Websites are acquiring a second audience: software that researches, compares and acts for customers. As that audience expands, businesses face a design question beyond mobile responsiveness or search rankings: can an AI agent actually use our website?
The traffic shift is substantial. Cloudflare reported in September that daily requests from AI agents across its network had grown more than 1,700% over the preceding year. HUMAN Security separately measured 7,851% year-over-year growth in traffic from agents and agentic browsers during 2025. Those figures describe different measurement populations, not interchangeable estimates of the entire internet.
There is also an important distinction between machines reading and doing. Training crawlers collect content; task-oriented agents navigate toward outcomes. In HUMAN’s dataset, 77% of agentic activity occurred on product and search pages, while 2.3% reached checkout pages—evidence of commercial activity, but not wholesale replacement of human shopping.
From Attractive Pages to Usable Capabilities
The emerging response is agentic-first design: making information and business capabilities discoverable and usable by software, rather than requiring every visitor to interpret a visual interface. Ora and enso illustrate complementary approaches to this transition.
Ora provides agent-readiness assessment, scoring how agents find, read, and use websites. Its homepage reports 136,388 sites scanned, an average score of 43 out of 100, and just 6% earning a B or above. These are Ora’s proprietary assessments, not an independent industry benchmark. Its “Journey” feature tests an agent against a specified task and records where it succeeds or stalls.
Enso, meanwhile, markets AI-operated growth services spanning search, outbound sales, communities, newsletters and social media. Its own website also provides an AI-facing site overview, extended Markdown documentation and a canonical sitemap alongside its human-facing sales experience. That is a concrete example of adding machine-readable entry points without abandoning conventional web design. Those resources alone, however, do not establish that agents can execute transactions.
Ora’s research offers a caution against treating special files as magic. Agents in its observed runs used documentation and homepages far more frequently than dedicated AI files. Its practical lesson: link structured resources from familiar pages, preserve navigation without JavaScript, and keep reachable information accurate. Agent-friendly infrastructure is ineffective when agents cannot discover it.
MCP Versus API: Complement, Not Replacement
An API exposes a system’s data or operations to other software. Model Context Protocol, or MCP, standardizes how AI applications discover and invoke tools and obtain contextual resources. An MCP server can expose capabilities that call existing APIs; the two therefore occupy complementary roles rather than representing competing replacements.
For a retailer, an underlying API might supply inventory or update a cart. An MCP-facing tool could present that operation with a discoverable description and structured inputs, reducing bespoke integration work for compatible AI applications. MCP standardizes the interaction layer; it does not itself guarantee reliable business outcomes or safe authorization.
For digital teams, the implication is straightforward: preserve the human website, but build a clearly signposted machine entrance—and test whether visitors using either can finish the job.
Top image courtesy of Pexels.com.
![Dallas Market Center ArcLight Summit banner: Call for Speakers, June 22–23, 2027 — Apply Now (logo with colorful arc.”] ,](https://i0.wp.com/www.lightnowblog.com/wp-content/uploads/2026/10/2610_LightNow_Archlight_Banner-Ad_900x400-e1792264459229.png?fit=900%2C400&ssl=1 "2610_LightNow_Archlight_Banner Ad_900x400 | LightNOW")
You must be logged in to post a comment.