AI Search Attribution Model

TL;DR: Most AI traffic does not arrive with a clean referrer, so a workable attribution model blends four signals — direct referrals (Perplexity passes them, ChatGPT often does not), UTM-tagged owned links, brand-search uplift, and citation-correlated traffic — then assigns fractional credit. This guide explains each signal, when it is reliable, and how to combine them into a single AI-attributed value.

Why AI search attribution is different

Traditional search attribution leans on one strong signal: an organic referrer from a known search engine. AI search breaks that pattern in three ways.

• Referrer behavior varies by platform. Perplexity typically passes a perplexity.ai referrer that GA4 captures under traffic acquisition. ChatGPT, Gemini, and most agent integrations frequently strip or never set a referrer, so visits land in (direct) / (none).
• Citation does not equal click. A user can read your snippet inside the AI answer, never visit, and still convert later through a brand search or a direct visit.
• Multi-touch journeys are common. Someone may discover your brand in ChatGPT, validate it in Perplexity, then convert through Google a week later.

These dynamics mean attribution must be modeled, not directly measured. The goal is not perfect accuracy. It is a defensible estimate that improves over time and stays consistent across reporting periods.

Four signals that make up the model

A mature model uses all four and reconciles their estimates into a single AI-attributed value.

Step 1 — Track AI referral traffic

In GA4, create a custom channel group that includes the hostnames AI platforms send traffic from. Common entries to monitor as of 2026:

• chat.openai.com, chatgpt.com (ChatGPT)
• perplexity.ai, www.perplexity.ai (Perplexity)
• bing.com/chat, copilot.microsoft.com (Microsoft Copilot)
• gemini.google.com (Gemini)
• you.com, phind.com, and other niche assistants

Group these into a custom channel — for example AI Search — so they appear alongside Organic Search and Direct in your acquisition reports. Re-validate the list every quarter, since platform domains and referrer behavior change frequently.

Step 2 — UTM-tag every owned link AI is likely to ingest

You cannot tag a citation an AI generates from your prose, but you can tag links you control on AI-facing surfaces:

• Avoid UTMs in canonical sitemap entries — they pollute organic data.
• Use UTMs inside llms.txt, ai.txt, structured-data fields like sameAs, and partner directories. These are AI-facing rather than primary human-facing surfaces.
• Use UTMs on press releases, podcast show notes, and syndicated content that AI may pass through verbatim.

A simple convention works well:

?utm_source=ai&utm_medium=citation&utm_campaign=

Use one consistent utm_source=ai so every AI-influenced visit lands in one bucket regardless of which platform sent it.

Step 3 — Measure citation-to-traffic correlation

Track citations and traffic side by side at a weekly cadence. The pattern below is illustrative — the absolute numbers are not benchmarks, only an example of the shape of the data:

If brand search and AI referrals consistently move in the same direction as citation volume across multiple weeks and across multiple articles, that correlation becomes the basis of a fractional attribution rule.

Step 4 — Estimate brand-search uplift

Brand-search uplift is the part of organic brand traffic that would not have occurred without AI exposure. To estimate it:

1. Establish a baseline brand-search volume in Google Search Console for a stable pre-AI period.
2. Compare it to the current period adjusted for seasonality.
3. Attribute the delta — minus any contribution from paid media, PR, or product launches — to AI exposure.

This is the noisiest signal, but it is often the largest contributor to AI-attributed value in B2B.

Putting it together: AI-attributed value

A defensible composite formula:

AI-Attributed Value =

(AI Referral Sessions × Conversion Rate × Average Order Value)

+ (UTM-Tagged AI Sessions × Conversion Rate × AOV)

+ (Brand Search Uplift × Brand Conversion Rate × AOV × AI Share)

+ (Modeled Direct Uplift × Direct Conversion Rate × AOV × AI Share)

Where AI Share is the fraction of uplift you assign to AI exposure based on your citation-to-traffic correlation. Start conservative — for example 25-40% — and refine the share as your dataset grows.

Reporting cadence

• Weekly: AI referrals, UTM-tagged sessions, citation count.
• Monthly: Composite AI-attributed value, blended conversion rate, top-cited articles.
• Quarterly: Re-validate referrer domain list, refresh AI Share, revisit baselines.

Common pitfalls

• Treating direct-traffic spikes as automatic AI wins without a baseline.
• Using a single signal — usually referral traffic — and missing the majority of AI-influenced sessions.
• Forgetting that platform behavior changes; ChatGPT has shipped and removed referrer behavior more than once.
• Overfitting the model to a single high-traffic article instead of validating across the catalog.

FAQ

Q: Does ChatGPT pass referrer data to my analytics?

ChatGPT's referrer behavior has been inconsistent, and many sessions arrive as direct traffic. Confirm current behavior in your own GA4 acquisition reports rather than relying on a static list.

Q: Does Perplexity pass referrer data?

Perplexity generally passes a perplexity.ai referrer that GA4 records under traffic acquisition. Always validate against your own data, since platforms can change behavior.

Q: Can I use UTMs to track citations directly?

No. You cannot UTM-tag a citation an AI generates from your prose. You can only tag links you control on AI-facing surfaces such as llms.txt, syndicated content, and partner directories.

Q: How accurate can AI search attribution be?

Expect a directional rather than precise model. The objective is consistency across reporting periods and the ability to detect lift, not deterministic last-click accuracy.

Q: What is the simplest first version of an AI attribution model?

A custom GA4 channel for known AI domains plus a citation tracker with weekly snapshots. That alone is enough to detect trends and build a baseline before adding UTM, brand-search, and correlation layers.