Machines describe you.
Whether or not you’ve told them how.
We build the dual identity — distinctive enough for a person to remember,
structured enough for a machine to get right.
the shift
One of your two audiences
can’t see your logo.
There’s a finding in creative testing that almost nobody has absorbed. When generated advertising is tested against the norms, it performs well on emotional response — often better than the average human-made ad. Most people don’t spot it. So the idea that audiences simply reject this work is not true.
But the same testing shows it consistently underperforms on distinctiveness and brand recognition — the two things that build memory over time. A model is trained to produce the middle of everything it has seen. A brand is, by definition, not the middle. Which means brand codes get more valuable as output volume rises, not less.
Brands are finding themselves described by AI assistants using products they discontinued years ago and prices that are long gone. You don't fix that inside the model. You fix it at the source: two audiences, two artefacts, one strategy.
Challenge
one
“An AI assistant describes us using a product we discontinued years ago.” We find the sources it’s actually reading, correct them at origin, and maintain a fact base so there’s something current to retrieve next time. Entity architecture Machine fact base Source correction
Challenge
two
“When AI compares us to competitors, it gets the comparison wrong.” Comparison sets are assembled from whatever is retrievable. If your category position and differences aren’t structured, the model infers them. Category definition Disambiguation Comparative positioning
Challenge
three
“Everything we generate looks like everything everyone generates.” Brand codes, written as constraints rather than as a mood board. Specific, unusual, and enforceable on a system that will otherwise smooth them away. Distinctive assets Brand codes Creative constraints
Challenge
four
“Our guidelines are a sixty-page PDF and nobody applies them at volume.” Guidelines become machine-enforceable constraints plus a feedback loop, where every approval and rejection teaches the system. Brand model Governance Feedback design
Challenge
five
“We have no idea what a model would say about us if asked.” Sampled across the major assistants, asked enough times to mean something, with the sources each answer drew from. Uncomfortable and productive. Perception audit Model sampling Source mapping
Structured interventions
What we build.
The way in to this capability is the Perception Audit. Everything below is what happens after it — defined, bounded commercial engagements.
Machine Perception Audit
The way in. What the major assistants actually say about you today, sampled enough times to be meaningful, with the sources behind each answer and a plain comparison against what you believe your positioning to be. Deterministic where it can be, statistical where it can’t, and always shown with the working.
Inquire about this scopeDual Identity Definition
The core deliverable. Two artefacts from one strategy: a human-facing positioning and expression system, and a machine-facing definition of who you are, what you sell, what you are not, and what you’re comparable to — written as structured, retrievable, unambiguous fact.
Inquire about this scopeDistinctive Brand Codes
The specific, defensible assets — visual, verbal, sonic, structural — that a generative system must reproduce and must never average out. Written as constraints a machine can be held to.
Inquire about this scopeEntity & Source Architecture
Disambiguation across the sources a model actually reads. Consistent references, structured data as hygiene, first-party fact sources, and a homepage rebuilt as a definition of the organisation rather than a brochure for it.
Inquire about this scopeMachine Fact Base & Experience Build
A maintained, structured source of truth at brand level: positioning, policies, claims, comparisons, category definition, corporate facts. Sites and digital experiences built to the dual standard: distinctive for a person, and structurally legible for a machine.
Inquire about this scope
where we draw lines
Three things
we won’t do here.
We build machine identity to reflect reality, not to manipulate it. These are the three lines we don’t cross.
Encyclopaedia entries.
REPUTATIONAL & POLICY INTEGRITY
Paying for or writing your own entry on a community-edited reference site carries real policy and reputational risk, and it matters far less to AI answers than people assume.
We’ll help you become the kind of organisation somebody writes about independently. We won’t write it ourselves.
Selling structured data as a citation lever.
TECHNICAL TRUTHFULNESS
It isn’t one. Structured data is genuinely useful — for disambiguation, for classic search results, for making a page parseable — and we use it for exactly that.
We’d rather tell you which job it does than let you believe it does a different one.
Manufacturing mentions.
SUBSTANCE OVER METRICS
There’s a version of this work that means placing real stories, real data and real expertise where they’ll actually be read. And there’s a version that means creating mentions whose only purpose is to be counted.
The test we apply: would this still be worth publishing if no model ever read it? If not, we don’t do it.
our standard
If a tactic wouldn’t stand up in front of your customers or your board, it doesn’t belong in your machine identity.
We build systems designed to compound over years, not tricks that break in the next model update.
frequently asked
Questions
& answers.
What leaders usually ask when establishing their dual brand and machine identity.
Often not. In most cases the human-facing brand is fine and the machine-facing one simply doesn’t exist. We build the missing half and tighten the codes on the existing half.
You do, and it’s in the contract. If you leave, it leaves with you — the reference sets and the accumulated approval history included. We think that should be standard. It currently isn’t.
Distinctiveness can be measured. We baseline recognition and attribution before, and re-test after. If the codes aren’t holding at volume, that’s a finding, not a failure.
Yes, and often that’s the right shape. They hold the human half. We build the machine half and make sure the two say the same thing.
next step
Find out
what the models say about you.
Two to three weeks, fixed fee, and the sources behind every answer.