Google Patents Search Scraper

Google Patents Search Scraper

Type what you would type into Google Patents — a keyword, an inventor, a company, a classification code, a date range — and get the results back as a table, one row per patent. Publication number, title, assignee, inventor, priority and filing and publication and grant dates, whether it was granted, whether the family is still active, and links to the page and the PDF.

Read this part first: these are search results, not patent documents. You get what the results list shows plus the abstract snippet. There are no claims, no full description, no citation lists, no family tree and no downloaded PDF — just the link to each one. And Google hands over at most 1,000 results for any single search, however many it says it found. If either of those is a problem, this is the wrong tool and you should not buy it.

• The same search grammar the website has: inventor, assignee, patent office, grant status, utility vs design, language, priority/filing/publication date ranges, litigated-only, and sort order.
• Already built a search on patents.google.com? Paste the address bar in and it runs that, filters and all.
• Every row is checked against the filters you asked for before it is delivered. A row that does not match is dropped and not charged.
• No API key, no Google account, no browser, no quota to top up.
• Empty input returns one labelled sample row, free, so you can see the columns before spending anything.

Price

$1.85 per 1,000 patents, plus a $0.0005 start fee per run.

This is a flat rate on every plan, free or paid. There are no volume tiers, no minimum spend, no subscription and no add-on fees. What you read here is what you pay on day one and on day four hundred.

A search that matches nothing costs the start fee and no more. So does a search Google refuses. You are billed for patent rows and for nothing else.

What is actually charged

• One patent-scraped event per patent row written to the dataset. Nothing else is metered per row.
• Free: the sample row an empty run returns, and every diagnostic row — a blocked target, a dead URL, a search that matched nothing. Those rows all carry "charged": false.
• Searches that match nothing — you get a diagnostic row saying so, not a bill.
• Duplicate patents already returned earlier in the same run.
• Rows that do not match the filters you asked for; they are dropped before they are charged.
• The notice telling you a search hit the 1,000-result wall.
• A run that finds nothing costs the start fee and nothing more.
• Rows never leave the dataset without a charge, and are never charged without a row. The billed event is a named one, so there is no price quietly attached to apify-default-dataset-item — the trick that makes some scrapers bill you for their own error messages.

Input

{
  "queries": [
    "solid state battery",
    "lithium metal anode"
  ],
  "countries": [
    "US",
    "EP"
  ],
  "status": "GRANT",
  "dateType": "priority",
  "dateFrom": "2020-01-01",
  "maxItems": 200
}

Run it with empty input and you get one clearly labelled sample row, free, so you can see the output shape before you spend anything.

Output

One row per patent. A real row from a real run:

{
  "ok": true,
  "charged": true,
  "recordType": "patent",
  "query": "solid state battery",
  "publicationNumber": "US11942620B2",
  "title": "Solid state battery with uniformly distributed electrolyte, and methods of …",
  "assignee": "GM Global Technology Operations LLC",
  "inventor": "Yong Lu",
  "priorityDate": "2020-12-07",
  "filingDate": "2021-12-06",
  "publicationDate": "2024-03-26",
  "grantDate": "2024-03-26",
  "isGranted": true,
  "legalStatus": "ACTIVE",
  "patentUrl": "https://patents.google.com/patent/US11942620B2/en",
  "pdfUrl": "https://patentimages.storage.googleapis.com/f8/02/22/a9e3950cab3340/US11942620.pdf",
  "snippet": "Accordingly, it would be desirable to develop high-performance solid-state battery materials and methods that improve contacts between the solid-state active particles and the solid-state electrolyte particles in the electrodes …",
  "thumbnailUrl": "https://patentimages.storage.googleapis.com/90/b8/0e/ed6c106ab89ebf/US11942620-20240326-D00000.png",
  "figureCount": 12,
  "resultRank": 0,
  "resultPage": 0,
  "totalResultsEstimate": 3799,
  "scrapedAt": "2026-09-20T08:29:14.306Z"
}

Field notes

• publicationNumber — the full publication number including the kind code — US11942620B2. Stable, and the right thing to use as a primary key when you re-run.
• title — the title as Google shows it, with the search-term highlighting stripped out.
• assignee — the owner at publication. It is not updated when a patent is later sold, so treat it as "who filed it", not "who owns it today".
• inventor — the first named inventor only. The results list does not carry the rest.
• priorityDate — when the idea was first claimed anywhere in the family. Usually the earliest of the four.
• filingDate — when this particular application was lodged.
• publicationDate — when it became public.
• grantDate — when it was granted, or null if it has not been.
• isGranted — true when a grant date exists. The quick way to split grants from pending applications.
• legalStatus — ACTIVE if any member of the family is still in force, otherwise NOT_ACTIVE, and null when Google publishes no status. It is a family-level summary, not a jurisdiction-by-jurisdiction answer, and it is not legal advice.
• patentUrl — the Google Patents page. Open it for the claims and the description.
• pdfUrl — the original document. Sometimes null — not every record has one published.
• snippet — an extract from the abstract, chosen by Google around your search term. It is not the whole abstract and it is not the whole patent.
• thumbnailUrl — the first drawing, when there is one.
• figureCount — how many drawings the results list mentions. Often 0 on foreign-language records.
• totalResultsEstimate — Google's own count of matches for the search, copied through untouched. It is an estimate and it wobbles: the same term with a country filter can report a *higher* number than without one. Use it for a sense of scale, never as a population count.

Every real row carries "charged": true. Sample rows carry "_sample": true and diagnostic rows carry "_diagnostic": true with an errorCode you can filter on, and neither is ever billed.

How it works

• It calls the same search endpoint the Google Patents website calls when you type in the box, and asks for 100 results a page. No page rendering, no headless browser, no login, nothing to install.
• Requests leave through a large pool of rotating addresses, and the run moves to a new one every few requests rather than waiting for the throttle to arrive.
• Each page is checked for the fields it should contain before anything is delivered. A response that is not an answer — a throttle page, an interstitial — is treated as a failed request and retried from a different address, not billed as data.
• Rows are then checked against your own filters. If you asked for US grants since 2020 and something arrives that is not one, it is dropped and you are not charged for it.
• Publication numbers already seen in the run are skipped, so a patent matching two of your terms is charged once.

What people use it for

• Prior-art sweeps before filing: run the phrase four ways, dedupe on publication number, and read the shortlist instead of clicking through result pages.
• Watching what another company's R&D is filing — set assignee, sort newest first, schedule it weekly and diff on publication number.
• Mapping a technology area by classification code, then counting filings per year from the dates.
• Finding who actually works in a field: pull a few hundred rows on the topic and sort by assignee or inventor.
• Due diligence on an acquisition: everything a company has filed, with grant status and whether the family is still active, in one table.
• Feeding a spreadsheet or a database. The rows are flat and typed, so they load without cleaning.

The 1,000-result wall, and how to get round it

Google Patents will tell you a search matches 106,876 things and then hand over the first 1,000. That is the endpoint's limit, not a setting, and no amount of paying changes it.

The way round it is to cut the search into pieces that are each under 1,000 and run them all:

• By year. Set dateType to priority and run 2020, then 2021, then 2022. This is usually the cleanest split, because filings spread out fairly evenly over time.
• By office. Run US, then EP, then CN. Useful when you only care about some of them anyway.
• By assignee. If you already know the twenty companies that matter, twenty searches beat one.
• By status. Grants and applications, separately, is a free doubling.

When a search does hit the wall you get a free notice row saying so, with Google's match count on it, so you know to split rather than quietly receiving a truncated answer.

Reading the output

Every run writes three kinds of row and they are easy to tell apart:

• Real rows carry "charged": true and "recordType": "patent". One billed event each.
• The sample row carries "_sample": true and "charged": false. There is exactly one, it only appears when you ran with nothing filled in, and it exists so you can see the columns before spending anything.
• Diagnostic rows carry "_diagnostic": true, "charged": false and an errorCode you can switch on: NO_RESULTS when a search matched nothing or hit the 1,000 wall, BAD_INPUT when Google rejected the search, NETWORK when it could not be reached, TIME_BUDGET when the run ran out of time first. Each carries a plain-English error and the query it belongs to.

If you only want the data, filter on charged == true. That count always equals the number of events you were billed for, so the dataset is its own invoice.

Limitations

• Search results only. No claims, no full description, no citation list, no family tree, no downloaded PDF — every row carries the link, and you open it for the rest.
• At most 1,000 results per search, whatever the match count says. Split the search by year, office or assignee to reach more; there is a section above on how.
• The snippet is an extract Google picks around your search term, not the full abstract.
• Only the first named inventor is in the results list, so that is all a row can carry.
• assignee is the owner at publication and is not updated when a patent changes hands later.
• legalStatus is a family-level summary — ACTIVE means in force somewhere, not everywhere. It is not a legal opinion and should not be used as one.
• Titles and abstracts of foreign-language patents are Google's machine translations, and inventor and assignee names often stay in the original script.
• totalResultsEstimate is Google's estimate and it moves. Adding a filter can make it go up. Do not report it as a population count.
• Sorting by relevance, newest or oldest changes which 1,000 results you see on a broad search. There is no way to page past them in any order.
• Google throttles per address. A very large run slows down as it rotates; it does not fail, but it is not instant either.
• The hard ceilings are 5,000 rows and 20 searches per run. Split bigger jobs across runs.
• Design patents carry sparse metadata — often no assignee and no snippet — because that is what the source publishes.

Questions

Do I need a Google account or an API key?

No. Nothing to sign up for, nothing to install, no quota to top up. You pay Apify per row and that is the whole arrangement.

Can I search by CPC classification code?

Yes — put it in queries like any other term, for example H01M10/0525. It is not a separate field because Google Patents does not treat it as one; the code works as a search term and it works well.

I asked for 5,000 rows on one term and got 1,000. Why?

That is the endpoint's ceiling for a single search and it applies to everyone. You will also have a free notice row saying so. Split the search by year, office or assignee — there is a section above explaining how.

What happens if a search matches nothing?

One uncharged diagnostic row with errorCode: "NO_RESULTS", and the run carries on to your other searches. You are never billed for a search that returned nothing.

Can I paste a search I already built on the website?

Yes. Put the address bar into searchUrls. Its filters are used exactly as they are, which is the easiest way to run something complicated — build it where you can see the results, then paste it.

Why is the assignee sometimes in Chinese or Japanese?

Because that is how it was filed. Google translates titles and abstracts but leaves names in the original script, and inventing a transliteration here would be making data up.

Will the run fail if something goes wrong?

No. A refused, empty or broken search produces an uncharged diagnostic row explaining what happened and the run still finishes as succeeded. A failed run would still bill you the start fee, which would mean paying to be told something went wrong.

Can I run this on a schedule?

Yes. Nothing is held between runs, so the same input is safe to repeat. Diff on publicationNumber to see only what is new, and sort newest first so the new filings arrive at the top.

How do I get exactly the rows I paid for?

Filter the dataset on "charged": true. Sample and diagnostic rows are always false, and the number of charged rows always equals the number of billed events.