How to Attribute an AEO Deal When the Buyer Researches on Mobile and Converts on Desktop

The pattern is now the default B2B journey, not an edge case. A buyer asks ChatGPT on their phone for the best tool in your category, reads an answer that names you, taps through, and glances at your site on a small screen. Nothing happens that day. Two days later, at a desk, they type your name into a browser, land as Direct traffic, and start a trial. Your analytics records a desktop Direct signup with no source. The AI answer that actually created the deal is invisible, and the mobile session that proved it is gone. Here is why the split defeats standard attribution and how to rebuild the chain.

The problem is not that AEO fails to drive pipeline. It is that the strongest AEO journey, discovery inside a mobile AI app followed by a considered desktop purchase, is exactly the journey your measurement stack is worst at seeing. If you have not first confirmed the pipeline is genuinely flat rather than hidden, start with the leak audit in why your AI citations are not converting into pipeline and how to find the leak before you conclude anything from a Direct-heavy dashboard.

Why the mobile touch disappears

Mobile AI apps open your link inside an embedded browser component, and those components drop the referrer header during the handoff from app to web. On iOS this is WKWebView behavior; on Android it is WebView or Chrome Custom Tabs. This is how the operating system works, not a deliberate choice by any single platform, which is why you cannot fix it by asking anyone to change a setting.

The result is measurable and lopsided by engine. On mobile app traffic, Seresa's teardown of zero-referrer AI sessions reports Perplexity passing a referrer roughly 70 percent of the time, Gemini around 9 percent, and ChatGPT, Claude, and Copilot arriving stripped except where a UTM parameter is baked into the citation link itself. Across all AI traffic, a large share lands referrer-less and falls into Direct. GA4's AI Assistant channel, which launched in mid-2026, only classifies a session when a recognized AI referrer header actually reaches it, so mobile app sessions bypass the channel entirely.

So the very first touch, the one that did the persuading, is recorded as an anonymous Direct visit from a phone. That is loss number one.

Why the desktop conversion cannot find its way back

Even if you caught the mobile touch, the second break is the device change. The person who read the answer on their phone converts on a laptop. There is no shared cookie between the two browsers, no shared GA4 client ID, and often days between them. Cross-device journeys are the norm now, not the exception, and B2B buying cycles stretch the gap wider: research on one device, a return visit on another, a purchase decision that can sit ninety days out from the first touch.

Cookie-based, single-device attribution has no mechanism to connect these. It sees two unrelated strangers: an anonymous mobile bounce and a desktop Direct conversion. That is loss number two. Stack the two losses and the highest-intent journey in your funnel becomes the one your dashboard is structurally blind to, which is the same undercounting problem described in why your AEO attribution undercounts AI-sourced revenue and how to correct it.

The three-layer recovery

You cannot patch this inside GA4's default model. You rebuild the chain with three independent layers, each of which survives a break the others cannot.

Run all three. The self-reported field is the backstop that works even when both technical layers fail, and the identity stitch is what turns two capture points into one journey.

Layer one: capture the mobile touch server-side

Read the request at the server before GA4's client-side classification runs. Three signals survive there that do not survive in the browser. First, user-agent fingerprints: some mobile AI clients carry distinctive strings, for example Gemini's iOS client identifying itself in the agent. Second, landing-page pattern matching: cross-reference incoming Direct sessions against the specific URLs you know AI engines cite, because a phone hitting a deep answer-page directly is a strong AEO tell. Third, UTM extraction: ChatGPT's utm_source=chatgpt.com survives the mobile handoff when it is embedded in the link itself, so it shows up in the landing URL even when the referrer header is gone. Server-side capture routinely recovers well above half of the mobile AI volume that GA4 files as Direct. The full setup sits alongside how to track AI referral traffic in GA4 and tie it to pipeline.

Layer two: ask at the moment of conversion

The layer that survives everything is the buyer's own testimony. Add a "How did you hear about us?" field to the trial signup or demo request, and make it an open text box, not a dropdown. Growth Method's work on self-reported attribution makes the case plainly: a dropdown tests whether the respondent knows your channel taxonomy, while someone influenced by ChatGPT simply types "chatgpt told me about you." In deployments through 2026, AI assistants show up at eight to twenty-two percent of self-reported source across clients, a share you would never see in referrer data alone. This field is device-agnostic by definition. It does not care that discovery was on a phone and conversion on a laptop, because it asks the one entity that was present for both. The playbook for building it is in how to capture AI-sourced deals with a self-reported attribution survey.

Layer three: stitch the two devices with a login

Deterministic matching beats guesswork. Any product with an account system, which is nearly every SaaS, can tie a mobile session and a desktop session to one person the moment they log in on both, or the moment the same email appears. When a buyer starts a trial on desktop and the email matches a newsletter opt-in or gated-download from the mobile visit two days earlier, you have joined the journey with certainty, not probability. Feed that stitched identity back into your CRM so the AI first-touch travels with the record all the way to closed-won. The reason this matters for finance is covered in how to prove AEO pipeline when the buyer leaves no referrer.

Reading the recovered data honestly

Once the three layers run, resist the urge to treat the number as precise. Self-reported data has recall bias, server-side capture has false positives, and the identity stitch only fires when the buyer authenticates. What you get is a triangulated estimate that is directionally right and defensible, which is far better than a Direct bucket that is confidently wrong. Lead with the self-reported share as your headline AI number, use server-side capture to size the traffic, and use the identity stitch to prove specific deals. When the three roughly agree, you can state an AI-sourced pipeline figure to finance without overclaiming.

This is also why publishing pages that AI engines actually cite matters more than volume. If the mobile answer never names you, none of this measurement has anything to record. A continuously updated, citable AI Feed Engine of answer-first pages is the supply side that puts you into the mobile answer in the first place, and a basic llms.txt file you can generate for free removes the most preventable reason a crawler never reaches those pages. The FastTrackr AI case study is a worked example of citations on buying-intent prompts turning into signups the team could actually count.

What to do this week

Ship the self-reported field first, because it is a one-day change that starts collecting the most durable signal immediately. Add server-side landing-page and UTM capture next. Wire the identity stitch into your CRM last, since it depends on the other two feeding it. Within a month you will have a device-agnostic view of AI-sourced pipeline that no longer loses the buyer at the moment they switch from the phone that persuaded them to the desktop that pays you. Teams that would rather score the mobile answer continuously than reconstruct it after the fact can see what that measurement layer covers in what OnlyAEO's plans include.