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.
- Index retrievalThe backend retrieves and scores hundreds of results, with Ascorer, the main ranker.
- Predoc twiddlersThey work on the whole list, results still thin, without snippets.
- ReorderingThe framework reconciles their recommendations and reorders the list.
- 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.
- AssemblyTangram assembles web, image, video and other results into one page.
What a twiddler can recommend
Boostmove a result up or downBoostAboveResultplace a result above anotherfilterremove a resultmax_totalcap the number of results of one kindstridespace out results of the same categoryAnnotatingleave 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 leak | Today |
|---|
Image Host CategorizerImageHostCategorizer2018 document | Host diversity in the image viewer |
Official Page TwiddlerOfficialPageTwiddler2018 document: an entity’s official page in position 1 on the query for its name | The 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 document | YouTube deduplication |
DMCA FilterDMCAFilter2018 document | DMCA site, at site level |
Site Boost TwiddlerSiteBoostTwiddlerleak 2024 | SiteBoost |
Freshness TwiddlerFreshnessTwiddlerleak 2024 | Freshness, in a second version, and date score |
Web Image Boost TwiddlerWebImageBoostTwiddlerleak 2024 | Image boost |
Skin Tone Twiddlerleak 2024 | Demographic diversity of images |
quality Twiddler Domain ClassificationqualityTwiddlerDomainClassificationleak 2024 | Domain 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.