Dark traffic in agentic commerce: how to find the AI shoppers your analytics miss

Dark traffic is website traffic your analytics tool cannot attribute to a clear source. In ecommerce, the problem is getting harder because shoppers now discover products inside ChatGPT, Gemini, Perplexity, Claude, Copilot, and other answer engines before they ever reach your site.

A shopper can ask an assistant for "the best trail running shoe under $150 for wide feet," click a product link, and land on a product detail page. If the app or browser does not pass a referrer or tracking tag, GA4 (Google Analytics 4) may classify that session as Direct. To the business, it looks like someone typed the URL or came back from memory. In reality, an AI shopping experience may have created the demand.

That is the measurement gap behind dark agentic commerce traffic, often shortened to DACT. The goal is not to prove every Direct session came from AI. The goal is to build a defensible estimate, compare it with answer-engine visibility, and improve the product data that helps assistants recommend the right products in the first place.

What dark traffic means#

Dark traffic is traffic that arrives without enough source data for analytics software to classify it accurately. Google Analytics describes (direct) / (none) as traffic that does not have a clear referral source. That bucket includes real direct visitors, but it also includes traffic where the original source was lost.

Common causes include:

• untagged email, SMS, affiliate, influencer, or paid campaigns
• links from PDFs, documents, chat apps, or private communities
• redirects or URL shorteners that drop UTM tracking parameters (the source, medium, and campaign tags added to URLs)
• ad blockers and privacy tools that interfere with tracking
• app-to-browser handoffs where referrer data does not survive
• AI assistant clicks that arrive without a recognizable referring domain

So Direct is not a clean channel. It is partly a channel and partly an unknown-source bucket. Dark traffic is the part of that bucket where the visitor did not really come "directly" in the everyday sense, but the analytics system cannot see the path they took.

Why agentic commerce creates a new dark traffic problem#

In agentic commerce, product discovery moves from keyword search into AI-assisted decision paths. A shopper does not only search, click, and compare. They ask an assistant to narrow choices, explain tradeoffs, compare specs, check fit, and recommend a product.

That creates a new attribution problem because analytics depends on signals that may not make it through that journey. The HTTP Referer header is one of the signals servers use to understand where a request came from. Referrer policies can reduce or omit that information for privacy and security reasons; MDN's Referrer-Policy reference notes that no-referrer omits the header entirely.

AI assistants add more handoffs. A link may open inside a mobile app, a webview, a privacy-controlled browser, or a copied URL. UTM parameters may never exist, or they may be stripped before the visitor reaches the site. When that happens, GA4 cannot reconstruct the original source after the fact.

Retailgentic's Dark Agentic Commerce Traffic article put a useful name on this ecommerce-specific version of the issue: answer engines can send high-intent shoppers to product detail pages while the analytics system credits the visit to Direct or another last-click source.

This is not the same as bot traffic. Bot traffic has its own user-agent and behavior patterns. DACT is about human shoppers whose source data is missing.

Why the missing traffic can matter#

AI referral traffic is still small for many sites, but the quality and growth patterns are hard to ignore.

Adobe reported that AI-driven traffic to retail websites jumped 12x between July 2024 and February 2025. Adobe also found that retail AI referrals had lower bounce rates by February 2025, and that the conversion gap between AI and non-AI traffic narrowed sharply during the same period.

Search Engine Land reported a Visibility Labs analysis of 94 ecommerce brands where ChatGPT referral sessions converted 31% higher than non-branded organic search, at 1.81% versus 1.39%. The same analysis found ChatGPT visits grew 1,079% during 2025, but it also kept the channel in perspective: non-branded organic was still far larger by volume.

The dark portion is harder to size. Loamly's 2026 benchmark claimed that 70.6% of observed AI traffic in its database arrived without referrer headers and was classified as Direct. Treat that kind of benchmark as a signal, not a universal constant. Your site, category, device mix, and customer journey will be different.

The practical takeaway is simpler: visible AI referrals are probably not the whole story. If answer-engine visibility is rising while Direct product-page traffic rises in the same places, there may be a hidden AI-assisted demand signal worth investigating.

The DACT fingerprint: what to look for#

You cannot label a single no-referrer session as AI-driven with certainty. You can look for a pattern that is unlikely to be explained by normal Direct traffic alone.

A useful dark agentic commerce traffic fingerprint often includes:

The important word is "fingerprint." This is evidence for an estimate, not proof for every session. A good DACT model should produce a range and show its assumptions.

How to estimate dark agentic commerce traffic in GA4#

Start by separating traffic you can see from traffic you can only estimate.

1. Build a known AI referrals channel#

Create a custom GA4 channel or exploration for known AI referrers before you estimate the dark portion. Guides from Yotpo and Orbit Media recommend grouping sources such as:

• chatgpt.com and chat.openai.com
• perplexity.ai
• gemini.google.com
• claude.ai
• copilot.microsoft.com
• meta.ai
• other AI or answer-engine domains relevant to your market

Put this group above generic Referral or Organic rules if your GA4 setup processes channel rules in order. Otherwise, visits from AI domains can be swallowed by broader buckets.

2. Create a normal Direct-to-product-page baseline#

Pick a baseline period before the AI visibility change you want to measure. For some brands, that may be before a major AI shopping feature launched. For others, it may be before a product was picked up by answer engines or before AI referral traffic started showing in GA4.

Measure Direct sessions that land on product detail pages or other buying-intent pages:

• product detail pages
• category or collection pages
• comparison pages
• buying-guide pages

Exclude pages that naturally attract real direct traffic, such as the homepage, login pages, account pages, support pages, careers pages, and store-locator pages.

3. Build a likely DACT segment#

A reasonable starting segment is:

• session source / medium = (direct) / (none)
• new users or first sessions
• landing page is a product detail page, product collection, comparison page, or buying guide
• landing page is not the homepage, blog hub, support, account, login, or careers
• session shows meaningful engagement, add-to-cart, checkout, or purchase behavior

For ecommerce, the landing page matters. A new visitor who lands directly on /products/black-wide-fit-trail-runner-size-10 is different from a returning customer who lands on /.

4. Compare the segment with known AI visibility#

Now compare the likely DACT segment with:

• known AI referrals
• non-branded organic traffic
• branded organic traffic
• paid search and paid social campaigns
• email/SMS sends
• influencer or affiliate campaigns
• answer-engine visibility for the same products and queries

If a product begins appearing in ChatGPT, Gemini, Perplexity, or Google AI answers and then Direct product-page sessions rise above baseline, that is a stronger DACT signal than Direct growth by itself.

5. Report a range#

Do not report "AI drove 12,000 hidden sessions" as if the number is exact. Report a range such as:

• conservative estimate: Direct product-page sessions above baseline with high engagement
• likely estimate: conservative estimate plus direct product-page sessions tied to known answer-engine visibility lifts
• aggressive estimate: likely estimate plus branded/product-name search lifts that coincide with AI visibility

This keeps the analysis useful without pretending GA4 can see something it did not receive.

Use server logs and product-page patterns to cross-check GA4#

Server logs can make the estimate more trustworthy, but they do not magically reveal a missing source.

Use logs to check:

• whether the request had a referrer header at all
• whether redirects preserved query strings and UTM parameters
• whether suspicious spikes are bot or crawler traffic
• which product pages received no-referrer sessions
• whether traffic changed by device, browser, geography, or time of day
• whether checkout or add-to-cart events followed quickly after the product-page entry

Logs are especially useful for ruling things out. If a spike came from a bot, a broken redirect, an untagged email send, or a tracking-code issue, it should not be counted as DACT. If the traffic is human, no-referrer, product-level, and concentrated around products where answer-engine visibility rose, the case gets stronger.

Measure answer-engine visibility beside traffic#

Traffic reports show what reached the site. Answer-engine visibility shows what may have created the demand before the click.

Track prompts that match real buying questions, such as:

• "best [product category] for [use case]"
• "[competitor] alternatives for [need]"
• "which [product] should I buy for [constraint]"
• "compare [brand] vs [competitor]"
• "best [product category] under [price]"

For each prompt, monitor:

• whether your brand or products are mentioned
• which products are recommended
• whether the answer links to you, a retailer, a marketplace, or a third-party source
• which competitors appear above you
• which sources the model cites
• whether the cited page is a product detail page, category page, blog post, review, marketplace page, or documentation

This is where AI visibility for ecommerce and ChatGPT product visibility connect to attribution. If your products are appearing more often in answer engines and your Direct product-page traffic rises at the same time, you have a stronger explanation for the movement.

Visibility monitoring also catches influence that never becomes a referral. A shopper may ask an assistant for recommendations, see your product, then search Google for your brand or product name and buy through branded organic. Search Engine Land noted this attribution gap in the Visibility Labs analysis: GA4 referral data can understate ChatGPT's influence when the final click comes from branded search.

Fix what you can, estimate what you cannot#

Some dark traffic can be cleaned up. Some cannot.

Fix the controllable issues first:

1. Add UTMs to owned campaigns, emails, SMS, PDFs, influencer links, affiliate links, and partner links.
2. Check that redirects preserve query strings and do not strip campaign parameters.
3. Create a known AI-referrer channel group in GA4.
4. Keep product URLs canonical and stable so assistants, shoppers, and reports point to the same destination.
5. Annotate major AI visibility changes, product launches, media hits, and campaign sends in your reporting calendar.
6. Audit Direct spikes by landing page before treating them as brand strength.

Then estimate the uncontrollable portion. You cannot force every AI app to pass a referrer. You cannot tag a link an assistant creates from a third-party source you do not control. You cannot see every influence path where a shopper asks an assistant, remembers the product, and comes back later through branded search.

That is why DACT measurement should be a model, not a single report.

Why structured product data is part of the attribution answer#

The measurement problem starts upstream. AI assistants can only recommend, compare, and link to products they can understand.

For an ecommerce brand, that means product data needs to be structured, current, and machine-readable. Useful product data includes:

• canonical product URLs
• titles and descriptions written for real product distinctions
• attributes such as material, size, color, fit, compatibility, and use case
• variant-level price, availability, and identifiers
• shipping, return, and warranty policies
• review and rating signals
• images and product specifications
• category and collection context

Better product data helps assistants choose the right product for the right query. It also creates cleaner measurement patterns. If an assistant links consistently to the canonical product detail page, the traffic is easier to analyze. If product data is fragmented across duplicate URLs, stale feeds, marketplace pages, and mismatched variants, the traffic signal gets noisier.

This is the Catalog angle. Catalog is the product data layer for AI commerce: it helps brands make product data structured, current, and assistant-readable without replacing the existing storefront. That matters for DACT because visibility, recommendations, links, and attribution all depend on product-level data being clear enough for machines to use.

The brands that handle this well will not only ask, "How much AI traffic did we get?" They will ask:

• Which products did assistants understand well enough to recommend?
• Which product pages received unexplained high-intent Direct traffic?
• Which product attributes or sources caused the assistant to choose a competitor?
• Which canonical URLs should answer engines cite?
• Which product-data gaps make the traffic harder to earn or measure?

That is a stronger operating loop than waiting for GA4 to solve the problem on its own.

Common mistakes to avoid#

Treating every Direct session as AI traffic#

Direct traffic has many causes. Some of it is real direct demand. Some of it is untagged campaigns. Some of it is dark social. Some of it may be AI-assisted. A DACT model should isolate likely patterns, not relabel the whole Direct bucket.

Looking only at visible ChatGPT referrals#

ChatGPT may be the largest visible AI referrer for many sites, but Gemini, Perplexity, Claude, Copilot, Google AI Overviews, and retail-specific assistants can all influence discovery. Visible referral reports are the floor, not the ceiling.

Ignoring branded organic influence#

A shopper can discover a product in an answer engine, then search for the brand or product name in Google. That conversion may be credited to branded organic even though AI did much of the qualification.

Forgetting mobile and app behavior#

Known AI referral data can skew desktop because browser sessions preserve source data more often. Mobile app traffic can be harder to classify. Segment device and browser patterns before drawing conclusions.

Waiting for perfect attribution#

Perfect attribution is unlikely. Waiting for it means you underinvest in the product data, monitoring, and measurement habits that make AI commerce legible.

FAQ#

Is dark traffic the same as direct traffic?#

No. Direct traffic is the analytics bucket. Dark traffic is the portion of traffic inside that bucket, or sometimes inside other under-attributed buckets, where the true source was lost. Some Direct traffic is genuinely direct. Some is unattributed traffic from campaigns, apps, documents, privacy tools, or AI assistants.

Is AI referral traffic always hidden?#

No. Browser-based visits from domains such as ChatGPT, Perplexity, Claude, Gemini, or Copilot can show up as referrals when source data is passed. The hidden portion usually comes from app handoffs, privacy controls, copied links, missing UTMs, redirects, or assistant-influenced journeys that end through branded search or direct visits.

Can GA4 measure DACT exactly?#

No. GA4 can measure known AI referrers and it can segment likely dark traffic patterns, but it cannot identify an original source that was never sent. The right approach is to combine GA4 segments, server-log checks, product-page patterns, and answer-engine visibility data into a range.

What should ecommerce teams do first?#

Create a known AI-referral channel, then build a baseline for normal Direct traffic to product and category pages. After that, monitor answer-engine visibility for the same products and compare visibility changes with Direct product-page movement, add-to-cart rate, conversion rate, and branded/product-name search.

The takeaway#

Dark traffic is not new. What is new is the way AI assistants can create high-intent ecommerce demand before analytics ever sees a source.

Treat DACT as an operating signal. Separate known AI referrals. Estimate the dark portion with GA4 and logs. Monitor answer-engine visibility. Improve the structured product data that helps assistants recommend and link to the right products.

The brands that do this will have a clearer view of AI commerce than teams that treat Direct traffic as a mystery bucket and move on.