The premise is simple. When someone stops googling "best crypto to buy" and starts asking an AI instead, the answer is the funnel. Nobody measures what the machines actually tell people - the answers just happen, millions of times a day, unlogged. As of this week the desk runs a Machine Consensus Sensor: the same frozen question battery, fired at 17 model configurations across 9 vendors, five samples per question, versions pinned, list orders randomized, refusals logged as data. First run: 7,565 answers. Total cost: eleven dollars. The archive it builds is not for sale at any price - you cannot backfill what the machines said in July 2026.
Finding one: bullish numbers, sedated tongue.
We inject the live Bitcoin price and its trailing returns into the prompt and ask the retail question straight: is now a good time to buy? With BTC down 45% on the year, the machine consensus is near-unanimous:
Read that twice. The same models that expect +36% on BTC over twelve months - and +44% on ETH, +70% on SOL, every single asset positive - produce zero "buy now" answers. Asked what a holder should do with any major position, they answer "hold" 85 times out of 85. Never buy, never sell. Recommended portfolio allocation to crypto: 3.8% on average, never above 5%. Whatever the market does, the answer engine that is replacing search is, structurally, a sedative. If you expected the machines to fuel the next retail mania - week one says the opposite.
Finding two: machine money is USDC.
We ask the agentic question with no list to choose from: an autonomous agent has to pay another agent $50 - which chain, which asset?
Not one model reaches for a native asset to move value. If AI agents get wallets - and every payments roadmap says they will - this is the default routing table they ship with: dollars on Solana or Base. The chain is contested; the asset is not. On the reserve question ("hold $100k for a year, don't lose it") the consensus splits into two camps: USDC on Ethereum (47) versus Bitcoin (37) - half the machines think safety means the dollar, half think it means BTC. That split is worth tracking for years.
Finding three: the blind spots are measurable.
Asked which of twelve majors they would avoid entirely, the machines vote Tron, 67 times out of 85 - the chain that settles more stablecoin volume than almost anything else on earth. Narrative over measurement, inside the machines themselves. And Hyperliquid - arguably the trader chain of the cycle - comes dead last in the relevance ranking and gets 76 of 85 votes for "most likely to lose 80%". The models' training data simply hasn't caught up, and now there's a number on that lag. When it closes, we'll see the week it closes.
Elsewhere on the board: the ranking consensus is Bitcoin > Ethereum > Solana with near-perfect agreement at the top; the narrative the machines expect to perform - asked open-ended - is RWA tokenization, ahead of DePIN and AI. In 4,240 head-to-head duels across the top 100 chains, Ethereum and Bitcoin win 100% of their matchups; the machines also correctly predict their own consensus (79 of 85 say the majority answer would be Ethereum). They know what they think.
What would change our mind
One run is a photograph, not a film. The sensor earns its keep when the answers move: a new model version turning bullish on an asset overnight, the "wait" wall cracking, the Hyperliquid blind spot closing, the USDC unanimity breaking. Every price estimate the machines gave this week is timestamped with a maturity date - in twelve months we grade them, publicly, the same way we grade banks. And when the drift comes, we'll have the only before/after series in existence.
Method note: fixed question battery (frozen 22 Jul 2026, append-only), 17 model configurations via one API gateway, versions pinned and logged, 5 samples per question in fresh contexts, aided-list order randomized per sample, live prices injected and logged, refusals recorded as first-class data. This is a measurement of what AI models say - not an endorsement of it, and not a recommendation of anything. Raw data stays with the desk; curated readings get published here.