When the Measure Becomes the Target

In 1975, a British economist named Charles Goodhart presented a paper to the Reserve Bank of Australia. He was writing about monetary policy, about what happens when central banks start targeting specific financial indicators to control inflation. His observation was precise: "Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes."

He was not writing about SEO. He was not writing about Google. Google would not exist for another 23 years. But he described, with uncomfortable accuracy, everything that has happened to search since 2003.

The popular version of his principle is cleaner: when a measure becomes a target, it ceases to be a good measure. The moment you tell people what you are measuring, they optimise for the measurement. The correlation between the metric and the underlying thing you actually cared about breaks down. The indicator detaches from the reality it was supposed to track.

Goodhart identified a structural dynamic: any measure used to make decisions will attract pressure to game it, and once gamed, it stops measuring what it was meant to measure. Google has been fighting this dynamic for twenty years without being able to escape it.

Google has been living this problem for two decades. Every signal it has ever used to approximate quality has been reverse-engineered, systematised, and industrialised by an entire profession whose economic survival depends on doing exactly that. The updates, the penalties, the framework changes, the algorithm rewrites: all of it is a response to the same dynamic playing out over and over again. The measure becomes a target. The target gets gamed. The measure stops working. Google finds a new measure. Repeat.

This is not a story about bad actors, though bad actors feature in it. It is a story about an unwinnable structural problem, and understanding it properly changes how you think about every framework Google has ever published, including the current one.

One caveat before continuing. Google does not publish its ranking signals. What follows is a pattern argument built on 20 years of documented updates and the economic principle that explains them. The history is on the record. The framework is the author's.

2003: The First Target

PageRank was Google's founding insight. Links as votes. The web was a democratic system of citations, and a page that earned citations from other well-cited pages was, in principle, more authoritative than one that had not. The logic was sound. The measurement was elegant. And for a few years, it worked.

Then it became a target.

By the early 2000s, exact-match anchor text had become a primary ranking lever. SEOs understood that a link pointing to a page with specific keyword anchor text was a direct ranking signal, so they built links with specific keyword anchor text. Link farms, reciprocal link directories, mass link exchanges. The web graph that had served as a proxy for authority became a proxy for whoever could manufacture the most keyword-rich links at the lowest cost.

The Florida update in November 2003 was Google's first serious response. It targeted keyword stuffing, manipulative anchor text patterns, and doorway pages, and it hit hard enough that the backlash was severe. Many innocent sites were caught in false positives. The SEO industry, still young enough to be surprised, spent months working out what had changed.

Florida established the pattern. Google identifies a signal that correlates with quality. The signal becomes known. SEOs systematise it. The signal degrades. Google penalises the degraded version. New signals emerge. The cycle restarts.

2011 to 2014: The Content and Link Wars

The era that followed produced some of Google's most consequential updates. Each one targeted a different surface. Each one followed the same logic.

Panda launched in February 2011. Its target was content farms: operations that had worked out that Google rewarded pages matching keyword queries, so they produced vast quantities of pages matching keyword queries, with no regard for whether those pages were useful. Demand Media, the most visible offender, was producing content at industrial scale through its eHow and Livestrong properties. The measurement Google was using, topical relevance signalled by keyword presence and volume, had become the target, and the target had been farmed into uselessness. Demand Media lost $6.4 million in the fourth quarter of 2012 after successive Panda refreshes compounded.

Penguin followed in April 2012. If Panda addressed what was on the page, Penguin addressed what pointed to it. Link building had become link manufacturing. Purchased links, private blog networks, keyword-stuffed anchor text at scale. The correlation between links and authority, PageRank's entire premise, had degraded to the point where Google could no longer trust its own founding signal without layering algorithmic detection on top of it. Penguin initially impacted 3.1% of English searches and underwent ten updates between 2012 and 2016.

September 2012 brought the Exact Match Domain update. Domains like bestlaptops.com and cheapflights.net had been ranking well for their target keywords regardless of content quality, because the domain name itself was a ranking signal. Google confirmed the update affected less than 0.6% of English searches, a modest figure that understates the strategic disruption: an entire approach to site acquisition and niche targeting was rendered unviable overnight.

January 2014 brought perhaps the most quotable moment in the history of Google's public communications. Matt Cutts, then head of Google's webspam team, published a post titled "The decay and fall of guest blogging for SEO." His conclusion was unambiguous: "Stick a fork in it: guest blogging is done; it's just gotten too spammy." Guest posting had emerged as the link-building tactic of choice after Penguin. It became a target. It was gamed. Google killed it as a reliable signal.

The Repeating Structure

It is worth pausing on the pattern before continuing, because it is not incidental. It is structural.

Google cannot measure quality directly. Nobody can. Quality is not a number. It is an emergent property of a page being genuinely useful to a specific person at a specific moment, and that is not something you can quantify in a ranking formula. So Google measures proxies: things that correlate with quality in the absence of gaming pressure.

Links correlate with quality because, in an unmanipulated environment, people link to things they find valuable. Keyword relevance correlates with quality because pages that actually address a topic tend to use the words associated with it. Domain names that match a query correlate with relevance. Guest posts on reputable sites correlate with expertise. First-person language correlates with genuine experience.

Every one of these correlations holds until it becomes a target. The moment Google signals, through its guidelines or through observable ranking behaviour, that a particular factor matters, the SEO industry begins optimising for that factor independently of the underlying quality it was supposed to represent. The correlation degrades. Google patches. A new proxy is introduced. The cycle continues.

This is Goodhart's Law operating at web scale, with billions of dollars of commercial incentive driving the optimisation pressure.

Campbell's Law, articulated by the American sociologist Donald T. Campbell in 1979, adds a dimension: "The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more it will distort and corrupt the social processes it is intended to monitor." Two researchers, working independently in the same decade, arrived at the same conclusion about what happens when you try to govern complex systems with simple metrics. SEO is a case study in both.

2018 to 2022: The Expertise Arms Race

The years between 2014 and 2018 saw continued refinements: Hummingbird's semantic understanding reduced the value of keyword density as a standalone signal, RankBrain in 2015 introduced machine learning to query interpretation, and the industry adjusted accordingly. Then came the Medic update in August 2018.

Medic, named by the SEO community for its disproportionate impact on health and finance sites, was the first major enforcement of what Google called E-A-T: Expertise, Authoritativeness, and Trustworthiness. Sites in YMYL categories, Your Money or Your Life, that lacked credible authorship, transparent sourcing, and editorial oversight were heavily penalised. Some sites in health and finance lost 40 to 80% of their visibility overnight. The message was clear: Google was trying to measure whether content came from people who actually knew what they were talking about.

The SEO industry's response was predictable. Author bios multiplied. Credential sections appeared. About pages were rewritten to hit every trust signal Google's quality raters were trained to look for. The measure became a target before the ink on the framework was dry.

In December 2022, Google added the second E, for Experience, upgrading E-A-T to E-E-A-T. First-hand, lived experience was now explicitly part of the framework. A review written by someone who actually used a product should outrank a compilation of other people's reviews. The intent was sound. The problem was that first-hand language is not the same as first-hand experience, and the former is trivially producible.

2022: The Framework Becomes a Checklist

The Helpful Content Update, launched in August 2022 and integrated permanently into Google's core ranking system in March 2024, was a direct response to scaled content production gaming topical relevance signals. Google's stated aim for the March 2024 integration was to reduce low-quality, unoriginal content in search results by 45%. That figure is a goal, not a measured outcome. The distinction matters, and not just for accuracy: it reflects how difficult it is to quantify what the update was trying to address.

E-E-A-T gaming is already well-established. Author bios written to specification rather than from biography. Credential claims that cannot be verified. First-person language added to content that was not written from first-hand experience. About pages optimised to hit every trust signal Google's quality raters are trained to look for. The same dynamic that turned PageRank into a link farm and keyword relevance into a content farm is now turning E-E-A-T into a compliance checklist.

Tom Capper, Search Science Lead at Moz, identified a likely mechanical trigger for the Helpful Content Update penalties in his analysis of the 2023 and 2024 updates. His data, summarised via Hobo Web, found that sites penalised by HCU consistently had high Domain Authority relative to Brand Authority: strong link profiles without commensurate real-world recognition. The ratio of approximately 2:1 appeared to flag sites as artificially amplified. This is external pattern analysis from observed ranking behaviour, not a confirmed Google signal. It is the best available hypothesis from the available data. It is also, now that it is known, a target if it is right.

Google Is Not Playing a Winnable Game

The important thing to understand is that Google is not incompetent for letting this keep happening. It is playing a structurally unwinnable game, and it has been playing it honestly.

Any signal transparent enough to be used as a ranking factor is transparent enough to be reverse-engineered. The moment Google publishes its quality guidelines, or the moment ranking behaviour reveals what it is rewarding, that information becomes available to everyone, including people who want to manufacture the appearance of quality without producing it. There is no version of a public ranking system that does not have this property.

The alternative would be a ranking system nobody knows anything about, which creates its own problems around accountability, transparency, and the ability of legitimate publishers to understand how to be found. You cannot have both: a system transparent enough to be usable by good-faith publishers and opaque enough to resist gaming by bad-faith ones. Goodhart's Law does not allow it.

What Google can do, and does, is raise the cost of gaming. Penguin made link manufacturing expensive and risky. Panda made content farming economically unviable at the previous margin. Helpful Content attempted to do the same for scaled content production. Each intervention pushes the cost of appearing to have quality higher, even if it cannot make that cost exceed the cost of actually having quality. The gap narrows over time. It does not close.

The AI Acceleration

AI has changed the economics of this dynamic in a way that deserves its own section.

Every previous iteration of the cycle was constrained by the cost of producing the gamed signal. Link manufacturing required infrastructure and coordination. Content farming required writers, even cheap ones. E-E-A-T manufacturing required believable author profiles and at least some effort toward the appearance of expertise. These costs acted as a partial brake on gaming pressure.

Generative AI has removed most of those brakes. Content that includes first-person language, that structures itself around experiential claims, that hits the topical depth signals Google rewards, that maintains a consistent author voice across hundreds of pages, is now producible at effectively zero marginal cost. The E-E-A-T framework was designed in an environment where manufacturing convincing expertise signals was expensive. It is being deployed in an environment where it is not.

This does not mean E-E-A-T is worthless. It means it is earlier in its Goodhart cycle than most practitioners recognise. The Google API leak of 2024 suggested the existence of author-level scoring signals — though the leaked data was widely reported and widely misread, and what it actually confirms remains contested. If those signals exist in the form described, the next response will target entity-level trust rather than page-level signals. The cycle does not end. It moves to a new surface.

What Resists the Cycle

If every signal gets gamed eventually, the natural question is whether there is anything that does not.

The answer is not nothing. But it is narrower than most SEO strategy accounts for.

Genuine differentiation is harder to manufacture than any signal Google has ever tried to measure. Not because Google will always detect the difference, but because real expertise produces outputs that behave differently in the world. Real expertise generates citations from other practitioners who have enough knowledge to recognise what they are citing. It produces content that people share in private Slack channels and Discord servers that AI cannot index. It earns links from people who found something genuinely useful, without being asked.

These signals are not Goodhart-proof. Nothing is. But they are more expensive to fake than any single on-page or off-page factor, because they require producing something that knowledgeable people find valuable. That bar keeps rising as AI lowers the cost of everything else.

Every tactic that can be reduced to a checklist will eventually be reduced to a checklist and then gamed into obsolescence. The E-E-A-T checklist, the topical authority framework, the structured content models: all of them will follow the same arc as keyword density, PageRank sculpting, exact match domains, and guest post outreach. Not because Google will fail to respond, but because response and gaming are two sides of the same cycle. Goodhart describes a law, not a tendency.

Reading the Current Moment

The Goodhart framework gives practitioners something more useful than a history lesson. It gives them a position in a cycle.

E-E-A-T is mid-cycle. The signals are known, the gaming playbook exists, and AI has dramatically compressed the timeline. Google's response will likely move to author entity level, targeting the ratio between real-world brand recognition and accumulated link authority, or some successor to the mechanism Capper identified. The API leak confirms this direction of travel.

AI citation signals are pre-cycle. Being referenced in AI Overviews is currently a meaningful visibility advantage, the mechanisms are not yet fully understood by practitioners, and gaming pressure is minimal. This is approximately where anchor text was in 2001: a signal that matters, that has not yet become a target, that will.

Brand authority signals, navigational search volume, real-world entity recognition: these are later-cycle. They have survived longer as reliable signals because they are harder to manufacture than link profiles or content depth. They will not survive indefinitely either.

Knowing where you are in the cycle changes what you should be investing in. Tactics at peak gaming pressure have short and declining returns. Signals that have not yet become targets have longer windows. And the only genuinely durable strategy the framework identifies is producing something worth reading by people who know enough to tell the difference.

Charles Goodhart was writing about interest rates. He described search engine optimisation more precisely than most people who have spent careers doing it.