Google Patents Search

Purpose

Search Google Patents (patents.google.com) — the public index of ~140M+ worldwide patent publications — and return matching patents with their publication number, title, priority/filing/grant/publication dates, inventor, assignee, language, snippet, and canonical landing-page + PDF URLs. Read-only; never authenticates, never submits forms, never opens Google Account.

When to Use

• Free-text patent search ("quantum computing", "lithium battery").
• Filtered patent search by assignee, inventor, country, status (granted vs application), language, or date range.
• Boolean / phrase / CPC queries against the title, abstract, claims, or full document.
• Bulk patent discovery feeding downstream PDF download (the pdf field plus the patentimages.storage.googleapis.com prefix yields a stable direct link).
• Anywhere you'd otherwise scrape the Google Patents results page — the public XHR endpoint is faster, structured, and avoids JS-rendering overhead.

Workflow

The Google Patents web UI is a thin client over an undocumented but stable public JSON endpoint at https://patents.google.com/xhr/query — no auth, no cookies, no CSRF token, no anti-bot challenge on the endpoint itself. The endpoint takes one parameter, url=, whose value is the URL-encoded form of the entire query string that would normally appear after patents.google.com/? (i.e. double-encoded relative to the outer URL). Lead with the API; the browser path is a slow fallback that pays a ~100× turn-cost premium because the search page is fully JS-rendered (browse snapshot returns no useful refs until after the SPA hydrates).

Stealth note: the Google Sorry interstitial ("your computer or network may be sending automated queries", HTTP 503) trips after ~5–10 sequential requests from the same datacenter IP. Always fetch through residential proxies (browse cloud fetch ... --proxies) — every call gets a fresh exit IP and the rate-limit never fires. A bare session (no proxies) works for a handful of one-off queries but is not safe for any sustained workload.

1. Build the inner query string

Construct the query exactly as you would type it into the URL bar of patents.google.com, then URL-encode the entire string and stuff it into url=.

Inner-query parameters (these go inside the url= value):

2. Fetch the endpoint

INNER='q=quantum+computing&num=25&page=1&sort=new'
URL_ENC=$(node -e "console.log(encodeURIComponent(process.argv[1]))" "$INNER")
browse cloud fetch "https://patents.google.com/xhr/query?url=${URL_ENC}&exp=" --proxies

The &exp= trailing parameter is required and always empty. The response wraps a JSON envelope around the page payload; parse .content as JSON to get the real result object.

3. Parse the result envelope

The inner JSON has shape:

{
  "results": {
    "total_num_results": 127800,   // total hits matching the query
    "total_num_pages": 100,        // pages available (capped — see step 1)
    "many_results": false,
    "num_page": 0,                 // 0-indexed page number of THIS response
    "cluster": [                   // exactly one cluster for ungrouped searches;
      {                            // empty result[] when zero hits
        "result": [ { "id": "...", "rank": 0, "patent": { ... } }, ... ]
      }
    ],
    "summary": "...",
    "landscape": { ... }
  }
}

For each item in cluster[0].result, the patent sub-object carries the fields you actually want: title, publication_number, priority_date, filing_date, grant_date (may be absent for applications), publication_date, inventor, assignee, language, snippet, and pdf (a relative path).

Strip HTML highlighting tags — title and snippet contain <b>…</b> around query-term matches and HTML entities (…, "). A simple replace(/<[^>]*>/g, '') plus entity decode produces clean text.

4. Construct canonical URLs

• Landing page: https://patents.google.com/patent/{publication_number}/en (the /en suffix forces the English-translated view for non-English publications).
• PDF: https://patentimages.storage.googleapis.com/{patent.pdf} (the pdf field is the relative path, e.g. 85/df/92/f5ac6f65349817/JP7446622B2.pdf). Empty pdf field means no PDF is hosted (occasionally true for very recent applications).

5. Paginate

Increment page= until num_page >= total_num_pages - 1. Because Google caps accessible results at 1000, choose num=100 if you need depth (10 pages max) and num=10 if you only need the relevance-top results.

Browser fallback

If the XHR endpoint is unreachable (extended outage, regional block — not observed in our trace) drive the search UI:

sid=$(browse cloud sessions create --keep-alive --proxies --verified | grep -o '"id": "[^"]*"' | head -1 | sed 's/.*"\([^"]*\)"$/\1/')
browse open "https://patents.google.com/?q=quantum+computing&num=25" --remote --session "$sid"
browse get markdown body --remote --session "$sid"

Parse the rendered markdown for the result blocks (titles are h2 links, the publication number appears immediately under the title with a link to patentimages.storage.googleapis.com/.../{pubnum}.pdf). This costs ~3 browser turns per query vs. one HTTP fetch, so only fall back if the API actually fails.

Site-Specific Gotchas

• Double URL-encoding is required. The url= parameter wraps an already-encoded query string. If you single-encode (treating url=q=quantum computing as one flat query string) the endpoint silently returns the global-relevance landing page with total_num_results from an empty query. Always: build the inner query string first, then encodeURIComponent the whole thing and assign to url=.
• &exp= is mandatory. Omitting the empty exp parameter returns a 400 in some response paths. Always include &exp= (no value).
• 1000-result hard ceiling. Google reports total_num_results: 127800 but total_num_pages × num never exceeds 1000. To enumerate beyond the top-1000, narrow the query (date slices, CPC codes, assignee filters) and re-page each slice. Identical-page requests beyond the ceiling return an empty cluster[0].result with the same total_num_results — there is no error, just silence.
• Rate-limit interstitial is HTML, not JSON. After ~5–10 rapid sequential requests from a single datacenter IP, the endpoint returns a 503 with an HTML body titled "Sorry..." ("your computer or network may be sending automated queries"). Parse defensively — check Content-Type or check that the first byte is {. Use --proxies (residential rotation) and the rate-limit never fires across hundreds of queries.
• Inventor / assignee names are not Latin-script-normalized. A patent originally filed in Japan returns its inventor and assignee in Japanese characters (e.g. プーリ，シュルティ, イェール ユニバーシティー); the Chinese-language equivalents appear in Simplified Chinese (王鑫, 北京百度网讯科技有限公司). The /en landing page exposes the English transliteration but the XHR API does not. If you need a Latin form, you must fetch the landing page.
• HTML in title and snippet fields. Search-term matches are wrapped in <b>…</b> and the snippet ends in …. Decode entities and strip tags before display.
• grant_date is sometimes missing. Applications that have not been granted have only priority_date, filing_date, and publication_date. Treat grant_date as optional.
• pdf field can be empty string. Most patents have a PDF; very recent applications, withdrawn entries, and some non-US jurisdictions don't. Test for empty string, not null.
• Default operator quirk. q=safety OR seat belt is parsed as (safety OR seat) AND belt, not safety OR (seat belt). Use explicit parens or phrase quotes ("seat belt") to disambiguate.
• num= ignored above 100. Setting num=500 returns 100 results and shifts the cap accordingly. Do not assume larger pages give you more depth.
• Don't bother with GraphQL discovery. Network inspection shows no GraphQL surface — the XHR endpoint is the entire API. Don't waste turns hunting /graphql or /api/v1/; they don't exist.

Expected Output

JSON envelope for a normal search (10 results / page, default sort):

{
  "query": "quantum computing",
  "total_num_results": 127800,
  "total_num_pages": 100,
  "num_page": 0,
  "results": [
    {
      "title": "Quantum information processing using asymmetric error channels",
      "publication_number": "JP7446622B2",
      "priority_date": "2018-06-29",
      "filing_date": "2019-06-28",
      "grant_date": "2024-03-11",
      "publication_date": "2024-03-11",
      "inventor": "プーリ，シュルティ",
      "assignee": "イェール ユニバーシティー",
      "language": "en",
      "snippet": "For example, it is known that certain computational problems can be solved more efficiently using quantum computing rather than traditional classical computing...",
      "pdf_url": "https://patentimages.storage.googleapis.com/85/df/92/f5ac6f65349817/JP7446622B2.pdf",
      "canonical_url": "https://patents.google.com/patent/JP7446622B2/en"
    },
    {
      "title": "Quantum computing service with local edge devices supporting multiple quantum computing technologies",
      "publication_number": "US11650869B2",
      "priority_date": "2019-11-27",
      "filing_date": "2019-11-27",
      "grant_date": "2023-05-16",
      "publication_date": "2023-05-16",
      "inventor": "Jeffrey Paul Heckey",
      "assignee": "Amazon Technologies, Inc.",
      "language": "en",
      "snippet": "a photon based quantum computer. 6. The system of claim 1, wherein the one or more computing devices that implement the quantum computing service are configured to...",
      "pdf_url": "https://patentimages.storage.googleapis.com/28/a0/9d/323bb84148dd61/US11650869.pdf",
      "canonical_url": "https://patents.google.com/patent/US11650869B2/en"
    }
  ]
}

Zero-results envelope:

{
  "query": "xyzqzx99asdfblahNoMatchPossible",
  "total_num_results": 0,
  "total_num_pages": 0,
  "num_page": 0,
  "results": []
}

Rate-limited (interstitial) sentinel — agent should retry with --proxies or back off:

{
  "query": "lithium battery",
  "error": "rate_limited",
  "http_status": 503,
  "message": "Google 'Sorry...' interstitial returned. Retry through residential proxy or wait ~30s."
}