Why AI traffic disappears into “direct”

When ChatGPT cites your URL, the user clicks through. The browser sends no referrer in many cases (depending on link-click implementation, browser, user setting). GA4 treats the session as direct/none. The AI citation that drove the click is invisible in the report.

Across our portfolio, before instrumentation, 60–80% of AI-driven traffic lands in direct/none. After instrumentation, that drops to under 15% — the rest gets correctly attributed to the LLM, the prompt cluster, and (with proper analytics integration) the eventual conversion.

If you do not instrument, you cannot prove the engagement is working. We have lost two early engagements because the client could not see the pipeline lift in their default GA4 dashboard.

The UTM strategy

Every AI-cited URL on the site needs a unique UTM-tagged variant that the LLM can pick up. Two patterns:

Pattern A: tag the canonical URL. Add ?utm_source=ai-search&utm_medium=organic&utm_campaign=[cluster] to the canonical URL form most pages. Risk: pollutes shareable URLs.

Pattern B: use server-side rewriting. When a request comes in without UTMs, sniff the referrer (or headers indicating LLM origin where available) and write the UTM into the session. Cleaner; requires server-side capability.

We use Pattern B on Scale and Enterprise engagements (the engineering capacity exists), Pattern A on Growth.

UTM source per LLM: chatgpt, perplexity, claude, gemini, copilot. UTM medium: ai-search. UTM campaign: prompt cluster name (mica-licensing, crypto-vasp-eu, observability-kubernetes).

Server-side dedup against direct

Pattern B catches most of it. The rest — sessions with no referrer, ad-blocker scenarios, mobile clicks where browsers strip everything — get caught by a server-side endpoint that dedups direct sessions against active prompt-cluster URLs.

Implementation: when a /services/[slug] page loads with no UTM and no referrer, fire a server-side check that asks “is this URL currently being cited in any tracked LLM answer?” If yes, mark the session as ai-search-attributed. If no, leave it as direct.

That last 5–10% recovery is what makes the dashboard fully accurate. Without it, the executive sync still shows 15–20% of AI-driven sessions as direct, and the C-level discounts the engagement.

The dashboard that closes the case

Three numbers, one chart, one decision:

  • AI-attributed sessions per week — total visits cited from any LLM
  • AI-attributed conversions per week — leads / signups / purchases from those sessions
  • AOV (or LTV proxy) of the AI cohort — average order value or signup-to-paid conversion rate of the cohort, vs paid baseline

The chart: weekly AI-attributed conversions, six-month rolling. The decision the C-level makes from this chart: scale up, hold, or pause.

A good engagement at month six shows AI-attributed conversions trending up with AOV lift over baseline. A bad engagement shows flat conversion despite rising visibility — that means the prompt cluster is wrong and the targets need re-ranking.

What “good” looks like in numbers

Across our portfolio at month six on a Scale engagement:

  • 8–22% of organic-attributed sessions come from AI sources
  • AI-cohort AOV typically 1.4–2× the paid-search baseline (proven on the D2C luxury case at $487 vs $312)
  • Conversion rate on AI-cohort 0.8–1.3× organic (slightly lower than direct organic because the buyer is earlier in the funnel)

The AOV lift is the lever the C-level cares about. It is also the reason Performance pricing works for SaaS and premium e-commerce — the AI cohort is the higher-LTV cohort, and the bonus is paid against citations that drive that cohort.

What you should do this quarter

If you have an active AEO programme without instrumentation: stop everything, ship the UTM strategy this week. The engagement is already running but the data is invisible.

If you do not have an active AEO programme yet: Starter does not include attribution work. It is a Growth-and-up deliverable. The reason: attribution work without an active citation pipeline produces nothing measurable, and we will not bill for it on Starter.

If you are at Scale or Enterprise on the right tier: attribution architecture is included by default. The dashboard is in your weekly report from week four.