Publishers Take 43.5% of Agentic Browser Traffic — Comet Alone Is 47%

Media and publisher sites took 43.5% of the agentic browser traffic HUMAN Security observed in July 2026, edging past ecommerce at 42%, with travel a distant third at 13.4%. HUMAN published the breakdown on August 6, 2026.

One caveat matters more than the headline, and most coverage buried it: these percentages describe traffic across properties HUMAN protects with its bot-defense platform. It is customer telemetry, not a census of the web. Treat it as a well-instrumented sample of large commercial sites, which is exactly why it is useful for anyone running a content site on Next.js and Vercel.

What HUMAN Security Actually Measured

HUMAN measured agent-driven browser sessions on properties that already run its detection stack, then split them by vertical and by originating agent. That sampling frame shapes every number in the report.

Three limits are worth stating plainly before you plan around this data:

• Customer mix drives vertical share. HUMAN sells heavily into media, retail, and travel. A vertical split from that base tells you where agents go *among instrumented enterprise sites*, not where agents go overall.
• Only identified agents count. Detection depends on an agent being distinguishable. Anything that presents cleanly as ordinary Chrome and behaves like a person is either missed or counted as human.
• Sessions are not requests. Agent-driven browsing bursts. A single user prompt can trigger a dozen page loads, so session-level shares and request-level load are different curves.

A small first-hand note on how hard this ecosystem is to observe: when we fetched the HUMAN report page for this piece, the URL returned a press-and-hold human-verification challenge rather than the article. The report about machine traffic is itself gated against machines. That is the entire industry problem in one round trip.

Which Agentic Browsers Send the Traffic

Perplexity Comet generated 47.13% of agent traffic in the July data, followed by Claude in Chrome at 24%, OpenAI Atlas at 15.5%, and ChatGPT Agent at 6.1%.

Two of those four account for roughly 71% of everything. That concentration is the single most actionable fact in the report, because it means a detection and policy rollout that handles Comet and Claude in Chrome correctly covers most of your exposure on day one.

The four also differ in ways that matter for infrastructure, not just marketing:

The split between device-side and cloud-side agents is what breaks naive blocking. Blocking datacenter ranges stops ChatGPT Agent and misses the 86.6% that Comet, Claude in Chrome, and Atlas contribute, because those three arrive from the same residential IPs as your readers.

Why 76% of Agent Requests Hit Product and Search Routes

76% of agent activity landed on product and search routes. Authentication routes drew 7%, account interactions 6.4%, and payments just 2.6%.

That funnel tells you what agents currently are: readers and comparison engines. They gather, then hand off. The drop from 76% to 2.6% is not shyness, it is the auth wall. Agents can read anything a logged-out visitor can read, and they mostly stop where credentials, MFA, or a card form begins.

For a publisher the translation is direct. Agent load concentrates on article pages, tag and category listings, and internal search. Those are the same routes that carry your slowest queries and your most cache-hostile personalization. Your least defended surface is your most visited one.

Why Agent Traffic Is Invisible in Most Publisher Analytics

Standard analytics miss agent sessions because they are built on assumptions agents break: that a real browser runs the JavaScript, that a referrer survives the hop, and that one session equals one person.

The specific blind spots we see repeatedly:

• Client-side collection. Products that fire on page view need JS execution. Some agents render fully and get counted as humans, others fetch server-rendered HTML and never appear at all.
• Referrer loss. Agent navigations frequently arrive with no referrer, which dumps a growing slice of real demand into the Direct bucket. Anyone still judging AI visibility by referrer strings is undercounting, a problem we covered in the Bing, ChatGPT and Perplexity referral capture strategy.
• Bot filters that disagree with each other. Default filtering drops obvious crawlers while classifying Chromium-based agent browsers as human Chrome. Two dashboards, two truths, no reconciliation.
• Session math. One prompt fanning out to fifteen article fetches reads as a highly engaged visitor with an impossible pages-per-session number and a zero-second dwell time.

If your traffic looks flat while your origin CPU and cache-miss rate climb, you are probably serving agents you cannot see.

How to Detect Agent Traffic in Next.js and Vercel

Detection works best as request classification in middleware, using several weak signals together rather than one user-agent string.

Signals that hold up in production:

• Declared identity first. Some agents send honest tokens in the user agent or client-hint headers. Read them, log them, do not trust them alone.
• Header shape. Missing Accept-Language, absent Sec-Fetch-Site on a navigation, or unusual header ordering separates automation from a person on the same IP.
• Subresource behaviour. People pull images, fonts, and CSS. Extraction-oriented agents often take the HTML and leave, so a document request with no matching asset requests from the same connection is a strong tell.
• Cadence. Ten article routes in four seconds from one client is a fan-out, not a reader.
• Network origin. Datacenter ASNs identify cloud-side agents such as ChatGPT Agent. Device-side agents will not show up this way.
• Cryptographic agent identity. Signed-agent proposals based on HTTP message signatures are the only durable answer, and adoption is early enough that you should log verification results now and enforce later.

The implementation pattern that keeps this maintainable: classify once in Next.js middleware, attach an internal agent-class request header, then let logging, rate limiting, and analytics all read that one value. Two rules go with it. Never fork content by class, because serving agents different bytes than people is cloaking and Google still treats it as such. And never put classification in the critical path of a cached response without accounting for how it affects your cache key.

How Caching and Rate Limits Should Change for Agents

Agent fan-out is a cache problem before it is a security problem, so the first fix is making sure agent requests hit the edge instead of your origin.

What actually moves the needle:

1. Serve stale aggressively. Incremental static regeneration with stale-while-revalidate turns a fifteen-page burst into fifteen edge hits and one background render.
2. Kill cache-key noise. Agents append tracking and query junk that fragments the cache. Normalize or ignore unknown query parameters at the edge.
3. Do not vary on the class header. If the bytes are identical for people and agents, keep one cached object. Varying doubles your cache footprint for no benefit.
4. Rate limit by class, not globally. A token bucket keyed on agent class plus network origin protects the origin without throttling a newsroom on shared office IPs.
5. Answer with 429 and Retry-After. Well-built agents back off and return. Silent drops and infinite challenges teach them the site is broken, and you lose the citation.

What a 2026 Bot Policy Should Say

A current bot policy has to distinguish three kinds of machine traffic, because one robots.txt line cannot govern all of them.

• Training crawlers collect corpora for model training. This is a licensing decision.
• Retrieval crawlers build the index an assistant cites at answer time. Blocking these removes you from AI answers entirely.
• User-driven agents browse because a specific person asked. Comet, Claude in Chrome, and Atlas sit here, and blocking them is functionally blocking a reader who chose an unusual browser.

robots.txt was written for the first two. It has no defensible claim over the third, and pretending otherwise mostly produces broken sessions for paying readers. The practical stance we run: allow user-driven agents at the same rate ceiling as humans, allow retrieval crawlers and keep llms.txt and schema markup current, and make training access an explicit commercial decision rather than an accident of your default config.

How Publishers Can Monetize Agent Traffic

Direct monetization of agent sessions barely exists today, and the 2.6% payment-route share is the proof.

Display advertising is the clearest casualty. An agent-rendered page view has no viewable impression, and if those impressions reach your ad stack they register as invalid traffic, damaging the metric that sets your CPMs. Counting agent sessions as audience is worse than not counting them.

What is left is indirect but real: citation share in AI answers, affiliate value when the agent hands a purchase back to a person, licensing deals with model providers, and per-crawl pricing at the CDN layer. All of those pay on attribution rather than impressions, which puts the zero-click shift from SEO to GEO upstream of revenue. To sanity-check what those answer-layer sessions cost providers to generate, use our LLM API pricing tracker.

Our Analysis: Winning Agentic Share Is a Cost Line, Not a Win

Publishers leading the vertical split is not a victory lap. It mostly measures which sites left the most content readable without an account, and the reward for that openness is origin load, distorted analytics, and ad inventory that cannot be sold.

Three readthroughs we would defend:

The 43.5% moves when auth walls move. This share reflects access, not preference. As commerce sites push more catalogue behind sign-in and publishers extend metering, the vertical mix reshuffles within a few reporting cycles. Do not build a 2027 strategy on a July 2026 split.

Comet at 47.13% is a temporary monopoly. Distribution decides this, not model quality. Atlas at 15.5% and ChatGPT Agent at 6.1% sit on an installed base that has barely started converting, and a default-browser deal or OS-level assistant would rewrite this table faster than any product launch.

Attribution beats blocking, but only if you can measure it. Teams that cannot separate agents from humans in their own logs keep making both mistakes at once: blocking readers and selling ads to machines.

The Agent Traffic Measurement Checklist

Run this before your next reporting cycle:

• [ ] Log an agent-class field on every request at the edge, not in client-side analytics
• [ ] Split dashboards three ways: human, user-driven agent, crawler
• [ ] Compare origin cache-miss rate against reported sessions to find unseen load
• [ ] Track Direct-bucket growth month over month as a proxy for referrer-stripped agent arrivals
• [ ] Verify your ad stack excludes classified agent sessions from impression counts
• [ ] Set per-class rate limits and confirm you return 429 with Retry-After, not a hard block
• [ ] Audit robots.txt and llms.txt for the three-class distinction, then record the reasoning
• [ ] Start logging signed-agent verification results now so enforcement is a config change later

Key Takeaways

• 43.5% of agentic traffic went to media and publishers in July 2026, ahead of ecommerce at 42% and travel at 13.4%.
• Perplexity Comet accounted for 47.13% of agent sessions, with Claude in Chrome at 24%, Atlas at 15.5%, and ChatGPT Agent at 6.1%.
• 76% of agent activity hit product and search routes, against 7% on auth, 6.4% on account interactions, and 2.6% on payments.
• This is HUMAN customer telemetry, not all web traffic — vertical shares reflect the properties HUMAN protects.
• For developers: classify agent traffic in middleware, serve identical content from cache, and rate limit per class instead of blocking datacenter ranges that carry only 6.1% of the load.
• What to watch: the next quarterly split after auth walls tighten, and whether signed-agent verification gets meaningful adoption before 2027 ad-fraud audits force the issue.

Related Reading

• Bing, ChatGPT and Perplexity referral traffic capture strategy
• Google AI Overviews and the zero-click shift from SEO to GEO
• llms.txt, schema and internal links for AI citation
• Google Search Live and AI Mode citation impact
• LLM API pricing tracker

Sources

• HUMAN Security, State of Agentic Traffic July 2026: Publishers Claim Highest Share of Agentic Traffic, August 6, 2026 (the page serves a human-verification challenge to automated fetches)