Act 2 · 11 min

What I Found

The gold mine, told through actual work.

2.1 · Prompts Are Intent

 Heatmap · /pricing2,184 sessions
acme.com/pricing
Plans & pricing
Pick the plan that fits your team.
StarterFreeGet started
Pro$29Choose Pro
Business$99Contact us
How many seats are included?
Can I upgrade later?
Do you offer team discounts?
Tells you what they clicked. Not why.
 Prompt · same usersingle message
"Is this worth it for a small team or should we just stick with the free tier?"
Tells you what they were thinking. In their own words.
price sensitivity team size: small comparison shopping objection: cost alt: free tier decision pending

I'd spent years inferring intent from click patterns. Now I was just… reading it.

2.2 · They Know What Their Users Want

Claude usage Anthropic Economic Index
Coding37%
Writing25%
Research16%
Education10%
Business ops8%
Other4%
Source: anthropic.com/economic-index
ChatGPT usage OpenAI · NBER paper
Practical guidance29%
Writing24%
Seeking info22%
Technical help13%
Self-expression7%
Other5%
Source: NBER w34255 · How people use ChatGPT

They publish whole roadmaps from this. Do you have this view of your own users?

…and we're at peak human internet.

65% 57% 50% 42% 35% 2013 '17 '21 '22 '23 '24 '25 '26 Bots overtake humans Humans 46%* Bots 54%*
Imperva · 2024
51% bots
First time automated traffic surpassed human visitors. Up from 43% in 2013.
imperva.com · 2025 Bad Bot Report
Imperva · 2013 → 2024
57% → 49%
Human share of web traffic. Declining for the sixth straight year.
businesswire · Imperva 13-yr trend
Cloudflare · 2025
+305% YoY
GPTBot crawler growth in one year. 82% of AI crawls are for training, only 2% for user actions.
cloudflare · Googlebot to GPTBot

And the prompts? They're not even on this chart. They're inside Claude, ChatGPT, Gemini — places your tags can't reach.

2.3 · "We thought everyone was doing X.
They were actually doing Y."

 On paper · documented
Green-energy legal review
1Receive contract
2Tag clauses by category
3Compliance checklist
4Sign-off & archive
 In the prompts · reality
What people actually asked the AI
1Receive contract
"What changed vs last year's template?"
"Who normally signs this kind of deal?"
"Is the indemnity clause unusual?"
"Draft me a redline I can send back"
4Sign-off & archive
was: personal · invisible now: every interaction · traceable company brain

Update the AI to match reality, not expectation — and every question since becomes a node of shared institutional memory.

2.4 · Analysing Text at Scale

How do you actually analyse thousands of prompts? Text, not numbers.

01
Embedding clusters
BigQuery ga4_prompts → embeddings
"pricing" "how-to" "bugs"
= segmentation
02
LLM-as-judge
"third time I've asked — just give me the refund"
intentrefund sentimentfrustrated resolvedno
= custom dimensions

Sound familiar? It's funnels and segments. Just different raw material.

2.5 · Langfuse Demo

Same analytics discipline — radically better data.

Langfuse fleet dashboard — traces, model costs, eval scores
Langfuse session view — multi-turn conversation with per-turn scores
Langfuse single trace — nested retrieval, embedding, generation spans
Langfuse trace with LLM-as-judge eval scores
Langfuse Gantt-style span timeline showing parallel execution
01 · fleet Traces · cost · scores, in one place your GA4 home screen, for agents
02 · session Multi-turn conversation, end-to-end every turn scored
03 · trace One trace, opened up retrieval → embed → generate
04 · judge LLM-as-judge scores attached satisfaction · coherence · fluency · feedback
05 · timeline Every span, every parallel call execution shape, not just the outcome
From 2.4 · now running
Judge scores attached
Human signal
Feedback on the trace
New vs GA4
See parallelism & latency shape
Coming up tonight · Peter Meyer
Using end-to-end testing to verify tracking implementation
"Peter will show you how to verify your GA4 tags actually fire. The same discipline applies to agent traces — except now you're verifying reasoning, not just events."
1 / 1
←→ navigate