It's running right now. An autonomous station at eigentrace.ai is ingesting live news feeds, measuring what five frontier AIs drop from every story, and broadcasting the findings — scripts, cover images, and titles generated unattended, around the clock, since April 2026. Go watch it work, then come back and read what it becomes when it's pointed at your client's industry instead of the news.
By Sean Adams · Boston · SEO at Wayfair, AMP Agency, and Simply Business · builder of EigenTrace
Every industry produces a river of content — trade press, competitor blogs, analyst notes, arXiv releases, regulatory updates. Everyone in your client's space reads the same river and writes the same takes. The consensus middle gets restated ten thousand times, and AI summarizers flatten it further.
This engine does the opposite. It ingests the river and measures its negative space: the concepts the coverage orbits but never names, the claims every summary dropped, the specifics that got softened into mush. Then it generates commentary anchored to those measurements — anti-editorial takes built on what the industry's content omitted, each one citing the measurement that found the gap.
Your client's competitors' blind spots become your client's content. Not their words — their omissions. That's transformative by construction, original by definition, and it compounds: every week the industry publishes, the engine finds new gaps to own.
The engine's own research is the strategy. Months of measurement across five frontier models produced one finding that matters commercially: AI summarizers drop the generic and keep the specific. The consensus middle — the content everyone writes — is exactly what gets erased when an AI answers a customer's question. First-hand analysis, contrarian angles, claims nobody else makes: that's what survives.
Anti-consensus content isn't just a brand voice. It's engineered to be the survivable kind — which means the same pieces work twice:
This isn't a theory of survivability. In the instrument's published demo run, a concept the source document never contained — surfaced by the geometry, woven in by one disciplined rewrite — was carried into eight of ten independent AI summaries of the result. A word entered the model-mediated version of a topic where, before, it did not exist. That is territory, measured once, in public, with the receipts one click deep.
One machine, two fronts — and the second front is the one almost nobody is fighting on yet.
The inputs: your client's river — trade publications, the competitor blogroll, arXiv categories relevant to their space, regulatory feeds, earnings coverage. The ingestion layer is feed-agnostic today; swapping the news feeds for an industry's feeds is retargeting, not rebuilding.
The measurement: the same instrument running the public broadcast — void detection, the VF-IDF omission ranking, dual geometric derivations of the concepts coverage circles but won't name, claim-level survival scoring. Every take the engine produces starts from a measurement, and the measurement is citable in the piece. And the vocabulary the instrument speaks its findings in is a declared parameter, not a constant: the public broadcast reads out against a 253,000-entry general dictionary; your deployment reads out against vocabulary derived from your client's own industry corpus — so the gaps it names are the industry's gaps, in the industry's words.
The outputs, chosen per client:
The content isn't the moat. The instrument is. Anyone can prompt an AI to write industry takes; the output is the consensus middle wearing a trench coat, and the AI layer flattens it accordingly. This engine's takes are downstream of a measurement system — deterministic, reproducible, publicly documented, with its own withdrawals page — and a competitor who wants to match the content has to first match the measurement.
Before this page shipped, the pitch was reviewed cold by a frontier model with no access to the code, the archive, or the receipts. It set three conditions for taking any system like this seriously: a dataset competitors can't copy, a long-term public evaluation record, and methods that survive scrutiny. Fair conditions. In order:
Roughly 26,000 measured segments across some 5,400 classified stories — five frontier models, fifteen measurement layers, timestamped, spanning multiple silent model retrains. Every retrain destroys the previous behavioral surface: there is no way to go back and measure how a spring-2026 model summarized live news, because that model no longer exists. A competitor who clones the pipeline tomorrow starts collecting — tomorrow. The methodology is public. The past is not for sale.
Pre-registered tests with the nulls committed before the run — the entity-swap control landed at p = 0.0085, d = 0.47. A standing withdrawals page of claims this project killed itself, each with the control that killed it. And a prediction loop that scores its own forecasts on air, misses included. The instrument that never delivers bad news cannot be trusted with good news.
VF-IDF — the omission-ranking metric — is named, formally defined, deterministic, and verified on two independent channels; a complete run is published with its receipts: five frontier writers under an evidence-ledger discipline, a five-judge panel, and a mechanical claim audit in which every positive fact traced — thirty of thirty, panel-wide — and every contest is quoted with its evidence. And when a pre-registered ablation showed most of the original synthesis objective was decoration, the decoration was retired in public and three interpretable terms shipped. What's left is only what survived.
The demo is not a deck — it's two live artifacts. The station has run continuously since April 2026: unattended ingestion, measurement, script generation, image generation, self-auditing, publication — through model updates, hardware contention, and the ordinary chaos of live news. Warts and all, on purpose: if you want to know what an autonomous content system actually looks like after months of real operation rather than a demo reel, it's live right now. And the instrument page is one complete run frozen with its audit trail — the same measurement engine that produces the AI Visibility Audit for agencies. One instrument, two commercial surfaces.
The feed swap is genuinely light — the pipeline doesn't care whose river it drinks. The client branding, voice tuning, scheduling blocks, and title strategy are a scoped build on top, and the editorial layer deserves human review before anything carries a client's logo: the engine's takes are labeled interpretation anchored to measurement, and the measurement is the part that's bulletproof. I won't pretend the commentary layer is. That division — measured versus argued — is the house rule of the whole project, and it ships with the product.
This is a build plus operations relationship, scoped per client: the retargeting and branding build up front, then a monthly operations retainer for the running channel — feed tuning, title strategy, editorial review cadence, and the quarterly measurement of what territory the content has claimed. If your agency wants the capability fully in-house instead, that's the practice-build conversation, and I'm glad to have it.
Start with the demo you don't have to schedule: watch the station, pick the client whose industry has the loudest consensus and the biggest blind spots, and write me.
eigentraceproject@gmail.com → point it at an industryMeasured: the archive, the survival numbers, the audits, the withdrawals. Argued: that a channel built on an industry's measured omissions wins search and AI-layer territory. Plausible, untested — nobody has that outcome data yet, because nobody else is reading the layer. Measure first.