Research Lab
Google and AI are data science problems.
Most agencies are not data science companies.
We built WLDM as a data science company that happens to do SEO. Seven research programmes run inside the lab. Every claim on this site comes out of one of them. This page is the working.
The lab, in numbers
Where a figure comes from someone else's research, we say so on the programme it belongs to. If we can't show our working, we don't put a number on the page.
01. The programmes
Seven programmes. Open one to read the methodology.
No other agency runs research at this scale. That is why no other agency can tell you whether a placement will work before it buys it.
R01 Methodology
Two models, scored separately, so the comparison holds.
From August to November 2024, working with DataForSEO, we pulled citations across ChatGPT and Gemini and ran every URL through the same normalise-and-enrich pipeline: authority metrics, real traffic data, schema presence, content type, freshness scores, Harmonic Centrality taken from the Common Crawl graph, and three cosine similarity measurements per citation.
From there we ran two parallel analyses. A Spearman correlation checking citation rank against each individual feature, and a cited-versus-not-cited classification tested against SERP-controlled negatives. Keeping ChatGPT and Gemini scores independent was deliberate: collapsing them would have hidden the exact differences we were looking for. A new study is running now with Search Atlas and Surfer SEO.
R02 Infrastructure
600,000 domains. Half a million vectorised.
WLDM runs a proprietary link database of roughly 600,000 domains, about half a million of them vectorised for semantic scoring. It is the largest agency-side database of its kind.
A keyword list goes in. Extraction, cleaning and enrichment come back out in about 24 hours. The ML pipeline pre-classifies every domain before it reaches a targeting list, so the list that arrives for outreach has already had the weak candidates stripped out.
R03 Tooling
Two are public. Seven are internal.
The two public tools run free in the browser on the same models as the internal pipeline. The AI Citations Audit forces live ChatGPT searches across buyer-intent queries and hands back a prioritised placement plan. The Backlink Relevance Cosine Scorer scores semantic alignment entirely on-device. Nothing leaves your browser.
The seven internal tools cover the ML backlink classification pipeline, a Harmonic Centrality analyser running against both the link graph and the co-citation graph, opportunity extraction and enrichment, and the citation data pipeline built on our 11M+ study dataset.
R04 Machine learning
Five models. 18,000 training domains. 3-of-5 voting.
The classifier was trained on 9,000 link-farm domains and 9,000 quality domains, all curated by hand, with an 80/20 train/test split and k-fold cross-validation. Nothing is deployed below 90% accuracy.
Four BERT processes run in parallel on every candidate: named entity extraction, keyword extraction, sentiment analysis and feature engineering. The classifier reads the entity cloud that comes out. A quality site's entities cluster tightly around one theme. A link farm's scatter, with no centre at all.
R05 Acquisition
Same acquisition engine. Different selection math.
A backlink and an AI citation are the same page. The page that links to you for Google is usually the page AI pulls from when it builds an answer. Win the placement once, win both surfaces.
What changes is the scoring. Instead of the Google link graph, we score against the co-citation graph and Harmonic Centrality, which show us the authority clusters LLMs already treat as trusted in a category. We vary the wording across placements on purpose, because the retrieval model rewards genuine diversity, not the same sentence copied across every site.
R06 The finding
AI cites the most connected brand, not the loudest.
This held up across multiple models and signals, with statistical output at both the correlation and the classification level. Only 10-20% of AI citations settle into a stable, recurring cluster. The rest cycle in and out with no real pattern. In one client's blackjack query cluster, 127 of roughly 3,000 cited domains showed up in almost every answer we tracked. Brie called it the authority bubble, and the name stuck.
Personalisation matters far less than buyers assume. We ran a live room test: everyone typed the same hotel query into ChatGPT, logged in or not, and got the same two results every time. And the Google and ChatGPT result sets barely overlap, roughly 12%. Ranking number one on Google tells you almost nothing about whether you exist in AI answers.
R07 Seven-month corpus study
Most of the web marks up its furniture, and it does a little more of it every month.
WLDM processed every monthly Common Crawl snapshot from January to July 2026, CC-MAIN-2026-04 through CC-MAIN-2026-30: 700,000 WARC files holding 14,784,536,534 HTML pages, for roughly $368 of compute. Extraction covered JSON-LD and microdata. A weighted mean of 58.32% of pages carry structured data, 8,620,770,647 of them, and across the corpus we recovered 127,990,580,002 individual type occurrences.
The vocabulary is used unevenly, to put it mildly. The top 100 types account for 97.87% of all occurrences, while more than 62% of distinct types appear exactly once. Grouped by function, navigation and structural markup takes 43.5% of top-200 occurrences in July. A further 811,946,273 occurrences, 0.63% of the total, are invalid @type values: the kind strict parsers throw away and AI training scrapes keep.
Coverage rises every month from March onward, 55.27% in February to 60.05% in July, and it is the same wherever you look inside a crawl. The standard deviation of coverage across each month's 100,000 WARC files stays below 1.25 percentage points, which makes this a property of the web rather than an artefact of our sample.
Latest paper · Working paper, September 2026
Seven months of schema.org, across 14.78 billion web pages
700,000 WARC files. 14,784,536,534 HTML pages. A weighted mean of 58.32% of them carry structured data, rising every month from March onward, and 43.5% of what the top 200 types describe is navigation rather than content.
02. Why it matters to you
Research is only worth something if it changes a decision.
Every rule below comes out of one of the programmes above. Our delivery team applies them before we spend a dollar of your budget.
| Research | Rule | Effect |
|---|---|---|
| From R04 | 3 of 5 models must agree | No placement proceeds on a single model's opinion. Candidates that split the vote are dropped rather than argued over. |
| From R01 + R06 | CS 0.20 is the floor | Every candidate page is scored for cosine similarity against your topic before outreach. Below 0.20 it is rejected, whatever the domain metrics say. |
| From R02 | Page traffic, not domain traffic | Every linking URL is checked for real page-level traffic. A dormant page on a strong domain sends nothing, so we don't buy it. |
| From R05 | One placement, two surfaces | Targets are chosen where the Google link graph and the co-citation graph overlap, so a single acquisition earns ranking and citation value. |
| From R06 | Dominant tier first | Harmonic Centrality decides sequencing. The clusters AI already trusts are entered before the peripheral ones, because that is what compounds. |
| From R03 | Check our working | The two public tools run the same models as the pipeline, so you can reproduce our scoring instead of taking a dashboard on trust. |
Research partners
We do not mark our own homework.
- DataForSEO: citation data partner on the 11M+ study
- Common Crawl: source of the link graph used for Harmonic Centrality, and of the corpus behind R07
- Search Atlas: partner on the citation study running now
- Surfer SEO: partner on the citation study running now
What we don't claim
The limits of the work.
The 2024 Google API leak confirms which fields exist and are active in the live ranking pipeline. It does not publish Google's weights. Our scoring model is built on those confirmed fields. The weighting is ours, derived from our own data, not extracted from Google's.
Where a finding isn't ours, we say so. The structured-data retrieval figures on the TurboQuant page come from WordLift's research, not ours, and they're cited that way.
TurboQuant Optimisation is built on a verified Google mechanism. No client result is tied specifically to it yet, and we're not going to pretend otherwise.
R07 measures deployment, not effect. It counts what the open web marks up and how consistently. It does not prove that any single @type earns a ranking or an AI citation on its own.
We'll run the research on your domain.
The same pipeline described above, pointed at your site and your category. You get the findings whether or not you work with us.
If we can't show you something your current agency missed, we'll say so.
Start with your data
Run the pipeline on my domain
Two fields. You get your classifier scores, your citation clusters and the gaps we find.