RESONEO ·

Twiddlers: what still moves a result after ranking

Between ranking and display, twiddlers, small programs, rework the list of results: they boost one result, demote another, filter out a third or space out those from the same site. In 2018, an internal Google document counted more than 65 of them in production for web search alone. We found 120 twiddlers across web, images and video, among them E-E-A-T expertise, topic authority, a filter named after Bard and the plan of searches to run for a question, the basis of AI Mode’s fan-out.

Constructed examplerobot vacuum
Ranking order Page shown
  1. An affiliate siteTop 10 robot vacuums (2026 comparison)↓ Spam demotion
  2. A home magazineRobot vacuums: how to choose one
  3. A manufacturer’s siteOur range of robot vacuums
  4. An encyclopediaRobotic vacuum cleaner
  5. A 2021 comparison2021’s robot vacuums put to the test↓ Freshness
  6. A test labWe tested 24 robot vacuums for three months↑ Expertise
  7. An appliance guideRobot vacuum buying guide, October edition↑ Date score
  8. A deals siteRobot vacuums: this week’s deals↑ Deals
  9. A DIY forumYour robot vacuum after a year of use?↑ Forums
  10. A specialist retailerRobot vacuums: 48 models in stock↑ Specialty merchants
Seven twiddlers move seven results; two others move down as a knock-on effect.

Between the score and the page shown, a second pass

Google’s ranking scores hundreds of results. Before display, Superroot, the system that sends the query to Google’s indexes and then assembles their answers, runs that list through the twiddlers. Each one recommends a change (move up, move down, filter, space out) and the twiddler framework reconciles these recommendations. The twiddlers’ names say along which lines a position still moves at this stage: AI, quality, spam and copyright, freshness, user preferences.

120twiddlers across web,
images and video
6AI and ranker twiddlers,
RankBrain and a Bard filter among them
111with no forerunner in the 2018
document or the 2024 leak
★

Google’s AI goes through twiddlers too: the fan-out plan, a Bard filter, a relevance oracle

Six twiddlers belong to Google’s AI and rankers. None appears in the 2024 leak. In the second pass, models rework the list, a full ranker scores the results again and a filter bears the name of Google’s conversational AI.

The fan-out search plan

The query fan-out Google describes for AI Mode (May 2025) starts from a plan: the searches to run for a question. That plan travels with the query into the second pass and sits among its ranking signals. The twiddlers know the searches planned for the question when they rework the list.

In the schemas

  • Planned Retrieval AnnotationsPlannedRetrievalAnnotationsplanned retrieval annotations

The sources of AI Overviews and AI Mode, part 4 of the series

A filter named after Bard

Bard became Gemini in February 2024; the filter keeps its name. Its settings arrive with the query and it leaves a trace on each result it examines: Google’s conversational AI has its own filter in Search’s second pass.

In the schemas

  • Bard Filter Twiddler ParamsBardFilterTwiddlerParamsfilter settings, sent with the query
  • Bard Filter Twiddler DataBardFilterTwiddlerDatafilter data on each result

A relevance oracle

In machine learning, an oracle is the model that gives the reference answer. Its inputs, for the query and for each result, sit among the ranking signals, and it scores each result: after ranking, a reference model measures relevance a second time.

In the schemas

  • Relevance Oracle Twiddler Result SignalsRelevanceOracleTwiddlerResultSignalsthe oracle’s score on each result
  • Relevance Oracle Twiddler Query Input SignalsRelevanceOracleTwiddlerQueryInputSignalsinputs taken from the query

RankBrain, partly applied after ranking

RankBrain, launched in 2015, is Google Search’s first deep learning system (Google, February 2022). Part of what it produces goes through a twiddler and is applied in Superroot, which sends the query to Google’s indexes and assembles their answers, on an already scored list.

In the schemas

  • RankBrain Query FeaturesRankBrainQueryFeaturesquery features
  • RankBrain Twiddler Result SignalsRankBrainTwiddlerResultSignalstwiddler signals on each result

A full ranker in the second pass

The For Loop Ranker (FLR), with its web-specific settings, scores the results in turn: a full-fledged ranker runs as a twiddler, next to simpler adjustments.

In the schemas

  • FLR Output Result SignalsFLROutputResultSignalsFor Loop Ranker scores on each result
  • Query Param SignalsQueryParamSignalsweb-specific settings

Relevance feedback: rescoring from the good results

Relevance feedback is a classic information retrieval technique: results are scored again by how much they resemble results judged good, often the top ones in the list. Google gives it a twiddler for web search, one for images, built on Starburst, Google’s image embedding already cited in the 2024 leak, and a third for video.

In the schemas

  • Relevance Feedback Result SignalsRelevanceFeedbackResultSignalssignals on each web result
  • Starburst Local Relevance Feedback Twiddler Ranklab SignalsStarburstLocalRelevanceFeedbackTwiddlerRanklabSignalsthe image variant, with Starburst

On the image side, models judge the image itself

  • Multimodalcombines image and text to score an image
  • Image aestheticsscores an image’s aesthetic quality, for the query and for each result
  • Lens sensitivityspots sensitive content in Lens
01

Expertise, authenticity, topic authority: Google’s quality notions each have their own twiddler

Google talks about E-E-A-T and topic authority in its guidelines and on its blogs. Each of these notions has a twiddler: they weigh in the score, then can still move a page after it.

Expertise

Expertise, one of the four criteria of E-E-A-T in the quality rater guidelines, has its twiddler, with features computed for each document. Expertise is therefore judged page by page.

In the schemas

  • Expertise Twiddler Doc FeaturesExpertiseTwiddlerDocFeaturesexpertise features of each document
  • Expertise Twiddler ParamsExpertiseTwiddlerParamstwiddler settings

Authenticity

Authenticity is measured for the query and for each result. An approximate version of Ascorer, the main ranker, also computes an authenticity boost: the notion counts during ranking, then after it.

In the schemas

  • Authenticity Twiddler Query Level SignalsAuthenticityTwiddlerQueryLevelSignalsquery-level signals
  • Authenticity Twiddler Result Level SignalsAuthenticityTwiddlerResultLevelSignalsresult-level signals

Topic authority

Google described its topic authority for news in May 2023: a local outlet rises for the news events it covers in its city, a specialist outlet for its field (health, politics, finance). The twiddler has features for the query and for each document: a site’s authority is measured on the topic of the query. Another twiddler attaches authority data to each result.

In the schemas

  • Topic Authority Twiddler Doc FeaturesTopicAuthorityTwiddlerDocFeaturesfeatures of each document
  • Topic Authority Twiddler Query FeaturesTopicAuthorityTwiddlerQueryFeaturesquery features

Web health

Web health signals come with each result. At Google, web health also covers citation-based originality, which gives weight to the outlet others cite for a piece of news.

In the schemas

  • Web Health Twiddler Result SignalsWebHealthTwiddlerResultSignalsweb health signals on each result

The same family also includes a “hero” document, the domain classification already named in the leak, and two twiddlers known by code names, Natural Pearls and Starbow.

02

The 120 twiddlers, family by family

Each of these 120 twiddlers reached production after experiments and a launch review: at Google, every proposed change goes before “our most experienced engineers and data scientists”, who decide whether it launches (How Search works). Once launched, a twiddler is there to last.

Web results: quality, authority, trust 9

  • Expertise

    Expertise Twiddler Doc FeaturesExpertiseTwiddlerDocFeatures Expertise Twiddler ParamsExpertiseTwiddlerParams

  • Authenticity

    Authenticity Twiddler Query Level SignalsAuthenticityTwiddlerQueryLevelSignals Authenticity Twiddler Result Level SignalsAuthenticityTwiddlerResultLevelSignals

  • Topic authority

    Topic Authority Twiddler Doc FeaturesTopicAuthorityTwiddlerDocFeatures Topic Authority Twiddler Query FeaturesTopicAuthorityTwiddlerQueryFeatures Topic Authority Twiddler ParamsTopicAuthorityTwiddlerParams

  • Authority data

    Authority DataAuthorityData

  • Web health

    Web Health Twiddler Result SignalsWebHealthTwiddlerResultSignals

  • Natural Pearls

    Natural Pearls Twiddler Query SignalsNaturalPearlsTwiddlerQuerySignals Natural Pearls Twiddler Result SignalsNaturalPearlsTwiddlerResultSignals

  • Hero document

    Hero Doc DataHeroDocData

    In web design, the hero is the large block featured at the top of a page; the visual digest SERP layout has its own “hero” block too.

  • Domain classificationleak 2024

    Domain ClassificationDomainClassification

    It follows each result and the whole result set. The leak already named it (qualityTwiddlerDomainClassification).

  • Starbow

    Starbow Query SignalsStarbowQuerySignals Starbow Result SignalsStarbowResultSignals

AI and rankers 6

  • RankBrain

    RankBrain Query FeaturesRankBrainQueryFeatures RankBrain Twiddler Result SignalsRankBrainTwiddlerResultSignals RankBrain Twiddler Result AnnotationRankBrainTwiddlerResultAnnotation

  • Relevance oracle

    Relevance Oracle Twiddler ParamsRelevanceOracleTwiddlerParams Relevance Oracle Twiddler Result SignalsRelevanceOracleTwiddlerResultSignals Relevance Oracle Twiddler Query Input SignalsRelevanceOracleTwiddlerQueryInputSignals Relevance Oracle Twiddler Result Input SignalsRelevanceOracleTwiddlerResultInputSignals

  • Search plan (fan-out)

    Planned Retrieval AnnotationsPlannedRetrievalAnnotations

  • Bard filter

    Bard Filter Twiddler DataBardFilterTwiddlerData Bard Filter Twiddler ParamsBardFilterTwiddlerParams

  • For loop ranker

    FLR Output Query SignalsFLROutputQuerySignals FLR Output Result SignalsFLROutputResultSignals Query Param SignalsQueryParamSignals

  • Relevance feedback

    Relevance Feedback Result SignalsRelevanceFeedbackResultSignals

Spam, removals and safety 12

  • SiteBoostleak 2024

    Site Boost Spam Query InfoSiteBoostSpamQueryInfo Site Boost Kemtek InfoSiteBoostKemtekInfo

  • Spam demotions

    Heavily Spam Demoted ResultsHeavilySpamDemotedResults Demoted ResultDemotedResult

    The list of results heavily demoted for spam, kept for the whole page, and a record on each demoted result.

  • DMCA site2018 document

    Dmca Site Twiddler SignalsDmcaSiteTwiddlerSignals

  • Crime demotion

    Crime Demotion ParamsCrimeDemotionParams Crime Demotion Query SignalsCrimeDemotionQuerySignals Crime Demotion Result SignalsCrimeDemotionResultSignals

  • Bad URLs

    Bad URLs AnnotationBadURLsAnnotation

    Google checks bad URLs query by query, and this twiddler keeps their annotation.

  • Document filter

    Doc Filter InfoDocFilterInfo

  • SafeSearch

    SafeSearch Twiddler Optt ParamsSafeSearchTwiddlerOpttParams Image Result SignalsImageResultSignals

    The SafeSearch twiddler itself, with image signals on each result.

  • SafeSearch blur

    SafeSearch Blur Twiddler Optt ParamsSafeSearchBlurTwiddlerOpttParams

  • SafeSearch adaptive violence

    SafeSearch Adaptive Violence Optt ParamsSafeSearchAdaptiveViolenceOpttParams

    The handling of violence in SafeSearch, depending on the query.

  • SafeSearch EDSA porn

    SafeSearch Edsa Porn Optt ParamsSafeSearchEdsaPornOpttParams

    SafeSearch settings for pornography.

  • Revenge porn filter

    Revenge Porn Filter ParamsRevengePornFilterParams Revenge Porn Filters SignalsRevengePornFiltersSignals

    Revenge porn is intimate imagery shared without the person’s consent. Google’s ranking systems guide provides for demoting sites that receive a significant volume of removals of such imagery, and Google announced ranking changes against explicit deepfakes in July 2024.

  • Kid-friendly search

    Kid Friendly Search Twiddler ParamsKidFriendlySearchTwiddlerParams

    For supervised children’s accounts, which get a restricted search.

Freshness, repeat queries and preferences 14

  • Topic affinity

    Topic Affinity Twiddler Query SignalsTopicAffinityTwiddlerQuerySignals Topic Affinity Twiddler Result SignalsTopicAffinityTwiddlerResultSignals

  • Freshnessleak 2024

    Freshness Twiddler Doc AnnotationsFreshnessTwiddlerDocAnnotations Freshness Twiddler Query AnnotationsFreshnessTwiddlerQueryAnnotations Freshness Twiddler 2 ParamsFreshnessTwiddler2Params Freshness Twiddler 2 HuVo ParamsFreshnessTwiddler2HuVoParams

  • Date score

    Date Score Twiddler ParamsDateScoreTwiddlerParams

    Its settings travel with each query.

  • View demotion on repeat queries

    View Demotion Twiddler ParamsViewDemotionTwiddlerParams

    When the user repeats a query, the results they have already seen move down. The Discover feed and the image feed also push back what has already been seen.

  • Freshest on repeat queries

    Freshest Twiddler AnnotationsFreshestTwiddlerAnnotations Freshest Substrate Twiddler ParamsFreshestSubstrateTwiddlerParams Freshest Substrate Twiddler AnnotationsFreshestSubstrateTwiddlerAnnotations Freshest Substrate Result Twiddler AnnotationFreshestSubstrateResultTwiddlerAnnotation

    On a repeated query, the freshest content can move ahead; one variant relies on the stateful layer.

  • Stateful substrate

    Stateful Substrate Twiddler ParamsStatefulSubstrateTwiddlerParams Stateful Substrate Annotation SignalsStatefulSubstrateAnnotationSignals

    A layer that keeps state from one repeated query to the next.

  • Intent personalized boost

    Intent Personalized Boost Twiddler ParamsIntentPersonalizedBoostTwiddlerParams

    A boost by the user’s intent, built on the stateful layer.

  • Recall boost

    Recall Boost Annotation SignalsRecallBoostAnnotationSignals

    Tied to repeat queries; its signals travel with each result.

  • Revisitable results

    Revisitable Results AnnotationRevisitableResultsAnnotation

    The results a user comes back to, on repeated queries.

  • Sirius suggestions

    Sirius Suggest Annotations Twiddler ParamsSiriusSuggestAnnotationsTwiddlerParams Sirius Suggest Annotations SignalsSiriusSuggestAnnotationsSignals

    Sirius, one of the sources of a query’s interpretations, provides suggestions here for repeat queries.

  • Domain preference

    P13n Rank Domain Preference Twiddler ParamsP13nRankDomainPreferenceTwiddlerParams P13n Rank Domain Preference Twiddler Result AnnotationP13nRankDomainPreferenceTwiddlerResultAnnotation

  • Pref

    Pref Twiddler ParamsPrefTwiddlerParams

    The user’s preferences.

  • For you

    For You Result SignalsForYouResultSignals

  • ULP

    Ulp Query SignalsUlpQuerySignals Ulp Result SignalsUlpResultSignals

SERP layout and verticals 26

  • Optionalization filter

    Optionalization Filter Twiddler Query AnnotationsOptionalizationFilterTwiddlerQueryAnnotations Optionalization Exported SignalsOptionalizationExportedSignals

    Optionalization is the rewriting step that makes some words of the query optional. This filter annotates the query and exports its signals.

  • Post-results QRef

    Post Results Qref Twiddler ParamsPostResultsQrefTwiddlerParams

    QRef, the annotation of the query’s entities, run again after the results.

  • Host path

    Host Path Twiddler ParamsHostPathTwiddlerParams

  • Web clusters

    Web Cluster AnnotationWebClusterAnnotation Potential Cluster AnnotationPotentialClusterAnnotation

    Groups of web results, with a cluster annotation and a potential cluster annotation.

  • Dynamic SERP

    Dynamic Srp Twiddler Query SignalsDynamicSrpTwiddlerQuerySignals Dynamic Srp SignalsDynamicSrpSignals

    A SERP whose shape changes with the query; its schemas track web results, location, each result’s ranking and similar shopping products.

  • About this result

    About This Result Twiddler ParamsAboutThisResultTwiddlerParams

    The panel that opens from the icon Google placed next to most results in February 2021 says where the result comes from and describes the site behind it; it has its twiddler.

  • Forums

    Forum Twiddler ParamsForumTwiddlerParams

    Google added a Discussions and forums block in September 2022. This forum ranking twiddler receives its settings with the query.

  • Nearby

    Nearby Twiddler Query SignalsNearbyTwiddlerQuerySignals Nearby Twiddler Result SignalsNearbyTwiddlerResultSignals Nearby Twiddler Doc AnnotationsNearbyTwiddlerDocAnnotations

  • Hashtags

    Hashtag Twiddler Result AnnotationHashtagTwiddlerResultAnnotation Hashtag Twiddler Response AnnotationHashtagTwiddlerResponseAnnotation

    Hashtag search has its own results, and Discover its trending hashtags.

  • Hashtag firehose

    Hashtag Firehose Twiddler ParamsHashtagFirehoseTwiddlerParams

    A firehose is a continuous real-time stream of posts.

  • Hashtag repeat query

    Hashtag Repeat Query Twiddler ParamsHashtagRepeatQueryTwiddlerParams

    The repeat-query logic, for hashtags.

  • Visual digest: social media

    Social Media Twiddler ParamsSocialMediaTwiddlerParams

    Visual digest is a SERP layout, with a hero block and web, image, video, news, fact and social media results; this twiddler handles its social media part.

  • Travel: several destinations

    Multidestination Twiddler ParamsMultidestinationTwiddlerParams

    Short travel videos that cover several destinations.

  • Seek volume

    Seek Volume Twiddler ParamsSeekVolumeTwiddlerParams

  • Specialty merchants

    Specialty Merchants Query SignalsSpecialtyMerchantsQuerySignals Specialty Merchants Result SignalsSpecialtyMerchantsResultSignals

    Measured for the query and for each result.

  • Deals

    Deals Result SignalsDealsResultSignals Deals Twiddler ParamsDealsTwiddlerParams

    In November 2023 Google opened a deals destination in Search that gathers on-sale products from across the web.

  • Shopping cluster annotator

    Shopping Cluster AnnotationShoppingClusterAnnotation Shopping Cluster Annotator ParamsShoppingClusterAnnotatorParams

    It groups shopping results. The 2018 document described annotator twiddlers, which prepare information for the stages that follow.

  • Shopping result tags

    Shopping Result Tag AnnotationShoppingResultTagAnnotation Shopping Result Tag Annotator ParamsShoppingResultTagAnnotatorParams

    Another annotator, which puts tags on shopping results.

  • Shopping personalization

    Shopping P13n Twiddler Result AnnotationShoppingP13nTwiddlerResultAnnotation P13n Result Annotation For ShoppingP13nResultAnnotationForShopping

    A personal adjustment of shopping results.

  • Scubed Spark

    Spark Twiddler InfoSparkTwiddlerInfo

    Scubed is a Superroot backend serving varied verticals (retail, real estate, restaurant dishes, movies and TV, audiobooks) and producing embedding annotations; Spark and UCS are its two twiddlers.

  • Scubed UCS

    Ucs Twiddler InfoUcsTwiddlerInfo

    The other Scubed twiddler; the schemas also name UCS embedding annotations.

  • UCP

    UCP Twiddler ParamsUCPTwiddlerParams

    Users’ contributions in Search, such as the discussions Google hosts about a match or a series; each contribution stays tied to its author’s account.

  • Search by entity

    Search By Entity Optt ParamsSearchByEntityOpttParams

    After a click on an entity, the next SERP lines up around it: the state passed from one search to the next can switch this twiddler off and pushes back results with no link to the chosen entity.

  • Signal outliers

    Signal Outliers Twiddler ParamsSignalOutliersTwiddlerParams

  • Date filter

    DATEFILTER TWIDDLERDATEFILTER_TWIDDLER

    Named among the reasons for putting forward a search tools option, with prices, boosts (of the tool, geographic, cross-language), results already visited and the recipe tool. The date score has its own twiddler.

  • Prices

    PRICES TWIDDLERPRICES_TWIDDLER

    One of the reasons for putting forward a search tools option, like the date filter.

Images and Lens 20

  • Image boostleak 2024

    Image Boost Twiddler Doc SignalsImageBoostTwiddlerDocSignals Image Boost Twiddler Query SignalsImageBoostTwiddlerQuerySignals

    The leak named WebImageBoostTwiddler, which reads, for the images of web results, the EQ and TQ scores, the image’s emotional and technical quality.

  • Demographic diversityleak 2024

    Demographic Diversity Twiddler ParamsDemographicDiversityTwiddlerParams Demographic Diversity Twiddler Query SignalsDemographicDiversityTwiddlerQuerySignals Demographic Diversity Twiddler Result SignalsDemographicDiversityTwiddlerResultSignals

    The leak named a Skin Tone Twiddler. In May 2022, Google presented the Monk Skin Tone scale, to show a wider range of skin tones in image results; the image viewer lets users report that a group is over- or under-represented.

  • Image aesthetics

    Image Aesthetics Twiddler Ranklab Query SignalsImageAestheticsTwiddlerRanklabQuerySignals Image Aesthetics Twiddler Ranklab Result SignalsImageAestheticsTwiddlerRanklabResultSignals

  • Commercial images

    Image Commercial Twiddler Ranklab Query SignalsImageCommercialTwiddlerRanklabQuerySignals Image Commercial Twiddler Ranklab Result SignalsImageCommercialTwiddlerRanklabResultSignals

  • Image viewer commercial

    Image Viewer Commercial Twiddler ParamsImageViewerCommercialTwiddlerParams

    The commercial side of the image viewer, the panel that opens on an image.

  • Low-relevance images

    Image Low Relevance Twiddler ParamsImageLowRelevanceTwiddlerParams

  • Image viewer host diversity2018 document

    Image Viewer Host Diversity Twiddler ParamsImageViewerHostDiversityTwiddlerParams

    The 2018 document describes ImageHostCategorizer, which keeps a single host from placing too many images in the same image block.

  • Inspiration diversity

    Image Inspiration Diversity Twiddler ParamsImageInspirationDiversityTwiddlerParams

  • Implication association

    Implication Association Twiddler ParamsImplicationAssociationTwiddlerParams Implication Association Triggering InfoImplicationAssociationTriggeringInfo

    With its settings and triggering conditions. WebRef, the Google system that recognizes entities in a text, also links entities through an implication.

  • More like these

    More Like These Twiddler ParamsMoreLikeTheseTwiddlerParams More Like These Result AnnotationMoreLikeTheseResultAnnotation More Like These Response AnnotationMoreLikeTheseResponseAnnotation

  • Repeat query (images)

    Repeat Query Twiddler ParamsRepeatQueryTwiddlerParams

    The repeat-query logic, for image search.

  • Guided chips

    Image Online Guided Chips Creator ParamsImageOnlineGuidedChipsCreatorParams

    The refinement buttons above image results.

  • Inappropriate content filterleak 2024

    IU Inappropriate Filter Twiddler ProtoIUInappropriateFilterTwiddlerProto

    It removes inappropriate images from the image block of web results (Images Universal). The leak already named this Superroot filter.

  • Multimodal

    Multimodal Twiddler Ranklab Result SignalsMultimodalTwiddlerRanklabResultSignals

    It combines several modalities, such as image and text.

  • Starburst relevance feedback

    Starburst Local Relevance Feedback Twiddler Ranklab SignalsStarburstLocalRelevanceFeedbackTwiddlerRanklabSignals

    Starburst is an image embedding, already cited in the 2024 leak: images are scored again by how much they resemble the query’s top results.

  • Image shopping personalization

    Image Shopping P13n Twiddler ParamsImageShoppingP13nTwiddlerParams

    A personal adjustment of shopping images.

  • Sensitive shopping images

    Image Sensitive Shopping Twiddler ParamsImageSensitiveShoppingTwiddlerParams

  • VSS

    Vss Ranklab Twiddler ResultVssRanklabTwiddlerResult

    In the leak, VSS is the image understanding stack: logo recognition, OCR, deep tags of shopping images.

  • Lens sensitivity

    Lens Sensitivity Twiddler Ranklab Query SignalsLensSensitivityTwiddlerRanklabQuerySignals Lens Sensitivity Twiddler Ranklab Result SignalsLensSensitivityTwiddlerRanklabResultSignals

    Sensitive content in Lens.

  • Lens non-exact match filter

    Non Exact Match Filter Twiddler ParamsNonExactMatchFilterTwiddlerParams

Video and YouTube 33

  • Toledo

    Toledo Twiddler Query SignalsToledoTwiddlerQuerySignals Toledo Twiddler Result SignalsToledoTwiddlerResultSignals

  • YouTube deduplication2018 document

    Youtube Deduplication Twiddler Result AnnotationYoutubeDeduplicationTwiddlerResultAnnotation

    The 2018 document described YoutubeDuplicatesRemovalTwiddler, which puts the original upload above its copies.

  • YouTube clickbait

    Youtube Clickbait Twiddler Query AnnotationsYoutubeClickbaitTwiddlerQueryAnnotations

  • Shorts authenticity

    Youtube Shorts Authenticity Twiddler Result AnnotationYoutubeShortsAuthenticityTwiddlerResultAnnotation

  • Nonsensical YouTube videos

    Youtube Nonsensical Twiddler Result AnnotationYoutubeNonsensicalTwiddlerResultAnnotation

  • Channel diversity

    Youtube Channel Diversity Twiddler Result AnnotationYoutubeChannelDiversityTwiddlerResultAnnotation

  • Playback speed

    Youtube Playback Speed Twiddler Result AnnotationYoutubePlaybackSpeedTwiddlerResultAnnotation

  • Seeks

    Youtube Seeks Twiddler Result AnnotationYoutubeSeeksTwiddlerResultAnnotation

    Viewers’ jumps (seeks) from one point of a video to another.

  • Video passages

    Video Passages Twiddler ParamsVideoPassagesTwiddlerParams

  • Segment topicality

    Video Segment Topicality Twiddler ParamsVideoSegmentTopicalityTwiddlerParams

    Topicality measured for each segment of the video.

  • Creator affinity

    Creator Affinity Twiddler ParamsCreatorAffinityTwiddlerParams

  • Brand affinity

    Brand Affinity Twiddler ParamsBrandAffinityTwiddlerParams

  • Video engagement

    Video Engagingness Twiddler ParamsVideoEngagingnessTwiddlerParams

  • Video thumbnail

    Video Thumbnail Twiddler ParamsVideoThumbnailTwiddlerParams

  • Video aspect boost

    Video Aspect Boost Twiddler ParamsVideoAspectBoostTwiddlerParams

    A boost by the ratio of image width to height (aspect), vertical or horizontal.

  • Short duration demotion

    Short Duration Demotion Twiddler ParamsShortDurationDemotionTwiddlerParams

  • Pulsar relevance feedback

    Video Pulsar Relevance Feedback Twiddler ParamsVideoPulsarRelevanceFeedbackTwiddlerParams

    The video variant of relevance feedback; the image one bears the name Starburst.

  • SVU duplicates

    Svu Duplicates Twiddler ParamsSvuDuplicatesTwiddlerParams

  • SVU diversity

    Svu Diversity Twiddler ParamsSvuDiversityTwiddlerParams

  • Local personalization (video)

    Local Personalization Twiddler ParamsLocalPersonalizationTwiddlerParams

  • YouTube localization

    Youtube Localization Twiddler Query AnnotationYoutubeLocalizationTwiddlerQueryAnnotation

    Adaptation to the user’s language and country.

  • Playlist seeking

    Youtube Playlist Seeking Twiddler Doc AnnotationsYoutubePlaylistSeekingTwiddlerDocAnnotations

  • QPrime boost

    QPrime Boost Twiddler Result AnnotationQPrimeBoostTwiddlerResultAnnotation

    In the leak, QPrime is a query classifier that assigns shopping categories; this boost depends on it.

  • YouTube shopping

    Youtube Shopping Twiddler Response AnnotationsYoutubeShoppingTwiddlerResponseAnnotations

  • Shopping categorical refinements

    Shopping Categorical Refinements Ranking Twiddler Result AnnotationShoppingCategoricalRefinementsRankingTwiddlerResultAnnotation

    The ranking of refinements by product category, in video results.

  • Shorts ranking

    Youtube Shorts Ranking Twiddler Response AnnotationYoutubeShortsRankingTwiddlerResponseAnnotation

  • YouTube safety

    Youtube Safety Twiddler Result AnnotationYoutubeSafetyTwiddlerResultAnnotation

  • YouTube demographics

    Youtube Demographics Twiddler Doc AnnotationsYoutubeDemographicsTwiddlerDocAnnotations Youtube Demographics Twiddler Response AnnotationsYoutubeDemographicsTwiddlerResponseAnnotations

  • RWJ

    Youtube Rwj Twiddler Result AnnotationYoutubeRwjTwiddlerResultAnnotation

    The leak spells RWJ out as Real World Journey, Knowledge Graph data under governance rules; this YouTube twiddler bears the same acronym.

  • Beehive

    Beehive Signals Twiddler ParamsBeehiveSignalsTwiddlerParams

  • Aquila

    Aquila Twiddler Result AnnotationAquilaTwiddlerResultAnnotation

  • CatNorvi

    CatNorvi Twiddler Result AnnotationCatNorviTwiddlerResultAnnotation

  • Shorts efficiency

    Youtube Shorts Efficiency Twiddler Result AnnotationYoutubeShortsEfficiencyTwiddlerResultAnnotation

03

Spam, copyright, crime: what pushes a well-scored result down

A page can come out of ranking in a good spot and move down afterwards. Spam still weighs at this stage: the results page keeps the list of results heavily demoted for spam, and SiteBoost, named in the leak, spots user-generated spam pages and relies on the site’s main category to judge whether a result matches the query.

Copyright targets the whole site. Since August 2012, Google has demoted sites that receive many valid removal notices. In 2018, the DMCAFilter removed the targeted results; today the twiddler works at site level: past a certain volume of notices, the whole domain goes down.

On a crime-related search, pages can move down after ranking: this demotion is measured for the query and for each result. Google’s ranking systems guide describes a related demotion, for sites that charge to remove personal information.

SafeSearch has its twiddler and variants: blurring of explicit images, announced in February 2023, violence according to the query, pornography, revenge porn. Kid-friendly search for children’s accounts has its own twiddler, and another keeps the annotation of bad URLs, which Google checks query by query.

04

Repeat queries, preferences, proximity: what the user moves

Take a query typed again the next day. A whole family of twiddlers handles it: results already seen move down, the most recent content moves ahead, results the user comes back to are spotted, and a stateful layer remembers from one search to the next. A recall boost is tied to the same repeat queries.

For you, domain preference, topic affinity

For you carries personal re-ranking, and personalized ranking’s domain preference annotates each result. Topic affinity brings results closer to the topics the user cares about. For news, Google launched Preferred sources in August 2025: users pick the sites they want to see more often in Top Stories.

Preferred languages and proximity

ULP, the user’s preferred languages in the leak, has signals for the query and for each result. Nearby measures how close each result is to the user: a nearby business can gain places.

Freshness and date score

The leak’s FreshnessTwiddler reads a page’s dates and the date Google estimates for its content, from the text, the anchors and related pages. It has a second version, and a date score receives its settings with each query. Google describes these “query deserves freshness” systems in its ranking systems guide.

05

Images and YouTube: clickbait, inauthentic Shorts, diversity of results

Images: aesthetics, diversity, inappropriate content

Images have 20 twiddlers. Among them: the image boost and demographic diversity, both named in the leak, aesthetics, commercial images, low-relevance images, guided chips above the results and a filter for inappropriate content. In the leak, the image boost for web results reads an image’s emotional (EQ) and technical (TQ) quality.

YouTube: clickbait, inauthentic Shorts, video types

Video results have 33 twiddlers. Two of them echo announcements. In December 2024 YouTube announced stronger action against egregious clickbait, starting in India, and in July 2025 it renamed its repetitious content monetization policy “inauthentic content”. Both topics have their twiddler: clickbait, which annotates the query, and the authenticity of each Short. Others judge nonsensical videos, channel diversity, playback speed and viewers’ jumps within a video.

Toledo sorts YouTube results into types: learning, information seeking, entertainment, dubbed videos, audio match.

06

For SEO: a position is also decided after the score

  1. Work on expertise page by page, on authority topic by topic

    Expertise, authenticity and topic authority each have their twiddler. Expertise is judged document by document; topic authority is built on a topic and over time.

  2. Treat DMCA notices as a domain-wide risk

    Past a certain volume of valid notices, the whole site goes down, including its flawless pages.

  3. Date your content for real

    Freshness twiddlers read a page’s dates and the date Google estimates for its content. A real update, clearly dated, helps a page move ahead of an old comparison.

  4. Get chosen as a source

    For you and domain preference apply the user’s habits and choices after ranking. In Top Stories, Preferred sources lets everyone pick their outlets: a site its readers add shows up there more often for them.

  5. Reckon with forums, deals and specialists

    Forum ranking, deals and specialty merchants each have their twiddler. On a shopping query, a discussion, a deals page or a product specialist can move ahead of a better-scored page: a brand gains from being present there.

07

How the twiddler framework works

An internal Google document from 2018 describes the mechanism; we presented it in 2024 in Google leak, part 2: understanding the twiddler framework. The framework is part of Superroot, the system that sends the query to Google’s indexes and then assembles their answers. It reworks the results of a single corpus, web, images or videos, once the backend has scored them.

  1. Index retrievalThe backend retrieves and scores hundreds of results, with Ascorer, the main ranker.
  2. Predoc twiddlersThey work on the whole list, results still thin, without snippets.
  3. ReorderingThe framework reconciles their recommendations and reorders the list.
  4. Lazy twiddlersSnippets for the top results arrive; lazy twiddlers work on that head of the list. If they push some out, the framework fetches the next ones.
  5. AssemblyTangram assembles web, image, video and other results into one page.

What a twiddler can recommend

  • Boostmove a result up or down
  • BoostAboveResultplace a result above another
  • filterremove a result
  • max_totalcap the number of results of one kind
  • stridespace out results of the same category
  • Annotatingleave information for later steps

Isolated twiddlers, a framework that reconciles

Each twiddler works alone, unaware of the others’ decisions: it recommends, the framework decides. The document also noted that a ranking experiment is easier to run with a twiddler: a few Superroot jobs are enough, where the alternative is building an Ascorer section or bringing up 1,400 jobs. That isolation and that ease explain their number.

RankLab, the system that records Google’s ranking signals, keeps track of every pass: the settings, the response and the results of the twiddlers, with a log and a debug table. Every adjustment can thus be analyzed, query by query.

The twiddlers of the 2018 document and the leak, and their counterpart today

Named in 2018 or in the leakToday
Image Host CategorizerImageHostCategorizer2018 documentHost diversity in the image viewer
Official Page TwiddlerOfficialPageTwiddler2018 document: an entity’s official page in position 1 on the query for its nameThe same idea travels in the state one search passes on to the next: pin the entity’s “about” part at the top, or stop forcing its knowledge panel into position 1
Youtube Duplicates Removal TwiddlerYoutubeDuplicatesRemovalTwiddler2018 documentYouTube deduplication
DMCA FilterDMCAFilter2018 documentDMCA site, at site level
Site Boost TwiddlerSiteBoostTwiddlerleak 2024SiteBoost
Freshness TwiddlerFreshnessTwiddlerleak 2024Freshness, in a second version, and date score
Web Image Boost TwiddlerWebImageBoostTwiddlerleak 2024Image boost
Skin Tone Twiddlerleak 2024Demographic diversity of images
quality Twiddler Domain ClassificationqualityTwiddlerDomainClassificationleak 2024Domain classification

The knowledge panel placed in position 1 ties this second pass to the Knowledge Graph: how Google feeds, maintains and grows its Knowledge Graph, part 5 of the series.

Upstream, twiddlers also adjust the meaning Google settles on for a query

The twiddler mechanism also works upstream, on what the query means. KScorer, the Google Search component that scores a query’s interpretations, that is, its possible meanings (jaguar the animal or jaguar the car), runs its scores through twiddlers twice: as soon as the query arrives, then again once results are in, based on the meaning most retrieved pages cover. Among these twiddlers, a freshness boost favors the meaning that is in the news, and a periodicity boost the one that comes back on a fixed date, such as a yearly tournament or Black Friday. The adjusted score, the k-score, is the one that counts from then on. Above 0.60 or 0.65, two signals mark it as good, and other systems take it up, including Google Assistant: a comment from Google’s engineers states that its k-score “comes from twiddled_ranking_scores in query interpretations”. A twiddler can therefore flip the meaning Google settles on, and the whole SERP with it: on the day Jaguar unveils a new model, the car can move ahead of the animal.