EEAT Mechanics

Twiddlers: The 120 Re-Ranking Rules Behind Google's Final Order

An agency analysis counts 120 twiddlers that move results after scoring, three of them named after Expertise, Authenticity and Topic Authority. The names are evidence that something exists. They are not evidence of a weight, and only nine of them can be traced back to the 2018 document or the 2024 leak.

By Thomas Wawra· Published · Updated · Version 1.0· Systems referenced: Mustang / SuperRoot etc., chard (+ YMYL/Hoax), Topic-Embeddings (Focus/Radius), OriginalContentScore, SpamBrain, FreshnessTwiddler / RealTimeBoost, RankBrain / DeepRank / FastSearch

01What the count rests on

The SEO agency RESONEO has published a list of 120 twiddlers across web, image and video search; SEO Südwest reported on it on 5 November 2026. A twiddler is a small component that reworks the result list after the main ranker has scored it and before the page is assembled. The analysis builds on three layers: an internal Google document from 2018 that listed 65 twiddlers for web search, the Content Warehouse leak of 2024, and the authors' own reading of Superroot and RankLab schemas, the parameter and signal names Google's systems pass around.

The third layer carries almost the whole list. By RESONEO's own count, 111 of the 120 twiddlers have no forerunner in either the 2018 document or the 2024 leak. Their names, such as ExpertiseTwiddlerDocFeatures or BardFilterTwiddlerParams, come from schemas we cannot inspect ourselves. That is why almost every statement in this article about a specific twiddler carries the code [C]: it reports an industry analysis, not an artefact we could check word for word.

Four of the names can be checked against the leak itself, and they check out. The leaked documentation (version 0.4.0) describes date information as "Used in FreshnessTwiddler", sub-image data as "consumed in web image boost twiddler", an entity list as "Used by the Skin Tone Twiddler", and a PerDocData field as a lookup key in "one of the superroot's PRE_DOC twiddlers". That confirms the framework, its home in Superroot and the predoc stage in Google's own wording. A search of the same documentation found no entry for SiteBoost or a domain classification twiddler. A search that finds nothing is not proof that the name is absent, but these two names stay at [C].

Source per claim (3)
C
RESONEO counts 120 twiddlers across web, image and video search; 111 of them have no forerunner in the 2018 document or the 2024 leak.[1]
C
An internal Google document from 2018 listed 65 twiddlers for web search, among them the DMCA filter and YouTube duplicate removal. Its provenance has not been independently verified.[2]
B
The leak names twiddlers in its own wording: PerDocData is a lookup key in "one of the superroot's PRE_DOC twiddlers", date info is "Used in FreshnessTwiddler", and other models feed the web image boost and Skin Tone twiddlers.[3]

02How a twiddler moves a result

The twiddler framework sits in Superroot, the system that sends a query to Google's indexes and puts their answers together. The backend retrieves and scores a few hundred results, with Ascorer as the main ranker. Predoc twiddlers then work on the whole list while the results are still thin and have no snippets. Lazy twiddlers follow and work on the top of the list once snippets have arrived. Tangram assembles web, image and video results into one page only after that.

No twiddler decides anything by itself. Each one works in isolation and recommends an action: move a result up or down, place it above another result, filter it out, cap how many results of one kind appear, space out results of the same category, or leave an annotation for later steps. The framework reconciles these recommendations. According to the 2018 document, a ranking experiment is easier to run as a twiddler than as a change to Ascorer, which goes some way to explaining why there are so many of them.

One of these rules is public. Google's site diversity system usually shows no more than two listings from the same site in the top results. Our article on Mustang and SuperRoot describes it as host-crowding. It is a display rule, not a quality judgement: a site that drops from two listings to one has not been demoted.

Source per claim (2)
C
Twiddlers run in Superroot after Ascorer has scored the results: predoc twiddlers work on the whole list, lazy twiddlers on its top once snippets exist. They only recommend (boost, BoostAboveResult, filter, max_total, stride, annotate), and the framework reconciles.[4]
O
"We generally won't show more than two web page listings from the same site in our top results."[5]

03Expertise, Authenticity, Topic Authority: E-E-A-T after the score

This is the part that matters for E-E-A-T. RESONEO places nine twiddlers in a family for quality, authority and trust. Three of them carry the names of concepts Google uses in its guidelines and blog posts. The Expertise twiddler has features per document, so expertise would be judged page by page, which puts it closest to the page-level quality model chard. The Authenticity twiddler has signals for the query and for each result, and an approximate version of Ascorer computes an authenticity boost as well. If that reading is right, authenticity counts both during and after scoring.

The Topic Authority twiddler has features for the query and for the document, so a site's authority would be measured on the topic of the query, not across the whole site. Google itself described a topic authority system for news in May 2023: it helps identify which expert sources are useful for a news query in specialised fields. In our system map, topic authority sits closest to Topic-Embeddings and siteFocusScore. A fourth twiddler, Web Health, includes citation-based originality, which gives weight to the outlet that others cite for a story. That is the same idea behind OriginalContentScore.

This conflicts with our own earlier description. The Mustang and SuperRoot article says the infrastructure computes rankings but does not judge quality. That still holds for the framework, which only reconciles recommendations. It no longer holds for every component that runs inside it. If the schema names mean what they say, part of the E-E-A-T assessment takes effect after the score, as a correction to the order. We have added a note to the older article. The names show that these twiddlers exist. They do not show how much a twiddler moves a result or how often it fires.

Source per claim (6)
C
An Expertise twiddler exists with features computed per document (ExpertiseTwiddlerDocFeatures), so expertise is assessed page by page after scoring as well.[1]
C
An Authenticity twiddler has query-level and result-level signals, and an approximate Ascorer also computes an authenticity boost, so the notion counts during and after scoring.[1]
O
Google describes a topic authority system that helps determine which expert sources are helpful for news queries in specialised topics such as health, politics or finance.[6]
C
The Topic Authority twiddler has query features and document features, so authority is measured against the topic of the query rather than for the site as a whole.[1]
C
The Web Health twiddler includes citation-based originality, which favours the outlet others cite for a piece of news, the same idea as OriginalContentScore.[7]
C
The twiddler framework itself does not judge, but some twiddlers running inside it carry E-E-A-T judgements. Part of the E-E-A-T assessment thus works as a correction to the order after scoring, with unknown strength.[8]

04Demotions, freshness, preferences: the other families

Twelve twiddlers handle spam, removals and safety. SiteBoost, which RESONEO says the leak names (our own search of the leaked documentation found no entry), detects user-generated spam pages and uses a site's main category to judge whether a result fits the query. Another twiddler keeps a list of results heavily demoted for spam. Copyright works at site level: once a site passes a certain number of valid removal notices, the whole domain moves down. That matches what Google documents for removal-based demotion. In our system map these belong with SpamBrain and the QualityBoost demotions. They are trust signals that work as a deduction.

Fourteen twiddlers deal with freshness, repeat queries and user preferences. The FreshnessTwiddler from the leak now has a second version and a date score. Google describes these as "query deserves freshness" systems, which is our system 24, fed by date triangulation. Other twiddlers, such as For you, domain preference and topic affinity, apply the user's own habits after scoring. Since August 2025, users can pick Preferred Sources for Top Stories, and those outlets then appear more often for them. Part of the RankBrain output also passes through a twiddler and is applied to the list that has already been scored.

For practitioners, this changes where to look, not what to do. A page can rank well and still move down afterwards: through a spam demotion, through DMCA notices against the whole domain, or because a more recent page or a forum is pulled ahead. The E-E-A-T work stays the same: expertise page by page, authority topic by topic, real and visible dates. What is new is how many places there are in which a well-scored result can still lose its position. The 52 twiddlers for images and video, from YouTube clickbait to inauthentic Shorts, are outside the scope of this site's system map.

Source per claim (5)
O
"If we process a significant volume of such removals involving a particular site, we use that as a signal to improve our results."[9]
C
Twelve twiddlers cover spam, removals and safety, including SiteBoost (which RESONEO says the 2024 leak names; our search of the leak found no entry), a list of heavily spam-demoted results and a site-level DMCA twiddler that demotes the whole domain past a volume of valid notices.[1]
O
"We have various 'query deserves freshness' systems designed to show fresher content for queries where it would be expected."[10]
O
Since 12 August 2025, users in the U.S. and India can select Preferred Sources; their articles then appear more often in Top Stories when they have fresh, relevant content.[11]
C
Part of RankBrain's output goes through a twiddler and is applied in Superroot to an already scored list, so RankBrain also acts after the main scoring.[1]

05References

This Deep Dive is Layer 2 content. It interprets the material and always points back to Layer 1, the reference layer. Every claim is mapped to a source, labeled by type: [A] Sworn testimony · [B] Verbatim from the leak · [O] Documented by Google · [W] Peer-reviewed study · [S] Checked on the page · [C] Our assessment · [P] Patent.

© Thomas Wawra · Senior SEO Manager · wetter.com — a Funke Digital company