Your brand is starting to show up in AI answers โ now what? When someone sees you in ChatGPT, their next move is usually to search your brand name or type your URL directly. In your analytics, those two behaviours get filed under "organic ยท branded" and "direct" โ and the AI's contribution gets credited to other channels.
That's the attribution gap of AI-driven traffic. There is no perfect fix, but there are three layers you can start watching in your own tools today. Each has hard limits, and this guide spells them out โ knowing what you cannot see matters as much as what you can.
Layer 1: AI referrers are starting to appear
Some AI products' web versions pass a referrer when users click links in answers. In GA4's acquisition reports, filter by session source โ or build a segment โ for these domains:
- chatgpt.com / chat.openai.com (ChatGPT web)
- perplexity.ai (Perplexity)
- gemini.google.com (Gemini)
- copilot.microsoft.com (Microsoft Copilot)
- chat.deepseek.com (DeepSeek web)
- kimi.moonshot.cn (Kimi web)
- doubao.com (Doubao web)
- tongyi.aliyun.com (Qwen/Tongyi web)
Practical GA4 setup: create a free-form exploration with "session source" as the dimension, add a regex filter that groups the domains above, and name the group "AI engines". Then glance at that group's monthly session trend.
The limit of this layer (important): Chinese AI engines are used mostly inside native apps, and in-app link clicks usually go through embedded browsers that strip the referrer. So this layer works reasonably well for Western engines (heavier web usage) and badly undercounts Chinese ones. A small "AI engines" number doesn't mean AI sends you no traffic โ the referrer may simply not survive the trip.
Layer 2: a UTM convention for content you publish
What AI cites is often not your homepage but content you've published elsewhere โ industry-media articles, Zhihu answers, reviews. If the links in that content pointing back to your site carry consistent UTM parameters, the path "AI cited that article โ reader followed the link" becomes measurable.
A minimal convention that's good enough:
- utm_source: the platform (zhihu, mediaX, reddit)
- utm_medium: always earned-content (to separate from paid)
- utm_campaign: topic or batch (2026q3-geo-faq)
Three rules that matter more than the parameters:
- Tag at publish time โ it cannot be retrofitted. Bake it into the publishing workflow, whether yours or your execution partner's.
- One lookup table for the whole team โ utm_source spellings must be consistent.
- Only tag links you control. You can't control how third parties link to you organically; this layer measures the return path of content you deliberately published.
The limit of this layer: AI often paraphrases content without driving a click โ the user reads the answer and leaves. UTM captures those who clicked through, not those who saw you and didn't. The latter may well be the majority.
Layer 3: the branded-search echo
The most common action after seeing a brand in an AI answer is searching for it. In Search Console, watch impressions and clicks for your branded queries: when AI visibility rises, branded search usually follows, lagging by days to weeks.
The limit of this layer: branded search responds to all marketing โ ads, PR, events all push it. It is an echo of AI visibility, not attribution. On its own it proves little; laid on the same timeline as visibility monitoring data, it becomes meaningful.
Even combined, this is indirect evidence
Honestly: all three layers together still cannot answer "how many customers did AI bring me" with precision. The recommendation happens inside the AI's chat window, and no tool sees all of it.
The complete evidence chain aligns both ends: your analytics show who arrived and from where; visibility monitoring shows what the AI says and cites. When "content published โ AI starts citing it โ UTM-tagged return traffic appears โ branded search rises" line up on one timeline, every link has its own data behind it โ more credible than any single number, and more honest than any promise.
For the measurement end, see how we score. Reading a vendor's results report? Run it through the five questions first. And to see what AI says about your brand today, the free checkup takes thirty seconds.