Your next buyer may never see your homepage. Their AI assistant will. Personal agents, agentic browsers and procurement bots are already reading pricing pages, filling quote forms and opening support tickets on behalf of real people. That traffic is not a bot problem or a marketing opportunity. It is both, and IT Directors and CXOs need one policy that treats it that way.
Inbound AI traffic stopped being one category in 2026, and the tooling finally caught up. Cloudflare said on July 1, 2026 that it was splitting AI traffic into three categories, Search, Agent and Training, so site owners could treat each differently. From September 15, 2026, Cloudflare said new domains would get defaults that block Training and Agent bots on pages with ads while still allowing Search.
Then, on November 4, 2026, Cloudflare announced per-agent analytics and access controls for personal AI agents. Based on secondary coverage of the announcement, it appears to have extended its Agents policy (renamed "AI Crawler and Agent access policies") so owners can set different terms for each supported personal agent, and it added a Personal Assistants category to bot analysis in Security Analytics.
Industry coverage this fall has also reported that agentic AI traffic is growing quickly. The exact figures vary by source and by site, so measure your own. The direction is not in question: a growing share of your visitors are software acting for a human.
Classify inbound automated traffic into four buckets, because each one deserves a different answer.
The hard part is telling bucket three from bucket four. That is an identity problem, not a content problem.
Do not trust the user-agent string; verify cryptographically wherever you can. A user-agent header is a label anyone can type. Scrapers routinely claim to be well-known crawlers or popular assistants.
The better pattern is signed agent identification. Web Bot Auth, built on HTTP message signatures, lets an agent operator sign requests with a published key so your edge can confirm the request really came from that operator. Pair it with what you already have:
Treat unverified traffic that claims to be an agent as anonymous at best and hostile at worst. This is the inbound mirror of the zero trust model for AI agents: identity first, least privilege, continuous verification.
Decide per surface, not per site. A blanket "block AI" rule hides you from buyers' shortlists, and a blanket "allow" invites abuse on your most sensitive flows. Use a matrix like this as a starting point.
| Surface | Search crawlers | Training crawlers | Verified user agents | Unverified / suspicious | Hand to a human when |
|---|---|---|---|---|---|
| Marketing and blog pages | Allow | Business decision (often restrict) | Allow | Rate limit | Rarely needed |
| Product and pricing pages | Allow | Restrict | Allow, rate limited | Challenge or block bulk access | Custom or negotiated pricing |
| Quote and contact forms | No access | No access | Allow with validation | Challenge | Large deals, unusual terms, regulated buyers |
| Support and knowledge base | Allow public articles | Restrict | Allow, ticket creation with confirmation | Rate limit | Complaints, escalations, safety issues |
| Logged-in customer portal | No access | No access | Read-only with user session | Block | Anything beyond viewing |
| Checkout and orders | No access | No access | Allow with step-up confirmation | Block | High value or first-time buyer |
| Account changes (email, bank, MFA, users) | No access | No access | Human only | Block and alert | Always |
The right-hand column matters most. Agents can gather, compare and draft. Changes to payment details, contact info, authentication or user access should always land with a human through a channel the agent does not control.
Set rate limits per verified agent and per surface, and keep them tight where money or credentials are involved. A personal assistant comparing three vendors needs a handful of page loads, not thousands.
Watch for the abuse patterns agent traffic makes easier:
Make your important pages easy for machines to read accurately, because agents increasingly build the vendor shortlist before a human visits. If an assistant cannot find your services, service area or pricing model, it will recommend a competitor it can read.
This connects directly to how buying is changing. Our look at agentic AI in B2B sales covers the seller side; here the point is simply that being legible to agents is now part of being findable.
Log agent traffic as its own segment so you can see what it costs and what it earns. Most analytics tools still blend it into "bots" or drop it entirely.
Edge analytics, such as the Personal Assistants view Cloudflare reportedly added, help, but join them with CRM and support data so you can answer the business question: are agent-sourced leads real? Security vendors are also building agent observability (CrowdStrike, for example, announced plans to acquire Fiddler for AI agent observability, according to November 2026 reports), though most of that work targets agents you run, not agents visiting you.
Security owns the controls, marketing owns visibility, sales owns the handoff, and one executive owns the tradeoffs. Without a named owner, marketing will open everything and security will block everything.
Note the contrast with outbound policy. Our guide to agentic browsing controls covers your employees' agents browsing the web. This policy covers everyone else's agents browsing you. You need both.
If you want help, Infonaligy's AI security team builds and runs these controls, our AI DevOps practice wires verification and logging into your edge and apps, and CRM and sales AI work makes sure agent-sourced leads land cleanly. For the policy itself, start with AI consulting.
AI agents are now visitors, buyers and attackers all at once. Classify them, verify them cryptographically instead of trusting headers, decide access surface by surface, keep account changes and payments with humans, and make your public pages easy for good agents to read. One named owner, one matrix, one 90-day plan.
It is automated traffic from AI systems visiting your website, portal, forms or support flows. It includes search crawlers, training crawlers, personal assistants and agentic browsers acting for a real person, and malicious bots posing as agents.
Not across the board. Blocking everything can keep you off the shortlists that buyers' assistants build. Allow verified search and user agents on public pages, restrict training crawlers as a business decision, and keep logins, checkout and account changes tightly controlled.
No. User-agent headers are easy to spoof. Prefer signed agent identification such as Web Bot Auth with HTTP message signatures, verified-bot status from your CDN or WAF, and published IP ranges as a secondary check.
Agents should not change account email, bank or payment details, MFA settings or user access. Those actions should always require the human, confirmed through a channel the agent does not control.
Name one executive owner. Security runs verification, rate limits and logging, marketing decides what is agent-readable, and sales and support define how agent-originated leads and tickets are handled.
Infonaligy delivers AI security and agent governance remotely nationwide from our home base in Dallas–Fort Worth; see all locations we serve on site.
We will map your surfaces, review your edge settings and agent traffic, and deliver a decision matrix and 90-day plan your security, marketing and sales teams can all sign off on.