gtmskills.com

Reference file

Evidence map

evidence-map.md

Evidence map — verified patents and leaked fields

A starter set that ties common on-page actions to the ranking mechanism behind them, with a Google patent and/or a field from the leaked Search API documentation as the receipt. Not exhaustive — extend it with the verification method below. For each candidate quick win, find the matching row, cite the ID/field, and confirm it before it goes in the deliverable.

What each kind of evidence proves

  • A patent shows how Google has described thinking about a problem. It does not prove the approach is live in ranking, or how heavily it is weighted.
  • A leaked field comes from Google's internal "Content Warehouse" Search API reference — a catalog of roughly 2,596 modules and 14,014 fields describing what Google can store for documents, sites, links, and interactions. A field shows Google stores or can consider a signal. It does not reveal ranking weight, scoring, or even that a field is in active use — no weightings appear anywhere in the corpus.

Say which one you hold. Never let a patent or a stored field become a claim about live weighting.

How to verify (and the anti-fabrication rule)

  • Patent: open a patent search, load the number, and confirm it resolves to a Google-assigned patent with the stated title. Several SEO-famous patents are mis-cited (see the last section) — confirm, don't trust memory.
  • Field: find the field on the public Content Warehouse API documentation mirror (module pages named GoogleApi.ContentWarehouse.V1.Model.<Module>) and read its description. Field naming is inconsistent — some snake_case, some camelCase — so search the module page rather than trusting a remembered spelling.
  • If a number does not resolve, or a field is not in the docs, do not cite it. Present it as a hypothesis or drop it. Never approximate a patent number or invent a field name.

1. Match the title to the query

  • Action: front-load the core query term in the title, keep it specific, drop boilerplate.
  • Mechanism: the title is what earns the click, and title-to-query match is a stored signal.
  • Field: titlematchScore (module QualityNsrNsrData) — "Titlematch score of the site, a signal that tells how well titles are matching user queries." Described at the site level, not strictly per-page.
  • Patent (supporting): US8938463B1 — the click a good title earns feeds ranking (see row 2). There is no clean, standalone Google title-matching patent; do not invent one.

2. Earn and keep the click

  • Action: write titles and descriptions that set accurate expectations (fewer pogo-stick returns to the results page); make the above-the-fold answer fast and complete so the visit becomes the satisfying, last, longest click.
  • Mechanism: click behaviour adjusts ranking, discounted for the fact that higher positions get more clicks regardless of relevance — so genuine post-click satisfaction is what the model isolates.
  • Patent: US8938463B1 — "Modifying search result ranking based on implicit user feedback and a model of presentation bias" (Google LLC).
  • Fields: goodClicks, badClicks, lastLongestClicks (module QualityNavboostCrapsCrapsClickSignals, the NavBoost / "Craps" click system). The field names are present in the module; the mirror carries no description text for them, so the good-vs-bad and last-longest reading is analyst interpretation (iPullRank, SparkToro), not a doc quote. Cite it that way.

3. Make thin content original and high-effort

  • Action: on thin or short pages, add genuinely original material — first-hand analysis, unique data, original media — and visible effort, not restated aggregate text.
  • Mechanism: a site-level quality factor demotes low-quality resources; originality is scored most pointedly where content is sparse.
  • Patent: US8682892B1 — "Ranking search results" (Google LLC; inventor Navneet Panda — the "Panda" quality work).
  • Fields: OriginalContentScore (module PerDocData) — "The original content score is represented as a 7-bits, going from 0 to 127." (iPullRank adds that only pages with little content carry this field). contentEffort (module QualityNsrPQData) — "LLM-based effort estimation for article pages."

4. Cover the topic, not just the keyword

  • Action: cover a topic with its naturally related terms and subtopics; stop repeating one exact-match keyword.
  • Mechanism: ranking works on meaningful phrases and their co-occurring related phrases, not isolated keywords.
  • Patent: US7536408B2 — "Phrase-based indexing in an information retrieval system" (Google LLC; inventor Anna Lynn Patterson).
  • Fields: chard_score_encoded and chard_score_variance (module QualityNsrNsrData) — "Site-level Chard (encoded as an int)." / "Site-level Chard Variance for all pages of a site." "Chard" is a described content-quality predictor exposed through these fields; there is no field literally named chard.

5. Use honest, descriptive, well-placed links

  • Action: place key internal links prominently (in the body, higher up, meaningful anchor text), not buried in boilerplate; keep anchor text — internal and inbound — descriptive and topically aligned with the destination.
  • Mechanism: links are weighted by how likely a reader is to click them, and anchor text matched to the query contributes to the target's ranking; mismatched anchors are demoted.
  • Patents: US7716225B1 — "Ranking documents based on user behavior and/or feature data" (Google LLC; the "reasonable surfer" — links weighted by click likelihood via position, font size, anchor characteristics). US7260573B1 — "Personalizing anchor text scores in a search engine" (Google LLC).
  • Field: anchor_mismatch_demotion (module CompressedQualitySignals) — "converted from QualityBoost.mismatched.boost."

6. Signal genuine freshness, not fake dates

  • Action: keep the on-page date, structured-data date, and displayed date consistent; signal real recency in the body (in-text dates, current references, updated facts and links), not just a changed timestamp.
  • Mechanism: time-based signals — inception date, update frequency and amount, link/anchor growth — factor into scoring, and the content date is estimated from the content itself, not merely declared.
  • Patent: US7346839B2 — "Information retrieval based on historical data" (Google LLC; inventors include Matt Cutts and Jeff Dean).
  • Fields: semanticDate (module PerDocData) — "estimated date of the content of a document based on the contents of the document (via parsing), anchors and related documents." bylineDate (module QualityTimebasedSyntacticDate) — set only when the byline date differs from the main date field (per iPullRank, it is the date shown in snippets).

7. Build site-level authority and topical focus

  • Action: concentrate topical content, internal links, and quality on one domain rather than fragmenting across microsites; keep new pages semantically close to the site's core topic. Page-level wins accrue to a stronger site.
  • Mechanism: a site-level quality/authority signal modifies page scores, and topical focus is measured as how far pages drift from the site's centre.
  • Patent: US8682892B1 — "Ranking search results" (site-level quality modification).
  • Fields: site_authority (module CompressedQualitySignals) — "converted from quality_nsr.SiteAuthority, applied in Qstar." siteFocusScore — "Number denoting how much a site is focused on one topic." and siteRadius — "The measure of how far page_embeddings deviate from the site_embedding." (both in QualityAuthorityTopicEmbeddingsVersionedItem, alongside siteEmbedding — "Compressed site/page embeddings."). chromeInTotal (module QualityNsrNsrData) — "Site-level Chrome views." (a behavioural signal on-page work alone cannot manufacture).

8. Mind domain and new-site cautions

  • Action: do not lean on a keyword-stuffed exact-match domain; for a new host, expect a slow start and prioritise clean, high-effort content and links over aggressive tactics.
  • Fields: exact_match_domain_demotion (module CompressedQualitySignals) — "converted from QualityBoost.emd.boost." hostAge (module PerDocData) — "The earliest firstseen date of all pages in this host/domain. These data are used in twiddler to sandbox fresh spam in serving time." This one field is the only in-corpus evidence for a "sandbox"; do not assert a dedicated sandbox score.

Do not mis-cite (checked, and wrong in the wild)

  • US7739277B2 — "System and method for incorporating anchor text into ranking search results" is Microsoft, not Google. It circulates in SEO posts as Google's anchor-text patent. It is not — cite US7260573B1 for anchor text instead.
  • US6285999B1 — "Method for node ranking in a linked database" (PageRank; inventor Lawrence Page) was originally assigned to Stanford, with Google holding an exclusive licence; it has since expired. If you cite it, get the Stanford origin right — and note it is link-graph ranking, not an on-page quick win.
  • The Panda patent US8682892B1 is titled only "Ranking search results" — the "Panda" association is through inventor Navneet Panda, not the title. Use the granted number, not the application US20120143789A1.
  • For historical data, use the granted US7346839B2, not the application number US20050071741A1.

Naming caveats carried from the corpus

  • panda_demotion exists in CompressedQualitySignals, but its description references baby_panda_demotion ("converted from QualityBoost.rendered.boost."), and separate baby-Panda fields also appear — the Panda / Baby-Panda naming is muddled in the corpus. Report the field you actually find.
  • Some fields render snake_case (site_authority) and some camelCase (titlematchScore, hostAge) across the mirror. Always confirm the exact spelling on the module page.