AI Assisted Conversion Data: What GA4 Can’t Prove

Most teams reporting on "AI-driven conversions" are quietly doing one of two things: describing a pattern in referral data, or making a causal claim their analytics stack cannot support. The two get presented identically in a slide — a number, a channel label, a trend line — and the distinction disappears. This piece is not another explainer on why attribution is hard in general. It's a working boundary: for each layer of evidence GA4 and standard analytics tooling can produce about AI-referred traffic, what specifically it proves, and where the defensible claim has to stop.

What GA4 referral data actually tells you

Direct answer: As of a rollout that reached most properties by June 2026, Google Analytics added a native "AI Assistant" channel to the Default Channel Group. When GA4 detects a referrer matching a recognized AI chatbot, it tags the session medium as ai-assistant, groups it under the AI Assistant channel, and assigns a reserved (ai-assistant) campaign value automatically — no configuration required (Search Engine Journal).

That's a real, usable signal. It is not a complete one. Three limitations matter for what you can claim:

• Referrer dependence. The channel only catches sessions where the AI surface passed a readable referrer header. Native mobile apps and some in-chat browsers strip that header, and those sessions land in Direct instead — indistinguishable from someone typing your URL from memory. Google's own guidance on building supplemental custom channel groups acknowledges this gap and is why many teams still layer a manual regex-based channel (matching chatgpt|perplexity|claude|gemini|copilot, per Google's documented method) on top of the native one (Analytics Help: Custom channel groups).
• Coverage gaps. The native channel's recognized-referrer list is not published in full, and third-party trackers have documented specific platforms (Perplexity has been reported as inconsistently included) landing in Referral rather than AI Assistant depending on rollout stage.
• Not retroactive. GA4 processes and classifies traffic daily; the new channel does not reclassify historical sessions. Any "AI traffic grew X% year-over-year" claim made before the channel existed is comparing two different measurement definitions, not two periods of the same one.

What this evidence supports: "Sessions with a detectable referral from an AI assistant reached this landing page N times this month, per GA4's channel classification." What it does not support: any total volume claim ("AI is now our #2 channel") without the caveat that a meaningful and currently unmeasured share of AI-influenced traffic arrives as Direct or Organic Search and isn't in that number at all.

Reading landing-page patterns without over-interpreting them

Direct answer: Once you've isolated AI-referred sessions (native channel plus a supplemental custom channel group, ordered above Referral so it evaluates first), the landing-page report tells you where that traffic lands and how it behaves — engagement rate, pages per session, session duration, and downstream key events, all segmented by AI Assistant channel and, ideally, by individual referring platform.

This is genuinely diagnostic. A GA4 session counts as engaged if it lasts 10+ seconds, includes 2+ pageviews, or triggers a key event; everything else is a bounce (Analytics Help documents the current channel and engagement definitions). A landing page with a low engagement rate specifically from AI-referred sessions — but normal engagement from organic search — is a legitimate, actionable finding: it usually means the page doesn't deliver what the AI's summary promised, a "promise gap" between how a chatbot framed your content and what the visitor actually finds when they click through.

Where this goes wrong is treating engagement-rate differences as proof of why AI visitors behave differently. AI-referred sessions are self-selected: someone who typed a specific, already-refined question into a chatbot and clicked a citation is a different population than someone who typed two words into a search box. Higher engagement or conversion rate on AI-referred sessions is consistent with the traffic being more pre-qualified — but it's also consistent with AI assistants simply citing pages deeper in your funnel (comparison pages, pricing pages) more often than they cite your homepage. Both explanations produce the same landing-page numbers. GA4 cannot distinguish them.

What this evidence supports: "AI-referred sessions on this landing page show a 12-point-lower engagement rate than organic search sessions on the same page" — a content/UX diagnostic. What it does not support: "Users who arrive from AI assistants convert better because they trust AI recommendations" — that's a behavioral-causal claim about why, and the landing-page report only shows what.

What assisted-conversion path data can and cannot support

Direct answer: The deeper layer is multi-touch attribution: GA4's default data-driven attribution (DDA) model, which evaluates up to 50 touchpoints across the 90 days before a key event and uses a machine-learning model — trained partly on data from randomized controlled trials of Google ad exposures — to estimate each touchpoint's contribution to the converting probability (Analytics Help: Get started with attribution). This lets you build an "assisted conversions" view where AI Assistant appears as a touchpoint earlier in a path even when a different channel gets last-click credit.

This is useful for internal prioritization — it tells you AI-referred touchpoints appear in paths that eventually convert, at some measured rate, relative to paths that don't include them. It is not evidence of causation, and this is where the industry's own standards bodies are explicit. The IAB's November 2025 Guidelines for Incremental Measurement in Commerce Media draws the line precisely: incrementality is "the causal impact of marketing... compared to what would have occurred in the absence of marketing activity," and the guidelines explicitly distinguish this from attribution and ROAS calculations, which "show outcomes rather than causality" (IAB Guidelines for Incremental Measurement). The Media Rating Council's finalized outcome-based measurement standards likewise require that non-experimental attribution results — which includes every GA4 DDA number — "be treated as estimates," not measured facts, unless validated by a controlled experiment (MRC standards coverage).

Concretely: if a customer's converting path includes an AI Assistant touchpoint, a direct visit, and a branded search click, DDA will split credit among them based on historical patterns of what touchpoint combinations preceded conversions elsewhere in your account. That split is a statistical estimate of correlation strength, not a measurement of what actually caused that specific customer to buy. The customer might have converted anyway after finding you through branded search; the AI citation might have been incidental exposure rather than the deciding factor. Nothing in GA4's assisted-conversion or DDA output can tell you which.

What this evidence supports: "AI Assistant appears as an assisted touchpoint in N% of converting paths, more often for [segment] than for [segment]" — a pattern useful for content and channel-investment prioritization. What it does not support: "AI assistants are worth $X in assisted revenue" or "removing AI visibility would cost us $X in conversions" — both require a counterfactual (what would have happened without that touchpoint) that only a holdout test or geo experiment can produce, per IAB and MRC guidance.

Track vs. don't-claim: a working reference

Why the dollar-figure claim specifically fails

"AI drove $X in revenue" is the claim most likely to end up in a board deck, and it's the one with the least support. Beyond the methodological gap, there's a substantiation standard worth knowing: the FTC's advertising substantiation doctrine requires that an advertiser have a "reasonable basis" for objective, quantified claims at the time the claim is made, and the required level of proof scales with how definitively the claim is stated — an unqualified dollar figure implies a higher standard of proof than a hedged one (FTC Policy Statement on Advertising Substantiation). That standard governs external claims to customers or investors, but it's a useful discipline internally too: if you wouldn't want to defend "$X in AI-driven revenue" to an auditor, don't put it in a stakeholder report as a bare number. State it as what it is — assisted-path revenue, correlational, not causally isolated — or don't state a dollar figure at all.

A reporting framework that survives scrutiny

For each AI-conversion claim you report, attach three things: the evidence layer it came from (referral, landing-page, or path), the specific metric (not a derived dollar estimate unless it's clearly labeled as path-correlated revenue), and the coverage caveat (native channel misses stripped-referrer and unrecognized-platform traffic). Reserve causal or incremental-revenue language for cases where you've actually run a holdout — geo-based or audience-based — comparing outcomes with and without AI visibility, per the IAB's incrementality framework. Everything else stays in the language of pattern and correlation. That's not a weaker report; stakeholders who've seen enough inflated attribution numbers actually trust the hedged version more.

Sources:

• Google Analytics Adds AI Assistant As Default Channel Group — Search Engine Journal
• Custom channel groups — Google Analytics Help
• Get started with attribution — Google Analytics Help
• What's new in Google Analytics — Analytics Help
• IAB Guidelines for Incremental Measurement in Commerce Media
• Media Rating Council releases finalized outcome-based measurement standards — StreamTV Insider
• FTC Policy Statement Regarding Advertising Substantiation — Federal Trade Commission