Google may be about to widen the SEO playing field

SEO has always been a fight for the first page of Google. Every toolchain, audit, and content brief assumes that Google’s ranking systems evaluate a relatively fixed set of roughly 20 to 30 candidate pages before final rankings are determined.

Google has kept that set small because evaluating more pages is computationally expensive.

Google’s VP of Search acknowledged the constraint in federal court. The company’s CEO later confirmed the hardware bottleneck behind it. Google’s research division has now published a technique designed to reduce those costs.

If the candidate set widens, the rules of the last decade stop working.

Why the ranking window is 20 to 30 results wide

Here’s the exchange that matters from Day 24 of United States v. Google in October 2023. DOJ counsel Kenneth Dintzer cross-examining Pandu Nayak, Google vice president of Search, from transcript page 6431:

Q: RankBrain looks at the top 20 or 30 documents and may adjust their initial score. Is that right?
A: That is correct.

Q: And RankBrain is an expensive process to run?
A: It’s certainly more expensive than some of our other ranking components.

Q: So that’s, in part, one of the reasons why you just wait until you’re down to the final 20 or 30 before you run RankBrain?
A: That is correct.

Q: RankBrain is too expensive to run on hundreds or thousands of results?
A: That is correct.

Four consecutive confirmations. The deep-learning component of Google ranking that SEOs have built a decade of theory around is deliberately withheld from the bulk of the index because Google can’t afford to apply it more broadly.

The architecture feeding that reranking window is equally revealing. Earlier in the same testimony, at transcript page 6406, Nayak described classical postings-list retrieval to Judge Mehta: 

• “[T]he core of the retrieval mechanism is looking at the words in the query, walking down the list, it’s called the postings list… [Y]ou can’t walk the lists all the way to the end because it will be too long.” 

The corpus gets culled to “tens of thousands” of pages before ranking begins, and from that pool only the top 20 to 30 results reach the deep-learning layer.

That runs against how most SEO commentary describes Google. The industry treats RankBrain, BERT, and other deep learning components as the definition of how Google ranks. Under oath, Nayak described them as expensive optional layers applied to a narrow window that classical retrieval has already culled.

Every practice in this industry that treats the top 20 to 30 as the competitive surface assumes it’ll stay that size. The testimony makes clear that the assumption is contingent, not foundational. The number could have been 50 or 500. It landed at 20 to 30 because that’s what Google’s hardware budget would support, and the constraint has held.

The constraint that held the number there is now in public view, and Google has published what comes next.

The wall and the algorithm that climbs it

On April 7, Sundar Pichai sat down with John Collison and Elad Gil on the Cheeky Pint podcast and described a set of hard supply constraints that no amount of CapEx will solve in the short term. The operative line: 

• “To be very clear, we are supply-constrained. We are seeing the demand across all the surface areas.”

Pichai named five specific bottlenecks: wafer starts at the foundries, memory, power and energy, permitting for data centers, and skilled labor. Of the five, he pressed hardest on memory: 

• “There is no way that the leading memory companies are going to dramatically improve their capacity.” 

For the 2026 to 2027 horizon, Google can’t buy its way past the memory bottleneck. Higher prices won’t create more capacity.

That matters because nearest-neighbor vector search, the mechanism behind modern semantic retrieval, is memory-bound. The wider the set of candidate pages a system can consider, the more memory it needs. The tight coupling between memory supply and retrieval breadth is what sets the cost boundary Nayak testified about.

On March 24, two weeks before the Cheeky Pint episode, Google Research published a blog post describing a technique called TurboQuant. The corresponding arXiv paper, “TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate,” was authored by researchers at Google Research, Google DeepMind, and NYU.

The headline claims:

• 4x to 4.5x compression of vector representations with performance “comparable to unquantized models” on the LongBench benchmark.
• Nearest-neighbor search indexing time reduced to “virtually zero.”
• Outperforms existing product quantization techniques on recall.

The paper covers two applications: KV-cache compression inside Gemini, and nearest-neighbor search in vector databases. Most coverage has focused on the Gemini application. The search-stack application is the nearest-neighbor-search half, and it’s the one relevant to the cost boundary Nayak described. 

If indexing is virtually free and memory per vector drops by 4x, the economics that held RankBrain at 20 to 30 candidates no longer apply. A system running on the same hardware could plausibly evaluate a candidate set several times larger.

TurboQuant hasn’t been confirmed as deployed in Google Search. TechCrunch reported at the time of announcement that it remained a lab breakthrough, and the March 2026 core update carried no public commentary from Google linking it to retrieval efficiency or vector quantization. Google has published the algorithm but hasn’t yet deployed it.

Google has been running quantized vector search in production for years through ScaNN. TurboQuant extends that approach rather than introducing it.

The question has shifted from whether the cost boundary can be moved to what SEOs do before it moves.

What to do before the boundary moves

Waiting for SERPs to confirm that retrieval has widened before adjusting is the losing strategy. The competitive surface is shifting. By the time it’s visible in rank-tracking tools, the positioning work of the next cycle is already done.

Three practical shifts are worth making now.

1. Measure whether your pages enter candidate sets

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, and in a wider RankBrain-style window once it arrives, the distinction is the entire game.

The fastest check is server log analysis. Two classes of retrieval user agents matter. 

• Search index crawlers build the corpus AI systems pull from. Some examples include:
  • OAI-SearchBot (ChatGPT search).
  • Claude-SearchBot (Claude search).
  • PerplexityBot.
  • Applebot (which also feeds Apple Intelligence). 
• User-driven agents fetch pages on demand when someone asks an AI model about a topic your page covers: ChatGPT-User, Claude-User, and Perplexity-User.
  • These don’t execute JavaScript, so they’re invisible to GA4 and any analytics tool that depends on client-side tags. If the pages you care about aren’t appearing against either list, they aren’t in the candidate sets those systems construct, and ranking work can’t put them there.

---

2. Audit pages for retrieval-friendliness separately from ranking-friendliness

Ranking and retrieval reward different properties. The ranking signals you already know include topical authority, link equity, and query-intent match. Retrieval systems look for something else: a clear, self-contained, citable claim that can be extracted and evaluated without reading the whole document. 

A page written for ranking often buries its main claim under context-setting, caveats, 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, and surrounded by evidence worth citing. Most sites we audit fail this test even when they rank well.

3. Stop treating the top 20 to 30 pages 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 for this query” is briefing against a snapshot of a window that’s narrower than it needs to be because of hardware economics. 

When the economics change, the window will widen. Content built to compete inside a narrow set will face broader competition once it expands. The margin goes to content that was strong enough to enter a wider candidate set from the start.

None of the three requires predicting when TurboQuant or its descendants ship to production. They require acknowledging that retrieval economics is moving and positioning for what lies on the other side of the move, rather than for the current snapshot.

2026 is a year of change for SEO

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 move that number.

The competitive surface is shifting under you. The rest follows.

View the full article