EEAT Mechanics

Starburst relevance feedback: rescoring images against the top results

Starburst image embeddings appear in the 2024 leak. RESONEO says this twiddler rescores images by their similarity to the best results for the query; that mechanism is RESONEO's description, not leak text.

By Thomas Wawra· Published · Version 1.0· Systems referenced: Mustang / SuperRoot etc.

01What the schema and the leak show

The schema object is StarburstLocalRelevanceFeedbackTwiddlerRanklabSignals. Starburst is the name of an image embedding: a model that turns an image into a list of numbers, so that similar images end up close to each other.

The 2024 leak confirms that Starburst embeddings exist and come in versions. One object stores them at frame level, which suggests they are computed not only for still images but also for individual frames of a video. The leak says nothing about this twiddler.

Source per claim (3)
C
The schema object is StarburstLocalRelevanceFeedbackTwiddlerRanklabSignals in the image_twiddler namespace.[1]
B
"Each supported Starburst embedding version will have a pair of AmarnaSignalsBlobInfo and …"[2]
C
The leak object stores Starburst embeddings at frame level, which suggests they are also computed for video frames; the leak does not mention this twiddler.[3]
Source: eeat-mechanics, own reading · Mustang / SuperRoot etc.

02How relevance feedback would work

RESONEO reads the twiddler as follows: after the first ranking, images are scored again by how close their embedding is to those of the best results for the query. "Relevance feedback" is an old information retrieval term for refining results with what the first results reveal; "local" would mean the feedback comes from this one result set, not from the whole index.

If that is right, an image that looks like the top results moves up, and an odd one out moves down, even if its text matches. That is a plausible design, but it rests on RESONEO's description and the name. Like every twiddler, it only recommends; the framework in Superroot, the layer that merges results from Google's indexes before the page is built, makes the call.

Source per claim (2)
C
RESONEO describes the twiddler as rescoring images by their similarity to the best results for the query, using Starburst embeddings.[4]
C
If RESONEO's reading holds, visual outliers in a result set would lose ground even when their text matches the query.[3]
Source: eeat-mechanics, own reading · Mustang / SuperRoot etc.

03E-E-A-T and the system map

This is a relevance mechanism. It compares images with images and says nothing about who published them, so it has no E-E-A-T bearing. On our map it belongs to the twiddler layer. A site owner cannot steer it beyond the obvious: an image that clearly shows what the query asks for has the best chance of resembling the other strong results.

Source per claim (1)
C
Starburst relevance feedback compares images with images and has no E-E-A-T bearing.[5]

04References

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