Most writing about generative engine optimisation stops at the interesting part: how to get cited. The harder and less discussed question is what happens next. How much traffic actually arrives, whether it is worth anything, and why the number in your analytics is probably wrong in a predictable direction.
The Channel Changed Shape in 2026
For most of 2025, AI referral traffic was a rounding error for the majority of sites, largely for a structural reason: AI answers rarely contained links. That began to change during 2026. According to analysis published by Similarweb, when ChatGPT started rendering brand names as links, the share of responses containing a URL rose sharply and referral traffic to brand websites increased substantially over a very short period.
Google moved on the measurement side as well, adding a native AI Assistant channel to GA4 in May 2026. AI traffic now has somewhere to live in default reporting instead of being scattered across Referral and Direct.
Neither development means the channel is solved. Default reporting remains incomplete in several ways, and each of them causes an undercount rather than an overcount.
Why the Numbers Understate the Channel
The native channel does not cover every assistant. GA4's AI Assistant channel captures the major ones but not the full landscape. If your audience skews technical or research-heavy, some of your AI traffic is being classified elsewhere.
Many AI sessions arrive without a referrer. Depending on the platform and whether the user is in an app or a browser, a large share of AI referral sessions carry no referrer header at all and land in Direct, indistinguishable from someone typing your domain. Published estimates of how large that share is vary widely, and I have not seen a figure I would treat as authoritative. Treat it as material rather than as a precise percentage.
The tagging is recent. Consistent source tagging from AI platforms only appeared during 2025. Any comparison reaching back further is comparing a tagged period against an untagged one.
The practical consequence is worth stating plainly: whatever figure GA4 shows for AI traffic, the real number is higher. This is the opposite of the usual analytics problem, where channels flatter themselves through over-attribution. Here the channel undersells itself, which makes it easy to dismiss something that is quietly working.
Setting the Measurement Up Properly
The fix is a custom channel group rather than reliance on the default. Build a rule that matches the assistant domains you care about on the session source dimension. At the time of writing that list covers the main consumer assistants - the ChatGPT domains, Perplexity, Gemini, Copilot and Claude - and it will need revisiting as the landscape changes, which is an argument for owning the rule rather than inheriting one.
Two details decide whether this works. GA4 evaluates channel rules from the top down, so the AI rule has to sit above Organic Search and Referral or those rules will claim the sessions first. And a custom channel group applies going forward rather than retroactively, so the sooner it exists the sooner the comparison becomes meaningful.
Then add the part most setups skip: watch Direct. If AI traffic is growing and much of it arrives unattributed, Direct sessions will rise with landing-page patterns that do not look like brand traffic - deep pages, long articles, no homepage entry. Direct traffic landing on page seven of your blog is not someone typing your URL from memory.
What the Traffic Is Actually Worth
Here the volume-first instinct fails. Similarweb's 2026 analysis found that AI platform usage grew substantially year on year while AI referrals to external sites stayed close to flat. On a dashboard, that looks like a channel going nowhere.
The behaviour of the sessions that do arrive tells a different story. In the same analysis, visitors arriving from ChatGPT spent roughly 15 minutes on site against about 8 minutes for visitors from Google, generated around 12 pageviews per visit against 9, and converted on transactional sites at approximately 7 percent against 5 percent.
That gap is not marginal, and the explanation is straightforward. Someone who has already had a multi-turn conversation about their problem, and who arrives having been told you are a credible answer, is not the same person as someone three seconds into a results page. They arrive pre-qualified and behave accordingly.
Small, high-intent and underreported is an awkward combination, because each property individually pushes you toward ignoring the channel.
The Metric That Should Replace Clicks
The deeper point is that referral clicks are the wrong primary metric here, and they will stay wrong.
Much of the influence never produces a click. Someone asks which consultants work on paid media in Athens, receives a few names with a sentence of context each, and forms an impression. No session is recorded. Weeks later they arrive through a branded search, and analytics attributes the whole thing to organic or direct.
The metrics that describe performance here are share of mentions across platforms, how often you are cited, and the context of those mentions. What gets cited in one assistant is often not what gets cited in another, so presence is worth checking per platform rather than assuming from a single one.
Referral traffic still belongs in the reporting. It just belongs in the supporting cast rather than as the headline.
What This Means in the Greek Market
Greece has approximately 8.64 million internet users at 87 percent penetration, from a population just under 10 million, according to DataReportal's Digital 2026 report. That ceiling cuts in two directions.
On the opportunity side, the body of authoritative Greek-language content on most professional topics is thin. When someone asks an assistant a specific question in Greek about paid media, measurement or hiring, there are comparatively few strong sources to draw on. Becoming one of them takes a fraction of the effort the same position would require in English.
On the other side, the absolute numbers will always look unimpressive. A Greek B2B or services business might see a few dozen AI-attributed sessions a month. Judged as a traffic channel, that is easy to dismiss. Judged on quality, a small number of sessions from people who arrive already convinced can be worth considerably more than a much larger number of low-intent visits - and the clicks cost nothing.
The mistake to avoid in a small market is applying large-market volume thresholds when deciding what deserves attention.
The Honest Summary
AI traffic is currently a small channel, measured imperfectly, understated by default, and unusually valuable per session. None of those things is likely to remain true. Volume will grow, measurement will improve, and whatever advantage exists for establishing authority early will compress.
The reasonable position is to measure it properly now, judge it on quality rather than volume, and stop expecting a channel built on conversations to report like one built on clicks.
Sources
Similarweb 2026 Generative AI Visibility Index; Google Analytics Help documentation on default channel groups; DataReportal, Digital 2026: Greece (Kepios analysis, October 2025 reference data). Figures are as published by those sources at the time of writing and may change.