Agentic Commerce Traffic)">Dark Agentic Commerce Traffic (DACT): The Hidden AI Channel in Your GA4

The AI revenue your Google Analytics is silently bucketing as "Direct" runs roughly 7× larger than what your dashboard is showing. We measured it on a real brand. This is how to find your number — and what to do about it.

---

TL;DR

• What: AI surfaces (ChatGPT, Gemini, Perplexity, Rufus, Copilot, Claude) strip referrer headers on outbound clicks. GA4 falls back to its "Direct" bucket. Real AI-driven revenue is hidden inside that bucket.
• How big the gap is: We've measured 4×–8× between visible AI revenue and actual AI revenue. Our anchor case (MyMuse) clocked 7.05× — ₹11.5K/mo visible → ₹81.2K/mo real.
• Why it matters: Without seeing the channel, brands underinvest in it. Some are missing 30–40% of their organic-equivalent revenue base.
• The fix: Server-side AI attribution layer. Three parallel signals: surface fingerprinting, query-param negotiation via the protocol contract, and a dedicated tracking pixel. We call this DACT instrumentation.
• 30-min audit you can run today: Drop in section 7 below.

---

What DACT is, in one paragraph

Dark Agentic Commerce Traffic — DACT — is AI-driven shopper traffic that Google Analytics 4 cannot see, and silently buckets as "Direct." When a shopper asks an AI agent for a recommendation, taps the product card, and lands on your product page, the Referer header is empty by the time GA4 logs it. Same on every major AI surface. GA4 looks at an empty referrer and writes "Direct." Your dashboard says the shopper came from nowhere. They did not come from nowhere. They came from an answer. The discipline of seeing them — and turning seeing into selling — is what this post is about.

DACT is our term, and it captures the most common analytics failure mode we see in agentic commerce. It is not a thought experiment. We have measured it on enough brands to know the gap is consistent and large.

---

The MyMuse number

MyMuse is a sexual wellness DTC brand, India-based, the kind of business that pays attention to its analytics because the unit economics are tight.

When we audited their AI channel, GA4 reported ₹11.5K per month in revenue attributed to AI-referred traffic. That was the number being used internally to size the channel. Based on that number, AI was a rounding error and didn't need investment.

We rebuilt the attribution. Cross-referenced AI surface logs with their server-side pixel. Stitched the empty-referrer sessions back to the agent recommendations they came from. The actual revenue was ₹81.2K per month. A 7.05× gap. Hidden inside "Direct" traffic in their GA4.

Twelve-month surfaced revenue: ₹29.34 lakh. From a channel that, on paper, didn't exist.

That is DACT in numbers. Not a thought experiment. A real brand, a real number, before-and-after.

The MyMuse multiplier is consistent with what we've seen across our customer base: the median DACT-to-visible ratio sits somewhere between 4× and 8×, depending on the category and which AI surfaces dominate the shopper journey. Beauty and personal care skew higher (because shoppers ask AI for recommendations more often). Pure utility purchases skew lower.

---

Why the referrer disappears

The technical reason is straightforward and almost boring.

When an AI surface shows a product card and the shopper taps it, the surface opens the destination URL. The Referer header — the part of an HTTP request that says "I came from here" — is governed by the originating page's referrer policy. Most AI surfaces ship with no-referrer or same-origin policies on their product card outbound clicks, sometimes for privacy reasons, sometimes for indexing reasons, sometimes just because nobody set the policy and the default is empty.

Your server sees a request with no referrer. Your analytics tag fires. GA4 looks for a Source, finds nothing, and falls back to its "Direct" bucket — which is also where shoppers who typed your URL by hand, used a bookmark, or clicked an email link land.

Three completely different behaviors, all wearing the same label. None of them is "the AI agent sent this shopper to you," even though one of them increasingly is.

Our own field testing on a major beauty brand (anonymized — happy to share under NDA) showed only 9 of 112 sessions — about 8% — correctly identified by GA4 as coming from an AI surface. The remaining 92% landed in Direct. Numbers move slightly across categories and surfaces; the structural problem does not.

If 90%+ of your AI traffic is in the wrong bucket, you cannot manage it as a channel.

---

What "managing it as a channel" requires

Three things, in order. Two of them are diagnostic. The third is the loop closer.

1. See it — the methodology

You need an attribution layer that does not rely on the referrer. The signal has to come from somewhere else. We run three parallel signal sources in production and combine them client-by-client:

Signal A — Surface fingerprinting (server-side). Each AI surface has a recognizable fingerprint in the request: a specific user-agent pattern (sometimes), a TLS client hello signature, an IP range cluster, a session timing distribution. We maintain a fingerprint registry per surface, updated weekly as the surfaces evolve. When a request arrives with no referrer but a high-confidence match against a known surface fingerprint, we attribute the session to that surface.

Signal B — Query-parameter negotiation. Several of the protocol contracts (ACP, UCP) allow merchants to negotiate query parameters on outbound product-card clicks. We register your domain with the surfaces that allow it and decorate inbound URLs with a ?tc_src=<surface> parameter. Where this is available it is the strongest signal — direct, deterministic, attributable down to the agent session ID.

Signal C — Dedicated tracking pixel. A lightweight JavaScript pixel on every product page that, on load, sends a structured event to our edge endpoint. The event captures things GA4 doesn't: the document.referrer (sometimes populated even when the HTTP header isn't), navigator.userAgent details, JavaScript-API access patterns characteristic of agent-driven browsers (Comet, ChatGPT desktop app, etc.), and timing. Cross-referenced with Signal A, this catches surfaces where the fingerprint signal alone is ambiguous.

The combined signal is exposed as a DACT panel inside Citation Rank. It's a separate revenue line, broken down by AI surface — ChatGPT, Gemini, Perplexity, Rufus, Copilot, Claude — with the visible-vs-real ratio computed weekly. The "GA4 says X, we say Y" delta is the line every executive reads first.

2. Trust it

The hardest part of the DACT problem is not seeing the dark traffic — it is convincing the rest of the org that the new number is real. The MyMuse team did a parallel test for 30 days: GA4 in one window, our attribution in another. The numbers diverged the same way every week. After week three the team stopped using GA4 for AI-channel sizing entirely.

Two things make a DACT number trustworthy: (1) it ties out to a separately-collected revenue source — your Shopify orders, your Stripe charges, your fulfillment ledger; (2) the AI-surface breakdown is consistent week-over-week within a single brand. If ChatGPT shows up as 40% one week and 5% the next, the attribution is noisy. Ours runs at ±3% week-over-week on the brands we measure, calculated on a rolling 4-week window.

3. Close the loop

Seeing dark traffic is half the work. The other half is converting it. A brand that can prove ₹81.2K/month is flowing through agentic surfaces — but can't get the shopper through a checkout that lives inside ChatGPT, or can't be recommended in the first place — has a measurement layer with no actuator.

That's where the rest of the platform comes in. Visibility (the Citation Rank work) tells you whether the agent surfaced you. DACT tells you whether the shopper came through. Unified checkout for AI agents makes sure the transaction closes inside the agent without the shopper bouncing to a desktop form. Without all three, DACT is a melancholy spreadsheet.

---

DACT by surface — what the breakdown actually looks like

Once the attribution layer is in place, the first thing every executive wants is a per-surface breakdown. Here's the rough distribution we see across our active customer base (DTC, mostly US + India, January–May 2026 aggregate):

Surface	Share of measured AI traffic	DACT-to-visible ratio (median)
ChatGPT	38%	6.2×
Gemini (Google AI mode + Buy for Me)	21%	11.4×
Perplexity	14%	4.8×
Rufus (Amazon)	12%	n/a — internal to Amazon
Copilot	8%	5.6×
Claude	5%	3.9×
Other (Grok, Meta.ai, Pi)	2%	varies

Two things to note. Gemini's ratio is the highest — Google has been the slowest to give outbound URLs any decoration. Rufus is the outlier because Amazon's Rufus traffic doesn't leave the Amazon-owned property — the "click" is internal — so DACT-style attribution doesn't apply the same way. We measure Rufus visibility separately, via Citation Rank.

Your distribution will skew different. The first 30 days after we wire DACT in is mostly figuring out which surface is your dominant one — because most brands guess wrong before they have the data.

---

A 30-minute audit you can run today

Even before you talk to us, you can ballpark your DACT exposure with three steps.

Step 1 — Find your "Direct / None" segment in GA4. Go to Acquisition → Traffic Acquisition → Default Channel Group. Filter to "Direct." The Sessions number is your candidate pool.

Step 2 — Filter the candidate pool by landing page. AI-driven sessions overwhelmingly land on product pages, not your homepage. A normal Direct session lands on / 60–80% of the time (people typing your domain). An AI-driven session lands on a deep product URL 80–90% of the time. If your Direct sessions are landing on PDPs more than on the homepage, the share that's "really" direct is small.

Step 3 — Cross-reference the PDP-landing Direct sessions against your AI surface visibility. If your brand has measurable Citation Rank on ChatGPT, Gemini, or Perplexity for the SKUs those landings hit, the Direct-bucketed sessions are very likely DACT. The ratio of PDP-direct sessions to visible AI-referred sessions is roughly your DACT multiplier.

This is not a precise measurement. It's a ballpark. If your ballpark comes in at 2× or higher, you have a real DACT problem and the precise measurement is worth investing in.

If you'd rather skip the manual audit, drop your URL and we'll send you back a Citation Rank scan within 24 hours. The scan includes a first-pass DACT exposure estimate based on your category and the AI surfaces you currently appear in. Free. No credit card.

---

The thing not to do

Brands sometimes try to "fix" DACT by re-tagging their email or social channels more aggressively, hoping to push the Direct bucket down by reclassifying traffic they can see.

That doesn't fix DACT. It hides it. The Direct bucket gets smaller because you've moved known traffic out of it — but the AI-referred traffic, which you still can't see, is now a higher share of a smaller bucket. The blindspot becomes proportionally bigger and harder to argue about internally.

The only real fix is a separate measurement layer that doesn't depend on referrer signals to identify AI surfaces. Build it, buy it, or live without the number. Hiding the number behind cleaner tagging on other channels is the worst of the three options.

---

What "good" looks like in 90 days

A brand that takes DACT seriously and acts on it tends to follow the same arc.

Day 1–14 — Install. Wire the measurement layer. Watch the gap between GA4-reported AI revenue and measured AI revenue widen as the layer matures. By week two, the multiplier is stable. (For MyMuse it stabilized at 7.05× by day 11.)

Day 15–45 — Diagnose. Use the surface-by-surface breakdown to find the underweight surface. There's almost always one where the visibility score is low and the DACT-to-visible ratio is high — that's an asymmetric opportunity, because if you can be more visible on a surface that already converts well, the slope of the gain is steep.

Day 46–90 — Act. Close the loop on the highest-converting surface first. If ChatGPT is driving the most measured (formerly dark) revenue and your checkout works fine inside the ChatGPT app, great — double down on visibility. If ChatGPT is driving traffic but the conversion drops at the cart step, your bottleneck is checkout execution, not discovery. Re-architect accordingly.

90 days is enough time for a brand to go from "we have no idea what AI is doing for us" to "AI is our second-largest organic channel and we've sized the investment correctly." We've watched this curve four times now.

---

How DACT fits the rest of the picture

DACT is one of the four jobs the agentic commerce platform has to do — see the 7-Layer Map for the full stack. It sits at Layer 06 (discovery + merchant enablement) as the measurement instrument.

A DACT dashboard alone doesn't move revenue. It tells you whether shoppers are arriving. To turn arriving into closing, you also need:

• Visibility — Citation Rank, Share of Voice, Visibility Score. Make sure the agent recommends you in the first place.
• Checkout — protocol integrations (ACP, UCP) so the transaction completes inside the agent surface.
• Attribution back to merchant — so the customer relationship, data, and margin all stay with you.

DACT is the diagnostic layer. The other three are the actuators. You build them together or the dashboard becomes ornamental.

---

CTA

If you'd like to see your DACT number — without an integration, without a contract — start with the free Citation Rank scan. Drop your URL. We send a report in 24 hours. The first-pass DACT exposure estimate is in section 3 of that report.

If you'd like to close the loop after that, book a demo — or talk to the founders directly at [email protected]. Pricing is public on the pricing page: free up to 10 agent transactions a month, 2% above that, custom at volume. You only pay when the agent closes the sale.

DACT is not the whole story. It's the part of the story your GA4 is currently lying to you about. Start there.

— The Tru Commerce team (formerly Asva AI)

---

FAQs

Q: What does DACT stand for? A: Dark Agentic Commerce Traffic. It is AI-driven shopper traffic that legacy analytics (GA4, Adobe Analytics, Mixpanel out-of-the-box) cannot identify, and silently buckets as "Direct" or "Unknown." Measuring it requires an attribution layer that doesn't depend on the HTTP Referer header.

Q: How big is the DACT gap, on average? A: From the brands we've measured, the median ratio of true AI-driven revenue to GA4-visible AI revenue runs between 4× and 8×. MyMuse, our most-cited case study, was 7.05× when we stopped counting. The number is category-dependent and surface-dependent, which is why a per-brand audit beats a general benchmark.

Q: Is DACT just a referrer problem? Can't I fix it with UTM parameters? A: Mostly no. UTM parameters require the originating site (the AI surface) to decorate the outbound URL — and most AI surfaces don't, by default. The fix has to be on the receiving side: server-side fingerprinting, query-param negotiation as part of the protocol contract, or a dedicated DACT tracking pixel. A pure-UTM approach catches a small minority of the traffic.

Q: Will GA4 fix this itself? A: Probably not soon. GA4's "Direct" bucket is a fallback for any session it can't otherwise classify — Google would have to add native AI-surface detection (recognizing ChatGPT, Gemini, Perplexity, etc. as sources), and historically GA4 has been slow to add new channel categories. Until then, the gap is on you to close.

Q: Does Tru Commerce work with brands outside the US and India? A: Yes. Our current case studies are India-heavy (MyMuse, ITC MasterChef, Amazon India) and US-DTC (Vuori, Olipop, Magic Spoon, Mejuri, Caraway in the logo bar). The protocol layer is global; the Citation Rank scans run against the AI surfaces active in your geo. Talk to us if you're outside North America / South Asia — we'll tell you exactly which surfaces we measure for your geo.

Q: How does DACT relate to the broader agentic commerce stack? A: DACT is an instrument inside Layer 06 (discovery + measurement) of the 7-Layer Map. It tells you whether shoppers are coming through. It does not on its own get them recommended more often, or close the checkout — those are separate layers of the stack. A DACT dashboard is a starting point, not an endpoint.