siteAuthority: What One Line of the Leak Proves, Suggests, and Cannot

The whole case for “Google lied about domain authority” rests on one line of a protocol buffer definition:

siteAuthority (type: integer(), default: nil)
  site_authority: converted from quality_nsr.SiteAuthority, applied in Qstar.

A name, a type, a default, the system it comes from and the system it goes to. Nothing else. The line sits in the CompressedQualitySignals message of the documentation that surfaced through a public client-library repository in March 2024, and since then it has carried more argument per character than anything else in the corpus.

Shaun Anderson at Hobo wrote the long version of what the line does and does not establish, and he is more careful with it than most. What follows is his argument compressed and re-sorted into the three grades I already use for external evidence in a teardown: proved, suggested, unprovable. The standard is the one from the patents post: a leaked schema is evidence of field names, types and comments at one moment. It is a vocabulary, not a program. Where I read the line differently from Hobo, I say so.

Where the line lives

CompressedQualitySignals holds per-document signals that are compressed and shipped to two serving systems, Mustang and TeraGoogle. For TeraGoogle the message is embedded in perdocdata, and the comment says it is used in preliminary scoring. The same documentation adds that for TeraGoogle “this data resides in very limited serving memory (Flash storage) for a huge number of documents.”

The neighbours are the useful context. Next to siteAuthority the message holds nsrDataProto, authorityPromotion (“converted from QualityBoost.authority.boost”), unauthoritativeScore, pandaDemotion, scamness and anchorMismatchDemotion (“converted from QualityBoost.mismatched.boost”). The one field with a link in it is about anchor text mismatch, which is a link quality attribute rather than a count. Nothing in the message looks like a PageRank accumulator. It is a store of pre-computed quality, demotion and trust integers that get applied downstream in Q*, Google’s query-independent site-level quality score.

Placement alone tells you two things. The score is pre-computed, because Flash-resident per-document data is not recalculated per query. And it is available early, because preliminary scoring happens before the expensive query-dependent work. Whether it is used early as a filter is a separate question, and the text does not answer it.

The quotes it supposedly contradicts

Every accusation pairs the field name with a public denial. Here are the denials with dates, because the wording is the whole argument.

• Matt Cutts, 2013: “If you’re an authority in the medical space, we want to know that and to push you up higher whenever a medical query comes along … if you are a topical authority, keep writing about it.”
• Gary Illyes, 2016: “We don’t really have overall domain authority.” And, the same year: “I don’t know of anything in ranking that would translate to ‘domain authority’.”
• John Mueller, 2016: “Google doesn’t use ‘domain authority’.”
• Paul Haahr, 2017: “We have no one signal that we’ll say, ‘This is authority.’ We have a whole bunch of things that we hope together help increase the amount of authority in our results.”
• Mueller, 2018: “I don’t know if I’d call it authority like that, but we do have some metrics that are more on a site level … some of those site-wide level metrics might kind of map into similar things.”
• Mueller, 2019, to a fertility site that lost half its traffic after an update: “In general Google doesn’t evaluate a site’s authority. So it’s not something where we would give you a score on for authority.”
• Mueller, 2022, on Reddit: “You do not need DA for Google Search. Google doesn’t use it at all … I don’t think I’ve ever looked up the DA for a site in the 14 years I’ve been doing this.” And, paraphrased by Seroundtable the same year: links don’t add up to some overall authority metric on the domain level.

Read in order, this is a different story from a decade of denial followed by exposure. Cutts confirmed site-level topical authority in 2013. Mueller confirmed site-level metrics that “map into similar things” in 2018. What every spokesperson refused, consistently, was two narrower things: Moz’s capitalised Domain Authority product, and the idea that links sum into a domain score.

Hobo calls this the Trademark Trap. SEOs asked about “Domain Authority”, a Moz metric built on an external link graph. Google answered the question as asked, with an engineer’s literalism, and let the industry hear a broader denial than the one given. The leaked line is consistent with every quote above. It is converted from quality_nsr, a site quality system, not from a link graph, and it is one integer among many in a quality store rather than “one signal” called authority.

Technically true and misleading at the same time is a fair description of those answers. “Lied” is not supported by the text.

Grading the claims

Proved

A persistent, pre-computed, site-derived integer named siteAuthority exists in the per-document index data and is applied in Q*. The 2018 position, published in Search Engine Journal, that site authority “does not exist as a metric that is in use at Google” does not survive this.

It is derived from quality_nsr, the New Site Rank framework that evaluates sites in chunks. Site-wide evaluation is wired into preliminary scoring; it is not a discarded idea from the HITS era.

It is a typed field with a nil default, which is what an automated pipeline output looks like. A manual override for brands Google dislikes would not look like a nullable integer converted from another system, so “Google uses this to tank companies by hand” has no support in the line.

default: nil is not 0. A site the pipeline has not assessed has no value, which is mechanically different from having a low one. “New sites start at zero authority” and “small publishers are hardcoded to zero” are refuted by the one piece of concrete data in the line.

Suggested

Placement in Flash-resident preliminary-scoring data suggests the score is slow-moving and structural, and does not react to this week’s publishing. The static reading is plausible. The text states no refresh cadence.

The neighbourhood suggests the field measures something closer to quality or trust than to link volume. Hobo goes further and reads it as “trustworthiness, or popularity, or satisfaction.” I would stop at: this is not where link counting happens. What the field aggregates is not in the text.

An early-stage filter is a reasonable hypothesis, since the field is available during preliminary scoring. It stays a hypothesis. Multiplier, tie-breaker and late-stage filter are all equally consistent with “applied in Qstar.”

Unprovable

What the integer measures. The line says integer(). Any sentence of the form “siteAuthority assesses trust, credibility and expertise” is an industry gloss on a type annotation.

What its inputs are. Nothing in the line names backlinks, user behaviour, Chrome data or topical focus. “Proves Google uses clicks for authority” and “proves it is link-based” are both read off a blank page.

Whether it is one value per domain. The name says site and the lineage says NSR sitechunks, but the line does not say the integer is a single value for a whole host. Persistent across the domain is an assumption.

Its sign. “Low authority causes a demotion” assumes the field can hurt. It may be a boost for high-tier sites that leaves the median site untouched. The text does not distinguish the two.

What happens on migration. There is no mapping from sitechunk to server, IP, CMS or URL structure anywhere in the line. “Changing hosts resets your authority” and “changing hosts preserves it” are both unsupported.

Whether buying links moves it. The field is converted from a quality system and shares a message with demotions, not with PageRank. That does not prove links are irrelevant to it. It proves the line is silent on the question, and the only safe claim is that the text gives no mechanism by which a purchased link reaches this integer.

Anything about topical focus. siteFocusScore and siteRadius are separate fields. “A tight topical radius raises your siteAuthority” conflates three attributes that the schema keeps apart.

The 2019 fertility site, revisited

The question Mueller got was “how could our authority drop more than 50% overnight?” The informative part of his answer was the redirection, not the denial: he pointed the site at quality and at asking users for hard feedback. A site losing half its traffic in a YMYL category after a core update has the signature of a quality threshold moving, and in that category the threshold moves one way. Whether the integer called siteAuthority was involved cannot be known from outside. That some site-level quality assessment was involved is the most economical reading of the quote and the schema together.

What the site owner could have done is the thing nobody in that exchange suggested: decompose the drop into impressions and click rate per query cohort before asking anyone what “authority” means. A 50% loss that is all impressions, across every cohort, on the date of an update, is a site-level event. A loss concentrated in a few cohorts is not, whatever the field is called.

The caveat over all of it

The leak is schema. It names fields and their lineage. It contains no weights, no thresholds, no curves and no indication of which fields are live in production and which are superseded. The DOJ record has a Google engineer saying as much: the documents name components of the ranking system without going into the specifics of the curves and thresholds. A neural reranker has no readable formula to leak in the first place.

So the honest summary of one of the most-cited lines in the corpus is short. A site-level, quality-derived integer exists, is pre-computed, and is applied in a query-independent quality score. Google’s public statements, read literally, never denied that. Everything else attached to the word siteAuthority, in either direction, is inference, and the better pieces say so.

If you have a quote I missed, or a line of the schema that moves a claim from one grade to another, that is the interesting case. Corrections get their own entry, linked from this one.