Act 3 · 7 min

What I Think Is Coming

Speculation grounded in experience.

3.1 · The web evolved in eras

What changed isn't the data — it's the path to it.

1991
Era 01 Web 1.0
GET
Static HTML
read-only documents
SQL
Read
2004
Era 02 Web 2.0
POST
HTML + forms
pretty UIs to read & write
SQL
Read · Write
2010s
Era 03 Cloud / SaaS
REST
SaaS portals
multi-tenant collaboration
API
Share
2025
Era 04 · now Web for Agents
×
HTML
bypassed
×
Transact

The database stayed The wrapper changed Agents skip the wrapper

3.2 · Agents need their own protocols

The page never loads. The JS never fires. Your tag never runs.

 The bypass
BrowserPage load
JS / GTMAnalytics
replaced by
AgentProtocol
Tool callAPI / Data
MCP Anthropic · 2024
Model Context Protocol — agents call tools and read context from any compliant server.
A2A Google · 2025
Agent-to-Agent — agents discover and delegate to each other across vendors.
AG-UI CopilotKit · 2025
Streams agent output into UIs — the new "frontend" for an agent.
ACP IBM / BeeAI · 2025
Agent Communication Protocol — standardised request/response for orchestration.
Coming up tonight · Serge Shkvarnytskyi · Stape
Server-side tracking: From audit to reporting
"Serge talks server-side later — same trust and control problem from a different angle. We're both moving collection off the page."

3.3 · Payment Required is back

HTTP status
402
Payment Required
Reserved 1991. Unused for 30 years. Revived by AI agents in 2025.
Agent hits a paid endpoint → server returns 402 + price → agent signs a stablecoin payment → retries with payment receipt header → gets the resource. One round trip. No accounts, no API keys, no human in the loop.
x402
Coinbase · Cloudflare · Stripe
HTTP 402 revival. Pay-per-request stablecoin micropayments. Cloudflare uses it for pay-per-crawl.
AP2
Google · 60+ partners
Agent Payments Protocol. Authorization & audit framework. A2A x402 extension is production-ready.
TAP
Visa
Trusted Agent Protocol. Identity & merchant trust signals for agent-initiated card payments.
ACP
Stripe · OpenAI · Mastercard
Agentic Commerce Protocol. Open spec for agent checkout in third-party storefronts.

Galaxy estimates agentic commerce reaches $3–5 trillion by 2030.

3.4 · Websites become research surfaces

Crawled for what's not in the training set. Visited by humans almost not at all.

 Why AI bots crawl
Training79%
Search17%
User actions3.2%
 Crawl-to-refer ratio · July 2025
Anthropic38,000:1
OpenAI3,700:1 peak
Perplexity~100:1
Pages crawled per page referred. Source: Cloudflare · The crawl-to-click gap

Your website used to be the destination. It's becoming a freshness layer — queried only for what isn't already in the model's head.

3.5 · The competitive surface

Everyone has the same models now. The moat lives somewhere else.

Pillar 1
Unique data
yours alone
Your customers' prompts. Your domain knowledge. Your private corpus. The only thing competitors can't copy.
← this is what Act 2 was about
+
Pillar 2
The model
commodity
GPT, Claude, Gemini, Llama, DeepSeek, Qwen. Open weights, hosted APIs, drop-in swaps. Everyone has the same access — you, me, your competitor.
no edge here anymore
+
Pillar 3
Harness & UX
your surface
The agent loop, the tool catalogue, the website, the chat interface, the eval harness. How your data and the model become a product.
the visible part of the moat
Your website is the visible surface of your harness — the part the world sees of how well you've combined your unique data with a commodity model. It's not the destination anymore. It's the showcase.

3.6 · The new customer journey

Not a funnel anymore. A stack — with humans on top and protocols at the bottom.

Tier 1
Human
Intent Judgement Trust Approval Satisfaction
Tier 2
Agent
Plan Retrieve Tool use Reason Retry Hand-off
Tier 3
Protocol
MCP A2A x402 / AP2 OpenAPI OTel
Coming up tonight · Caroline Vidal · DFDS
Agentic Lead Enrichment & Attribution
"Caroline goes deeper on agentic enrichment after me — she'll show you how to attribute and act on these stratified journeys. I'm here to ask whether you're capturing them at all."

3.7 · A new class of data

Some metrics rename. Most are genuinely new.

Web analytics
Agent analytics
Pageviews
Tool calls
Sessions
Traces
Conversion rate
Task success rate
Bounce rate
Hallucination & retry rate
Cost per acquisition
Cost per task (tokens × rate)
Page load time
Time-to-first-token / latency budget
Funnels
Trace trees
Custom dimensions
LLM-as-judge scores
BigQuery
Still BigQuery

The tools changed. The thinking didn't.

1 / 1