AI Optimization
What we refuse to score
The measures deliberately left out, and why leaving them out is a scoring decision as significant as any included one.
Why an exclusion list matters
Every scoring model is a claim about what matters. The included signals get argued over; the excluded ones are usually invisible, and they shape the result just as much. Publishing what was deliberately left out is part of publishing the method.
Anything we cannot observe
Internal ranking signals, personalisation, and whatever a system does that has not been disclosed. These plainly affect outcomes and cannot be measured from outside. Inferring them from results is speculation, and scoring a speculation produces a number that looks like evidence.
Anything that moves on its own
A signal that returns a different value on two identical runs cannot be scored as a fact about a business. Where a variable signal is genuinely the thing being measured, it is sampled repeatedly and reported as a rate with its sample size, never as a single value.
The consistent principle: a score should only contain things that would be the same tomorrow if the business did nothing. Everything else belongs in a trend line, or nowhere.
Anything a business cannot influence
Category competitiveness, market size, and the general prominence of an industry all affect outcomes and none is a property of the business. Including them produces a score that rewards being in an uncrowded field, which tells an operator nothing they can act on.
These belong as context alongside a result, not inside it.
Anything that rewards volume for its own sake
Page count, publishing frequency, word count, and number of citations without regard to their independence. Each is trivially inflatable, and scoring them would reward exactly the manufactured-content behaviour that AI optimization is supposed to identify.
Anything we would have to guess
Where a signal cannot be read, the result is unmeasured, named, and excluded from the total rather than scored as zero. A business that scores 60 out of a possible 80 with 20 points unmeasured has been told something useful. One scored 60 out of 100 has been told something false.
The thing this list is really for
An exclusion is where a thumb on the scale is easiest to hide, because nobody audits an absence. Writing the list down is the only way it stays honest, and it means an addition to the model has to be argued for in public rather than made quietly.