A New Era for Websites: When Your First Visitor Isn't Human

Your website has two kinds of customers

Picture Tuan, who runs a furniture store in Hanoi. Every morning he opens his dashboard and sees a few hundred visits to his website. But in 2026, there is a group of visitors he has never actually seen: they don't click ads, don't browse product photos, don't read promotional banners. They are AI agents — digital assistants hired by Tuan's real customers to find answers and complete tasks on their behalf.

This is the simplest way to understand what's happening to the web: websites have always been designed to be "hired" by humans. Now they need to be hired by agents — the human's representative. And these two kinds of customers need very different things.

People want beautiful images, warm copy, clear buttons. Agents want machine-readable content: a map of the site, clean data structures, callable APIs, error messages precise enough to self-correct. If an agent lands on Tuan's website and finds nothing but a heavy JavaScript page, it leaves — and sends his real customer to a competitor that's easier to talk to.

This article sells nothing. It simply recounts what's happening on the web, backed by source-verified data, and helps small businesses ask themselves: what should we be preparing?

Three eras of the web, in three minutes

To understand the agentic era, take a quick look at the two before it:

Web 1.0 (1990s–2000s): websites were electronic brochures. Static content, identical for everyone. The only "visitor" being optimized was the human eye.

Web 2.0 (2005 until recently): the web became interactive — social networks, e-commerce, applications. A second class of visitor appeared, though few named it: search engine crawlers. An entire industry — SEO — grew up serving exactly one bot: Googlebot. Businesses learned to talk to Google through sitemaps, schema markup, and page speed.

The agentic web (now): this third class of visitor doesn't just read content the way Googlebot does. It acts: compares prices, books appointments, fills forms, calls APIs, even pays. And unlike Googlebot, which only needed indexable text, an agent needs a website that exposes real capabilities: knowing what the site can do, which interface to call, how errors get handled.

Each era produced its own measurement industry. Web 1.0 had analytics logs. Web 2.0 had SEO audits. The agentic era is producing "agent-readiness scoring" — rating how ready a website is for agents. That's the signal that this has stopped being a theoretical trend: there is now a market measuring it.

Why websites must learn to "talk" to agents

A fair question: can't agents just read HTML like people do?

Technically yes, but it's expensive and error-prone. A modern page contains hundreds of kilobytes of markup, scripts, and tracking code. To extract "how much is that four-seat oak dining table," the agent has to guess from all that noise. Every wrong guess becomes a bad experience for the real customer behind it.

So the web is developing a parallel surface: same website, but exposing data in ways machines understand immediately. Here are the main pieces:

llms.txt — a markdown file at the site root summarizing what the site is and which links matter. Spec v2 (2026) added standard discoverability: pages declare their markdown version via rel="alternate" type="text/markdown". Think of it as "a sitemap for LLMs." One notable number: across a fixed panel of 219 well-known tech hosts, 53.7% now serve llms.txt (source: https://llmtxt.info/blog/state-of-llms-txt-2026/, Aug 2026). Across the Tranco top-10K of the whole web, the figure is 4.2% — up from roughly 0.3% in mid-2024 (source: https://crawlmind.ai/research/state-of-llms-txt-2026). Crawler logs confirm OAI-SearchBot, PerplexityBot, and ClaudeBot fetch llms.txt regularly. Honest caveat: Google says it does not use llms.txt, and no major provider has confirmed using it at inference time — this is still a growing convention, not a certified standard.

MCP (Model Context Protocol) — the protocol letting agents call a website's capabilities directly, as tools. The striking part: the July 28, 2026 spec moved the core to stateless, meaning MCP servers are now ordinary HTTP workloads that scale behind load balancers like any other web service. TypeScript + Python SDKs have passed 1 billion cumulative downloads (source: https://blog.modelcontextprotocol.io/posts/2026-07-28/). Anthropic donated MCP to the Linux Foundation in late 2025, with OpenAI and Block as co-founders — so no single vendor owns it.

A2A (Agent-to-Agent) — one layer up: your agent talks to my agent, across organizations and frameworks. A2A reached stable v1.0 in March 2026 with Signed Agent Cards (cryptographic verification of agent identity), more than 150 participating organizations, and production use in supply chain, finance, and insurance (sources: https://a2a-protocol.org/latest/announcing-1.0/ and https://www.linuxfoundation.org/press/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year).

WebMCP — a W3C draft developed jointly by the Chrome and Edge teams, letting websites declare capabilities as structured tools via a browser API; early preview in Chrome during 2026 (source: https://gravity.fast/blog/ai-agent-interoperability-standards-2026/).

The common thread: these turn websites from "documents to read" into "partners to collaborate with." Agents stop scraping — they consult the map (llms.txt), call tools (MCP), verify identity (Signed Agent Cards), and coordinate with other agents when needed (A2A).

Measuring readiness: the 80/20/+5 model

When something new appears, the first question is always: how do we measure it?

One interesting example of how this field is forming is is-agentic.com — a free service that scores how "agent-ready" a website is. Submit a public URL, the system scans it and grades results in three tiers (per the published methodology at https://is-agentic.com/methodology):

• Essential checks — an 80-point pool: foundations every agent needs. Server-side rendered content, correct HTTP behavior, clear document structure, errors detailed enough for the agent to recover on its own.
• Recommended checks — a 20-point pool: activated only when the scan detects the site actually has APIs, OAuth flows, MCP servers, or commerce surfaces. Sites without them have those checks excluded from scoring — not penalized for lacking what they never had. That design detail is worth learning from: fair evaluation creates incentives for genuine improvement.
• Emerging signals — up to 5 bonus points: formats too new to require. Missing them never lowers the score.

Another detail that captures the mindset of this era: the very site doing the scoring operates as a model agent-ready website — publishing llms.txt, an OpenAPI spec, an MCP server, and using content negotiation: the same report URL serves HTML to browsers, Markdown when an agent requests text/markdown, JSON to integrations. One URL, multiple representations depending on client type. It's the future website in miniature.

One honest clarification: this score measures "agent legibility" — how easily machines can read, call, and debug the site — not business quality or security. Those remain your problems.

The economic pressure: why this isn't just a fun experiment

If it were purely elegant engineering, small businesses could ignore it. But the economics underneath are taking shape fast.

Questions are leaving Google. AI Overviews trigger for roughly 48% of tracked Google queries (+58% year over year, BrightEdge 2026); zero-click searches rose from 56% to 69% within a year of AI Overviews launching (Similarweb; source: https://seoscaleup.com/blog/geo-aeo-statistics-2026/). For most first-touch queries, users now get their answer right there — without clicking anyone.

AI traffic is tiny, but converts differently. AI referrals currently make up about 1% of total visits, but Seer Interactive measured ChatGPT-referred conversions at 15.9% versus 1.76% for Google organic — nearly 9x (source: https://ranqo.ai/blog/generative-engine-optimization-statistics). Adobe (March 2026) likewise found AI traffic converts about 42% better than organic. The reason is simple: visitors arriving via agents come pre-qualified — the agent has already advised them close to a decision.

And who do agents need for good advice? Machine-readable sources. Princeton's GEO research (KDD 2024) showed that adding statistics, quotations, and clear structure boosts visibility in AI answers by 30–40%. The causal chain keeps linking: answer engines absorb queries → organic clicks erode → brands must optimize for citation rather than ranking → citations require machine-readable content → websites must expose agent surfaces → a market for measuring agent-readiness emerges.

There's also one big-picture number worth remembering: Adobe Analytics measured generative-AI traffic into US commercial sites at +4,700% YoY through July 2025, then another +393% in Q1 2026; HUMAN Security recorded AI agent / agentic-browser traffic growth of 7,851% YoY (compiled at https://asiangrade.com/agent-readiness). Absolute numbers are still small relative to the whole web — but the velocity is impossible to ignore.

A real-world analogy: this feels like 2005, when "should a coffee shop make a Facebook page?" sounded silly. Early movers set up cheaply; by the time everyone had one, the cost of competing had multiplied.

What SMBs should actually prepare

No rewrite required, no large engineering team needed. A pragmatic list, ordered by return per dollar spent:

1. Check whether core content renders without JavaScript. Open your website and view source. If essentials — products, prices, address, opening hours — aren't in the initial HTML, agents (and many crawlers besides) see nothing. This is an "Essential" item in virtually every scoring model, because everything else sits on top of it.

2. Write llms.txt. A simple markdown file: an H1 with your business name, a blockquote summarizing what you do, then a list of important links with one-line descriptions. For a small business this is an afternoon of work — and the cheapest way to give agents a map drawn to your intentions instead of their guesses. Tooling is already mainstream: Yoast SEO, AIOSEO, Wix, and GitBook all offer auto-generation (source: https://llmstxt.org/).

3. Serve a markdown version of each key URL. llms.txt spec v2 lets you declare rel="alternate" type="text/markdown" — your product page can serve both HTML and Markdown at the same address. Start with guides, pricing, and FAQ pages, since those are what agents get asked most.

4. Clean up structured data (JSON-LD schema). Products, prices, reviews, addresses — mark them up with standard schemas so both search engines and AI engines read the same source of truth. GeoReady's benchmark of 750+ sites found an average score of just 54.3/100, with 58% having llms.txt (source: https://geoready.dev/) — the overall bar is still low, so the opportunity is wide open.

5. Think about the error experience. When an agent triggers a function and hits an error, is the response clear enough to retry or route around? Generic messages ("Something went wrong") are a classic agentic-web anti-pattern — humans tolerate them; agents can't.

6. Don't wait for perfect standards. Some honesty here: Google denies using llms.txt, and the scoring tools on the market (AgentGEOScore, GeoReady, AEOlens...) weight things differently with no unified benchmark yet (sources: https://agentgeoscore.com/, https://geoready.dev/, https://aeolens.ai/). But web-standard history repeats itself: RSS, sitemaps, schema — developer-adjacent adopters led by 18–36 months before mainstream followed once CMSs shipped defaults. Moving early isn't about guaranteed wins; it's about paying the lowest possible tuition while the lessons are cheap.

The takeaway

Back to Tuan and his furniture store. What he needs to do this week is not build an MCP server. It's smaller: make sure his website states the truth in language both people and machines can read — clear prices, complete descriptions, cited sources, markdown versions of the most important pages, and one llms.txt file as the map.

Because the rule of this era is simple: your real customers will increasingly arrive through a digital broker. That broker only recommends businesses it can understand. Your website gets hired by two kinds of customers — make sure that for the second kind, the door is open.

Keep reading in the Agentic Web series

• DNS4ACID and Vietnam's open-source Agentic Web community — how a Vietnamese community is building agentic web infrastructure.

References

• is-agentic methodology & scoring model: https://is-agentic.com/methodology
• State of llms.txt 2026 (219-host panel): https://llmtxt.info/blog/state-of-llms-txt-2026/
• Tranco top-10K llms.txt study: https://crawlmind.ai/research/state-of-llms-txt-2026
• llms.txt spec v2: https://llmstxt.org/
• MCP stateless spec 2026-07-28: https://blog.modelcontextprotocol.io/posts/2026-07-28/
• A2A v1.0 announcement: https://a2a-protocol.org/latest/announcing-1.0/
• A2A adoption (Linux Foundation): https://www.linuxfoundation.org/press/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year
• WebMCP / interoperability standards 2026: https://gravity.fast/blog/ai-agent-interoperability-standards-2026/
• GEO/AEO statistics (BrightEdge, Similarweb): https://seoscaleup.com/blog/geo-aeo-statistics-2026/
• Conversion benchmarks (Seer Interactive, Adobe, Conductor): https://ranqo.ai/blog/generative-engine-optimization-statistics and https://www.conductor.com/academy/aeo-geo-benchmarks-report/
• GEO research (Princeton, KDD 2024): https://www.similarweb.com/blog/marketing/geo/what-is-geo/
• Agent readiness benchmarks (AgentGrade, GeoReady): https://asiangrade.com/agent-readiness and https://geoready.dev/

— 5ac.vn Editorial, 23/08/2026. The agentic web era for Vietnamese SMBs, powered by G-Company OS.