A search engine does not read. It measures. The distinction matters because reading produces a judgement about whether a piece of writing is any good—whether it is accurate, whether it is clear, whether it earns the attention it is asking for.
Measuring produces a score based on indicators that correlate, in the aggregate, with the kind of content that human readers tend to find trustworthy. The correlation is real. The indicators do track something. But what they track is the signal of trustworthiness, not trustworthiness itself, and the gap between the two is where most of what is interesting about the system can be found.
The indicators are well documented. Regular publication suggests an active, maintained source. Inbound links suggest that other sources found the content worth referencing. Structured headings and appropriate keyword density suggest that the content is organised around a topic rather than assembled from whatever was available. Page load speed suggests that someone has invested in the technical infrastructure that serious publishers typically maintain. Author credentials, where visible, suggest expertise. These are all reasonable proxies. A source that publishes regularly, attracts links, organises its content clearly, and is maintained by credentialed authors probably is, on average, more reliable than one that does none of these things.
The problem is not the proxies. The problem is what happens when the proxies become the goal.
A proxy is a measurement that stands in for something harder to measure. It works when the relationship between the proxy and the underlying thing is stable—when the signal reliably tracks the substance it is supposed to approximate. The relationship is never perfect, but if it is consistent enough, optimising for the proxy produces something close enough to optimising for the underlying thing to be useful. The signal does the work of the substance because the two move together.
The relationship destabilises when people learn that the proxy is being measured. At that point, optimising for the proxy becomes independent of the underlying thing. The person who wants to rank well in search results does not need to produce trustworthy content. They need to produce content that exhibits the indicators that the algorithm associates with trustworthy content. The indicators can be produced without the trustworthiness.
Regular publication can be achieved by producing content on schedule regardless of whether there is anything worth saying on schedule. Inbound links can be cultivated through exchange arrangements that have nothing to do with the quality of what is being linked to. Structured headings and keyword density can be optimised by writing around a structure rather than writing the structure from the substance. Author credentials can be attached to content the credentialed author did not write and has not reviewed.
Each of these is a signal without the substance that the signal was designed to approximate. Each of them, in isolation, might be explainable. The system cannot distinguish between them and the genuine article because the system is not reading. It is measuring. The measurement cannot see the difference between a heading that emerges from the logic of the content and a heading that was inserted to satisfy the algorithm’s structural preferences. Both produce the same signal.
This is the non-falsifiability Vorpel’s formulation identifies. The heuristics cannot be falsified because they do not make falsifiable claims about the content. They make claims about the content’s structure, consistency, and connectivity—claims that can be satisfied by manipulating the structure, consistency, and connectivity independently of the content’s substance. A claim that can be satisfied without reference to the thing it is supposed to measure is not a claim about that thing. It is a claim about the signal. The signal can be produced. The thing it was supposed to track need not be.
The human judgement that the system is trying to approximate would, in principle, catch this. A reader who examines a regularly published, well-linked, clearly structured source and finds that the content is inaccurate, repetitive, or assembled from other sources without synthesis will revise their assessment of the source’s trustworthiness. The judgement is self-correcting because the judgement is about the underlying thing rather than its indicators. The signal informs the initial assessment. The substance corrects it.
The algorithm cannot make this correction because it does not encounter the substance. It encounters the signals and records their presence or absence. The substance is behind the signals, accessible only to the reader who engages with it. The algorithm does not engage. It scores. The score reflects the signals. The substance does not affect the score except through the signals it produces—which means that content optimised to produce the signals without the substance will score as well as content that produces the signals through the substance. The scores are identical. The content is not.
The same structure appears in human systems wherever the measurement of a proxy has become the primary mechanism for assessing an underlying quality that is harder to measure directly. The performance review that measures attendance, output volume, and response times rather than the quality of judgement the person brings to their work. The academic citation count that measures how often a paper has been referenced rather than whether what it says is worth referencing. The customer satisfaction score that measures how the interaction felt immediately after it ended rather than whether the underlying problem was resolved. The compliance audit that measures whether the procedures are in place rather than whether the procedures are working.
In each case the proxy was chosen because the underlying thing it approximates is genuinely difficult to measure. Quality of judgement, worth of ideas, resolution of problems, effectiveness of procedures—these require engagement with substance, which is slow, expensive, and not easily scaled. The proxy is fast, cheap, and scales without difficulty. The proxy is a reasonable approximation in the early stages, when the relationship between signal and substance is still intact and the people being measured have not yet learned to optimise for the signal independently.
The relationship degrades as the measurement becomes known. People learn what is being measured. They produce what is being measured. The production of the measured thing becomes the work, alongside or instead of the underlying thing the measurement was designed to track. The performance review produces attendance and output volume. The academic incentive structure produces citations. The customer satisfaction score produces interactions that feel good immediately after they end. The compliance audit produces procedures that are in place. Whether the underlying things are present is a question the measurement system has stopped asking, because the measurement system can only ask about what it measures.
The search engine’s problem is a version of this at enormous scale, with an additional complication: the system’s heuristics are not fully disclosed. The signals being measured are partially known and constantly changing, which means the optimisation is always chasing a moving target while the target itself remains in the same place. The underlying thing—content that is genuinely accurate, useful, and worth a reader’s time—does not change. The signals the algorithm uses to approximate it do change, as the algorithm is updated to address the most recent forms of signal manipulation. The update addresses the manipulation. It does not address the underlying relationship between signal and substance.
The result is an escalating elaboration of proxy measures, each one designed to close the gap that the previous measure opened when it became known and manipulable. The elaboration is real work. The algorithm is genuinely attempting to approximate human judgements of trustworthiness more accurately. But each new measure is still a proxy, still a signal, still vulnerable to optimisation by people who learn what is being measured and produce the signal without the substance.
The gap between signal and substance is not closed by adding more signals. It is obscured by making the signal complex enough that optimising for it becomes expensive—expensive enough that the cost of genuine production and the cost of signal production converge, which is a different condition from the signals actually tracking the substance.
What both systems share—the search engine’s content heuristics and the human performance measurement systems—is the replacement of judgement with measurement, and the assumption that sufficiently refined measurement can substitute for judgement indefinitely. The assumption has a logic to it. Measurement scales. judgement does not. An algorithm can assess millions of pages. A human reader cannot. A performance system can score thousands of employees. A manager who knows each one well enough to assess their judgement cannot manage thousands. The measurement is necessary precisely because the judgement cannot be applied at scale.
But the measurement is not the judgement, and the signals the measurement tracks are not the substance the judgement would assess. The gap does not close as the measurement becomes more refined. It changes shape. A more refined measurement produces a different set of signals that can be produced without the substance. The signal becomes more elaborate. The substance remains what it was.
The guarantee is still there at the bottom of the search results page, promising that the content has been assessed for trustworthiness. The assessment has been performed. The assessment measured the signals. The signals were present. The substance was not tested, because the system does not test substance. It scores signals.
The word guarantee was stripped of its teeth before the page was ever written.
The teeth were the judgement.
The algorithm does not have teeth.
It has metrics.
