EEAT Mechanics

Relevance feedback: results re-scored by their likeness to good ones

One schema object holds relevance feedback signals per result. The technique itself is decades old in information retrieval. RESONEO names image and video variants; only the image embedding name, Starburst, can be found in the leak.

By Thomas Wawra· Published · Version 1.0· Systems referenced: RankBrain / DeepRank / FastSearch, Mustang / SuperRoot etc.

01What the schema shows

The web version has one schema object, RelevanceFeedbackResultSignals, in the twiddler.relevance_feedback namespace. It holds signals for each result. There is no object for settings or for the query, so the schema shows only that something is attached to every result.

RESONEO adds that Google runs three such twiddlers: one for web results, one for images built on Starburst and one for video built on Pulsar. Like every twiddler, a relevance feedback twiddler only recommends; the framework in Superroot, the layer that fans a query out to Google's indexes and merges what comes back, decides.

Source per claim (2)
C
The web relevance feedback twiddler has one schema object, RelevanceFeedbackResultSignals, with signals per result.[1]
C
According to RESONEO, Google has three relevance feedback twiddlers: for web, for images (Starburst) and for video (Pulsar).[2]

02What the technique is, and what the leak confirms

Relevance feedback is a standard method in information retrieval. Results that are known or assumed to be good serve as a model, and the remaining results are scored again by how close they come to that model. When no user judges the results, the top few are simply assumed to be good; this variant is called pseudo-relevance feedback.

Of the three names, one can be checked. The leaked API documentation describes Starburst as a family of embeddings, numeric representations of content, in several versions. That fits an image variant that compares results by similarity, but the leak does not link Starburst to relevance feedback. We could not confirm Pulsar in the leak, so its role rests on RESONEO alone.

Source per claim (3)
C
Relevance feedback re-scores results by their similarity to results taken as good; in the pseudo variant, the top-ranked results are simply assumed to be good.[3]
B
The leak names Starburst as a versioned embedding: "Each supported Starburst embedding version will have a pair of AmarnaSignalsBlobInfo and …"[4]
C
That Starburst feeds an image relevance feedback twiddler, and that Pulsar does the same for video, is RESONEO's reading; the leak does not state either link.[3]

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

Relevance feedback works on similarity, so on our map it sits next to system 23, RankBrain and DeepRank, which also judge fit rather than quality. It has no direct E-E-A-T bearing.

One property is worth knowing. If Google uses the pseudo variant, the twiddler reinforces whatever already ranks at the top: pages that resemble the leaders gain, pages that differ lose a little. Whether Google uses that variant, and how strong the effect is, is not known.

Source per claim (1)
C
Relevance feedback judges similarity, not quality, and has no direct E-E-A-T bearing; a pseudo-relevance variant would reinforce the existing top results, but whether Google uses it is unknown.[3]

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