SalespeakRead the full report →
The 2026 Agent Traffic Benchmark · Living Edition

Your B2B buyers sent someone else to evaluate you.

The first ongoing, cross-industry study of how AI agents are reading, comparing, and deciding on B2B vendors. Over 3 million agent events logged. Every major agent observed, continuously. Updated quarterly.

3M+ agent events/110 days/9 agents tracked/5 verticals
The headline findingYTD 2026
0%

of B2B content-page traffic now comes from AI agents — not humans.

Source · 3,060,582 agent events · Jan 1 – present · Salespeak monitored network

Trusted by high-growth B2B teams

LIDRAlkiraAnodotBeefreeCADDiConveyorCrystalKnowsCynomiFaros AIIONIXKubiya AISedaiStackGenZuoraLIDRAlkiraAnodotBeefreeCADDiConveyorCrystalKnowsCynomiFaros AIIONIXKubiya AISedaiStackGenZuora
Live · rolling 60-second window

Across our monitored network, right now.

847
Pages fetched
by AI agents in the last 60 seconds
412
Questions asked
of company content, in real time
31
Went unanswered
no page on the site had the answer
6
Agent platforms active
ChatGPT, Claude, Perplexity, Gemini, more

The audience changed. The stack didn’t.

On the B2B web, a larger share of content-page traffic is now non-human than human. Agents are reading the comparison page, the security brief, the pricing calculator, the integration doc. Every page your buyer used to read themselves.

The mechanics are new. Agents fetch specific pages, in milliseconds, without ever loading the JavaScript your analytics stack depends on. Fifty-nine percent of ChatGPT agent traffic in our dataset originates from a Postman-class HTTP client. One percent of Perplexity agent visits ever touch a homepage. They arrive at the deep page, parse it, and leave. Your stack was waiting for a session that will never start.

The audience is not just reading differently. It is deciding differently. The buyer on the other end of the agent receives a synthesis – a compressed verdict on your company – seconds after the question is asked. By the time a human shows up at your sales call, the shortlist is already written.

The consequence
You need a layer that speaks to them directly.

This is what the agent sees.

A real agent transcript, recorded during the benchmark. Names anonymized. Every question below was asked. Every answer is what the site returned.

agent://ChatGPT · evaluating acme-corp.com
>What is [company]'s pricing model for enterprise plans?Q
<Not stated. Pricing page directs to sales.Unresolved
>Does [company] support SOC 2 Type II?Q
<Yes. Referenced in /trust.Answered
>How does [company] compare to [competitor] on deployment time?Q
<Conflicting claims across blog and docs. Confidence low.Degraded
>Is [company] compliant with EU AI Act requirements?Q
<No page answers this. Falling back to [competitor].Lost
0
questions agents asked that no page could answer
Sample · rotating every 3 seconds
?What is your enterprise pricing model?
The consequence
You need a layer that knows what it cannot say.

There is no such thing as AI traffic.

Four separate channels, each with a distinct buyer profile. Treating them as one averages away the signal that matters most.

ChatGPT
The volume
Widest reach, lowest per-visit intent. Reads everything, sends fewer buyers than its volume suggests. Tags its own outbound traffic with UTM parameters – visible in GA4 if you look.
84%of agent crawl volume
Claude
The buyer
The pricing-intent channel. Clicks pricing pages at roughly five times the rate of ChatGPT. Disproportionately present on comparison, “best of,” and vendor evaluation content.
pricing click rate vs. ChatGPT
Perplexity
The citer
The only agent whose UX sends users out to sources at scale. Lowest crawl volume among the four, highest click-through per crawl. Citations are the product.
2.9×click-through premium
Gemini
The phantom
Near-zero visible crawl footprint in server logs. Rides Google's existing index. Drives substantial referral traffic with no corresponding bot visit. Copilot follows the same pattern.
1K+monthly referrals, ~0 crawls
The consequence
You need a layer built for all of them, not one.

The channel distribution is not settling.

Claude-agent traffic grew 26× in five months. The channel distribution in your 2026 plan is already obsolete.

Monthly AI agent events, monitored network · Nov 2025 → Apr 2026
ChatGPT
Claude
Perplexity
+26× Claudesince NovNovDecJanFebMarAprvolume
26×Claude agent growth in 5 months
+91%ChatGPT URL coverage in 2 months
~20dClaude's current doubling time

Two months ago, ChatGPT was the default synonym for “AI traffic.” In our most recent 31-day window, Claude agent traffic is doubling every twenty days. On the current curve, Claude crosses Perplexity in monthly volume by June. The distribution in your 2026 plan was drawn against a snapshot that no longer exists.

The consequence
You need a layer that learns from every conversation.

Three ways companies try to catch up. Two are on the wrong path.

Every category responds to a real shift. Most categorize the problem wrong, then ship tools that make it more legible without changing anything. The agent web is at that point now.

Read path · downstream

Watch what agents say.

  • Monitoring dashboards that track how you appear in ChatGPT, Claude, Perplexity
  • AI SEO tools that observe your content's ranking in agent answers
  • Brand-mention alerts for LLM outputs
  • Better human analytics – intent data, visitor ID

All downstream of the decision. Tells you what happened. Cannot change what happens next.

Write path · upstream

Answer the agent as it decides.

  • Detect the agent arriving at your site in real time
  • Respond with grounded, governed, on-brand answers
  • Capture every question – answered or not – as intent data
  • Compound the knowledge layer with every interaction

This is the Agent Interaction Layer. Salespeak operates here.

The consequence
Observe, or participate. There is no third option.

Four stages. We are in stage one.

The Agent Interaction Layer is not a static product. It is stage one of a transition that ends with agents transacting on behalf of both sides. The infrastructure decision made now determines whether you participate in every stage that follows.

01
Agent-readable
Now · 2026
Content is structured, accessible, and parseable by agents. Pages return useful answers. Most companies are not here yet. The gap is widening.
Where we are
02
Agent-answerable
Near · 2026–2027
The company can respond to any agent question, in real time, with a grounded and governed answer – including questions no page answers today.
Salespeak
03
Agent-negotiable
Mid · 2027–2028
Offers, terms, and configurations become a live interface. The RFP collapses. The quote collapses. The contract becomes the interaction.
Mid-term
04
Agent-transactional
Far · 2028–2030
Agent-to-agent commerce at scale. Buyer agent and seller agent transact. Human sign-off retained where it matters, automated where it doesn't.
End state

The analogy that matters: companies that built real APIs in 2010 became platforms by 2015. Companies that bolted fake APIs onto brochure sites had to rebuild from scratch. The decision made early was the decision that mattered.

Two ways to meet the agent web on your own site.

Open-source · Claude Code skill

Run buyer-eval on your site.

An open-source evaluation skill that acts like an AI buyer researching you. Runs locally in Claude Code, against any public URL. Gives you the transcript: what it asked, what your site answered, what it couldn’t find. Use it yourself, or share it with your team.

Get it on GitHub →
Free · CDN-level analytics

See agents on your site with Agent Analytics.

A free analytics product that captures every AI agent visit to your site at the CDN level. See which agents are reading which pages, in real time. Installs in minutes. No performance overhead. No JavaScript dependency.

Install Agent Analytics →
About this benchmark
Dataset3,060,582 agent events observed since Jan 1
CadenceContinuous capture · Updated quarterly
WindowJan 1, 2026 – present (year-to-date)
AgentsChatGPT, Claude, Perplexity, Gemini, GPTBot, Google-Extended, ClaudeBot, BingPreview, PerplexityBot
CaptureCDN-edge measurement, Salespeak monitored network
VerticalsB2B SaaS, FinTech, MarTech, DevTools, Cybersecurity