We wrote our own rules.
Following them led somewhere
that already had a name.
Nobody at Cast Iron LA had heard the term GEO. Stage30 was not built to enter a category. It was built by following a writing discipline further than was comfortable, and the place it arrived at turned out to be occupied.
We came to this from advertising
Almost everyone selling AI search visibility today learned search first. They arrived at the language model from twenty years of ranking pages, and they are troubleshooting it from where search left off. Keyword optimization became citation optimization. Backlink strategy became authority signals. The mental model never changed, only the vocabulary did.
Cast Iron LA is an advertising agency. John Otto Barbush and Luis Emilio Ramirez have been partners for twenty three years, most of them at RPA on Honda, Acura, Discovery and Hilti. Neither of them had a search framework to protect, which meant nothing had been archived and everything was still a variable.
A golfer of twenty years slices the ball and works on everything except the grip, because the grip was solved in year one and stopped being a variable. The problem space quietly shrank and nobody told him. The answer to AI search was sitting in the part of the map the field had already cleared.
We learned by designing, not by studying
Most method in this category is inferred backwards from watching search behavior. The Stage30 canon came from building an instrument and watching it break.
Every rule in the canon has a defect attached to it. A real page, a real run, a real thing that went wrong on paper before anyone caught it. That is why the rules are specific enough to test, and it is why there are 663 of them across 22 documents rather than a dozen principles on a slide.
The engine that enforces them runs to roughly twenty one thousand lines, and it was written by a copywriter who had never shipped software. That is not a claim about the code. It is the reason the order came out right. The canon was written first, to be true. The software was written second, to enforce it. Almost everywhere else it happens the other way around, and the methodology quietly bends to fit whatever was easy to build.
What we did not invent
None of the underlying mechanics are secret. Tokens, embeddings, next word prediction, and the reason structured writing dominates a training corpus are published science, and any machine learning engineer can walk you through them. Stage30 does not claim to have discovered any of it, and a reader should distrust anyone who does.
What had not been done was connecting that science to how a brand writes a sentence. Language models were trained on legal briefs, medical journals, patents, technical documentation and encyclopedias, which all share one skeleton: claim, then evidence, then conclusion. Repeated at that scale, structure becomes the shape a model reaches for. To a language model, human means structured, which makes the standard advice to write more conversationally for AI precisely backwards.
The science sat in one field and the writing sat in another. The people paid to close that gap were busy optimizing keywords. Stage30 recognized a structural constant in how humans communicate, matched it to how machines collapse, and formalized it into something that can be taught, measured and enforced. That is the difference between an observation and a methodology.
This was never a search problem
Making one message survive every surface it has to live on is the oldest discipline in advertising. A sixty second spot becomes a thirty, then a fifteen, then a six, then a banner two hundred pixels wide, then a thumbnail on a phone held at arm's length. The campaign was never the sixty. The campaign is whatever is still standing at one second.
A creative team that has done that for twenty five years has one instinct burned in above all others: find the part of the message that cannot be cut, and build everything else so that part survives the cut.
The surfaces changed. Four engines now read four different indexes. A question fans out into a dozen a buyer never typed. An answer arrives as two lines with no link attached, or spoken aloud with no screen at all, or as a chunk of a page pulled out and read cold. A search firm learned to get a page ranked. Stage30 came from learning to keep a message intact when it gets cut to one second by someone who does not care about it. The answer layer is the same compression problem at a new size.
What is built, and what is not
Stage30 publishes what it has measured and marks what it has not. Of the 663 canon rules, 483 are enforced in code, each one wired to the engine function that implements it, and the build fails if that stops being true. 102 are doctrine: definitions, vocabulary and case records that were never going to be code, and they are excluded from the build list rather than counted as work nobody has done. The rest are marked backlog, or marked as needing an input Stage30 does not yet have, and the classification is public in the canon rather than implied.
The instrument reads thirteen surfaces, from on-page writing and semantic HTML through the knowledge graph, institutional authority, community, reviews and the entity architecture that ties them together. It scores five weighted inputs and reports Resolution as their output, never as an input to itself. It reads both the answer an engine gave and the sources that answer was built from, and it names the sources.
Repeated measurement of one client domain across 35 dated runs on the same pinned question produced readings from 32 to 70, which is why a Stage30 reading is the median of repeated runs measured against that brand's own noise floor rather than a single draw. A methodology that logs its own open contradictions, dated and unresolved, is the only kind worth believing.