Somewhere in your traffic right now are people who found you through an AI assistant. They asked ChatGPT for a recommendation, or Perplexity for an explanation, and an answer pointed them at your site. This is a real acquisition channel with distinctive, usually favorable behavior - and in most analytics setups it is either mislabeled or invisible. Here is how to actually see it.
What an AI referral looks like when you can see it
When an assistant links to you and a human clicks through in a browser, the visit can carry a referrer header naming the assistant's domain. The signatures worth knowing:
chatgpt.com ChatGPT answers (also historical chat.openai.com) perplexity.ai Perplexity citations gemini.google.com Gemini claude.ai Claude copilot.microsoft.com Microsoft Copilot
Some assistants also append UTM parameters - ChatGPT has used utm_source=chatgpt.com on outbound links - which survive even when referrer policies would otherwise hide the source. When one of these appears, treat it the way you treated an organic search click in 2010: a high-intent visitor whose journey started with a question.
Why you see the domain but never the conversation
A detail that puzzles people the first time they look: the referrer says chatgpt.com, never the URL of the chat that sent the visitor. That is the web's referrer policy working as designed. The default policy most sites and apps use (strict-origin-when-cross-origin) sends only the origin - the bare domain - on cross-site navigation, deliberately stripping paths and query strings so browsing context does not leak between sites. For AI assistants this is doubly certain, since a chat URL can contain private conversation content. The practical consequence: you can measure that an assistant sent a visitor and which page they landed on, but never what was asked. The landing page is your only lens back into the question - one more reason the landing-page dimension matters more in this channel than in any other.
Why the channel undercounts itself
The visible referrals are the floor, not the ceiling. Plenty of AI-originated visits arrive dark:
- Clicks from native apps (the ChatGPT mobile app, desktop clients) often carry no referrer at all and land in "direct."
- Strict referrer policies on the assistant's side can strip the source even from browser clicks.
- The largest effect never produces a click: a user reads the assistant's answer - built partly from your content - and visits you later by searching your name or typing your URL. That shows up as branded search or direct, with the AI's role laundered out.
So measure the visible channel precisely, and treat its trend as a proxy for the larger dark channel moving underneath it.
Not all assistants can even send you traffic
Calibrate expectations per engine, because the assistants differ structurally in how much measurable traffic they are capable of sending. Perplexity cites sources prominently on every answer, which makes it the most generous click-sender relative to its size and the cleanest referrer to study. ChatGPT links out mainly in its search-grounded answers, so its referral volume tracks how often your topics trigger web search rather than how often you appear in training-informed answers. Claude's referrals are lighter still. And Google's AI Overviews are the big asterisk: clicks from them arrive attributed as ordinary Google organic traffic, so the AI-driven share of your Google clicks is folded invisibly into your existing search numbers. If your organic click-through drifts down while impressions hold, AI Overviews answering above your listing is one of the first hypotheses to check - and no referrer segment will confirm it directly. Measure each engine against its own mechanics, not against a single "AI channel" expectation.
The behavioral tell: deep pages, not homepages
AI referrals have a recognizable shape. Assistants cite the specific page that answered the question, so visitors land deep - a comparison page, a piece of documentation, a five-year-old technical post - rather than on your homepage. They arrive mid-journey, already briefed by the assistant's summary, which tends to make them decisive: they came to verify or act, not to browse. If your AI-referred visitors bounce from exactly one page, that page answered the question badly, or answered it completely - and which one is a question your engagement data can settle.
Segmenting the channel in GA4
GA4 will not break out AI sources for you, but the data is there. Build a custom segment (or an exploration filter) matching session source against the assistant domains - a regex like:
chatgpt\.com|openai\.com|perplexity\.ai|gemini\.google\.com|claude\.ai|copilot\.microsoft\.com
Then compare that segment against organic search on engagement rate, conversions, and landing pages. Two practical warnings: source strings change as vendors rebrand (chat.openai.com became chatgpt.com), so revisit the regex quarterly; and keep the segment separate from "referral" as a channel group, because default groupings will happily file perplexity.ai under generic referral noise.
If you want the segment to survive in everyday reports rather than living only in explorations, promote it to a custom channel group: define an "AI assistants" channel matched on those source domains, ordered above generic Referral so it wins the assignment. Once it exists as a channel, three views become routine. Landing pages by AI channel tells you which content earns citations. Engaged sessions and conversions per assistant tell you which engines send readers versus skimmers. And the week-over-week trend line tells you whether your AI visibility is compounding - the number a quarterly report actually wants. Resist the urge to benchmark against published industry averages; the channel is young, definitions vary wildly between tools, and your own baseline from last quarter is the only comparison that means anything yet.
What earns citations in the first place
Working backwards from the visible referrals teaches you something about the content itself. Assistants cite pages that resolve a question cleanly: a definition stated plainly, a comparison with actual criteria, documentation with steps, a dated answer to a specific problem. Homepage-style pages - broad, brand-forward, question-free - almost never appear, which is why AI referrals land deep. If you want more of this channel, the implication is uncomfortable but useful: write pages that answer one question completely, with the answer visible in the HTML rather than assembled by JavaScript, and keep them current. The crawl side of your data previews this before the referrals arrive - the pages retrieval bots fetch most are the candidates answer engines are considering citing.
The WordPress-side hygiene that protects the signal
Measurement is only as good as what survives the trip to your server, and a few WordPress habits quietly destroy AI referral data:
- Redirect chains eat parameters. If an assistant links your http:// or non-www variant and your redirect setup is sloppy, UTM parameters and referrers can be lost mid-hop. One canonical host, one clean 301, no chains.
- Consent-gated analytics undercounts the channel. A visitor who declines cookies never fires the GA4 tag, and AI referrals skew toward exactly the privacy-conscious audiences that decline. Server-side traffic data has no such gate - another reason to read both sources side by side.
- Keep answers in the HTML. Retrieval bots read your served markup. Key facts assembled client-side by JavaScript - prices in a widget, specs behind a tab - may simply not exist for the engine deciding whether to cite you. If a page is your answer to a question, the answer should be in the source.
- Date your content honestly. Answer engines weigh freshness, and visible, truthful updated-dates on your money pages are cheap credibility - both with the engines and with the pre-briefed human who clicks through to verify.
A monthly half-hour that compounds
The channel rewards a small ritual more than a big project. Once a month: pull the AI-assistant segment and note sessions, engaged sessions, and top landing pages. Put that beside your crawl view - which retrieval bots came, which pages they fetched most. Check Search Console for branded-query volume drift. Then ask the only three questions that matter yet: is the trend up, which pages earn the citations, and did any engine's behavior change sharply? Thirty minutes, four data points, written down. Within two quarters you own something almost nobody in your niche has: a baseline for the AI channel from before it mattered - which is exactly the artifact that makes next year's decisions easy.
Attribution honesty: what this channel will never tell you
Be clear-eyed about what even perfect segmentation captures. AI referral measurement is last-click measurement, and this channel's influence is unusually front-loaded: the assistant answers, the user absorbs, and the visit - if it ever comes - arrives later through branded search or a typed URL. Two proxies help you see around the corner. Watch branded search volume in Search Console alongside your AI referral trend; if both rise together while nothing else changed, the assistant channel is likely larger than its click count. And treat any conversion path that starts with an AI referral as evidence about the channel's quality, not its size - a channel that converts at double your average from deep landing pages is telling you the upstream answers are pre-qualifying your visitors. Illustrative, not a benchmark: a docs page cited by an assistant might show thirty referral clicks in a month - and a much larger, unmeasurable audience that read the assistant's summary of that page and never clicked. The thirty clicks are the tip you can measure; manage the iceberg by its tip, but do not mistake the tip for the iceberg.
Closing the loop: crawls in, referrals out
Referral data alone answers half the question. The full economics of the AI channel are an exchange rate: how much crawling do you tolerate (cost) against how many cited, clicking humans come back (return). That requires seeing both sides - server-side crawl volume per bot, and human referral sessions per assistant - in one place. This is exactly the loop Viz closes on WordPress: it classifies the crawlers hitting your server, and its GA4 integration ties the human sessions AI assistants send back to the other side of the ledger.
The sites that will negotiate well with the AI economy - whether that means blocking, licensing, or leaning in - are the ones that can state their exchange rate as a number. Most site owners cannot yet. It is an easy advantage to take.