AI traffic attribution: models and the metrics that prove ROI

By 2026, many buyers first meet your brand inside an AI answer — from ChatGPT, Perplexity, Gemini, Copilot or Google AI Mode — then arrive on your site later through a channel that looks nothing like the conversation that sent them. "How do I see ChatGPT and Perplexity traffic in GA4?" has become one of the most-asked analytics questions, and it exposes the real issue: the models most teams use to credit a conversion were built for clicks, not answers. This post walks the three attribution models — first-click, last-click and assisted — as they apply to AI referrals, then lists the KPIs that still hold up when the referral itself is half-invisible.

What AI traffic attribution is

AI traffic attribution is the practice of crediting AI answers and referrals for the traffic, leads and revenue they actually influence.

The raw material sits upstream of your analytics. An AI visibility tracker samples a fixed prompt set across ChatGPT, Perplexity, Gemini, Copilot and Google AI Mode on a rolling schedule, classifying every brand mention and source citation. Attribution then joins that upstream signal to the downstream sessions your analytics can see — referral visits, branded search, direct traffic — and asks which AI answers set them in motion. The models below are the different ways to assign that credit, and each one distorts AI in its own direction.

1. First-click attribution: credit the answer that started the journey

First-click attribution gives full credit to the first touch a buyer had, before any other channel entered the path.

For AI referrals the appeal is obvious: an AI answer is usually the discovery moment, where someone first learns you exist, and treating it as the origin credits AI for the awareness it creates. The catch is that buyers rarely click straight from an AI answer to a purchase. They read the answer, absorb it, then search your brand name a day later. So the "first click" your analytics records is often that branded search — not the AI conversation that planted it. First-click flatters AI in theory yet under-reports it in practice, because the true first touch happened somewhere your tags never fired.

2. Last-click attribution: credit the referral that converted

Last-click attribution hands all the credit to the final referral before a conversion, ignoring every touch upstream of it.

This is the model many reports still fall back on, and it is the harshest possible lens for AI. AI almost never earns the last click, because that final touch before a conversion is usually branded search or direct traffic. Under last-click, an AI answer that shaped the entire decision registers as doing nothing. Worse, where AI does earn a genuine last click — someone taps a source link inside a Perplexity answer and lands on your page — the referrer is frequently stripped or bucketed as direct or unassigned, so even that honest credit leaks away. Last-click makes AI look invisible, which is why teams relying on it conclude AI "sends no traffic."

3. Assisted (multi-touch) attribution: credit AI as an influence

Assisted attribution credits AI as one touch along a path, rather than as the single cause of a conversion.

This is the model that fits AI referrals most honestly, because AI's real role is influence, not closing. Multi-touch and data-driven models distribute credit across the touches a buyer had, so an AI answer that opened the journey keeps a share of the outcome even when branded search takes the final click. Assisted-conversion reports and path analysis are where AI referrals surface most truthfully, showing the channel doing what it actually does: shaping belief before anyone clicks. The remaining limit is that you cannot see the AI conversation itself, so even an assisted model depends on the downstream referral being tagged well enough to enter the path.

4. Citation share and share of voice: the leading indicators

Citation share and share of voice move before traffic does, which makes them the earliest honest signal of AI influence.

Because every click-level model is leaky, the metrics that hold up soonest are the upstream ones you can measure directly inside the answers. Citation share is how often your pages are cited as sources; share of voice is how often your brand is named relative to competitors. In Asva's live 30-day window ending 18 September 2026, the tracker logged 12,849 citations across the engines it samples — a leading indicator of referral potential that shows up weeks before anything lands in GA4. But citations are not demand: in one tracked category on 16 September 2026, a brand was cited as a source 142 times across 1,796 non-branded answers yet recommended as a vendor zero times. Citations prove your content is read; they do not yet prove it converts.

5. Referral sessions and assisted conversions: the lagging KPIs that hold up

Referral sessions and assisted conversions are the lagging KPIs that survive scrutiny once real traffic finally arrives.

Downstream, the KPIs worth reporting are the ones tied to an observable action: AI-referral sessions captured from the referrers you can still see, branded-search lift as the tell-tale of AI-driven awareness, and assisted conversions inside a multi-touch report. Vanity numbers do not hold up — a raw "AI impressions" count or a single blended visibility score cannot be tied to anything a buyer did. Branded search is usually where the measurable click lands: in one 28-day Google Search Console window, the same brand saw 229 clicks against 50,738 impressions, a 0.45% click-through rate, and 69% of those clicks were branded. The AI answer creates the intent, but branded search is the channel your report actually credits.

The attribution gap you cannot fully close

No model closes the gap completely, because the AI conversation that moved the buyer stays invisible to your analytics.

So the honest posture is layered, not a single dashboard number. Measure the upstream signal — citation share and share of voice — directly, since it is the part you can see cleanly. Tag every AI referral you can, and read downstream conversions through an assisted lens rather than a last-click one. Never claim one model captures AI's whole contribution, and never blend separate measurement windows into one headline figure. If you want the pieces in one place, Asva's AI traffic decoder works the downstream side, citation intelligence tracks the upstream citations, and the brand visibility tracker holds the share-of-voice line. To see where you stand first, the free AI visibility checker runs a first pass, and the best AI visibility tools guide compares how each tracker defines these metrics.

FAQs

What is the best attribution model for AI traffic? Assisted, or any multi-touch model, because AI is usually an influence rather than the closing click. First-click over-credits discovery and last-click erases it entirely, so a distributed model reflects AI's real role most faithfully. Pair it with upstream citation share as a leading indicator.

Why does AI traffic show up as direct or unassigned in GA4? Because the referrer is often stripped or not passed when someone clicks a source link inside an AI answer, and because many buyers read the answer, then return later through branded search. Last-click reporting then buckets that visit as direct or organic, hiding the AI origin.

Can you measure ROI from ChatGPT or Perplexity referrals? Partly, and only by layering signals. Combine AI-referral sessions, branded-search lift and assisted conversions downstream, and lead with citation share upstream. No single number captures it, so report the layers rather than one blended figure.

What KPIs prove AI traffic ROI? Downstream: AI-referral sessions, branded-search lift and assisted conversions in a multi-touch report. Upstream: citation share and share of voice, which move first. Avoid raw impression counts or a single blended visibility score, because neither ties back to an action.

Is citation share a real attribution metric? It is a leading indicator, not a click-level credit. It measures how often your pages are cited as sources across AI answers, and tends to move weeks before referral traffic does — the earliest honest predictor of AI-driven demand.

Key takeaways

• First-click over-credits AI as discovery yet misses it in practice; the recorded first touch is often branded search.
• Last-click is the harshest model for AI, which rarely earns the final click and whose referrers leak to direct or unassigned.
• Assisted, multi-touch attribution fits AI referrals best, since AI's real job is influence, not closing.
• Upstream metrics lead: in Asva's live 30-day window ending 18 September 2026, the tracker logged 12,849 citations — weeks before that shows in analytics.
• No model closes the gap alone — measure citations directly, tag every AI referral, read conversions through an assisted lens, and see how the metrics are defined.