Google to level the SEO playing field

For years, the SEO industry has operated under a quiet but critical assumption: Google’s ranking system evaluates only a small batch of roughly 20 to 30 candidate pages before settling on final results. This wasn’t a design choice rooted in algorithmic purity. It was a hardware constraint, plain and simple.

Now, that constraint is being challenged. A new technique from Google Research, combined with public testimony from the company’s top executives, suggests the rules of the SEO game are about to change.

The bottleneck has always been computational cost. In October 2023, during the United States v. Google antitrust trial, Google VP of Search Pandu Nayak confirmed under oath that RankBrain, a deep-learning ranking component, is only applied to the top 20 or 30 documents. Why? Because running it on a larger set is simply too expensive. “RankBrain is too expensive to run on hundreds or thousands of results,” Nayak stated, a point he reiterated four times during cross-examination.

This reveals a fundamental truth about Google’s architecture. The retrieval system first culls the entire index down to “tens of thousands” of pages using classical methods. From that pool, only the top 20 to 30 reach the expensive deep-learning layer. The industry has long treated RankBrain and BERT as the core of Google’s ranking. The reality, as sworn testimony shows, is that they are expensive optional layers applied only after a narrow window has already been carved out.

That window has remained small not because it’s optimal, but because Google’s hardware budget couldn’t support a larger one. The number could have been 50 or 500. It settled at 20 to 30 due to supply constraints. And those constraints are now in the spotlight.

On April 7, Google CEO Sundar Pichai acknowledged the problem on the Cheeky Pint podcast. He described a set of hard supply constraints that capital spending alone cannot solve in the short term. “We are supply-constrained,” Pichai said, pointing to bottlenecks in wafer starts, memory, power, and data center permitting. He emphasized that leading memory companies cannot dramatically improve their capacity through 2026 and 2027.

This matters because nearest-neighbor vector search, the engine behind modern semantic retrieval, is memory-bound. The wider the candidate set, the more memory required. You cannot simply throw money at the problem to create more capacity.

But Google Research may have found a workaround. On March 24, the company published TurboQuant, a technique that achieves 4x to 4.5x compression of vector representations with performance “comparable to unquantized models.” The paper claims indexing time is reduced to “virtually zero,” and it outperforms existing quantization techniques on recall.

While TurboQuant hasn’t been confirmed as deployed in Google Search, the economics it enables are clear. If memory per vector drops by 4x, the cost boundary that held RankBrain at 20 to 30 candidates no longer applies. A system running on the same hardware could plausibly evaluate a candidate set several times larger.

Waiting for SERPs to confirm a wider retrieval window is a losing strategy. By the time rank-tracking tools show the shift, the positioning work of the next cycle will already be done. Three practical shifts are worth making now.

1. Audit retrieval access, not just ranking position. Rank tracking tools measure position within the set. They say nothing about whether a page was eligible for the set in the first place. In classical search, the distinction matters less because the set is narrow. In AI-mediated retrieval, it is the entire game. The fastest check is server log analysis. Look for search index crawlers like OAI-SearchBot, Claude-SearchBot, PerplexityBot, and Applebot, as well as user-driven agents like ChatGPT-User and Claude-User. If your pages aren’t appearing in those logs, they aren’t in the candidate sets.

2. Separate retrieval-friendliness from ranking-friendliness. Ranking signals reward topical authority, link equity, and query-intent match. Retrieval systems reward something different: a clear, self-contained, citable claim that can be extracted without reading the whole document. A page written for ranking often buries its main claim under context and SEO-driven preamble. In a retrieval-ready page, the claim sits in the first 100 words, attached to an entity or statistic a retrieval system can verify, surrounded by evidence worth citing.

3. Stop treating the top 20 to 30 as a fixed target. The window is a hardware constraint that has held for years because no one at Google could afford to widen it. Briefing content against “what ranks in positions 1 to 10” is briefing against a snapshot of a window narrower than it needs to be due to hardware economics. When the economics change, the window will widen. Content built to compete inside a narrow set will face broader competition. The margin goes to content strong enough to enter a wider candidate set from the start.

None of this requires predicting when TurboQuant ships to production. It requires acknowledging that retrieval economics is moving and positioning for what lies on the other side of the move.

The test is simple. Pull your server logs for the last 30 days. Count the retrieval user agents that have hit the pages you care about. If the answer is zero, or close to it, no amount of ranking work will change that. The competitive surface is shifting. The rest follows.