Black Friday is over. The third-party panels are publishing their AI-traffic numbers. Most of them are underweighting the actual signal.
Not because the numbers are made up. They’re not. Because the panel methodology was designed to measure something it isn’t really measuring. Adobe Analytics counts agent visits when the User-Agent string identifies the agent. Salesforce Commerce Cloud counts agent referrals when the referrer header contains a known LLM domain. Both methods miss the agent traffic that doesn’t identify itself, which is a meaningful and growing category.
This post pulls out the patterns from BFCM 2025 that look most worth paying attention to. The shape of what happened over the holiday weekend matched some predictions, surprised us in others, and clarified what merchants should be measuring going forward. The patterns matter more than any specific percentage. The percentages will firm up as the category matures and more merchants instrument cleanly. The patterns are already actionable.
The traffic was bigger than the headlines
Headline AI-referred holiday traffic numbers from the major analytics providers came in lower than the actual signal behavioral classification work would suggest. The discrepancy isn’t conflicting truth. It’s a different measurement.
Public panels count identified agent traffic: User-Agent matches plus referrer signals. Behavioral classification adds another layer: viewport heuristics, mouse-event absence, CDP flags, the timing patterns of ChatGPT Operator and Claude Computer Use that look nothing like a human Chrome session even when the User-Agent does. Adding those signals catches the agent traffic that’s deliberately mimicking human browsers, which is a growing category.
The takeaway: the headline AI traffic share your analytics dashboard reports is almost certainly an undercount. The true number is meaningfully higher. It’s still a minority of your traffic. It’s also moving fast enough that “minority” stops being the right framing by next year.
The dominant query mode was comparison
The single most common agent intent during BFCM was comparison-shaped: “best X for Y,” “X vs Y,” “what should I buy if I want both A and B.” Constraint-stacking was a strong second (“under $200, ships before Dec 18, available in size M, men’s size large”).
This matters because comparison-shaped queries are the queries Google has structurally lost to LLM surfaces. The answer to “best wireless earbuds for running under $200” is a recommendation with justification, not a list of links. ChatGPT and Perplexity are better-fit for that question type. Buyers using agents on BFCM weren’t doing what they used to do on Google. They were doing what Google was never quite right for.
If your category gets a lot of comparison traffic, the agent share of that subset is higher than the headline AI percentage.
The trust signal worth watching
A pattern that’s emerging across the agent surface: agent referrals that end in a completed purchase are disproportionately likely to come from sessions where the agent quoted both a positive and a negative aspect of the product in its summary.
The mechanism is straightforward when you think about it. When an agent presents only positives, the buyer downweights the recommendation as marketing. When the agent presents balanced specifics (“highly rated for sound quality, runs small per consistent feedback”), the buyer reads it as a real evaluation and trusts the answer. The conversion happens because the agent surfaced the limitation, not despite it.
The catalogs that support these “trust-shaped” summaries have honest review distributions in their JSON-LD and product descriptions that name real product limitations. The catalogs that don’t get recommended less often by agents that find their marketing copy and go looking elsewhere for the negative data.
This is the catalog honesty thesis with a holiday’s worth of behavior behind it. We expect the effect to strengthen as agents get more sophisticated about cross-referencing claims.
The time-of-day flattening
Human BFCM traffic has a familiar shape. A midnight Thanksgiving spike, a deep midday Friday peak, the Cyber Monday morning surge. Agent traffic doesn’t follow it.
Agent visits across the BFCM window are more uniform, with mild dips during overnight hours and a Sunday-evening lift attributable to planning sessions (“agent, help me figure out what to get for [recipient] before tomorrow”). The Cyber Monday human spike is barely visible in the agent series.
What this means operationally: the surge planning for BFCM (extra inventory at peak, staffed-up customer service for Monday) needs a parallel plan for the always-on agent surface. The agent traffic isn’t going home for dinner. It’s running comparison queries at 3am. The site needs to be honest and fast and right at every hour, not just during human shopping windows.
The cohort that can’t be classified
A meaningful share of suspected agent visits can’t be confidently classified. The User-Agent looks like a human Chrome session. The behavioral signatures are ambiguous: too clean to be a person but not stereotyped enough to fingerprint a specific CUA platform. The working hypothesis is that a non-trivial share of this cohort is ChatGPT in a browsing mode that doesn’t identify itself, which OpenAI has acknowledged exists.
The conclusion worth working from: a non-trivial share of agent traffic is structurally invisible to traffic analytics. Which is another way of saying: the only way to optimize for agentic discovery is to make every page work well for an unidentified visitor who behaves like an agent. The “always-on” content strategy (JSON-LD on every page, honest meta descriptions, structured product data) isn’t a bonus. It’s the only thing that catches the agent visits you can’t see.
What this predicts for 2026
Black Friday 2025 was the rehearsal. The platforms that mattered (ChatGPT Search, Perplexity, Claude Computer Use, Google AI Overviews) all behaved the way the early signals predicted. The merchants who instrumented for agent traffic learned something. The merchants who didn’t, and that’s the majority, have to do the rehearsal next year on a surface that will be measurably larger.
Three things we expect to be true by BFCM 2026:
The agent share of holiday traffic will be meaningfully larger than 2025, with high-consideration categories (electronics, home, apparel-with-fit-complexity) leading. ChatGPT Instant Checkout will be GA and a measurable share of agent visits will end in a purchase that never touches the merchant’s site directly. The merchants who survived BFCM 2025 without an agent strategy will not survive BFCM 2026 without one. Or rather, they will survive in a smaller version of themselves.
The work happens now or it happens late.