they ever land, which is exactly what a citation in an AI answer does. It is indirect, but branded-search growth with no matching campaign or press is one of the cleaner signals that AI exposure is doing its job.
Self-reported attribution: just ask
The oldest tool outperforms the analytics here. Put a “how did you hear about us?” field on your contact form with an explicit AI option, and read the answers. Note when it comes up on sales calls. For now, the customer telling you directly is the most reliable attribution you have, and it costs one form field.
Visibility tools: are you in the answer at all
A category of tools checks whether the major models cite you for the prompts you care about, and tracks your share of voice against competitors. Treat them as a complement to the signals above, not a replacement. They tell you whether you are in the answer. Your logs and your forms tell you what being in the answer is actually doing for you.
Reading the signals together
None of these is proof on its own. Together they triangulate. Say your organic sessions are flat quarter over quarter, but your logs show ChatGPT-User fetches have tripled, branded search is up 30% with no campaign behind it, and three contact forms last month said “ChatGPT.” No report draws that line for you, but the pattern is hard to miss: AI is feeding you traffic that GA4 is filing under direct.
Keep the scale honest while you do this. Ahrefs found that 63% of sites get at least some AI-assistant traffic, but for most it still averages a fraction of a percent of visits, and part of that low number is the measurement gap itself: referrer-less visits from assistants like Copilot land in “direct” and never get counted. The trend line matters more than the absolute today.
The discipline is to baseline these signals now and watch the trend, not chase a precise number that does not exist yet. Set a monthly glance at crawler activity, branded search, direct traffic, and self-reported mentions, and watch whether they move together. Above all, do not cut content the models are citing because a dashboard shows it flat. That flat line is a measurement gap. Treating it as a performance verdict is the expensive mistake this whole exercise exists to prevent.
When to bring in help
Most of this is a competent team’s afternoon: a GA4 channel configured, a log filter saved, a Search Console view bookmarked, a form field added. Do it yourself if your stack gives you access to server logs and someone who can read them.
Bring in help when the logs sit behind a CDN that needs configuring to expose user-agents, when you want the signals pulled into ongoing reporting instead of a monthly manual check, or when you need someone to tell a real AI-referral trend from normal noise. That ongoing read is the work Monitor exists to do: baseline the signals, watch them over time, and separate the drift that matters from the drift that doesn’t.
The principle
Attribution always lags behavior. People found businesses through word of mouth for years before anyone could put a number on it, and an AI recommendation is word of mouth at machine scale. The tools will catch up. Until they do, the teams that win are the ones who measure the influence they can’t fully see and keep investing in the content that earns it, rather than cutting it because the dashboard stayed quiet.
If you want a read on how much AI search is already sending your way and what your logs are telling you, tell us what you’ve got.