1Start with what the free databases do superbly

Any honest comparison has to begin here, because the free collections are not a budget option. They are the reference layer everything else is checked against.

Espacenet offers “more than 150 million patent documents” from “more than 100 patent authorities around the world,” with simple and extended family views and legal event data drawn from INPADOC, which covers “legal events from over 50 international patent authorities worldwide.”^{8, 12} PATENTSCOPE indexes more than 128 million patent documents including some 5.5 million published PCT applications, with cross-lingual query expansion across 14 languages.^{13, 15} Google Patents indexes “over 120 million patent publications from 100+ patent offices around the world” alongside technical documents and books from Google Scholar and Google Books and the Prior Art Archive.⁵ USPTO Patent Public Search exposes Boolean and graded proximity operators over US text.¹⁷

None of that is a consolation prize. It is public infrastructure of remarkable scale, and the boundaries below are boundaries of purpose rather than of quality.

2Boundary one: coverage is not uniform, and the gaps are published

A patent count is a count of records, not of searchable text, and the difference decides what a search can find.

Google Patents indexes over 120 million publications from more than 100 offices and lists full text for 22 of them.⁵ USPTO Patent Public Search states that “at this time, Patent Public Search does not have access to foreign patent databases.”¹⁶ On the EPO side the shape of the underlying data shows in what it distributes: the worldwide bibliographic dataset DOCDB carries abstracts, citations and simple families with explicitly “no full text or images,” while machine-readable full text is published as separate products covering EPO publications since 1978 and character-coded national extracts for France, Spain, Switzerland and the United Kingdom.^{9, 10, 11}

Read together, that means full-text keyword search is deep for the two dozen or so offices that publish machine-readable text and shallower everywhere else, so a claim limitation buried in the description of a publication from an office outside that group is much harder to reach. This is not a criticism of the offices. It is what WIPO itself said about the system as a whole, writing in 2012 that “no single Office is capable of searching the whole of the PCT minimum documentation in its original language.”²⁰

What an agent changes here. A licensed corpus removes the seams between free services. Eureka searches “200M+ patents in 174 jurisdictions,” and its published description extends beyond granted patents to “published applications, and, where relevant, non-patent literature,” so one run covers ground that would otherwise mean four interfaces and four sets of export limits.¹

And what it does not. A wider corpus is not a complete one. No commercial index contains machine-readable full text for every office either, and WIPO’s point about language coverage applies to any searcher. What changes is how much of the reachable material one pass actually reaches, not whether the search becomes exhaustive.

3Boundary two: you have to know the right words before you start

A retrieval instrument matches what you type. That is the design, and the operators say so.

USPTO Patent Public Search states, in its own FAQs, that “unlike third party proprietary patent search databases, Patent Public Search does not use semantic searching.”¹⁶ Google Patents surfaces related records through a Similar Documents view that Google describes as based on text similarity, which is a different thing from an analysis of what a claim requires.⁶

The consequence is the failure mode nobody detects. Patent drafters choose broad, unusual vocabulary on purpose, so the document that destroys your novelty frequently describes the same mechanism in words you would never type. A keyword search returns nothing, and an empty result set looks exactly like an absence of prior art. You cannot audit a search for the terms you did not think of.

Worth noting that the offices themselves reached the same conclusion. The USPTO deployed an AI-based Similarity Search inside its internal examiner system, and a memorandum of 24 October 2025 states that “examiners are required to use Similarity Search during examination of a plant or utility application and record the conducted Similarity Search in the application file wrapper.”^{18, 19} Semantic retrieval is now part of the search your application meets, whether or not it is part of the search you ran.

What an agent changes here. The input stops being a query. Eureka’s published description is that it “turns an invention description into search strategies, ranks potentially relevant references” rather than “matching titles or abstracts,” which moves the burden of guessing the drafter’s vocabulary off the searcher.¹ You paste the invention summary; the strategy is derived rather than typed.

And what it does not. Deriving a strategy is not the same as guaranteeing the right one, and no retrieval method reaches text that was never indexed. The gain is that a term you would never have thought of is no longer a silent failure. The residual risk is that it may still be ranked too low to read.

4Boundary three: a ranked list is an input, not an output

This is the boundary that costs the most hours and gets discussed the least.

Every free database returns the same shape of thing: documents, ranked. What a novelty opinion or a clearance decision actually runs on is different. It is a mapping, feature by feature, between the claim or the product and the specific passage of a specific publication that discloses it, with the source attached so somebody else can check it.

None of the free databases described here produces that mapping, and none of them is trying to. Building it is the professional work. It is also where the hours concentrate, because retrieving a large candidate set is quick and reading it is not. Compressing retrieval time changes very little. Compressing the reading and the mapping changes the schedule.

What an agent changes here. This is the boundary agent platforms were built for, and the published output description is specific about the shape of it. Eureka returns “ranked prior art and feature-level evidence,” mapping “passages back to the disclosed features,” so a reviewer can “review source-linked findings before making a filing decision.” For clearance the equivalent output is “claim-level evidence with legal-status context” intended to “focus review on potential blockers before launch or expansion,” and the deliverable is described as “professional outputs with traceable patent & literature evidence.”¹ The first pass exists before anyone opens a document.

And what it does not. A first pass is not an opinion. Every mapping still has to be read against the passage it cites, which is precisely why source links matter more than scores: the check is opening a citation, not trusting a number. The task shrinks from construction to verification. It does not disappear, and the certification does not move.

That is the practical difference between the two generations. A database leaves the reading, the mapping and the write-up with you. An agent platform performs a first pass and hands you something to verify, which is a materially smaller task than construction from scratch.

5Boundary four: art from another field

Both generations struggle here, which is worth saying plainly rather than pretending AI solves it.

DAPFAM, a 2025 family-level benchmark, ran 249 controlled experiments across lexical, dense and rank-fusion configurations over 1,247 query families and 45,336 target families, and found that “OUT-domain performance remains roughly five times lower than IN-domain across all configurations.”²¹ A companion benchmark of 15 patent embedding tasks and 2.06 million examples reports its strongest model reaching 0.377 nDCG@100 on that dataset.²²

The reason this matters for the comparison is asymmetric, though. A keyword search cannot reach cross-field art unless you personally guess the other field’s vocabulary and classification. A semantic or multi-route system reaches it imperfectly. Imperfect and unavailable are different states, and cross-field art is exactly what a motivated challenger goes looking for later.

What an agent changes here, and what the numbers say. Against examiner-cited X references, Eureka’s novelty agent identifies at least one relevant X document in 85% of test cases and retrieves 37% of all of them within the top 100.² The first number says something usable surfaces in the large majority of cases. The second says a single pass does not recover the whole known-correct set. Both are published, with the dataset and the metric definitions beside them, which is what makes either figure checkable rather than a claim.

What follows practically. Treat any single search, by any method, as a first sweep rather than a closing argument, and load your own evaluation set with cases whose decisive reference came from an adjacent field. That subset is where tools actually differ.

6Boundary five: volume, export and automation

The free services publish operational limits, and they matter the moment a search becomes a process rather than a one-off.

Google Patents caps CSV export at 1,000 rows and states that the result count shown is an approximation.⁷ PATENTSCOPE limits result sets for logged-in users to 10,000 and explicitly prohibits bulk download and automated access.¹⁴ USPTO Patent Public Search loads 500 results at a time, holds 20,000 in a result set and exports 10,000 to CSV.¹⁶

These are entirely reasonable terms for a free public service. They are also the reason a free database cannot sit inside an automated screening gate that runs on every invention disclosure or every product specification. Screening at volume needs a licensed corpus and a supported interface.

What an agent changes here. The Patsnap Open Platform publishes 32 MCP servers exposing patent research, novelty and freedom-to-operate agents to Claude, Cursor or a custom client, with setup consisting of an API key and a generated connection link, so a screen can run inside a product gate rather than beside it.^{3, 4} The second thing that changes is what you can safely put in the box: an unfiled invention pasted into a public search interface is a different risk from one submitted under published terms. Eureka states that “your queries, saved patents, uploads, and analysis outputs are never used to train our models,” that data is “encrypted at rest and isolated within secure, enterprise-grade infrastructure,” and that enterprise controls include “SSO, audit logs, and role-based permissions.”¹

And what it does not. Metered access is still metered. Volume has a price, and an automated gate that screens everything will surface more candidates that need human judgment, not fewer. If review capacity does not move with it, the queue simply relocates.

7What changes on the other side of the boundary

An AI-native platform is not a better database. It changes what you supply and what comes back.

Taken together the changes above are one change: the unit of work moves from the query to the disclosure, and from the document list to the evidence set.

The part that makes this comparable rather than merely different is measurement. Patent search is one of the few AI applications with an objective answer key, because examiners publish the references they cited. On PatentBench, dated July 2026, ground truth is “X and Y references cited by examiners at different receiving offices,” collected, deduplicated and normalised by patent family across 340 cross-jurisdiction samples. The page glosses its two metrics plainly: within the top 100 results the agent “successfully identified at least one relevant X document” in 85% of test cases and “retrieved 37% of all relevant X documents.” It also reports that “Claude Opus 4.8 ranked second, with a 52.37% X Hit Rate and an 11.68% X Recall Rate.”²

Two boundaries in this article are not crossed by any of it. Coverage does not become complete, and recall on cross-field art remains the measured hard case for every method tested. An honest summary is that the agent generation changes what you supply, what comes back and how much of it you have to build yourself, while leaving the underlying difficulty of patent retrieval exactly where it was.

When a free database is the right answer

There are jobs where reaching for a paid platform is simply the wrong move.

• Verifying what a document actually is. Family structure and legal events tell you whether a hit is one right or twelve and whether it is still alive. Espacenet and the national registers are the authority for this, whatever tool found the document.
• A single known number. If you already have the publication number, Google Patents shows you the document and a Similar Documents view based on text similarity, with no account and no setup.⁶
• Jurisdiction-specific depth. USPTO Patent Public Search exposes Boolean and graded proximity operators over US text, which is precise control a general interface may not give you.¹⁷
• Non-English prior art on a budget. PATENTSCOPE cross-lingual expansion across 14 languages is free and genuinely useful when you suspect the art is Japanese or Chinese.¹⁵
• A record you can re-run. An exact Boolean string can be re-executed in five years and produce a comparable answer. That reproducibility is a real asset for a file, and it is a discipline worth keeping whichever tool does the searching.

What neither generation settles

• No search proves a negative. Coverage differs by office, by year, by full-text depth and by language. A search establishes what was found under a recorded method on a recorded date, not what exists.
• Novelty is not patentability. Novelty is assessed one reference at a time. Inventive step is a separate analysis under a named jurisdictional framework, and no retrieval score speaks to it.
• A percentage without a method is not a measurement. Whatever tool you assess, ask for the sample size, the definition of a correct answer, the cut-off and the metric before you accept a figure.
• The certification stays with the person who signs. The USPTO states that “simply relying on the accuracy of an AI tool is not a reasonable inquiry,” and under 37 CFR 11.18(b) whoever signs certifies that factual contentions have evidentiary support, to the best of their knowledge, information and belief formed after an inquiry reasonable under the circumstances.^{23, 24}