AI patent search solves one specific problem, and it is the problem that defeats most keyword searching.
Prior art is written in the vocabulary of its own field. You search for a "flow stabiliser"; the reference that anticipates you calls it a "damping baffle". No amount of keyword iteration finds it unless you happen to guess that phrase.
Semantic search matches on meaning, so the vocabulary gap stops mattering. That is a genuine advance and it is most of what these tools are for.
What it does not do is make the results self-interpreting. Recall improves; precision does not. You find more relevant documents and you also read more.
Keyword versus semantic
| Keyword search | Semantic search | |
|---|---|---|
| Matches | Exact terms and variants | Meaning |
| Finds different vocabulary | No | Yes |
| Foreign-language art | Only if translated first | Yes, with translation |
| Result volume | Lower | Higher |
| Precision | Higher per result | Lower per result |
| Requires knowing the field's terms | Yes | No |
| Reproducible | Exactly | Approximately |
The reproducibility difference matters for records. A keyword search can be rerun identically; a semantic search may rank differently as the index or model changes. Record the query and the date, and save the results you relied on.
Neither replaces classification searching. CPC codes group documents by what they are rather than what they are called, and that catches a third category of document again.
Writing a good query
Describe, do not name.
| Weak query | Better query |
|---|---|
| "smart irrigation" | "Controller that samples soil moisture at intervals, compares against a threshold derived from recent rainfall, and actuates a valve when the deficit persists across consecutive readings" |
| "folding chair hinge" | "Hinge mechanism that locks at a fixed angle using a spring-loaded detent and releases under axial rather than rotational force" |
| "AI patent tool" | — |
Length helps. Semantic matching works better with a paragraph than a phrase, because more of the concept is expressed for it to match against.
Include the problem, not just the solution. Documents solving the same problem by different means are relevant to obviousness even where they are not anticipatory.
Include materials, ranges and parameters where they are part of the invention. They sharpen matching considerably.
How semantic matching works, roughly
You do not need the technical detail, but the shape explains the behaviour.
| Step | What happens |
|---|---|
| Your description is converted to a numeric representation | Meaning becomes coordinates |
| Every indexed document has been converted the same way | One-time, at indexing |
| The system finds documents whose coordinates sit near yours | Similarity, not matching |
| Results are ranked by proximity | Closest first |
This is why vocabulary stops mattering. "Flow stabiliser" and "damping baffle" land near each other because their descriptions mean similar things, regardless of sharing no words.
It is also why precision does not improve. Proximity is a continuum, so there is no clean line between relevant and irrelevant — the system returns a ranked list and someone has to decide where to stop reading.
And why longer queries work better. More description means more signal to position against, which is the opposite of keyword search where extra terms narrow results.
What it still misses
| Gap | Why |
|---|---|
| Unpublished applications | 18-month blackout — absolute for every tool |
| Non-publication request filings | Invisible until grant |
| Non-patent literature | Coverage is uneven across tools |
| Conference papers and theses | Frequently unindexed |
| Product manuals and datasheets | Rarely indexed |
| Products on sale, never documented | Still prior art |
| Unusual conceptual framings | May fall outside the match |
The eighteen-month blackout is the hard limit. An application filed last year cannot be found by anything, and it may grant into something that covers your product.
Non-patent literature is where professional searchers still win. Conference proceedings, trade publications, standards documents and manuals are where damaging references frequently hide, and automated indexing of them is patchy.
A product on sale is prior art whether or not anyone wrote about it, which is why market searching sits alongside database searching rather than after it.
Verifying results
| Check | Why |
|---|---|
| Open the document | Summaries misstate content |
| Read the claims, not the abstract | Scope lives in the claims |
| Check the priority date | Determines what it is art against |
| Confirm it is a patent or an application | Different status entirely |
| Check the jurisdiction | Rights are national |
| Note where you found it | For the record, and for disclosure |
Never rely on a generated summary. Retrieval returns real documents, but any layer that summarises or comments can misstate what a reference teaches, and a misread reference is worse than a missed one.
Record what you found and when. If you are prosecuting an application, the duty of disclosure under 37 CFR 1.56 requires material references to be submitted regardless of how you came across them.
Worked example: the same invention, three searches
A mechanical damping mechanism for a portable device.
| Keyword | AI semantic | Professional | |
|---|---|---|---|
| Time | 3 hours | 20 minutes | 2 weeks |
| Cost | Free | Low | $1,500–$3,000 |
| Documents returned | 40 | 180 | 95 |
| Relevant after reading | 6 | 11 | 14 |
| Closest reference found | No | Yes | Yes |
| Non-patent literature | None | 1 item | 4 items |
| Foreign-language art | None | 3 items, translated | 5 items |
What each search actually did
The keyword search missed the closest reference because that document described a "compliant restraint" rather than a damper. Nothing about the invention suggested that phrase.
The semantic search found it in twenty minutes and returned 180 documents to read, of which 11 mattered. Better recall, unchanged precision.
The professional search found three more relevant items, all non-patent literature — a conference paper and two trade publications — plus two additional foreign references.
The sensible sequence is all three, in order of cost. Semantic search first because it is fast and eliminates most inventions; professional search before committing to a non-provisional.
Using it on someone else's patent
The same tool works in reverse, and this is the use most people miss.
| Goal | Query with |
|---|---|
| Invalidate a patent | The patent's own claim 1, as the search text |
| Find design-around space | The claim, then look at what sits nearby but outside |
| Find licensing targets | The claim, matched against product descriptions |
| Map a competitor's coverage | Each of their claims in turn |
Pasting a claim in as the query is the fastest validity check available. If semantic search returns something predating the priority date that reads on the claim, you have found prior art the examiner missed.
Check the dates carefully. A document is only prior art if it predates the patent's priority date, which may be years before its filing date where a continuation chain exists.
This is a screening tool, not a validity opinion. It surfaces candidates cheaply; a litigation-grade validity search is a different exercise. See patent invalidation.
Where the cheap search pays
Searching used to cost enough that people skipped it, and skipping it is expensive in a specific way.
| Discovery point | Cost of learning the invention was anticipated |
|---|---|
| Before filing | An hour |
| After a provisional | ~$2,000 and 12 months |
| After a non-provisional | $8,000–$20,000 and 2 years |
| At the first office action | The above, plus response fees |
| During litigation | The patent, and the case |
The gap between the first and last rows is the whole argument. A search that costs an hour and eliminates the invention has done more for you than a patent that issues and is later invalidated.
And it does more than eliminate. The reformulation that makes an invention patentable is frequently visible in the close prior art — what was tried, what failed, and where the real unsolved problem sits. See invention ideas.
Recording what you searched
Search records matter in two situations, and both arrive later.
| Situation | What the record supports |
|---|---|
| Prosecution | Disclosure of material art under 37 CFR 1.56 |
| An opinion | The opinion is accurate as of its search date |
| Litigation | What you knew, and when |
| Renewal decisions | Why a patent was kept or released |
Record the query, the tool, the date and the results relied on. Semantic rankings may not reproduce identically later, so the saved results are the record rather than the query.
Searching before renewal, not just before filing
Most patents are abandoned without anyone checking whether they mattered.
| Ipiry Patent Survival Curve v1.0 | Rate |
|---|---|
| Survive the 3.5-year fee (2022 cohort) | 85.8% |
| Survive the 7.5-year fee (2018 cohort) | 64.6% |
| Reach full term (2014 cohort) | 41.4% |
| Abandoned before full term | 58.6% |
Computed from 27,273,654 USPTO maintenance fee records covering 8,262,336 US utility patents — see the patent survival curve.
Cheap search changes the renewal decision as much as the filing decision. Checking whether anyone practises a patent's claims used to cost more than the maintenance fee, so the decision defaulted to paying or lapsing on instinct.
A semantic search against current products takes minutes per patent, which makes an evidence-based renewal decision affordable at portfolio scale for the first time. See patent portfolio management.
Combining the three approaches
Each search type finds documents the others miss, which makes them additive rather than competing.
| Approach | Finds | Misses |
|---|---|---|
| Keyword | Documents using your terms | Different vocabulary |
| Semantic | Documents meaning the same thing | Unusual conceptual framings |
| Classification | Documents in the same technical area | Things classified elsewhere |
Run all three on anything that matters. Twenty minutes each, and the overlap between their results is smaller than people expect.
Use one to seed the next. A relevant document found semantically carries CPC codes; browsing that subgroup finds documents neither of the other approaches reached.
Limits worth stating plainly
No search tool can promise clearance, and any that implies it is overselling.
| Limit | Absolute? |
|---|---|
| 18-month publication blackout | Yes, for every tool |
| Non-publication request filings | Yes, until grant |
| Undocumented products on sale | Yes |
| Non-patent literature coverage | Partial everywhere |
A clean result means you did not find anything. That is different from nothing existing, and the gap is why professional searching still exists.
AI patent search: the checklist
- Run it before spending anything. It is fast, cheap, and eliminates most inventions.
- Describe functionally, at paragraph length. Do not name the invention.
- Include the problem, the mechanism, and any materials or parameters.
- Run classification searching alongside it. CPC catches what neither keywords nor semantics reach.
- Expect more results, not fewer. Recall improves; precision does not.
- Open every document you rely on. Never trust a generated summary.
- Read the claims, not the abstract, and check the priority date.
- Search products and market listings too. A product on sale is prior art.
- Commission a professional search before a non-provisional, particularly for non-patent literature.
- Remember the eighteen-month blackout. A clean search is not clearance, and no tool can see an unpublished application.