What the Google leak and the antitrust documents revealed about links
Two sets of documents changed what we know about links in Google. Court evidence from the US antitrust case confirms how links are used. The May 2024 leak shows which link data Google stores, but not how much any of it counts.
In short
- Sworn evidence in the US antitrust case confirms that links work in two places: anchors feed a relevance score, and PageRank feeds a quality score.
- The leaked documentation published on 27 May 2024 covers 14,014 attributes across more than 2,500 pages, and contains no weights and no scoring formulas.
- The leak documents a site-level authority value, quality tiers for linking pages and measurements of sudden anchor text spikes. Whether each is in live use is unknown.
- The popular claim that links from pages with no clicks are ignored is an analyst's hypothesis, and an Ahrefs test found only a weak relationship between a linking page's traffic and ranking benefit.
Two bodies of evidence now sit behind most serious claims about how Google treats links. One is sworn testimony and exhibits from the US antitrust case against Google. The other is internal API documentation that leaked in 2024. They are not equally strong, and most of the bad advice written since comes from treating them as if they were. A statement made under oath by someone who works on ranking and a description of a data field with no context around it are different kinds of thing, and a reader who keeps them apart will avoid most of the overconfident rules that have circulated since.
This page sorts the link findings into four grades: confirmed, documented, inferred and speculative.
The two sources and how much weight each can bear
The antitrust case is the stronger of the two. The liability trial in US v. Google ran from September to November 2023, and the court ruled on 5 August 2024 that Google was an illegal monopolist. A remedies trial began on 21 April 2025, with the remedies opinion on 2 September 2025. Testimony was given under oath and the exhibits are Google’s own material, so this is primary evidence. When something on this page is graded as confirmed, it is because it was said or shown in that setting.
The leak is larger and weaker. Documentation for Google’s Content Warehouse API, more than 2,500 pages covering 14,014 attributes, was published to GitHub by mistake in March 2024. Rand Fishkin and Mike King made it public on 27 May 2024. According to press reports at the time, Google confirmed the documents were real within days and warned against drawing conclusions from information that was out of context, outdated or incomplete. The warning came from an interested party, but it is also a fair description of what the documents are.
The limit of the leak matters more than its size. It lists data fields and modules. It contains no weights, no scoring functions and no proof that any field is used in production. Some fields are marked deprecated and many carry no status at all. So the leak can tell you that Google has, at some point, stored a particular piece of information about links. It cannot tell you whether that information is still collected, whether any ranking system reads it, or how much it matters if one does. The four grades used on this page follow from that difference between the sources.
| Grade | What it means | Where it comes from |
|---|---|---|
| Confirmed | Stated in sworn testimony or a court exhibit | Antitrust case |
| Documented | A field exists in the leaked docs; use and weight unknown | The leak |
| Inferred | An analyst’s reading of what a field implies | Commentary on the leak |
| Speculative | No document supports it directly | Commentary |
What the court documents confirm
The clearest material comes from an interview with a Google engineer, entered as an exhibit and reported by Search Engine Journal on 15 May 2025. Our quotes come from that report, not from our own reading of the exhibit. With that caveat, the exhibit as reported makes five points that bear on links.
The first is that relevance is built from three things. Google’s topicality score combines anchors, body text and clicks, in what the exhibit calls a “relatively hand-crafted way”. The second is that PageRank feeds quality. The exhibit describes PageRank as “a single signal relating to distance from a known good source” and as “an input to the Quality score”. The third is that quality is mostly a property of the site. The quality score, described as “the notion of trustworthiness”, is “incredibly important” and “generally static across multiple queries”. Taken together, these three points place links on both sides of the system: the anchor on the side that judges whether a page fits a query, and PageRank on the side that judges whether the site can be trusted.
The remaining two points are about how the signals are made. The exhibit is reported as saying that “the vast majority of signals are hand-crafted”, so that engineers know what to fix when something breaks, and that Chrome data is used in a popularity signal. Separately, Pandu Nayak testified on 18 October 2023 that a system called Navboost uses 13 months of click data, down from 18, and is one of the important signals. Other testimony and an exhibit on the life of a click show that clicks are used heavily in ranking, which contradicts years of public messaging.
For link builders the confirmed picture is simple. A link acts in two places: its anchor text feeds relevance to a query, and PageRank, measured as distance from trusted seed sites, feeds trust in the site. Clicks are a third pillar that links cannot replace. That last part is easy to pass over, and it sets a limit on what link building can do: on the confirmed evidence, relevance draws on clicks as well as anchors, so a page can have the links and still lack one of the inputs.
What the leak documents
The items in this section exist in the leaked documentation. We describe what each field does and do not print field names, because the exact spellings have not been checked against a copy of the original files. Each item comes in two halves. The first half is what the documentation describes, and that half is documented. The second half is what analysts infer from it, and that half is inference. An inference can be sensible and still be wrong.
Several items concern the site as a whole. The documentation describes a site-level authority value stored with other quality signals, from which analysts infer that Google has a sitewide authority concept, despite denying that it uses “domain authority”. It describes the homepage’s seed-based PageRank being copied to every page on the site, which analysts read as meaning a new page starts with the homepage’s authority as a default. It also describes measures of a site’s topical focus and of how far a page sits from that focus, and the inference drawn is that topical fit between the linking site and the target matters.
A second group concerns the page a link sits on and where the link appears. The documentation describes a quality bucket on each anchor (high, medium or low) tied to the index tier of the linking page. Analysts infer that links from pages in the top tier are worth more and that links from the lowest tier may be ignored. It describes trust flags on link targets and counts of trusted anchors, read as meaning that trusted sources are protected and their links weigh more. It describes extra data for anchors from news sources, from which the inference is that links from news sites are handled separately. And it records font size and the text either side of an anchor, which analysts take to mean that prominent, in-context links carry more than hidden or boilerplate ones.
A third group concerns anchor text and spam. The documentation describes a demotion for anchor mismatch, and analysts infer that anchors that do not fit the target, or links from off-topic sources, can be demoted or ignored. It describes counts of spam-phrase anchors, how fast they appeared, the start and end of the spike and how many were demoted. The inference is that sudden bursts of similar anchors are measured and the anchors inside the burst can be discounted, and that the same mechanism would blunt negative SEO attacks. It also holds Penguin-era anchor statistics, including counts of dropped anchors, which analysts read as showing that Penguin logic still lives in the anchor pipeline, and that it drops anchors instead of punishing pages.
The last group concerns time. The documentation describes first-seen, last-seen and deletion dates for anchors, so analysts infer that link age and link churn are tracked. It describes a host age value used “to sandbox fresh spam in serving time”, read as meaning that new hosts can be held back when they look like spam. And it states that only the last 20 versions of a URL are kept, from which analysts infer that Google can see what a repurposed page or expired domain used to be. In every one of these pairs the stored data is the documented part, and the statement about what Google does with it is somebody’s reading.
What is inferred, and what is speculation
The best known inference is the claim that “links from pages with zero clicks are ignored”. This is Mike King’s hypothesis: click data helps decide a page’s index tier, the tier sets the anchor’s quality bucket, so links from pages nobody visits count for little. It is plausible and consistent with the documents. It is not confirmed, and an Ahrefs test found only a weak relationship between the traffic of a linking page and the ranking benefit of its link. The hypothesis joins a confirmed fact, that clicks matter, to a documented field, the quality bucket, with a step that no document supplies, which is why it sits in the inferred grade however reasonable it sounds.
A related idea is that clicks can outvote links. It is confirmed that Navboost and links both exist and both matter. No document states how they interact. The idea that a well-linked page sinks if users do not click it is an inference. The same gap applies to how much links contribute to site quality. The court exhibit confirms that PageRank is an input to the quality score. Its share is unknown, and so is the way the site-level authority value is calculated. Anyone who quotes a percentage for either is supplying it themselves.
Two further claims belong in the speculative grade. The first is a sandbox for all new sites. Only the spam-related host age value is documented. A general sandbox is speculation. The second concerns disavow data. Analysts noted that disavow files do not appear in the leaked modules and guessed that the tool is separate from core ranking data. That is speculation too, since the absence of something from an incomplete set of documents shows very little about where it lives.
How to use this evidence
The grades translate into five working rules, and each one says which grade of evidence it rests on.
- Treat the two-job model as settled: anchors for relevance, PageRank for trust. Both parts need attention, which is why anchor text gets its own lesson.
- Prefer links on pages that real people visit and that sit close to trusted sites. This is consistent with the confirmed and the documented evidence, even though the exact mechanism is inferred.
- Avoid sudden bursts of the same commercial anchor. The measurement of anchor spikes is documented, and the cost of caution is low.
- Check the history of any expired domain before you buy it. Google keeps past versions of pages. See expired domains.
- Do not turn a field name into a rule with a number attached. No leaked document supports a “safe” anchor ratio or link velocity threshold.
The reasoning behind the middle three rules is the same in each case. The underlying mechanism is documented or inferred and not confirmed, so none of them can be stated as a fact about how Google ranks. What makes them worth following is that acting on them costs little and is sensible on other grounds, so nothing is lost if the inference turns out to be wrong. The fifth rule guards against the opposite error: a document that contains no weights cannot be the source of a threshold, and any figure presented as coming from the leak has been added by the person presenting it.
Read how search engines evaluate links for the wider model, then the timeline of Google’s link spam updates to see how enforcement changed over the same years. Our methodology explains how we grade evidence across the site.
Common questions
Did the Google leak reveal the ranking algorithm?
No. It revealed the names and descriptions of data fields and modules. It has no weights, no scoring functions and no proof that any single field is used in live ranking.
Does Google have a domain authority score?
The leaked documentation describes a site-level authority value held with other quality signals. That is not Moz's Domain Authority, which Google has always said it does not use, and nobody outside Google knows how the value is calculated.
Is PageRank still used?
Yes. An exhibit from the US antitrust case, reported in May 2025, describes PageRank as a single signal about distance from a known good source, used as an input to Google's quality score.
Is there a Google sandbox for new sites?
Not as a confirmed general rule. The leak documents a host age value used to hold back fresh spam. A general penalty for all new sites remains speculation.

