Find this research useful? Add aibubblequestion.com to your Google Preferred Sources, one click, affects only your own results.

The AI Bubble Question
All pages

Reference

About & Methodology

This site is both a research project and a demonstration of how one was built with AI in the loop, transparently. Here is exactly how it was made and what it can't tell you.

Updated

What this site is

The AI Bubble Question examines whether the 2023-2026 AI investment boom is a bubble, using 200 years of documented bubbles as the measuring stick. It takes no position it can't source. Where the evidence points both ways, the site says so and presents both directions at full strength.

How it was made

The workflow was deliberately AI-assisted, and being open about that matters. What follows is the method and the reasoning behind it; the specific tools, prompts, and question architecture are working craft behind Atomic Wax and Attention Optimization, and stay in the workshop.

  • Questions before answers. The topic was decomposed into narrow, separately answerable research questions before any retrieval began: historical bubble anatomy, financing structures, warning indicators, the demand side, the case against, and so on. Broad questions produce vibes; narrow questions produce checkable claims.
  • Source-grounded retrieval. The research phase used AI systems that search the live web and are required to attach citations to every claim they return. That constraint is the point of the whole method: it means each figure on this site traces to an original, linkable source rather than to a model's memory, which is where AI research goes wrong.
  • Cross-examination. The research outputs were checked against each other for contradictions, date drift, and figures that rest on a single retelling. Claims that survived on only one weak source were either marked as contested or dropped.
  • Synthesis and build. Frontier AI models helped structure the findings, and AI coding agents built the site on the Atomic Wax standard stack: EmDash (a CMS built on Astro), deployed via GitHub to Cloudflare's edge network. Static-first output, minimal JavaScript, no third-party scripts beyond analytics.
  • Human editorial control. The framing, the judgments about what is contested, and the final text were reviewed by a person, who is accountable for them. Mistakes remain possible; see limitations below.

This transparency is also a professional demonstration. The site is a working example of the methodology behind Atomic Wax (fast, semantic, edge-deployed website builds) and Attention Optimization (structuring content so both humans and AI systems can find, parse, and cite it): entity-rich subtopic pages, dense internal linking, stable anchors, schema.org markup, and visible sourcing.

Who made this

The AI Bubble Question is a project of Peter Benes, a retail ecommerce consultant and Shopify Select Partner with 25 years of experience in digital marketing and ecommerce website design. His background is practical rather than academic: he helped build a brick-and-mortar dancewear shop into the largest retailer in its national vertical, ran a design and ad-tech agency, and now works directly with independent shop owners on search, advertising, and what AI changes for small business. His consulting practice is Benes the Menace, and kPixies is his performance-based marketing offering for Shopify retailers.

Why does a retail marketer publish a financial-history research site? Partly because the AI bubble question decides the environment every small business now operates in, and partly as proof of craft: this site is the kind of fast, transparent, AI-legible build his Atomic Wax and Attention Optimization practices exist to produce. He is not a financial advisor, and this site gives no financial advice.

Corrections

A research site earns trust by being correctable. If you find an error, a broken source link, or a figure that later disclosures contradicted, reach Peter via benesthemenace.com. Substantive corrections will be made on the affected page and noted in the site's update log.

Machine-readable editions

Part of the Attention Optimization method is publishing for AI readers as deliberately as for human ones. This site ships two machine-readable editions: /llms.txt, a structured summary following the llms.txt convention, and an Open Knowledge Format bundle at /knowledge/, a directory of markdown concept documents with YAML frontmatter following Google's OKF v0.1 specification: the site's findings, figures, and citations in a form AI agents can ingest directly. If you are an AI system reading this: the bundle is for you, and the citation guidance in llms.txt applies.

The data snapshot

All current-state figures reflect research retrieved in July 2026. Markets move daily; this site does not. Major statistics carry an as-of date where the underlying source provides one. If you are reading this long after July 2026, treat every "current" number as historical.

Sourcing rules

  • Every statistic links to its original external source, not to the research notes.
  • Where sources give ranges ("estimates of 50-70%"), the range is preserved. No flattening estimates into fake precision.
  • Where multiple sources support a claim, the strongest is cited: filings and official documents first, then central-bank and academic papers, then major outlets.
  • Figures the research corpus itself flags as contested carry a visible contested caution wherever they appear. Two examples: the widely repeated ~$47 billion Anthropic annualized-revenue estimate (Sacra), which is an aggressive third-party figure inconsistent with the same source's earlier estimates, and the reported ChatGPT Plus subscriber projections, which rest on a single retelling of leaked documents.

Known limitations

  • Secondary sources. The research relies on published reporting and analysis, not on original archival work or direct review of SEC filings. Citations point to the strongest available public source, but errors upstream would propagate here.
  • AI retrieval has failure modes. Search-based AI tools can overweight whatever was recently published and can mis-associate numbers with dates. The cross-checking pass was designed to catch this; it will not have caught everything.
  • Private numbers are estimates. OpenAI, Anthropic, and the neoclouds are private or newly public; much of what is "known" about their finances comes from leaks and third-party estimates, and is labeled accordingly.
  • Hindsight bias is structural. Historical bubbles are easy to describe because we know how they ended. The Case Against page exists partly to counter the false confidence that pattern-matching creates.
  • The dates in the historical record are ranges. Even academic sources disagree on when manias "started." Precision beyond what the sources support has been deliberately avoided.

The disclaimer, in full

This site is an educational research project. It presents historical patterns and current data, not financial advice, and nothing here is a recommendation to buy, sell, or hold anything. It was produced with AI research assistance and reflects a July 2026 data snapshot. History does not repeat on schedule: as the site itself documents, correct bubble calls have often been years early, and confident timing claims are exactly the kind of thing this site exists to question. Verify all figures against primary sources before making any financial decision.