Reviewed 9 November 2026: Reports of rapid AI agent growth are useful signals, but they do not demonstrate that agent requests now account for most visits to a typical website. Earlier versions of this article mixed global bot-traffic estimates, an individual security company’s agent observations and Google AI Mode click research as though they measured the same thing. They do not. This revision separates the datasets and turns the evidence into a practical decision for publishers.
The question is not whether every website needs a new practice labelled “agentic engine optimisation” or “agent experience optimisation”. The acronym AEO is also used for answer engine optimisation, so a label alone is not a technical specification. The useful questions are which automated clients can access the site, what tasks they attempt and whether those actions help people without creating avoidable security exposure. Ordinary search SEO still matters, and an automated request is not automatically a lead, referral or ranking signal.
Three Numbers, Three Different Populations
The figures cited in earlier discussions measure distinct phenomena. Compare their denominators before making a site-wide decision.
| Reported result | What the original research covers | What it cannot prove |
|---|---|---|
| 53% bot traffic | The 2026 Thales Bad Bot Report estimates that all bots, not just AI agents, represented 53% of measured internet traffic in 2025. | That a majority of your own visits came from AI search or autonomous AI agents. |
| 7,851% agentic growth | HUMAN Security’s 2026 benchmark reports growth in the agentic category across its observed customer network during 2025, from a small base. Agentic activity was 1.7% of its observed AI-driven traffic by December. | A global share of websites, customer sessions or business transactions involving AI agents. |
| 69% OpenAI operator share | HUMAN attributes approximately 69% of its observed AI bot traffic to OpenAI’s identified bots in 2025. | OpenAI’s share of all web traffic, Google search queries or referring clicks. |
| 92% to 94% AI Mode zero-click | A Semrush clickstream study of early 2025 US Google Search sessions found that share of Google AI Mode sessions had no external click. | A 2026 worldwide conversion forecast or proof of how a specific publisher’s traffic changed. |
Thales and HUMAN analyse security telemetry with different coverage and definitions. Semrush measures browsing behaviour in a particular time and market. These are useful observations, not parts of a single causal funnel. The proper next step is to examine a site’s own evidence.
Start by Identifying the Request, Not the Marketing Label
AI-related requests can be training crawls, search discovery crawls, user-triggered fetches, content scrapers or agents completing a workflow. The same platform may operate several clients for different purposes.
For example, OpenAI’s crawler documentation distinguishes OAI-SearchBot for ChatGPT search discovery, GPTBot for training-related crawling and ChatGPT-User for certain user-initiated browsing actions. These identities should not be grouped as one search-referral source. Google similarly distinguishes Googlebot for Search from Google-Extended, which governs certain other Google AI model uses.
Check robots.txt rules, server request logs, response codes and supported verified bot identities. A user-agent string alone can be spoofed; use a platform’s IP and identity guidance where available. A security decision to restrict an automated client should consider abuse prevention as well as discoverability. Do not approve every agent merely to increase request counts.
Where AI search discovery matters, compare the evidence behind AEO claims with platform documentation. The site should still satisfy core crawlability, helpful content and search policies; none of those fundamentals requires a new site architecture for every bot.
Google AI Search Does Not Require a Special Agent File
Google’s 2026 optimisation guidance states that AI Overviews and AI Mode use ordinary search fundamentals. It does not require llms.txt, skill.md, tiny content chunks or special schema markup for inclusion. Structured data may still help supported traditional rich results when it accurately matches visible page content, but it is not an AI ranking shortcut.
The Model Context Protocol (MCP) can be useful when an application connects to an authorised tool or service. That is a different technical question from whether a public webpage is eligible to be cited in Google Search. Introducing an API or MCP server without a real user task, security model and maintenance owner does not make a site automatically more discoverable.
For publishing teams, focus on accessible text, meaningful internal links, reliable navigation, clear business facts, descriptive media and well-maintained schema markup where it is supported. If an important page requires JavaScript to display its primary content, test the rendered output for the relevant crawler before treating it as a ranking defect.
A Practical Agent-Readiness Audit
Use the following checkpoints to decide whether a change is needed.
- Identify a real task. Is the request discovering a public answer, comparing a product, reading a document, completing a form or accessing an account? Choose the correct access policy for that task.
- Inspect the request evidence. Review a sample of paths, response statuses, frequency and bot identity information. Mark which requests appear verified, unverified or suspicious. Do not treat an arbitrary user-agent label as conclusive.
- Check content access. Make sure useful public content is readable and navigation works in normal browser conditions. For private services, verify authentication, authorisation, rate limits, input validation and user consent before exposing programmatic actions.
- Measure useful outcomes. Compare verified bot requests with actual human sessions, search clicks, qualified leads or completed user actions. Use separate dashboards for these different populations.
- Change one control at a time. Document the baseline, apply a scoped robots or technical fix only with appropriate approval and check whether its measured effect matches the goal.
Publishers with mostly static informational pages may need no agent-specific implementation beyond well-maintained SEO and bot security. Commerce or service applications with agent-mediated transactions may need a separate permission and fraud review. A security requirement should not be recast as an SEO growth guarantee.
What SEO Teams Should Watch Next
Product vendors are changing crawler policies, search interfaces and supported tool integrations quickly. Review official documentation when deciding which bots to allow, and identify whether a new feature concerns indexing, model training, user-triggered requests or authenticated actions.
MOCOBIN’s dated AI SEO vendor announcement offers a separate account of industry positioning, not platform rules or proof of results. Keep any claims from vendors or surveys labelled with their sample and dates. A higher number of AI requests should be a prompt to inspect quality and security, not an automatic reason to publish more generic AEO content.
- Thales, 2026 Bad Bot Report (2025 traffic scope)
- HUMAN Security, 2026 AI Traffic and Cyberthreat Benchmark
- Semrush, July 2025 AI Mode adoption and clickstream study
- OpenAI, crawler and user-agent purposes
- Google Search Central, AI features and your website
- Google Search Central, generative AI search optimisation guide









