A default GA4 setup scatters AI-driven sessions across Direct, Referral, and AI Assistant — meaning most dashboards are quietly reporting a fraction of the real traffic.
I found this problem the annoying way — by staring at a client’s GA4 dashboard for twenty minutes convinced I was losing my mind. This was earlier this year, one of five client sites where I personally rebuilt the GA4 setup specifically to measure AI referral traffic properly. The client had mentioned, almost in passing, that they were starting to see a noticeable bump in email signups from people who said they’d “asked ChatGPT” about them. I pulled up GA4’s AI Assistant channel to check the numbers, and the total looked suspiciously small next to what the client was describing anecdotally. So I went digging into the raw referral data underneath it, and that’s when I found the actual problem: GA4 wasn’t missing the traffic. It was there. It was just scattered across three different channels, none of which added up to the real total.
If you’ve set up GA4’s default channel grouping and assumed the “AI Assistant” bucket is giving you the full picture, I’d bet money you’re undercounting too. Let me walk you through exactly what’s happening, why it happens, and the fix I’ve now rolled out across five client accounts this year.
The Problem: GA4 Is Fragmenting a Single Traffic Source
Here’s the core issue. GA4’s default channel grouping logic tries to classify AI-driven traffic using referrer domain and UTM parameter matching, the same general approach it’s always used for classifying traffic sources. The problem is that AI platforms don’t send visitors to your site in one consistent, predictable way.
Someone clicking a link inside a ChatGPT response might arrive with a referrer of chat.openai.com, which GA4 correctly buckets into AI Assistant. Someone else, clicking a link from the same platform but through a slightly different flow — a shared conversation link, a mobile app in-app browser, or a link that’s been passed through an intermediate redirect — might arrive with no referrer at all, which GA4 defaults into Direct traffic. And a third visitor, arriving from a ChatGPT response that included a source citation formatted differently, might get picked up as Referral or even Organic Search if the underlying URL structure happens to match a pattern GA4 associates with search engines.
I saw all three of these happening on the same client site, for the same platform, within the same week. Same source of traffic, three different buckets in GA4, and only one of those buckets is labeled clearly enough that anyone would think to check it. That’s the undercounting problem in a nutshell — it’s not that GA4 fails to log the sessions, it’s that it mislabels a meaningful chunk of them, and most of us go straight to the AI Assistant channel and assume that number is the whole story.
What I Actually Found Across Five Client Accounts
Once I knew what to look for, I went back through the other four accounts I’d set up this year and checked each one the same way. The pattern held every time, though the split ratio varied a fair amount by industry and by how the site itself linked internally.
On the low end, one client — a fairly established SaaS company with strong brand recognition — had roughly 60% of their true AI-referred sessions correctly bucketed as AI Assistant, with the rest scattered mostly into Direct. On the high end, a newer eCommerce client had barely 35% landing in the correct bucket, with the majority showing up as Direct traffic and a smaller chunk misclassified as Organic Search because of how certain AI platforms were structuring outbound links at the time.
That’s not a rounding error. That’s the difference between telling a client “AI search is a minor traffic source, not worth much attention” and telling them “AI search is quietly becoming one of your top five channels, and we should build a real strategy around it.” I’ve had that exact conversation with two different clients this year, and in both cases, once we corrected the measurement, the real number changed the entire conversation about budget allocation for the next quarter.
Why This Actually Matters, Beyond Just Getting the Number Right
I want to be honest about why I care about this beyond simple reporting accuracy. If you’re undercounting AI traffic, you’re not just working with a smaller number — you’re making genuinely bad strategic decisions off bad data. Clients deprioritize content investments that are actually working. Budget gets pulled from initiatives that deserve more, not less. And when someone eventually asks “should we care about AI search visibility,” the honest data-backed answer gets buried under a channel report that’s quietly lying by omission.
I’ve also noticed that undercounted AI traffic tends to hide conversion patterns that matter a lot for strategy. AI-referred visitors, in my experience across these five accounts, tend to convert at a meaningfully higher rate than average organic traffic, because they arrive after already getting their initial questions answered elsewhere. If that traffic is scattered into Direct and blended with your actual direct/type-in traffic, that conversion signal gets diluted and disappears into an average that doesn’t tell you anything useful. Fixing the classification doesn’t just fix a number, it surfaces a genuinely important behavioral insight that was sitting there the whole time, hidden by a labeling problem.
The Fix I’ve Been Implementing: A Custom Channel Grouping
The most reliable fix I’ve found doesn’t rely on GA4 catching up and fixing its default classification logic — because honestly, I wouldn’t wait on that. Instead, I build a custom channel group specifically for AI referral sources, using a broader and more deliberate set of matching rules than GA4’s defaults apply.
The first step is expanding referrer matching well beyond just the obvious domains. Yes, chat.openai.com, gemini.google.com, and perplexity.ai are the obvious ones, but I’ve also had to add copilot.microsoft.com, claude.ai, and a handful of app-specific in-app browser signatures that show up in the user agent string rather than the referrer field. Every client’s traffic mix looks a little different depending on their audience, so this list isn’t universal — I build it out per account based on what I actually see in the raw referral data over the first few weeks.
The second step, and the one that catches the most previously hidden traffic, is landing page and session pattern analysis rather than referrer alone. A meaningful chunk of AI-referred sessions arrive with no referrer at all, especially from mobile app contexts where the referrer header gets stripped. What I do instead is look for sessions with zero referrer, combined with specific behavioral signatures that are unusual for genuine direct/type-in traffic — landing directly on a deep content page rather than the homepage, for instance, is a strong signal for someone who clicked a specific AI-generated citation rather than someone who typed your domain into a browser from memory.
The third piece is UTM tagging wherever you actually have control over it. This doesn’t solve the whole problem since most AI platforms don’t let you control the links they generate from your content, but for any AI-adjacent placements you do control — llms.txt referenced resources, structured data that AI platforms might pull citation links from, or any owned distribution you push specifically for AI visibility testing — tagging those consistently at least removes ambiguity for the traffic you can influence directly.
Once these rules are defined, I build them into a custom channel group in GA4’s Admin settings under Channel Groups, so the corrected classification sits alongside the default grouping rather than replacing it. This matters for client reporting, because I want to be able to show both numbers side by side: what GA4 says by default, and what the traffic actually looks like once you account for the fragmentation. That comparison alone has been one of the most effective ways I’ve found to get skeptical stakeholders to take AI search seriously, because the gap between the two numbers tells its own story.
What I’d Recommend You Actually Do This Week
If you’re managing GA4 for your own site or for clients, here’s the sequence I’d follow, based on what actually worked across these five accounts.
Start by pulling your raw referral traffic report, not the default channel grouping view, and manually scan for known AI platform domains showing up outside the AI Assistant bucket. You’ll likely find some in Referral and occasionally in Organic Search, depending on how a given platform structures its outbound links at the time you’re checking, since this does shift as platforms update their own systems.
Next, segment your Direct traffic by landing page and look specifically for sessions landing on deep, non-homepage URLs with no referrer. This won’t give you a perfectly clean number, but it will give you a realistic sense of how much AI-influenced traffic might be hiding inside what looks like ordinary direct traffic.
Then build the custom channel group. Don’t overwrite GA4’s default grouping — keep both visible, because the comparison itself is valuable context for anyone reviewing the reports later, including future-you six months from now trying to remember why the numbers shifted.
Finally, revisit this list every month or two rather than setting it once and forgetting it. I’ve had to update the domain list on every single one of these five accounts at least once this year already, because new AI platforms keep launching and existing ones keep changing how their referral traffic is structured. Treating this as a one-time setup task rather than an ongoing maintenance item is the single biggest mistake I see other agencies make when they do attempt to fix this.
How We Handle This at Backlinkgen
This is now a standard part of our GA4 setup process for every client we onboard, not an optional add-on. When we take on AI search visibility work for a client, step one is always fixing the measurement layer first, because there’s no point building an AISO strategy around traffic numbers you can’t trust. We build the custom channel grouping, cross-reference it against server log data where we have access to it for an even cleaner picture, and set up a recurring quarterly review to catch new AI platforms or referral pattern changes before they quietly start skewing the numbers again.
We also build this comparison directly into client reporting dashboards, showing the default GA4 AI Assistant number next to our corrected estimate, so nobody’s making budget decisions off a number that’s silently missing a third or more of the real traffic.
The Bottom Line
GA4 isn’t lying to you exactly, but it’s not telling you the whole truth either, and in a year where AI referral traffic is becoming genuinely significant for a lot of businesses, that gap matters more than it used to. I’ve now fixed this exact problem on five separate client accounts this year, and in every single case, the corrected number changed how the client thought about their AI search strategy. If you haven’t checked your own setup for this yet, I’d genuinely put it near the top of your list this week, because the traffic is very likely already there. You’re just not seeing all of it yet.
