USPTO Patent Public Search

The USPTO Patent Public Search is the United States Patent and Trademark Office's own window onto every issued US patent and every published patent application — more than 11 million granted patents plus several million pre-grant publications, with full text back to 1976 and bibliographic records reaching earlier. It is the successor to the retired PubEast/PubWest and PatFT/AppFT terminals, and within the Electronic Equipment & Instruments industry it is the canonical prior-art stop: when an electronics team needs to know whether a circuit topology, a sensor design or a display method was already claimed, this is the corpus the answer lives in.

The interface exposes a Quick Lookup for a single known number, a Basic Search that joins two fielded text boxes with AND/OR/NOT logic, and an Advanced mode for full query strings — a capable research surface. What it is not is an analysis environment: results come back as session exports, and turning eleven-digit application numbers, padded patent numbers and YYYYMMDD dates into join-ready rows is manual work. Datadory hands it to you shaped instead. Get a sample of this dataset and the rows come back cut to the assignees, inventors and date ranges you nominate, in the same schema the production feed uses.

The record anatomy below is the delivered shape — one row per patent or published application, with the identifier quirks already normalized:

# Delivered row shape -- electronics prior-art cut; slots fill in your sample
publication_number : 20210123456      # pre-grant publication: 4-digit year + 7 digits
patent_number      : 0123456          # issued grant, left-padded to 7 digits
title              : <invention title>
inventor_name      : <named inventor>
assignee_name      : <rights-holding entity>
publication_date   : 20210101         # YYYYMMDD

Two conventions in that block break naive pipelines when they go unhandled. Issued patent numbers are left-padded to seven digits — 123456 becomes 0123456 — while application publication numbers run to eleven characters once the four-digit year is attached: 2021123456 becomes 20210123456. Treat both as plain integers and joins silently drop records. Every extract ships with those keys typed and padded consistently, so deduplication against your own filing lists is a set operation rather than a cleanup project. Live values for the assignees, date ranges and technology classes you care about arrive with your sample.

Eight fields anchor the dictionary, each carrying one axis of the bibliographic spine — who filed it, who holds it, who invented it, when it published, and the number that keys the whole record.

Additional fields on request: classification symbols (CPC/IPC), citation links between documents, claims and abstract text, and inventor geography arrive as extended columns when your analysis calls for them. Tell us the question you are answering and we map that depth onto your sample rather than shipping columns nobody reads — which is also where column-level naming gets locked down for your pipeline.

Three chips summarize the footprint:

• Geography — United States grants and publications, the full national corpus held by the USPTO; foreign-origin filings appear wherever an international applicant entered the US system.
• Time — everything the office holds: full text from 1976 forward, bibliographic records stretching earlier, spanning issued patents and pre-grant publications in a single corpus.
• Granularity — one row per patent or per published application document, keyed on its publication number.

That document-level grain is the point for electronics work: component-level innovation moves in applications years before products ship, so a corpus covering publications alongside grants shows a rival's direction of travel roughly two to three years ahead of the storefront — the difference between reacting to a launch and anticipating it.

API, files, or your warehouse. Daily, weekly, or hourly.

You choose the channel and the cadence; extraction, normalization and schema stability are our problem. The identifier-padding and date-formatting conventions above travel fixed across all three channels, so a nightly warehouse load and an on-demand lookup agree row for row. When the office revises records or extends the corpus, the feed you consume stays normalized — same columns, same types, same join keys. A sample ships first either way, sized to test in your pipelines the same day.

Three of our persona groups lean on it hardest:

1. Competitive intelligence & product teams — pull every filing under a rival's assignee name, watch where those filings cluster technologically, and benchmark your own portfolio's shape against theirs. More in the competitive intelligence & product teams hub.
2. Data scientists & ML engineers — train prior-art retrieval, claim-language and technology-classification models on nearly five decades of full text, with assignee and inventor fields arriving as ready-made labels. Details on the data scientists page.
3. Journalists & academics — ground reporting and research in the primary record rather than press-release summaries, with every row traceable to its publication number. See journalists & academics.

The recurring use cases: competitor tracking built on assignee-scoped filing histories, ML model training over historical full text, and citation-grade research where a single wrong patent number would embarrass the whole piece.

Cards worth reading next: the rest of the Electronic Equipment & Instruments shelf, the structured patent-data siblings that complement this corpus with disambiguated inventors and machine-translated international text, and the comparison that sets it against a components-side neighbor.