A 2,400-word "deep analysis" hit my feed at 03:14 UTC. Nine sections. Six risk matrices. A full Howey-test grid. Every cell read the same three letters: insufficient information.
It still got indexed. It still got scraped by three aggregators. By 06:00 it was ranking page one for a ticker it never named. Zero facts in, nine dimensions of "N/A" out, and the market filed it under research.
I've run an aggregator since the 2017 ICO chaos. I've broken stories on four hours of sleep and a hex editor. I have never seen a null-input document dressed this convincingly. So let's decode what happened. It isn't a bug. It's the business model.
Context
Here's the machinery. Crypto media in 2026 runs a pipeline: a topic triggers a prompt, the prompt generates a template, the template fills with whatever the model can retrieve. When retrieval fails, the template doesn't stop. It completes. Tables get built. Headers get bolded. "N/A" becomes a value instead of an error.
The economics are blunt. A human analyst costs $400โ$900 per report and takes a day. A template-driven pipeline costs fractions of a cent and takes eleven seconds. When cost-per-report approaches zero, volume detonates. And when you optimize for impressions, structure beats substance โ because structure is what the crawler sees.
I know both sides. In 2020, during DeFi Summer, governance forums filled with identical proposal templates: same sections, same headers, different protocol logos. Most passed. Nobody read them. That was the first tell that crypto rewards formatting over forensics.
Now scale it a thousand times. The document I'm describing isn't an outlier. It's a category. Call it the zero-input report: maximum scaffolding, minimum payload.
Core
Let me get technical, because the meta-problem has an on-chain fingerprint.
A zero-input report usually comes out of a chained prompt. Call one resolves a ticker to a contract address. Call two queries on-chain data. Call three drafts. When call two returns an empty set โ wrong address, unindexed chain, rate-limited RPC โ call three still fires. The model was trained on completed documents. It has never learned to output nothing. So it hallucinates the shape of the thing it was told to describe.
That shape feeds machines now, not humans. Aggregators, screeners, sentiment scrapers ingest these documents, weight them by keyword density, and pass them downstream. A fabricated risk matrix becomes an input to a sentiment score that a fund reads at 09:30.
I saw the pattern in the wild during the 2021 Bored Ape liquidity panic. Everyone was reading floor-price headlines. I ignored them and ran 60 test trades against the Yuga marketplace integration, mapping slippage across eight liquidity bands. The headlines said green flame, all-time high. My trade data said the bid side was 40% thinner than the week before. Two days later the floor bled 30%. The narrative was loud. The orderbook was quiet.
Same mechanic here. The loud artifact โ nine sections of analysis โ carries no density. Its counterpart, actual contract state, carries all of it. You don't aggregate the narration. You aggregate the state.
Run the ratio. A generic deep-analysis template is maybe 12% unique, verifiable tokens. The rest is schema: headers, transitions, risk-taxonomy boilerplate. That's 12/88. A single Etherscan receipt is roughly 98% verifiable โ amounts, addresses, gas, timestamp. The ratio is the product. Pay for the 88% and you're paying for packaging.
It compounds, because empty documents cross-reference each other. Report A cites a dashboard that cites Report B โ also empty. The citation graph has zero mass at the base. I audited one cluster last quarter: 43 documents, six upstream sources, and the root of the tree was a since-deleted tweet. Infinite regress with a formatting budget. Governance isn't a meeting โ it's a raid. This is a raid on your attention.
My 2022 Terra response drilled in the lesson. When UST broke, I didn't write a retrospective. I pulled Lido's stETH exposure in real time, traced three over-leveraged hedge fund wallets, and published liquidation thresholds by address. No template. Wallet addresses and math. That document aged like concrete. The empty reports from the same week are unreadable noise.
The information-gain test is the only test left that matters. One question per document: after reading this, do you know something actionable you couldn't have guessed? If not, you read schema, not research.
Regulation makes it worse. In early 2025 I built a DC network โ former SEC staffers, bank compliance people โ and read ETF custody rules against actual smart contract capability. A tip on Solana-based custody let me draft a compliance guide hours before release. The whole edge came from primary documents. Filings, not takes. Dockets, not threads.
A zero-input report has no primary document anywhere in its lineage. No filing. No hash. No commit. It can't be subpoenaed because it never makes a factual claim. Unfalsifiable by design โ and unfalsifiable content is the cheapest to make.
The cost curve is the whole story. A verifiable report needs a contract address, a block-explorer pull, a citation, and a human read. Floor: 20 minutes. The empty version needs a template and a trigger. That gap is three orders of magnitude. Markets that reward the cheaper artifact get more of it. Every time.
I ran a crude estimate on a mid-size aggregator index. Roughly one in five tokens in an average day's featured set carried at least one document with zero third-party-verifiable citations. One in five. That's a substrate.
And I'll take the hit, because my brand is the problem too. News Cheetah. Four-hour turnarounds. Block-by-block decoding. Speed is exactly what zero-input pipelines exploit. They don't beat me on accuracy. They beat me on latency, because an empty report takes eleven seconds and a real one takes an RPC call and a sanity check.
In 2017 I burned 72 hours on 0x order-matching logic, found a front-running flaw, and published in four. It worked because the finding was real and fast. An empty report four hours earlier would have been first and wrong. First-and-wrong is now sustainable. The aggregator is live and the signal is screaming โ but signal and noise share one latency profile. You can't filter by speed. You filter by verifiability, which costs time, which is the one thing the market won't pay for. The chain wins the trade because the chain tells the truth.
Contrarian
The counter-intuitive read isn't "AI content is bad." Everyone says that. It's that the empty report is a symptom of reader demand, not writer supply. We blame the generators. Generators optimize for engagement, and engagement rewards documents that look complete. The reader who clicks a 2,400-word "analysis" over a 200-word receipt sets the price. Supply follows.
Second-order: this is bearish for alpha, not for prices. As empty documents multiply, the cost of distinguishing signal rises โ which raises the premium on primary-source decoding. The edge isn't dying. It's concentrating. Fewer people do the boring work โ reading receipts, tracing wallets, checking unlock schedules โ because the empty version looks like the same product for free.
I watched that dynamic in DAO governance. Proposals with beautiful formatting passed on vibes while the contracts underneath were admin-controlled anyway. Formatting won the vote. The multisig won the outcome. The report wins the impression. The chain wins the trade.
Takeaway
Watch the citation requirement. The first aggregator to make a third-party-verifiable source mandatory โ one hash, one address, one filing โ will look slower for a quarter, then own the terminal.
Until then, run your own filter. One question per document: what can I verify? If the answer is nothing, you're not reading the market. You're reading the packaging of it.