TurboQuant Optimisation
Has your GEO agency
told you about TurboQuant?
99% of GEO agencies have never heard of it.
Google released TurboQuant in March 2026: a compression algorithm that changed how AI retrieves and processes information. Publicly traded RAM company stocks dropped 5-10% on release day. The GEO industry has barely noticed. WLDM's delivery workflow was already built for it.
Same answer quality, a fraction of the memory footprint. That's what expanded the candidate pool from 5 sources to 20.
Drop in publicly traded RAM stocks on release day, March 2026
AI Overview answers sourced from the classical top-10, before and after TurboQuant
Google searches now answered by AI Overviews
01. What changed
Google stopped retrieving documents.
It started computing entities.
Before TurboQuant, AI search worked like this: a buyer asks a question, the model finds the closest matching document and summarises it. That process is gone. The model now extracts the entities from the question itself and computes an answer directly from structured facts, instead of retrieving and regurgitating.
If your site is built around keywords and long-form content, you're optimised for a process the model no longer runs.
A buyer types a question. The model finds the single most relevant document and summarises it. Long content and keyword coverage were the levers that mattered, because the model was reading pages one at a time.
Retrieval pipeline
One document in, one summary out. Whoever owned the best-matching page owned the answer.
The model extracts the entities from the question and computes an answer directly, building it from structured facts rather than summarising a single page.
Computation pipeline
Twenty candidate sources instead of five. Whether you're one of them depends on where you sit in the entity graph, not on page content.
When an AI model processes a sequence of tokens, it stores a Key vector and a Value vector for every token already processed. This structure is called the KV cache, the model's working memory. It is also the biggest cost driver in long-context inference.
Two structural problems made that cost spiral. Attention cost scales quadratically with sequence length (O(N²)), so doubling the context roughly quadruples the cost. And even with a large context window, models don't read evenly. They favour the start and end of a document and skim the middle. Throwing more context at the problem was never going to work. Compression was the only path forward.
TurboQuant solves this in two stages. PolarQuant (stage 1) re-encodes the cache vectors in polar coordinates: radius and angle instead of full Cartesian directions. For attention computation, direction matters more than magnitude, and angles cluster predictably across tokens, so they compress cheaply onto a small fixed codebook. QJL (stage 2, Quantized Johnson-Lindenstrauss) adds one extra bit per dimension that corrects stage 1's small directional bias, making the inner-product estimates the model uses for attention scores provably unbiased: mathematically guaranteed not to drift.
This is lossless compression of the model's working memory: on average, the compressed answer matches the uncompressed answer exactly.
| Metric | Result | Source |
|---|---|---|
| KV cache footprint | 4.5× smaller | Google Research |
| Attention speed (H100) | 8× faster (4-bit vs 32-bit) | Google Research |
| Evidence per query | 100K → 450K tokens | Google Research |
| Candidate sources per query | 5 → 20 | Google Research |
| Retraining required | None. Drop-in on existing models | Google Research |
The market repriced it before the GEO industry noticed.
If AI answers need a fraction of the memory they used to, the forecast demand for that memory changes with them. Memory stocks fell in the same week the TurboQuant details went public.
02. The proof
The mechanism is Google's. The numbers are documented.
TurboQuant is a published Google Research algorithm. The compression figures are verified. The case for building computation-grade entity coverage rests on those numbers, not on anything we're asserting ourselves.
AI Overview query coverage (%)
Top-10 pages share in AI citations (%)
Content structural lift to citations (%)
AI signal weight index (relative)
Conversion rate by traffic source (%)
TurboQuant impact score (vs. baseline 50)
Smaller working memory footprint per query
Google Research
Faster computation on current production hardware
Google Research
Retrieval accuracy from structured entity pages vs HTML content
WordLift · arXiv:2603.10700
Candidate sources per query
5 before. 20 after.
Fifteen of those twenty slots didn't exist before March. Whether your brand fills one comes down to reachability in the entity graph, not ranking position.
"It's no longer retrieving and regurgitating.
It's computing entities, creating new answers."
Prior compression algorithms required trade-offs that made them unsafe for production retrieval pipelines. TurboQuant is the first algorithm to combine all four properties required for reliable, high-performance deployment at web scale.
| Algorithm | Unbiased | Codebook-free | GPU-native | Data-oblivious |
|---|---|---|---|---|
| PQ / OPQ | no | no | no | no |
| ScaNN | no | no | partial | no |
| RaBitQ | no | no | yes | no |
| TurboQuant | yes | yes | yes | yes |
- 01Unbiased: attention estimates guaranteed not to drift.
- 02Codebook-free: no per-dataset calibration, so it deploys without retraining.
- 03GPU-native: runs at hardware speed.
- 04Data-oblivious: no prior exposure to the data distribution required.
Source: Google Research blog.
What the structure research shows
Three findings from WordLift (Andrea Volpini) make the empirical case for structured entity pages over prose content. The research is WordLift's, not WLDM's.
- +29.6%AI retrieval accuracy from dedicated entity pages with RDF structured data, over plain HTML and over HTML-plus-schema-without-RDF. (arXiv:2603.10700)
- 71%Win rate for graph traversal (RLM-on-KG) over Microsoft GraphRAG on complex, scattered-evidence reasoning, which is the real-world case for most brands. (arXiv:2604.17056)
- 7%Of AI citations go to pages outside the classical top-10 SERP, reached via entity links not rankings. Empirical proof that AI citation has already diverged from classical ranking.
Schema presence (has_schema) is also an independent variable in WLDM's own 11M+ AI citations dataset. WLDM measures schema as a citation signal, not just asserts it.
What we are not claiming yet
No WLDM client result is tied to TurboQuant Optimisation yet, and we're not going to manufacture one. Our own schema and TurboQuant data study are running now. Everything on this page is either a published Google mechanism or independent WordLift research, sourced line by line. Our own numbers will be published in exactly the same way once they exist.
03. The difference
Stop being built to be read. Start being built to be computed.
| Dimension | Current approach | TurboQuant Optimisation |
|---|---|---|
| What AI encounters | Prose it has to read and parse | Structured facts it can compute directly |
| What gets measured | Rankings and traffic | How often your brand appears in AI answers |
| What the work is | Content and keyword targeting | Fact coverage, structured data, consistency |
| What accountability looks like | Content published | AI retrieval accuracy, before and after |
Your agency's reports probably still track keyword rankings, backlink counts and time on page. None of that tells you whether an AI model can extract a clean fact from your site and cite it.
The gap between what your agency is measuring and what AI now rewards is the gap between your investment and your results.
The unit of optimisation has shifted three times. Each shift made the previous era's techniques not just suboptimal but counterproductive. Optimising for keyword density in the embedding era is noise. Optimising for vector proximity in the graph-traversal era misses the signal that now decides visibility.
| Dimension | GPT-3 era | GPT-4 era | GPT-5.4 era |
|---|---|---|---|
| Method | Pattern match | Semantic similarity | Graph traversal |
| Unit | Token | Embedding | Entity |
| What matters | Keyword density | Vector proximity | Structural connectivity |
| Content's job | Be present | Be similar | Be reachable |
| Failure mode | Not indexed | Low similarity | Disconnected from graph |
You can't optimise for GPT-5.4 using GPT-4 intuitions.Andrea Volpini, WordLift
04. The gap
The GEO industry is still optimising for document retrieval. The model moved on.
The standard GEO playbook (keyword targeting, content volume, embedding proximity) was built for a system that no longer runs. An agency that has never heard of TurboQuant has no way to prepare you for the environment it created, because it doesn't yet know that environment exists.
When Brie opens a pitch with "Have you heard of TurboQuant?", almost every agency and in-house team in the room says no. That gap is the advantage right now. It won't stay open forever.
The standard GEO playbook
Document retrieval
- Keyword targeting
- Content volume
- Embedding proximity
Candidate sources per query5
The model moved on
Entity computation
- Extract entities
- Traverse graph
- Compute answer
Candidate sources per query20
Before provably-unbiased compression, aggressive KV cache compression degraded retrieval accuracy without any visible signal: no errors, no ranking-drop notifications, no warnings. The right pages simply stopped appearing in AI answers, and no one knew why.
Bad embedding compression destroys rankings. Similarity drifts and rankings degrade silently. This is how content disappears from AI answers without anyone noticing.Andrea Volpini, WordLift
Because that failure was silent, AI systems stayed cautious and kept the candidate pool narrow, which is exactly why the 5-source budget held for as long as it did. TurboQuant's lossless guarantee removed that constraint: compression can now be aggressive without sacrificing accuracy, which is what opened the door from 5 candidates to 20.
The structural consequence is that if a brand is absent from the entity graph, the failure is invisible. No content quality metric surfaces it. The fix is always structural, never editorial.
05. The service
TurboQuant Optimisation structures your brand as entities AI can compute.
Under entity computation, the model faces a simple resource decision: parse a 6,000-word page, or compute 20 clean entities on a structured one. It picks the entities every time, because they cost a fraction of the compute.
TurboQuant Optimisation is the work of making your brand those 20 entities. It rests on two pillars: covering every question your buyers actually ask, and staying consistent everywhere your facts appear.
Option A · parse the prose
6,000 words
High cost. Uncertain extraction.
Option B · compute the entities
20 clean entities
A fraction of the resource. It picks this every time.
Four steps from audit to computation-ready coverage.
Step 01
Audit
Map every fact your brand should be known for against what AI can currently find. Most brands cover less than 20% of the questions their buyers actually ask AI.
Step 02
Identify gaps
Where coverage is missing, inconsistent, or in a form that AI cannot compute without additional parsing effort.
Step 03
Rebuild
Structured fact pages, stable identifiers, and consistent fact grounding across third-party sources, so the model can compute your brand cheaply and reach it reliably.
Step 04
Measure
AI retrieval accuracy before and after, independently verifiable. Not our dashboard. Your data.
The GEO program is the foundation. TurboQuant Optimisation is the specialised tier that takes your entity coverage to computation grade.
Four categories of work make a brand computable at computation grade.
- 01Schema skyscrapers: comprehensive structured-data pages built so the model computes the brand's entities without parsing prose at all. The goal is maximum computable density: every entity the brand should be known for, represented in structured markup with stable identifiers, so the model can answer a broad range of queries from a single page traversal.
- 02Article IDs and @id stability: every schema entity, and every reference to it anywhere on the site, resolves to one canonical @id, character for character. This is a computation-grade requirement, not a nice-to-have best practice.
- 03Wikidata entity engineering: the brand's Wikidata QID anchors its identity across all AI systems that read it. Most enterprise QIDs already exist but sit thin, missing category-level and relationship properties. The work here is filling in those missing properties and anchoring the on-page sameAs to the correct QID.
- 04Off-page entity stacks: consistent facts about the brand across third-party sources, with deliberate wording variation to validate the cluster rather than look duplicated. Coordinated directly with the AI Citations service.
Boundary with GEO
GEO delivers the broad on-page foundation: entity graph, @id consistency, crawler compliance, content structure. TurboQuant Optimisation is the specialised tier that pushes entity coverage to computation grade across the full question surface and enforces stability as a hard requirement.
Internal tooling
WLDM's TurboQuant simulators and TurboCon tool are in QA. The schema and entity work described on this page is production-ready today. The research tooling ships behind it.
06. Two pillars
Coverage, then consistency.
Pillar 01 · Coverage
Coverage across every question your buyers ask AI.
Entity breadth means your brand is computable for head terms, comparisons, methodologies and use cases: the full range of questions buyers ask before they ever reach your site.
Most brands only cover the obvious questions. TurboQuant Optimisation maps the entire question surface and builds structured fact coverage for the ones your brand currently doesn't answer at all.
Pillar 02 · Consistency
The same facts, everywhere they appear.
Entity stability is NAP consistency, applied to facts instead of addresses. If your brand resolves in three different ways across your own site, third-party sources and your structured data, the model reads that as three weak, conflicting signals rather than one strong one.
TurboQuant Optimisation locks your entity identity down, on-page and across the web. That's the level of consistency the model actually needs to compute your brand reliably.
The question surface
Typical enterprise coverage: under 20%.
Head termsComparisonsMethodologyLong tail
Entity breadth is the computation-grade extension of GEO's entity inventory. The principle: for every question a buyer asks AI in the category, the model should be able to compute a brand-relevant answer from structured facts without falling back to parsing prose.
Most enterprise brands cover head-term queries well (brand name, primary product, core use case) and cover almost nothing on the long tail: comparisons, methodology questions, use-case variations, integration queries. That long tail is where much of the computation-era citation concentrates, because those are the questions buyers ask AI before they've built a shortlist at all.
The audit maps the full question surface for your category, scores your current entity coverage against it, and flags exactly where a competitor or a generic third-party source is getting computed in your place. The rebuild prioritises the queries with the strongest buyer-intent signal and the weakest current brand coverage.
Schema skyscrapers are the architecture for breadth at scale: a single structured-data page designed to answer a cluster of related buyer questions from computed entities, so the model never needs to fall back to a 6,000-word article for questions in that cluster.
@id discipline: the computation requirement
For a model to compute with any confidence, every reference to the same entity has to resolve to one canonical identifier, character for character. "Acme Corp", "ACME Corp" and "Acme Corporation" are treated as three separate entities by a model performing entity computation, even though a human reader knows they're the same company. The correct canonical @id, applied consistently across every instance of the entity in schema markup, internal links and structured data, is what consolidates three weak signals into one strong one. This is the @id consistency rule from GEO raised from a best practice to a non-negotiable computation requirement.
Wikidata grounding
Wikipedia is the human version; Wikidata is the data version for LLMs.WLDM
A brand's Wikidata QID is the anchor point AI systems use to resolve brand references across the entire web. Most enterprise brands already have a QID that correctly resolves the brand name, but it's missing the category-level properties, relationship properties and industry context needed to make it computable for real buyer questions. The work here is property completeness: filling the QID in with the properties AI systems actually query, and anchoring every relevant entity page back to that correct QID through sameAs.
Off-page entity stacks
The same facts, names and properties need to appear in third-party sources across the web, with deliberate wording variation to prevent the cluster from collapsing to a single source. The variation validates the cluster; the consistency validates the entity.
We create entity stacks across the internet. Consistently we want to change the wording a little bit, then variate it. Once you start clustering those entities together with a little variation, you're validating.Brie Moreau, Founder, WLDM
This work is coordinated with the AI Citations service.
07. The window
The brands that build for TurboQuant now will own the positions that form over the next 18 months.
Full adoption of TurboQuant across AI systems sits on roughly an 18-month horizon. The entity graph you build today is what the model reaches for once that adoption completes.
This isn't a future problem to think about later. It's a present window.
Working with WLDM now means operating in the entity-computation environment while 99% of your competitors are still figuring out what actually changed in March.
Adoption horizon
MAR 2026 · RELEASEDMONTH 9~18 MONTHS · FULL ADOPTION
The floor is set. TurboQuant is arriving regardless of whether a brand prepares for it. The one compounding asset here is your entity graph, and unlike the model itself, that's entirely within your control to build.
TurboQuant is deployed now and its effect on retrieval behaviour is measurable now. The variable you control is not the compression algorithm, which runs inside the model. The variable you control is the quality, breadth and consistency of your entity graph.
Full adoption across all AI systems sits on approximately an 18-month horizon. The entity coverage built during that window is what the model reaches for as adoption completes.
Context windows will keep growing. Models will keep getting cheaper. The variable that compounds is your data connectivity.Andrea Volpini, WordLift
Talk to the team already working in the TurboQuant environment.
A strategy session with Brie takes 30 minutes. You'll leave knowing exactly where your brand sits in the entity-computation era, and what it actually takes to get ahead of the 99% still catching up to March.
Start with your data
See where AI is citing you, and where it is not.
Two fields. Your free backlinks and AI citations audit, read against the questions your buyers ask AI.
If we can't show you something your current agency missed, we'll say so.
Prefer to book directly? Strategy session
The window is open. Most of your competitors haven't heard of TurboQuant yet.