Table of Contents
Key Takeaways
- In 2024, Google accidentally published documentation for about 14,000 attributes it stores about pages, links and sites. It shows what Google measures, not how much any of it counts.
- 28 live quality signals sit in Google’s scarce serving memory. 13 of them push pages down and only 4 lift them up.
- Clicks, site-wide quality and the effort behind a page are all recorded.
- At least 16 popular “leak findings” have no support in the documents.
Most Google leak SEO advice traces back to a few summaries written in mid-2024. Those summaries were fast and useful, but some of the most-quoted fields in them do not exist in the leaked files.
This guide goes back to the source. SEO Optimizers parsed the leaked package itself and read the engineer comments behind each claim.
You will learn what the documents are, which signals Google treats as worth the most storage, what to change on your website, and which claims to stop repeating. Every field name used as evidence below is real and can be checked against the package.
What Was the Google Leak?
The Google leak is the accidental 2024 publication of documentation for Google’s internal Content Warehouse API. It describes about 14,000 attributes Google stores about web pages, links and sites, each with a short engineer note in many cases. It contains no ranking weights or formulas.
An automated bot committed the documentation to a public code repository around March 13, 2024. The files stayed public until May 7, 2024, and the story broke on May 27, 2024.
Google confirmed the documents were real on May 29, 2024. A spokesperson told The Verge, “We would caution against making inaccurate assumptions about Search based on out-of-context, outdated, or incomplete information.”
The package covers roughly 2,600 modules. SEO Optimizers’ parse of version 0.4.0, the latest leaked version, counted 14,031 attributes in 2,593 modules, and about 9,500 of those attributes carry an engineer-written description.
Does the Google Leak Show How the Algorithm Works?
The Google leak does not show how the algorithm works. It is a filing system, not a recipe: it lists the data Google keeps and, for many fields, what an engineer said each field is for.
Three limits apply to every claim built on it:
- No weights. Nothing in the documents says how much any field counts. Anyone quoting a percentage for a ranking factor is guessing.
- Existing is not the same as used. A field in a storage schema may never be read during ranking. Only 2.3% of fields carry a deprecation marker, and some, like an inp field in the Core Web Vitals module, have no description at all.
- It is a snapshot. The files date to early 2024, and Google has shipped many updates since.
The leak is also incomplete. The CompositeDoc record, for example, numbers its fields up to 193 but shows only 44 of them, about 23%. That is why “it’s not in the leak” is weak evidence that Google ignores something.
Which Ranking Signals Does Google Keep in Its Scarce Serving Memory?
The closest thing to a priority list in the leak is likely a module called CompressedQualitySignals. Its engineer comment says the data “resides in very limited serving memory (Flash storage) for a huge number of documents” and can be used in preliminary scoring.
Storage that scarce suggests the contents were chosen with care. Of the 38 fields in that module, 5 are deprecated and 5 are experimental, which leaves 28 live production signals.
| Signal group | Live fields | Share of the 28 |
|---|---|---|
| Product review quality | 6 | 21% |
| Click behavior (NavBoost family) | 5 | 18% |
| Panda-family content quality | 4 | 14% |
| Link, domain and UX demotions | 4 | 14% |
| Site authority | 3 | 11% |
| Page-level quality containers | 3 | 11% |
| Scam, user-generated content effort, topic embeddings | 3 | 11% |
Thirteen of the 28 fields act downward and 4 act upward, a ratio of about 3 to 1. That suggests most ranking value comes from avoiding demotion, not from chasing a boost.
Product reviews stand out. Six fields, over a fifth of the 28 live fields, concern product review pages alone.
The grouping and the up or down labels are SEO Optimizers’ classification. The field list itself is exact, and other analysts could group it differently.
What Does the Google Leak Mean for SEO?
The Google leak mostly supports slow, unglamorous work. Fix pages that disappoint searchers, clean up weak content, link from your strongest pages, and publish things nobody else has.
Each action below names the field behind it. Before changing anything, an SEO website audit shows which of these problems your site actually has.
1. Deliver the Title’s Promise Above the Fold
Two live serving fields, serpDemotion and navDemotion, are demotions. The serpDemotion comment says it is applied in Qstar, and navDemotion is converted from a QualityBoost value. Neither comment explains what triggers it, but the names point to results-page and navigation behavior.
Put a short, direct answer near the top of the page. Make the next step easy to reach from every landing page.
2. Improve Your Weakest Pages
The documents measure how much quality varies across a site, in addition to its average. Fields like siteQualityStddev and chardScoreVariance suggest thin pages can hold back strong ones, so regular content pruning is an active lever, not housekeeping.
3. Link From Your Strongest Pages
A field called onsiteProminence is described as “computed by propagating simulated traffic from the homepage and high craps click pages.” Links from your best-visited pages raise the pages they point to. A deliberate internal linking strategy costs nothing but attention.
4. Add Something the Top Results Lack
The contentEffort field is described as “LLM-based effort estimation for article pages.” A separate OriginalContentScore applies to pages with little content. Original data, real photos and first-hand detail are what creating original content really means here.
5. Edit AI Content Before Publishing
A field called racterScores is described as a “site-level AGC classification score.” AGC appears to mean auto-generated content, so heavy unedited AI publishing may weigh on the whole site.
6. Cut Clutter on Desktop Too
The clutterScore field targets sites “with a large number of distracting/annoying resources.” A separate field flags desktop interstitial violations, so popups matter on desktop as well as mobile.
7. Use Field Data for Core Web Vitals
The voltData module states that “the data is used for ranking changes” and stores 75th-percentile data from real visits. Lab scores are a guide, but real-visitor data is what gets stored.
8. Build Fewer, More Local Service-Area Pages
Google stores simhash fingerprints of page text, and the code names an ApproxDupsParseHandler, which suggests near-duplicate detection. Imagine ten city pages that differ only by place name: they risk looking like copies of each other. Three pages with real local detail give each page something of its own.
Does Google Use Clicks to Rank Pages?
The leaked documents store click data per page and per site, split by country, device and language. They do not say how click volume is weighted.
- voterTokenCount records a lower bound on the distinct users behind a click entry. Its comment ties it to privacy filtering, not to ranking.
- crapsUnscaledIpPriorBadFraction appears to record a bad-click fraction tied to IP. It has no engineer description, so that reading comes from the name.
- goodClicks and badClicks also have no description. What they mean is inferred from their names and from testimony by a Google vice president in the U.S. antitrust case against Google.
The safe conclusion is simple. Google records what people do after they click, and the honest way to improve it is a better title and a better page.
16 Google Leak Claims the Documents Don’t Support
Several widely shared leak “findings” have no supporting field or comment in the package. SEO Optimizers checked every attribute name, module name and comment across all three leaked versions (0.2.0, 0.3.0 and 0.4.0), and anyone can repeat the check in the Content Warehouse package documentation.
| Claimed field or system | What the documents actually contain |
|---|---|
| raterData / isTrustedRater | No quality-rater data field exists |
| QualityAuthorAuthorVectors | No author-vector module. Real author fields are isAuthor, isAuthorIndex and authorObfuscatedGaiaStr |
| PerDocData.author | The field does not exist on PerDocData |
| RankBrain | Not named anywhere |
| DeepRank | Not named anywhere |
| Tangram / Tetris | Tetris is not named. Tangram appears only as an undocumented tag, and nothing describes results-page assembly |
| RealTimeBoost | Only three undocumented indexing tags. No field or comment describes a real-time boost |
| spamRun | Real anchor-spam fields are spamProbability and phraseAnchorSpam fields |
| forwardingDupType | Real field is CompositeDocForwardingDup.purposes |
| isGoldStandardDocument | No gold-standard flag exists |
| original_media_score | Real image fields are nimaAva, nimaVq and deepImageEngagingness |
| Per-document time to first byte | Timing lives in crawl fields like ServerResponseTimeMs |
| MerchantListing | Real field is shoppingProductInformation |
| linkBeforeSitechangeTimestamp | Not present. The real field is linkBeforeSitechangeTaggedAnchors |
| “Freshdocs link multiplier” | No field says newer pages pass more link value |
| T* topicality score | T* is not named. A separate WebRef topicalityScore rates how related an entity is to a page |
Two more claims are half right. “Tofu” is a content-based site quality predictor, not “trust on first use.”
The homePageInfo field has four values, not three.
What the Leak Gives No Support For
Some common SEO habits find no support in the documents:
- Word-count targets. No field sets a minimum length, and the numTokens comment notes that documents are truncated at a maximum cap.
- Keyword density. The only related field is KeywordStuffingScore, which appears to work as a penalty, not a reward.
- A 60-character title rule. Title length is measured in tenths of an em over a line budget, not in characters.
- Schema as a ranking factor. No documented structured-data ranking field exists. One snippet feature, schemaOrgDescriptionBoostScore, carries no description. Schema earns rich results, not positions.
- Social shares. No field counts shares or likes as a ranking input.
- Domain age as a bonus. hostAge appears in the context of sandboxing fresh spam, not as a reward.
Frequently Asked Questions
How many ranking factors does Google use?
Google uses an unknown number of ranking factors, and the leak does not settle the question. The documents list about 14,000 attributes, but that is everything Google stores, not a list of ranking factors.
Many attributes describe images, crawling or other products. An analysis by SEO Optimizers found 533 that relate to ranking, and even those come with no weights, so a fixed count like “200 factors” is a guess.
Are social signals a Google ranking factor?
Social signals are not a Google ranking factor in the leaked documents. No field counts shares, likes or followers as a ranking input. The closest match, socialgraphNodeNameFp, is a fingerprint a Social Search twiddler uses as a lookup key.
Social posts can still bring visits and links, and those can matter. The shares themselves are not counted as votes.
What is Google NavBoost?
Google NavBoost is a system that uses click data to adjust search results. A Google vice president testified about it in the U.S. antitrust case against Google, and the leaked documents name NavBoost and its related CRAPS click modules many times.
Click data is split by country, device and language. A field such as voterTokenCount stores a lower bound on the distinct users behind a click entry, used for privacy filtering.
Where can you find the leaked Google documents?
You can find the leaked Google documents on Hexdocs, the documentation site for the Elixir package registry. Search for google_api_content_warehouse and open version 0.4.0, which is the latest leaked version.
Versions 0.5.0 and later are much smaller and appear to drop the search material. The pages are technical reference files, so expect field names and short comments, not articles.
Was the Google leak real?
The Google leak was real, and Google confirmed it on May 29, 2024. A company spokesperson cautioned against “inaccurate assumptions” based on “out-of-context, outdated, or incomplete information,” but did not dispute that the files were authentic.
That caution is fair. The documents are real, but they are a snapshot from early 2024 and they carry no weights.
The Bottom Line on Google Leak SEO
The leak did not reveal a shortcut. It confirmed that the unglamorous work is what Google built fields to measure: effort, originality, consistency across a whole site, and what people do after they click.
Three findings hold up under every caveat. Downward signals outnumber upward ones about 3 to 1 in the serving set, site quality is measured several ways including its spread, and click behavior is stored and reused.
SEO Optimizers keeps a full reference covering all 533 ranking-related signals with Google’s own wording. If you want it, or want help checking your site against it, contact SEO Optimizers and ask for the Content Warehouse reference.
Have a question about this? Ask me in Be The Answer, my free SEO community.







