Google patents 2024–2026: an index of patent families (bajsarowicz.pl)

Data up to September 17, 2026.
About the study: https://bajsarowicz.pl/en/blog/seo-strategy/google-patents-2024-2026/

Files
- google-patents-2024-2026.csv: 4,651 Google and DeepMind patent families with a publication from January 1, 2024 on
  and at least one CPC class G06F16, G06F40, G06N, G06Q30 or G06V. One row = one family.
- sample-60.csv: the validation sample (15 families from each score of the first model). second_model_blind_score_0_3:
  the second score (a language model from a different developer; title and full abstract, random order, without
  seeing the first score); the post relies on this one. second_model_open_score_0_3: a score of the same families based
  on the title and the beginning of the abstract, made with knowledge of the first score. expert_score_0_3: left empty
  for an expert score.
- scoring-instructions.txt: the instructions for the model (translated from Polish) and the rule for the language
  model flag.
- query.sql: the SQL query against Google Patents Public Data. Running it on a newer version of the table returns a
  different set.

Columns
- abstract: the abstract of the publication in the patent column, from Google Patents Public Data (may be empty).
- priority_date: the priority date of the publication in the patent column, in YYYYMMDD format: the date of the
  earliest application that the publication claims, or the date of its own filing. The post groups families into
  years by this date.
- model_score_0_3, model_area: the result of the AI model based on the title and up to 700 characters of the abstract.
  A pointer for choosing patents to read, not a quote from the patent. In the validation sample, the second model gave
  scores of 2–3 less often. The model's one-sentence notes are not published: they are untested hypotheses about
  visibility.
- language_model_flag: a term search in the title and abstract of the publication in the patent column (rule in
  scoring-instructions.txt). It also catches mentions in the background of an invention and misses patents that use
  other words.
- written_up: a match in the public Semgence catalog (as of August 17, 2026), in Olaf Kopp's write-ups (by title,
  approximate, some of them paid) or in Rafał Borowiec's posts on patentcore.digital.
- granted_patent: the family has a publication with a kind code starting with B (a granted patent, for example
  US B1/B2, CN B, TW B). This is not the current legal status.

License
Source data: "Google Patents Public Data" by IFI CLAIMS Patent Services and Google, CC BY 4.0.
This dataset (scores, areas, flags): Konrad Bajsarowicz, bajsarowicz.pl, CC BY 4.0.
When you use it, credit the source: bajsarowicz.pl, "Google patents 2024–2026", and Google Patents Public Data.
