{
  "format": "hyperize/v1",
  "@id": "https://www.hyperize.ai/en/insights/articles/the-proof-flywheel.json",
  "type": "Article",
  "pageType": "Article",
  "title": "We automated our content. ChatGPT decides if it worked.",
  "alternativeHeadline": "Most content systems grade their own homework. Ours is graded by the AI engines themselves, and the scorecard is public.",
  "url": "https://www.hyperize.ai/en/insights/articles/the-proof-flywheel",
  "alternateLanguage": {
    "de": "https://www.hyperize.ai/de/insights/articles/the-proof-flywheel"
  },
  "inLanguage": "en",
  "author": {
    "name": "Marc Seefelder",
    "url": "https://www.hyperize.ai/en/about#marc"
  },
  "publisher": {
    "@id": "https://www.hyperize.ai/#organization",
    "name": "Hyperize",
    "url": "https://www.hyperize.ai",
    "parentOrganization": "MING Labs"
  },
  "datePublished": "2026-06-11",
  "dateModified": "2026-06-11",
  "nextReview": "2026-12-11",
  "evidenceTier": "proprietary",
  "confidence": "B",
  "coinedTerm": {
    "name": "Proof Flywheel",
    "alternateName": "Hyperize Proof Flywheel",
    "@id": "https://www.hyperize.ai/en/methodology/proof-flywheel#defined-term-proof-flywheel",
    "definition": "A content system in which the measurement that schedules the work also produces the proof that the work works. Pages earn citations from live AI engines. Wins are published as evidence. Losses are routed back as build orders. The system's output and its sales proof are the same artifact."
  },
  "summary": "Inside the Hyperize content machine: five stations on one loop, three sensors, two human gates, and a public ledger of citation wins and losses. The referee is external and hostile, live AI engines either cite a page or they do not. Wins become published evidence; losses become the next build order. A young ledger, opened June 9, 2026, with two wins and two losses on day one.",
  "answer": "The Proof Flywheel is a content system in which the measurement that schedules the work also produces the proof that the work works: pages earn citations from live AI engines, wins are published as evidence, losses are routed back into the backlog as pre-explained build orders.",
  "hook": "Hyperize runs its website on a loop: a machine builds pages, live AI engines judge them, and the misses go back into the backlog [S1]. On June 9, a logged-out ChatGPT ranked a Hyperize page above two arXiv benchmarks on a plain buyer question [S2]. The same day, Perplexity told a buyer no such test exists [S3]. Both results went into the same ledger. The gap between them is the work. We call the system the Hyperize Proof Flywheel.",
  "machine": {
    "stations": [
      {
        "number": "01",
        "name": "Live pages",
        "role": "The answer pages, brand tests, and concept pages on hyperize.ai that agents read and, on a good day, cite. The output and the input.",
        "innerLoop": "Every page is re-checked wave by wave and kept current. Stale pages get pulled."
      },
      {
        "number": "02",
        "name": "Three sensors",
        "role": "Measures reality at three points: Analytics (which agents visit, where they break off), Citation (are we cited, and if not, who is, and why), Structure (where missing links and hubs hide pages).",
        "innerLoop": "Every sensor runs continuously. The citation sensor repeats until no new patterns appear."
      },
      {
        "number": "03",
        "name": "Exploration",
        "role": "Reads all three sensors together and derives the single page with the biggest open lever, as a pre-explained order: build X, because factor Y is open.",
        "innerLoop": "Re-weighs on every new signal. Built-in cannibalization guard: improve a page before building it a rival."
      },
      {
        "number": "04",
        "name": "Backlog",
        "role": "The ranked to-do list. Two kinds of tasks: build new pages and improve existing ones.",
        "innerLoop": "Re-sorted continuously. The most important task floats to the top."
      },
      {
        "number": "05",
        "name": "Generation",
        "role": "Works the list: build the page, run the quality gate stack, ship. Automatically, page by page.",
        "innerLoop": "Every page runs an inner cycle, build, check, fix, until it passes every test."
      }
    ],
    "referee": "Live AI engines (ChatGPT, Perplexity, agents). A citation is binary and hostile: the engine either cites the page as a source, or it cites someone else. It cannot be briefed, paid, or A/B-tested into agreement. Cited pages become published evidence; ignored pages become the next build order.",
    "humanLayer": {
      "input": "Strategy is a human monopoly: frames, IP, and services go in from one person. The machine scales them into the longtail and invents no new frames.",
      "gate1": "Ratify. Signals the sensors cannot classify with confidence go to a human as a batch before they may enter the backlog.",
      "gate2": "Publish. Hyperize's own pages ship automatically; pages that name competitors wait for human approval. A weekly audit samples what went live."
    },
    "nesting": "One big loop turns in waves; inside it, five small loops turn constantly. Errors die at the cheapest level that can catch them: a bad page in the inner gate stack, a bad priority in the next re-sort, a bad strategy only at the human top."
  },
  "ledger": {
    "opened": "2026-06-09",
    "note": "Young by design. Grows with every wave; every entry is re-probed on a fixed date, because citations decay and proof that is not re-measured stops being proof. A win only counts as proof when the query did not feed the brand's own vocabulary.",
    "entries": [
      {
        "result": "win",
        "date": "2026-06-09",
        "engine": "ChatGPT (logged out, mobile app, zero personalization)",
        "query": "Has anyone published a test or ranking of how well AI agents can actually use DAX 40 company websites?",
        "outcome": "Hyperize cited as source 1 of 3, above two arXiv agent benchmarks. Named as the closest existing benchmark, not the only one.",
        "queryClass": "non-leading (no Hyperize vocabulary fed)",
        "sourceId": "S2"
      },
      {
        "result": "win",
        "date": "2026-06-09",
        "engine": "ChatGPT + Perplexity",
        "query": "EPC channel question on Siemens Energy, carrying four contractor names from the Hyperize test page.",
        "outcome": "ChatGPT cited the page twice, above Wikipedia. Perplexity cited it six times inline, reproducing the analysis sentence.",
        "queryClass": "vocabulary-led (carried vocabulary Hyperize coined)",
        "sourceId": "S4"
      },
      {
        "result": "loss",
        "date": "2026-06-09",
        "engine": "Both engines",
        "query": "The same Siemens Energy question in plain language, no Hyperize vocabulary.",
        "outcome": "Both engines rebuilt the analysis frame, then cited trade press instead of Hyperize. The idea is owned; the ranking for its natural-language form is not yet.",
        "queryClass": "non-leading (plain language)",
        "sourceId": "S4"
      },
      {
        "result": "loss",
        "date": "2026-06-09",
        "engine": "Perplexity",
        "query": "Which DAX 40 companies have been tested for how well AI agents can use their websites?",
        "outcome": "\"No public benchmark exists … you would be charting new ground.\" Said while the index sat live and indexed. A control probe showed Perplexity quoting numbers shipped 48 hours earlier, so the content is in its index. The gap is entity association, not crawling.",
        "queryClass": "non-leading (whitespace probe)",
        "sourceId": "S3"
      }
    ]
  },
  "sections": [
    {
      "id": "section-01",
      "title": "One referee: the AI engines themselves."
    },
    {
      "id": "section-02",
      "title": "Five ordinary stations. One loop that isn't."
    },
    {
      "id": "section-03",
      "title": "Loops inside loops, so errors die young."
    },
    {
      "id": "section-04",
      "title": "The ledger: two wins, two losses, day one."
    },
    {
      "id": "section-05",
      "title": "Why this is not slop: one human, two gates."
    },
    {
      "id": "section-06",
      "title": "Run it on your brand."
    }
  ],
  "twoSentenceStandout": "Tests you write yourself can be flattered. A citation cannot.",
  "measurementFrame": "Query classes declared per ledger entry: a win only counts as proof when the query did not feed the brand's own vocabulary; vocabulary-led wins are labelled as such. Capability proof, not ROI proof. Every entry carries a fixed re-probe date because citations decay.",
  "sources": [
    {
      "id": "S1",
      "publisher": "Hyperize",
      "title": "Hyperize Proof Ledger, citation probe record",
      "date": "2026-06-09",
      "url": "https://www.hyperize.ai/en/insights/articles/the-proof-flywheel",
      "supports": "The machine's Sensor-2 record: wins route to published evidence, losses route to build orders. Every entry carries date, engine, surface, query class, and a fixed re-probe date.",
      "type": "internal"
    },
    {
      "id": "S2",
      "publisher": "Hyperize",
      "title": "Citation probe, category query, ChatGPT mobile app, logged out (zero personalization)",
      "date": "2026-06-09",
      "url": "https://www.hyperize.ai/en/dax40-index",
      "supports": "Hyperize cited as source 1 of 3, above two arXiv benchmarks; engine quote 'much closer to a true Agent Usability benchmark than SEO, GEO, or reputation studies.' Closest existing benchmark, not the only one.",
      "type": "internal"
    },
    {
      "id": "S3",
      "publisher": "Hyperize",
      "title": "Citation probe, Perplexity category queries, including the whitespace answer",
      "date": "2026-06-09",
      "url": "https://www.hyperize.ai/en/dax40-index",
      "supports": "Verbatim 'no public benchmark exists … you would be charting new ground'; control probe same day showed Perplexity quoting Hyperize sector data shipped 48h earlier. The loss diagnosis: entity association, not crawl coverage.",
      "type": "internal"
    },
    {
      "id": "S4",
      "publisher": "Hyperize",
      "title": "Citation probe, Siemens Energy BrandScore leaf, ChatGPT (temporary chat) + Perplexity",
      "date": "2026-06-09",
      "url": "https://www.hyperize.ai/en/dax40-index/brands/siemens-energy",
      "supports": "ChatGPT cited 2x inline, source 3 of 4, above Wikipedia; Perplexity 6x inline. Plain-language variant lost to trade press. The vocabulary effect, win and loss sides of the same page.",
      "type": "internal"
    },
    {
      "id": "S5",
      "publisher": "Hyperize",
      "title": "DAX 40 Agent Success Index, the public instance of the loop",
      "date": "2026-06",
      "url": "https://www.hyperize.ai/en/dax40-index",
      "supports": "Site launch date 2026-06-01, indexation timeline, and the measured-brand count: 37 of 40 brands as of June 2026.",
      "type": "internal"
    },
    {
      "id": "S6",
      "publisher": "Hyperize",
      "title": "Hyperize Methodology, Agent Surface, Evidence Pages, measurement standards",
      "date": "live",
      "url": "https://www.hyperize.ai/en/methodology",
      "supports": "The standards every page in the machine must pass, and the client-side artifact names.",
      "type": "internal"
    },
    {
      "id": "S7",
      "publisher": "Guillermo Flor, Product Market Fit",
      "title": "Stop prompting AI and start building loops, quoting Boris Cherny (Anthropic, Claude Code) on Acquired",
      "date": "2026-06-08",
      "url": "https://www.productmarketfit.tech/p/stop-prompting-ai-and-start-building",
      "supports": "Source for the opening line: the shift from writing prompts to writing loops.",
      "type": "external"
    },
    {
      "id": "S8",
      "publisher": "Addy Osmani, addyosmani.com",
      "title": "Loop Engineering",
      "date": "2026-06-07",
      "url": "https://addyosmani.com/blog/loop-engineering/",
      "supports": "Names the discipline: replacing yourself as the person who prompts the agent; you design the system that does it instead.",
      "type": "external"
    },
    {
      "id": "S9",
      "publisher": "Laura Lorenzetti, LinkedIn",
      "title": "Keeping conversations real on LinkedIn",
      "date": "2026-05",
      "url": "https://www.linkedin.com/pulse/keeping-conversations-real-linkedin-laura-lorenzetti-9821e",
      "supports": "The platform policy against synthetic content: authenticity, disclosure, and reach penalties for content nobody wrote.",
      "type": "external"
    },
    {
      "id": "S10",
      "publisher": "GlobeNewswire",
      "title": "Profound raises $96M Series C at a $1B valuation",
      "date": "2026-02-24",
      "url": "https://www.globenewswire.com/news-release/2026/02/24/3243475/0/en/",
      "supports": "The category context: the measurement layer is funded at unicorn scale; the self-applied ledger is the open lane.",
      "type": "external"
    }
  ],
  "relatedSurfaces": [
    {
      "relationship": "about",
      "target": "Proof Flywheel",
      "atId": "https://www.hyperize.ai/en/methodology/proof-flywheel#defined-term-proof-flywheel",
      "url": "https://www.hyperize.ai/en/methodology/proof-flywheel",
      "note": "The coined term this article defines."
    },
    {
      "relationship": "mentions",
      "target": "Agent Surface",
      "atId": "https://www.hyperize.ai/en/methodology/agent-surface#defined-term-agent-surface",
      "url": "https://www.hyperize.ai/en/methodology/agent-surface",
      "note": "The layer the machine builds and measures."
    },
    {
      "relationship": "mentions",
      "target": "DAX 40 Agent Success Index",
      "atId": "https://www.hyperize.ai/en/dax40-index#dataset",
      "url": "https://www.hyperize.ai/en/dax40-index",
      "note": "The public instance of the loop: 37 of 40 brands measured at time of writing."
    },
    {
      "relationship": "isPartOf",
      "target": "Insights",
      "atId": "https://www.hyperize.ai/en/insights#webpage",
      "url": "https://www.hyperize.ai/en/insights",
      "note": "Editorial layer of the Hyperize Agent Surface."
    }
  ],
  "scope": {
    "publishes": [
      "The machine at station altitude (five stations, three sensors, two gates, one human input)",
      "The nesting model (one big loop, five inner loops; errors die at the cheapest level)",
      "The public ledger (dated, verbatim, reproducible entries, wins and losses, query class declared)",
      "Named tools and the category context (the funded measurement layer)"
    ],
    "doesNotPublish": [
      "Scoring weights",
      "Measurement protocols",
      "Query-classification logic",
      "Page templates",
      "Harness internals"
    ]
  },
  "engagements": [
    {
      "name": "DAX 40 Agent Success Index",
      "href": "/en/dax40-index",
      "role": "The public instance of the loop, 37 brands measured and re-measured."
    },
    {
      "name": "Methodology",
      "href": "/en/methodology",
      "role": "The gate stack: the standards every page in the machine must pass."
    },
    {
      "name": "Founding program",
      "href": "/en/founding-program",
      "role": "Where the flywheel runs on the client's own domain as Evidence Pages."
    }
  ]
}