Earlier this month, analysts noticed what looked like a significant drop in ChatGPT-attributed traffic across several analytics datasets. Teams started asking whether ChatGPT's influence on search behaviour had declined, whether a model change had affected citation patterns, or whether they had lost visibility. The answer, it turns out, was none of the above. The traffic was there. The tracking was not.
According to reporting from Search Engine Journal, the apparent decline ties directly to ChatGPT links opening without tracking tags attached. When links arrive without UTM parameters or referrer data, analytics platforms either misattribute the session or drop it into direct traffic. The visits happened. The attribution chain broke.
Why This Matters Beyond One Platform
The ChatGPT tracking gap is not an isolated quirk. It is a specific example of a structural problem that affects every AI answer engine. AI platforms do not behave like traditional search engines. They do not consistently pass referrer strings. They open links in ways that strip or bypass the standard mechanisms analytics tools rely on. The result is that a meaningful portion of AI-referred traffic has always been under-reported, and most teams have not noticed because they were not looking carefully enough.
When a session arrives without a referrer and without UTM parameters, GA4 calls it direct. That session then gets attributed to whatever channel the attribution model favours - often paid search or organic, depending on prior touchpoints. The AI answer engine that triggered the visit gets no credit. Revenue that originated from an AI citation flows into the wrong bucket, and the programme that earned the citation looks like it contributed nothing.
This is not a minor rounding error. If your AEO programme is being evaluated on attributed revenue and a significant share of AI-referred sessions are being classified as direct or organic, you are almost certainly under-presenting the value of the work.
The Specific Mechanics of the Tracking Failure
The issue reported in the ChatGPT case is that links generated within the platform open without tracking tags. This can happen for several reasons. The AI may construct the link itself rather than passing through a tracked URL. The platform may use an intermediary redirect that drops the referrer. Or the session may begin in a context - such as a standalone app or a browser that enforces strict referrer policies - that prevents referrer data from being passed at all.
Each of these scenarios produces the same outcome in your analytics: the visit arrives, looks like direct traffic, and gets attributed accordingly. Without a deliberate effort to intercept and correctly label these sessions, they disappear into the noise. The sharp drop analysts observed this month was almost certainly traffic that had previously been tracked under looser conditions, now losing its label as platform or browser behaviour changed.
How to Build a More Accurate Picture
The first step is accepting that standard analytics configuration is not sufficient for AI search measurement. You need to build redundancy into the attribution chain. That means UTM-tagging every URL you actively promote in contexts where AI systems might cite it - press releases, structured data, knowledge panels, third-party listings. If your entity data points to a clean homepage URL with no UTM, every citation of that URL will look like direct traffic.
The second step is using server-side data to cross-reference what client-side analytics tells you. If you see a spike in sessions with no referrer arriving on pages that are commonly cited by AI platforms, that pattern is worth investigating. Segment your direct traffic by landing page and look for pages that are disproportionately represented in AI visibility monitoring tools. That gap between what the monitoring shows and what analytics records is your missing attribution.
The third step is connecting identified AI-referred sessions through the CRM. Sessions that arrive via AI citations often exhibit different behaviour to organic or paid sessions - they arrive on specific deep pages, they have different session lengths, and they convert through different paths. If you can identify a cohort of sessions with those characteristics, tag them, and follow them through to revenue, you build a more defensible attribution case than relying solely on referrer strings that the platform may not pass.
What This Means for Reporting AI Visibility Value
The ChatGPT tracking incident is a useful stress test for any AEO programme. If your only evidence that AI citations drive value is the referral traffic line in GA4, you are standing on fragile ground. A single platform change - or a shift in how browsers handle referrer headers - can make that line drop to near zero overnight, even if your visibility and citation volume are completely unchanged.
Programmes that have built multi-layered measurement - platform monitoring for citation frequency, analytics segmentation for likely AI sessions, CRM tagging for AI-assisted leads, and UTM infrastructure for controlled tracking - are far better placed to maintain a credible revenue attribution story when the tracking environment shifts.
The teams that panicked this month when their ChatGPT traffic appeared to collapse were, in many cases, looking at a measurement artefact rather than a real performance change. The lesson is not just to fix the tracking gap. It is to build reporting that does not collapse the moment one data signal becomes unreliable.
The Broader Principle for AI Search Measurement
AI answer engines are not going to standardise referrer behaviour to suit analytics platforms. Perplexity, ChatGPT, and Google's AI surfaces each handle link attribution differently, and that will continue to shift as products evolve. Expecting consistent, clean referral data from these platforms is an unrealistic baseline.
What you can control is the robustness of your own measurement infrastructure. That means not treating AI search attribution as something you read off a single report, but as something you triangulate from multiple signals - visibility monitoring, traffic patterns, CRM data, and controlled UTM tagging where it is feasible. The platforms will keep changing. Your measurement framework needs to hold regardless.
The ChatGPT traffic drop that alarmed analysts this month was, ultimately, a reminder that measuring AI search requires more deliberate architecture than measuring paid or organic. The attribution chain is thinner, more fragile, and more exposed to platform behaviour outside your control. Building redundancy into that chain is not optional if you want to make a credible case for the value AI visibility generates.