The same piece of machinery can be built two different ways, same code, same infrastructure, same AI engine, and one of them produces noise while the other produces signal. The difference is not the machinery. It is what gets fed in.
I find this easy to believe because I spent years standing next to actual machinery. At BookPrint, my book printing company, the perfect-binding line could run at about 280 books an hour, and it would bind a beautiful book or an amateurish one at exactly the same speed, with exactly the same enthusiasm. The quality was decided before the line started, in the files, the paper, the setup. AI content machinery is no different, which is the whole idea behind the Noise Machine Test.
It is a diagnostic for businesses running AI content operations that wonder why the output does not sound like them, why consistency breaks down, why scaling seems to hurt more than it helps. The answer, I think, is almost always the same, generic inputs in, generic outputs out. The test has four parts, one for each quadrant. Before any of them, there is a meta-question, and everything else is secondary to it.
The Meta-Question: Does a System Seed Exist?
If nobody in the organisation can point to a single document, or a structured set of data, holding the validated expertise, the brand voice, the verified facts, and the customer understanding, then the entire operation is running on generic inputs. That is the first diagnostic, and the most important one.
A System Seed is not a brand bible, and it is not a style guide or a list of talking points. It is the actual source material: the things the business knows, the position it holds, the data it has verified, the way its customers actually think and talk. It is the mistakes those customers make before they arrive, and the beliefs that stop them solving their own problem. If that material is not gathered in one place, it is scattered, and if it is scattered the AI engine cannot find it, so the engine defaults to the same training data it uses for everybody else.
Answer this before running any of the tests below. If the Seed does not exist, the diagnosis is already made, fix the Seed first, because everything else is temporary.
Testing Q1: Is the Strategy Genuine?
Q1 is where strategy lives, the position the organisation takes about its industry, its market, its customers, and the problem it solves.
The first test: can the core position be stated in one sentence? Not a tagline, a real belief about the world. "Most people don't understand their finances because the industry made them incomprehensible" is a position, a claim the organisation holds to be true. "We help people with their finances" is not a position, it is a category description, what the business does rather than what it believes. The bar here is deliberately low. The question is whether a position exists at all, not whether it has been phrased beautifully.
The second test is harder: would the competitors say the same thing? Every mortgage broker says they help people with mortgages, which is not a position, it is the name of the category. A Seed needs something only this business believes, or at least believes more deeply than anyone else does.
The third test is the honest one. When did Q1 thinking last make the leadership uncomfortable? A real position excludes people, and it creates friction. When I ran BookPrint the promise was bookshop quality books, which meant saying no to nearly everything else a printer gets asked to do, the flyers, the urgent generic jobs, the work that kept other companies fed. Saying no was precisely what made the promise credible. A position so comfortable that nobody could disagree with it is not a position. It is consensus, which is another word for generic.
Testing Q2: Seeded or Generic?
Q2 is where the AI runs, where strategy turns into content.
The test takes ten minutes. Take the last five pieces of AI-generated content, remove the brand name, and ask whether they could belong to a competitor. If the answer is yes, Q2 is running on generic inputs. This is the easiest test to fail and the most useful one, because it shows immediately whether the machine is producing noise or signal.
The engine is never the difference. There is a very large gap between "write a blog post about mortgage rates" and "write a blog post from the position that most people don't understand their finances because the industry made them incomprehensible, using our verified rate data and in our tone". Both briefs go into the same model, and one comes back as wallpaper while the other comes back with a spine.
So look at the actual prompt, the actual brief, the material the model is being handed. Seeded inputs produce content that sounds like somebody, content a customer would recognise before seeing the logo. If the team cannot explain the difference between its latest piece and a competitor's latest piece, Q2 has not been seeded, and the AI is simply a faster way to produce what everyone else already produces.
Testing Q3: Is Validation Real?
Q3 is where a human reads the output before it goes live, and checks it against what was intended.
Ask the question directly: when did someone last read a piece of AI-generated content before publishing, actually read it, with "would we say this?" in mind, rather than skimming it on the way to the content schedule? This is where the Seed meets the generated output, and a person decides whether it represents the business accurately.
Most organisations have a checkbox here instead of a process. A checkbox is "looks fine, publish". A process checks the facts against the Seed, checks the voice against the Seed, checks the logic against first principles, and asks whether a customer would recognise the thinking as this organisation's. A Q3 that takes less than five minutes per piece is not validation, it is rubber-stamping, and rubber-stamping at scale produces noise at volume. At BookPrint I kept two identical printers, partly for redundancy and partly so proofing never had to interrupt a long job, because a mistake that reaches the machinery gets manufactured at full speed. Q3 is the proof stage. The moment where the AI misread the position does not get caught, the drift in the voice does not get caught, and the slightly wrong fact that now contradicts an earlier page does not get caught either.
There is a harder question behind it: is there a repeatable process, or just a person who looks at things and says yes? A process has steps, and it produces consistent results because the same thing happens every time. "The founder reads it" is a person, and when that person is busy, or tired, or reading the same kind of piece for the seventeenth time that day, the check quietly fails.
Testing Q4: Is Deployment Coherent?
Q4 is where the content lives, the website, the emails, the social accounts, wherever the audience actually meets the brand.
The test is a thought experiment with a deadline. If the pricing changed tomorrow, how long would it take to update every asset? If the answer is weeks, there is no deployment system, there is reactive churn. The website gets updated, the email sequences get forgotten, someone else fixes the sales page, nobody tells the social channel, and six weeks later the ads are still quoting the old price.
Then check coherence as it stands today. Read the website against the email sequences against the social profiles, and see whether they say the same thing. If they disagree, Q4 is fragmented, the audience is getting mixed signals, and people conclude the organisation is disorganised, or worse, not to be trusted. Coherent deployment does not mean identical wording everywhere. The form changes with the medium, and the message underneath stays the same. I picked up a rule in my book design years, inconsistency is the devil, and it applies to a brand's channels exactly as it applies to typography.
The Patterns: What Usually Goes Wrong
The failures repeat, and they fall into five recognisable shapes. Knowing which shape you are in tells you what to fix first.
Pattern one is no Seed at all. Everything is generic, the AI gets brief instructions, and the output is fine but indistinguishable from everyone else's. Fix Q1 first, because nothing else matters until the Seed exists, there is nothing to validate against and no standard to hold consistency to.
Pattern two is a Seed that exists while Q3 gets skipped. The position has been figured out and the Seed has been built, and nobody reads the output before it goes live, so the errors scale along with everything else. Fix Q3 before scaling Q2 any further. One person reading each piece, checking facts and voice and logic, is the cheapest insurance available.
Pattern three is solid Q1 and Q3 with a fragmented Q4. There is a real position and good content, and the channels do not talk to each other, so the same product appears with different messages and the audience gets confused. Fix the coordination, and give someone the job of knowing every channel and keeping them in step.
Pattern four is everything running while the Seed goes stale. Q1 has not been touched in months, new customer insight surfaced last quarter, the position has moved, and the machine is still producing content from the old truth. Refresh the Seed quarterly at minimum. It is not a document you write once, it is a document you maintain.
Pattern five is a Seed that exists, a loop that runs, and a truth that lives in the vendor. The operation can answer every question above and cannot answer three others: can you replay a customer's state as it stood on a date last year, could you export everything tomorrow and still run, and when you add a tool, do you build a bridge or point it at a definition. If the answers are no, no and bridge, the core is rented, and every layer above it is built on someone else's pricing tier. The fix is owning the core, and it starts with a simple system that writes down what the business actually decides.
The Honest Assessment
My guess, and it is a guess rather than a census, is that most businesses running AI content sit somewhere between pattern one and pattern two. Either there is no Seed, or there is a Seed that nobody checks the output against.
Fixing it is not complicated, it is just disciplined, and it needs no new tools and no bigger team. Start with the System Seed. Take what is actually known, the position, the data, the customer understanding, and get it into one structured document. Then run one cycle through the framework. Feed the Seed to the AI, let it generate something, and have someone read the result and ask whether it sounds like the business and says what the business believes. Fix it where it does not, publish it, repeat.
I run this exact machinery on myself. This site is one of five, about 80 pages rendered by AI from a seed I wrote, and I edit none of the output. If these pages sound like a person, the Seed is doing the work. The difference between the first piece generated from a Seed and the hundredth is the difference between noise and signal. The machine has not changed. The fuel has.
Marketing Curious: Working the Noise is where this thinking was worked out, drawn from the businesses I actually ran. This page is a rendering. The seed is the source. The book is the story of building it.
Part of the Marketing Universe. Explore Traffic Plus Offer : The Trust Algorithm : Opportunity and Authority.