Google’s Antitrust AI Overviews Replace Links With Fast Search

Links are not used in the AI Overviews ranking algorithm, as demonstrated by the Google antitrust case.

An experienced search marketer figured out why spammy websites were displayed in Google’s AI Overviews. A line in the most recent Memorandum Opinion in the Google antitrust case provides a hint as to why that occurred and speculates on how it may be a reflection of Google’s decision to no longer use links as a major ranking element.

In the latest Memorandum Opinion, Ryan Jones, the founder of SERPrecon, drew attention to a section that demonstrates how Google grounds their Gemini models.

The Grounding Responses from Generative AI from the passage is found in the section on using search data to ground answers. The ranking of the webpages that an AI model obtains from a search query to an internal search engine is typically influenced by links. In other words, when a user asks a question to Google’s AI Overviews, the system searches Google and then uses the search results to provide a summary.

However, it seems that Google doesn’t operate that way. A different algorithm used by Google fetches fewer web documents more quickly.

It reads that, “Google uses a proprietary technology called FastSearch to ground its Gemini models.” Rem. Tr. (Reid) at 3509:23–3511:4. Based on a collection of search ranking signals called RankEmbed signals, FastSearch produces ranked, condensed web results that can be used by a model to generate a grounded answer. Id. Because it retrieves fewer pages, FastSearch produces results faster than Search, but the quality of the results is inferior to that of Search’s completely ranked web results.

Ryan Jones observed that this it is intriguing and according to her, this supports what many of us believed as well as what we observed in preliminary testing. What does it signify? This indicates that Google has a different search algorithm for grounding. They don’t care about as many signals, but they do require speed. All they need is language to support their claims.

Furthermore, she states that “possibly a lot of spam and quality signals are not calculated for fastsearch either.” That would explain why some spammy websites and even sites that had been penalised appeared in AI overviews in early versions.

He continues by expressing his belief that since the grounding relies on semantic relevance, linkages are irrelevant in this case.

The Memorandum further notes that FastSearch produces just a small number of search results such that FastSearch acts as a technology that rapidly generates limited organic search results for certain use cases, such as grounding of LLMs, and is derived primarily from the RankEmbed model.

So, what exactly is the RankEmbed model? RankEmbed has described as a deep-learning model in the Memorandum. And it puts it simply, as a deep learning model can recognise patterns in large datasets and can determine linkages and semantic interpretations, for instance. It merely recognises patterns and correlations; it does not comprehend anything in the same manner as a human.

Data that is “user-side” is used by RankEmbed. RankEmbed, the foundation of FastSearch, is described as follows in the Memorandum’s section on the type of information Google should give to rivals.

An Alternative View of AI Search, to be accurate is to say that links have no bearing on which websites are chosen for AI Overviews? Speed is a top priority for Google’s FastSearch. According to Ryan Jones’ theory, it might indicate that Google has many indexes, one of which is dedicated to FastSearch and consists of popular websites. That might be a reflection of FastSearch’s RankEmbed component, which combines “click-and-query data” with data from human raters.

It would be impossible for raters to manually rate more than a very small portion of the billions or trillions of pages in an index. Thus, quality-labeled training examples are obtained from the human rater data. A model is trained using labelled data, which makes it easier to see the patterns that are present when determining whether a page is high-quality or low-quality.