The 2026 agent-traffic landscape — who actually drives requests at your site
Eight AI products move meaningful agent traffic to websites in 2026 — ChatGPT browse, Atlas, Claude with web access, Perplexity, Gemini AI Mode, Operator, Mariner, Copilot. Each reads your site differently. This post is the public-information survey of what each agent reads, what each one buys from, and which kit files they actually probe.
The 2026 agent-traffic landscape — who actually drives requests at your site
If you're trying to decide whether agent-readiness is worth the install effort, the right question isn't "do agents exist" — they obviously do — but which agents, doing what, at what scale. Most site owners hear "AI agents" and picture ChatGPT browsing the web, full stop. The reality is eight major surfaces, each with different reading patterns, different transaction capability, and different shares of the agent-traffic pie.
This post is the public-information survey. What each major agent reads when it visits your site, what it can transact on, how it discovers your kit files, and where each one fits in the overall traffic mix as of mid-2026. The data is necessarily approximate — none of these vendors publish per-site request breakdowns — but the directional picture is clear enough to act on.
The shape of the landscape
Eight agent products move meaningful traffic in 2026:
| Agent | Vendor | Capability | Agent share (estimated) |
|---|---|---|---|
| ChatGPT browse mode | OpenAI | Reads + cites; limited transaction | ~35% |
| ChatGPT Atlas | OpenAI | Reads + transacts + checkout | ~15% (growing fast) |
| Claude with web access | Anthropic | Reads + cites; no native transaction | ~12% |
| Perplexity | Perplexity | Reads + cites + shopping comparator | ~10% |
| Gemini AI Mode | Reads + cites + product result | ~10% | |
| Operator | OpenAI | Reads + transacts (general-purpose) | ~7% |
| Mariner | Reads + transacts (general-purpose) | ~6% | |
| Microsoft Copilot | Microsoft | Reads + cites | ~5% |
Share estimates are directional, based on public usage figures (where vendors disclose them), traffic-tracker observations from analytics platforms that distinguish agent user-agents, and informed guesses from the BTA cohort. The picture in 12 months looks meaningfully different — Atlas and the transactional agents (Operator, Mariner) are gaining share against the read-only agents.
Two structural facts worth pinning down before the per-agent walkthrough:
- Read-only agents (ChatGPT browse, Claude web, Copilot) drive citation traffic — your content gets quoted in the agent's answer, and a share of users click through. The conversion shape is similar to organic search: traffic, attention, eventual purchase via the human flow.
- Transactional agents (Atlas, Operator, Mariner, increasingly Perplexity for shopping) drive direct revenue without a human visit. The user asks for a product, the agent finds it on your site, the agent transacts. Whether the user even sees your storefront depends on the agent's UI and the user's settings.
The transactional share is small in mid-2026 but it's the share that scales toward 50%+ over the next 18-24 months. Sites that are agent-readable but not agent-transactable miss the high-growth share.
ChatGPT browse mode (OpenAI)
What it reads. When a ChatGPT user enables web browsing or asks a query that triggers a search, the agent fetches robots.txt, then the search-result URLs that pass robots rules, then those pages' HTML. It does not deterministically probe /llms.txt or /agents.json paths today — discovery via <link rel="alternate"> tags is the spec-aligned route.
What it buys. Currently, browse-mode ChatGPT doesn't transact on your site. It links the user out, where they complete the purchase via the standard browser flow. The session-replay shape: agent cites your page, user clicks through, conversion attribution looks like organic search.
What you do. The kit's llms.txt and agents.json help when the agent picks which of your URLs to fetch. Without them, the agent samples randomly from search results. With them, the agent picks higher-signal URLs first.
ChatGPT Atlas (OpenAI)
What it reads. Atlas is the transactional successor to browse-mode. It reads the same things browse-mode reads, plus it probes for /agents.json to know which actions on your site it can call without simulating a human in a browser. Atlas's checkout-policy reading also pulls your /agent-instructions.md and treats stated rules ("refund policy at /policies/refund is canonical") as authoritative.
What it buys. Anything its target user asks for that maps to a declared action in your agents.json. The flow: user says "buy me X" to ChatGPT, Atlas reads your action manifest, sees a purchase or add_to_cart action with the right parameters, and executes it. No browser, no human pixel-pushing.
What you do. Typed agents.json actions are the load-bearing detail — Atlas can't call an action with parameters: null. The agents-json-actions-typed audit is what catches this; the per-audit fix reference walks the typing rules.
Claude with web access (Anthropic)
What it reads. Claude's web access (in Claude.ai or via API tool use) fetches URLs deterministically, follows redirects, respects robots.txt, and reads HTML body content plus <head> metadata. It picks up the auto-discovery <link rel="alternate"> tags when they're present, and fetches the kit files when the user's query suggests they'd be useful (questions about "how do I do X on this site").
What it buys. Claude doesn't natively transact. Its tool-use shape lets a separately-built application call your site's actions — for example, a custom Claude-powered shopping assistant could read your agents.json and execute purchases — but Claude itself doesn't ship Atlas-equivalent transactional capability in mid-2026.
What you do. Claude is a heavy citation surface. Pages with clean Schema markup and FAQ pages with FAQPage Schema get cited more often in Claude's answers. Schema density (the schema-org-density audit) is the highest-impact signal here.
Perplexity
What it reads. Perplexity is the agent that most aggressively crawls. Its PerplexityBot user agent fetches your URLs frequently and pulls structured data (Schema, JSON-LD, RSS feeds) for its knowledge layer. For shopping queries, Perplexity has product-comparator features that pull from Product Schema on retailer sites.
What it buys. Perplexity transacts via its shopping integration on a growing list of supported retailers. Direct transaction without leaving the Perplexity UI is in beta as of mid-2026, expanding.
What you do. Product Schema on product pages is the highest-leverage work for Perplexity-driven shopping traffic. The Cluster-2 supporting on schema-org-density on product pages covers what passes the audit and what gets cited.
Gemini AI Mode (Google)
What it reads. Gemini AI Mode pulls heavily from Google's existing index — which means your Schema, your sitemap, your robots.txt, your structured data all matter for Gemini-driven traffic. The AI Mode results panel cites specific URLs and quotes content from Schema-structured pages preferentially.
What it buys. Gemini links out, similar to ChatGPT browse-mode. Transactional Gemini (via Mariner) is separate.
What you do. Google's existing SEO surface plus the Schema work for Lighthouse Agentic Browsing covers Gemini AI Mode. There's no Gemini-specific install path — what works for Google Search works for Gemini.
Operator (OpenAI) and Mariner (Google)
What they read. General-purpose agents that browse the web like a human — read HTML, click buttons, fill forms. They don't deterministically read agents.json today, but they do read agent-instructions.md when they encounter a stuck flow (the runbook tells them how to handle exceptions). Both follow the auto-discovery tags when present.
What they buy. Anything available through standard browser interaction. The shape: user delegates a task, agent opens a virtual browser, agent navigates your site like a human would, agent transacts. Slower than Atlas but more general — works on any site, not just sites with agents.json.
What you do. agent-instructions.md matters disproportionately for Operator and Mariner. A well-written runbook covering "where the canonical refund policy lives, what to do if the user's input fails validation, how to reach a human" smooths their flow and reduces failed transactions. The /docs/agent-instructions-md reference has examples.
Microsoft Copilot
What it reads. Copilot's web browsing reads similar surfaces to ChatGPT — HTML, structured data, sitemap. It cites pages with clean structured data preferentially and links out for conversions.
What it buys. Copilot doesn't transact on third-party sites in mid-2026. It links the user out.
What you do. Schema markup and clean sitemap discovery. Copilot doesn't have a unique signal — what works for the rest of the stack works for Copilot.
What this means for the install priority
The kit files have different value depending on which agents drive your traffic:
- High-Atlas / high-Operator stores (e-commerce, shopping-heavy) →
agents.jsonwith typed actions is load-bearing. The actions-typed audit is the audit that gates whether Atlas can transact on your site. - High-Claude / high-Perplexity / high-Gemini traffic (citation-heavy) →
schema-org-densityandllms.txtare higher-leverage. The agents need to find quotable content and structured data. - High-Copilot / high-browse-mode (read-only) → auto-discovery links + sitemap discovery + Schema. The basics close most of the gap.
In practice, you don't know your agent traffic mix without instrumentation. Most sites discover post-install that their traffic shift comes from a different agent than they expected. The honest play is to ship all three files — the kit's $49 price point exists precisely because the marginal effort of shipping one more file alongside two is zero.
Why agent share matters for your scoring framework choice
Cloudflare's Agent Readiness Score weighs the file-presence audits roughly equally — passing on agents.json and llms.txt and agent-instructions.md is the requirement. Chrome Lighthouse Agentic Browsing weighs schema-org-density and agents-json-actions-typed highest — those are the audits that gate whether Atlas, Claude, and Perplexity can act on your site.
If your traffic mix is heavy on transactional agents, Lighthouse is the score that predicts your revenue better. If your traffic mix is heavy on read-only citation agents, both scores correlate similarly with outcome. The scoring frameworks reference compares the four major scoring frameworks (Lighthouse, Cloudflare, Bridge AI, The Optimisers) side-by-side on what each one weighs.
Related
- The agents.json gap in 2026 standards roundups → — Critical analysis of why mainstream press omits agents.json from the agentic-web narrative.
- agents.json vs WebMCP vs llms.txt — what every site needs in 2026 → — the file-by-file primer
- Lighthouse Agentic Browsing — every audit, every fix → — the per-audit fix reference
- Cloudflare's Agent Readiness Score vs Chrome Lighthouse → — which scoring framework predicts what
- agents-json-actions-typed deep-dive → — the audit that gates Atlas transactions specifically
- schema-org-density on product pages → — the audit Perplexity / Claude / Gemini weigh heavily
- Agent-readiness scoring frameworks reference → — four scoring frameworks compared
- Run the free readiness audit → — five-second check on your site