The Google leak crossed with Anthropic’s open code shows that ranking penalties are increasing. A single weak signal is enough to bring down a site.
When a site loses almost all of its visibility overnight, we look for the single cause. This is a model error. By crossing the 14,014 attributes of the ContentWarehouse leak with a real case, a regularity emerges: the penalties do not add up, they multiply. Here’s what this changes for your audits.
A site loses a large part of its traffic. The audit begins. We are looking for the cause.
A cause. In the singular.
This is the reflex of almost all SEO audits: isolate the culprit, the main problem, the lever that explains everything. We reason in addition. We pile up grievances: speed here, toxic backlinks there, duplicate content elsewhere. We totalize, we classify, we prioritize.
The Google leak from May 2024 shows that this reasoning misses the real mechanics. Google penalties do not add up. They multiply.
And when factors multiply, a single flaw is enough to collapse everything.
The additive reflex, and why it breaks down
When I cross-referenced the 14,014 attributes in the ContentWarehouse leak with Anthropic’s open source claude-code repo — 18 submodules documented in its PARITY.md file — a pattern emerged. Both systems describe quality in the same way: not a grade, a product of factors.
The additive model that most audits employ looks like this:
Score = baseline − (penalty A + penalty B + penalty C)
Intuitive. Fake.
Its weakness is seen in an extreme case. Take a site with a severe quality problem on just one dimension. In addition, you remove a share, the site goes down, but it remains in the race. The model predicts comfortable residual traffic.
The observed reality shows sites which do not go down: they fall. Suddenly. On a step.
The additive model cannot produce this walk. The multiplicative model, yes.
A string, not a stack
Here is the analogy that makes the difference obvious.
The additive model imagines quality as a pile of chips. Each good point adds a token, each penalty removes one. If the pile loses a few tokens, it remains high. You can remove a lot before the pile disappears. This is reassuring, and how most people think about their SEO.
The multiplicative model imagines quality as a chain. Each dimension is a link. The strength of the entire chain is never the sum of the links. This is the weakest link.
A chain with many steel links and a single cardboard link does not hold nearly its full load. She holds what the cardboard holds. Almost zero.
This is exactly how a product behaves. Multiply together factors all close to 1, but slightly lower, and repeat the operation on many dimensions: the product slides gently downwards, without any factor being catastrophic. Gentle accumulation. Now slip a single factor very low into the product. The whole result plunges with it. A link. The whole chain.
The battery is forgiving. The channel is unforgiving. The leak describes a channel.
The HouseFresh proof: 0.09
HouseFresh is the case that settles the debate. An air purifier testing site, collapsed in the SERPs, then in recovery. I looked for which formula reproduces his actual score.
The multiplicative formula:
Q = baseline × ∏(1 − δ_k)
Each δ_k is a degradation on a quality dimension: δ₁ for the first, δ₂ for the second, and so on. We do not subtract the damage. We multiply what remains after each: (1 − δ₁) × (1 − δ₂) × (1 − δ₃)…
With this formula, the HouseFresh score drops to 0.09. A walk, exactly as observed.
The additive model applied to the same degradations predicts a zero or negative score — a site that no longer exists. HouseFresh was still around and going back. Only the multiplicative corresponds to reality.
The difference is not cosmetic. Here is one landmark, and one only.
Purely illustrative values, not from the study: if three dimensions each lose half of their strength, the multiplicative model gives 0.5 × 0.5 × 0.5 = 0.125 — a damaged but alive site. The additive model subtracts three times 0.5 and falls below zero – a dead site.
On the ground, this is the first behavior we observe. The second does not exist.
And the 0.09 says something else valuable: there is a way back. This score is not a zero. Lift up the postman who pulled it down, and the product rises mechanically. HouseFresh’s recovery is not a miracle. It’s reverse arithmetic.
38 quality dimensions, not a grade
The leak names a specific structure: compressed_quality_signals. Inside, 38 distinct dimensions of quality.
Google does not calculate “a” quality score. It calculates 38, on different axes, then combines them. And the combination is multiplicative.
The word “compressed” in the name is not insignificant. These 38 signals are condensed to fit into a structure that can be used across the entire index. Google stores this quality product on each document, permanently, ready to be weighted at the time of ranking. Quality is not recalculated on the fly. It is carried by the document, like a label with 38 boxes.
These 38 axes do not measure the same thing. Some look at the content itself. Others look at the site that hosts it. Still others look at the author, the entities cited, the thematic coherence, the historical reliability of the field. Each axis is a different question asked of the same document. And the final result multiplies the 38 answers.
The system that orchestrates this has a name in the leak: NSR, Neural Site Ranking. It is based on 17 embedding modules — 17 ways to represent and compare the content, the author, the site, the entities. It’s an architecture, not a cursor. Each module transforms an aspect of the site into a vector, then these vectors feed into the evaluation of dimensions. Seventeen parallel representations of the same site, which converge towards the quality product.
Crossing with the Anthropic repo gives a measurable parity point: on operator OP_11, the correspondence between Google logic and Anthropic logic reaches 0.63. Two systems built by two different companies converge on the same weighting mechanics. This is not a coincidence of vocabulary. It’s a convergence of engineering. When two teams who don’t talk to each other reinvent the same structure, it’s usually because the structure is the right answer to the problem.
38 dimensions multiplied together: this is why a site can be excellent on almost all axes and collapse when only one falls to the lowest. A multiplicative factor close to zero pulls the entire product down. This is the weak link, translated into mathematics. And this is also why great content doesn’t save a site whose domain reliability has collapsed: the big factor is canceled out by the small one.
Three audit decisions that change
If the penalties multiply, the audit no longer runs the same.
1. Look for the lowest factor, not the sum of the problems. In addition, we prioritize by weight: we attack the problem that costs the most points. In multiplicative, we attack the dimension closest to zero. A very low axis weighs down the entire product; fixing it brings up the entire score. Multiple near-perfect axes cost you next to nothing. The priority is reversed. Concretely, the auditor no longer draws up a shopping list. He looks for the single point that multiplies the least, and he starts there.
2. Stop piling on patches without hierarchy. Correcting ten small things while one dimension remains on the floor does not change the result. The product remains crushed by its lowest factor. This is the most costly mistake in SEO: spending weeks polishing steel links while the cardboard link is still holding the chain. We win by first attacking the point that blocks multiplication, and only then the marginal gains. Hypothetical example: a site that invests its entire budget in loading speed when its lowest factor is editorial reliability will see nothing change — it optimizes an already high factor and ignores the one that crushes the product.
3. Read a collapse as a step, not a slope. When a site suddenly drops, it is not the slow accumulation of small defects. It is a dimension that has just fallen below a threshold. We’re looking for the event, the single factor that went close to zero — not a long list of minor grievances. The shape of the curve is the clue: a gentle slope tells of an accumulation, a step tells of a postman who has fallen. That’s exactly what HouseFresh’s 0.09 says. And this is good news for the listener: a march has an identifiable cause, therefore targeted repair.
These three decisions transform the audit. We move from “inventorying all the problems” to “identifying the factor that multiplies the least”. It’s faster, fairer, and it concentrates effort where it pays off.
The limit: saturation at 57%
An honest demonstration names its limits. This one has one, and it is quantifiable.
The leak doesn’t give everything. By applying a Shannon calculation to the truly usable information, the CRAPS system saturates at 57%. This is the informational limit of the leak: beyond that, the available data no longer allows the mechanics to be reconstructed with certainty.
What 57% saturation implies for a listener
This number is not a theoretical reserve. It changes the working posture.
First, it sets a boundary between what we affirm and what we suppose. The structure is solid: quality is a multiplicative product of dimensions, period. The details – exact weightings, precise thresholds, order of importance of the 38 axes – belong to the rest, outside the leak. A good listener knows in which zone he is speaking. He affirms the mechanics, he does not guess the coefficients.
Then, it imposes humility on the thresholds. We know that a factor close to zero crushes the product. We do not know the exact point where a dimension falls below the critical threshold. So we are not promising a quantified gain of places. They say: this factor is your weak link, raising it brings up the whole product. One is engineering. The other is marketing.
Finally, the 57% protects against overfitting. Reconstructing a model that fits perfectly with the leak would amount to inventing the missing part – therefore creating a fake. Saturation is the safeguard. She says: this is how far the data goes, and not one step further.
This is enough to change the way we audit. Not to pretend to know everything. And it is precisely this honesty that separates a demonstration from a promise.
What this changes for your site
Your site does not have a quality rating. It has a product of 38 factors.
As long as none collapse, you’re holding on. The day one hits rock bottom, the whole product follows — even if everyone else is perfect.
So the real question in an audit is no longer “how many problems do I have?” “. It’s: “What is my lowest factor, and what keeps it near zero?” »
Do you know what yours is?
