About · Methodology & standards
About the Lab: Methodology, Advisory Inquiries and Editorial Standards
A one-person research lab, split between New York and Lisbon, that treats search ranking as an engineering system with published mathematics rather than a marketing discipline with folklore.
Tom Belfort is a Principal Search Engineer and information retrieval researcher who has spent the better part of two decades on both sides of the ranking problem: building retrieval and link-analysis systems for large-scale crawlers, and later advising operators of very large websites on how those systems treat them. The monographs on this site are the written output of that second career, informed by the first. They are published under a single byline, without sponsorship, and are revised in place when the evidence changes.
What the lab does
The lab’s working premise is that modern web search is a legible system. Its foundations were published: PageRank in 1998, phrase-based indexing in 2004, historical link analysis in 2003, topic embeddings in a steady stream of filings since 2018. Its internal vocabulary surfaced in the 2024 Content Warehouse documentation. Its operating constraints, including the three-tier index and the NavBoost click system, were confirmed under oath in 2023. Taken together, these sources describe a machine with knowable physics. The lab’s job is to state that physics plainly, test it against measurements where measurement is possible, and translate it into architectural decisions a site operator can actually make.1Monographs are organised around one mechanism each rather than one tactic each. A tactic is a bet about a mechanism; stating the mechanism first makes the bet inspectable.
Methodology
Every monograph is built from three layers of evidence, and the text is explicit about which layer each claim rests on.
- Primary documents. Granted patents and published applications, cited by number; the Content Warehouse attribute names, quoted exactly; trial testimony and official documentation. These establish what the system is designed to do.
- Replicated measurement. Longitudinal crawls of client and reference sites, link-graph reconstructions, embedding analyses run with public models, and ranking observations across core updates. These establish what the system appears to do in practice. Sample sizes and time windows are stated in sidenotes.
- Working models. Where the production formula is not public, the monograph presents an explicit simplified model, labels it as such, and keeps it falsifiable. The calculator on the front page is one of these models. It is pedagogical, not forensic.
Claims that cannot be placed on one of those layers are not published. Where the literature is ambiguous, the monograph says so rather than resolving the ambiguity in whichever direction makes a better headline.
Advisory inquiries
The lab takes a small number of advisory engagements each year, typically with operators of sites between 5,000 and 500,000 URLs who face an architectural problem rather than a content problem: index coverage collapse, site-wide quality dampening after a migration or acquisition, topical dilution from programmatic expansion, or expired-domain reconstruction at scale. Engagements begin with a written diagnostic grounded in the site’s own crawl and link-graph data, and end with an architecture specification the operator’s engineering team can implement. The lab does not sell links, content or software, and does not accept engagements that depend on concealment from the engines being studied.2Corpus data from advisory work appears in monographs only in aggregate, with brands withheld, and only where the engagement agreement permits it.
Inquiries, corrections and replication notes all go to editorial@tombelfort.com. Correspondence is answered from New York or Lisbon depending on the season, usually within a week.
Editorial standards
- Attribution is mandatory. Every mechanism cites a patent number, a schema attribute or a dataset. Sidenotes carry the citation so the main text stays readable.
- Revisions are visible. Each monograph shows its publication and revision dates. Substantive corrections are noted in the front-page research log with credit to whoever caught them.
- Models are labelled. Any formula that is not taken verbatim from a primary source is marked as a working model in the note beneath it.
- No commercial interference. The site carries no advertising, no affiliate links and no sponsored placements. Nothing published here is syndicated or republished elsewhere.
- Prose over persuasion. The monographs are written to be read in order, by someone with time, and are typeset accordingly. There are no pop-ups, no gated downloads and no newsletter interstitials.
About the design
The site is set as a broadsheet monograph because the subject rewards slow reading. Headlines are in Playfair Display, body text in Karla, and formulas, patent numbers and attribute names in Cousine so that the three registers, argument, evidence and notation, are visually distinct. Margin sidenotes follow the convention Edward Tufte popularised: citations and asides sit beside the text that needs them rather than at the end of the page. The entire site is static HTML generated from a small build script, which is both an editorial choice and a demonstration of the crawl-efficiency principles the monographs describe.