How Patent Examiners Approach a Prior Art Search
A patent examiner's prior art search follows a structured, multi-step process, starting with analyzing the application's claims and ending with a professional judgement on novelty, using tools like search statements, classification codes, and dedicated platforms such as EpoqueNet.
Before building anything, the Cardiff University research team went straight to the source, the patent examiners who actually carry out prior art searches every day. In January and February 2019, academic researchers conducted a series of interviews with IPO patent examiners specialising across different technical sectors. The goal was simple, understand exactly how a real search is carried out, in enough detail to design an AI system that could genuinely fit into that workflow, rather than one built around assumptions.
The Examiner's Step by Step Search Process
The purpose of any prior art search is, at its core, to find the closest existing prior art that could affect whether a patent application is granted. But behind that simple purpose lies a surprisingly intricate, multi-step process. Based on their interviews, the researchers mapped out the process examiners actually follow, step by step. It begins with examining the patent's claims and identifying possible keywords and relevant terms. From there, the examiner distills what the truly defining, novel part of the invention actually is, and uses this to form what's called a "search statement." Next comes identifying the most relevant classification codes, drawing on both the keywords already identified and the examiner's own background knowledge of the field. At this point, examiners sometimes carry out an optional background search purely to identify the most suitable terms and synonyms before committing to a formal query.
With this groundwork in place, the examiner forms actual search queries, primarily using a professional tool called EpoqueNet, produced by the European Patent Office, combining keywords, classification codes, and Boolean logic. This produces a set of candidate patents judged likeliest to be relevant to the application in question. What follows next is often the most laborious part of the whole process, sifting through the retrieved documents within EpoqueNet, using tools like colour coded highlights, virtual "drawers," and sticky notes to work out which of the retrieved patents are genuinely the most relevant. From there, examiners narrow the results down further, often turning to the drawings included in patents and manually disambiguating overlapping or similar concepts, in order to identify genuinely close conceptual similarities. Some examiners will also carry out an optional search of published research or other online materials before finally reaching a conclusion, a professional judgement about whether the application in front of them is genuinely novel and inventive.
Why the Search Statement Is the Hardest Part
Among all these steps, one stands out as particularly critical, defining the search statement. This single step requires the examiner to develop a genuinely clear understanding of both the critical subject matter of the application and its potential novelty, arguably the intellectual heart of the entire search process. It's also worth noting that this isn't a one and done step. Examiners routinely revise and refine their search statement multiple times as their understanding of the relevant prior art, or of the application's potential patentability, develops over the course of the search. Interestingly, the final search statement often ends up including words that never actually appeared in the original patent claims at all, a reflection of just how much interpretive, expert judgement goes into framing the right question to ask.
Search Techniques and the Biggest Time Sink
Of all these steps, the researchers found that sifting through the sheer volume of retrieved patents is consistently the most time consuming part of the entire process. In terms of overall approach, examiners tend to work very systematically, a habit shaped by just how structured patent information itself tends to be. The actual search techniques currently in use are impressively varied: keyword searches, classification based filtering, Boolean logic, proximity operators (which look for terms appearing near each other), truncation operators (such as matching the right hand portion of a word), linking to full text documents and related patent families, connecting to both external and internal document repositories, careful selection of keywords and synonyms, combining previously saved search queries in sensible ways, iteratively refining stored search queries as new phrases and terminology come to light, citation based searching, and even navigating across multiple languages. Once documents are retrieved, examiners rely on a smaller but no less important toolkit for post search analysis, mainly colour coding, highlighting, and the drawer and sticky note features built into EpoqueNet.
This is exactly the kind of manual, time-intensive sifting that AI-powered tools aim to reduce. If you want to see how an automated system handles the ranking and highlighting work described above, you can try a free novelty search with WOIPS on your own invention.
What Examiners Actually Need From an AI Tool
From these observations, the researchers were able to distill a clear picture of what examiners actually need from any future AI-assisted tool. The key user requirements were: retrieving the closest matching documents, ranking relevant documents sensibly, suggesting useful synonyms, suggesting appropriate classifications, suggesting which passages to highlight, and visualising or clustering the retrieved documents and concepts in a way that makes sense at a glance. Beyond these core requirements, examiners also flagged a set of additional, more ambitious features that would be genuinely desirable but fell outside the scope of this particular feasibility study. These included the ability to retrieve image based patents (rather than relying purely on text), tools to help collect evidence for due diligence, features that could spot ambiguity within a document, tools to find contradictions between documents, support for disambiguating word senses, ways to visualise relationships across multiple documents, and even the ability to search directly through pictures and drawings. In terms of overall scope, the study was framed around searching and filtering granted patents and published patent applications across a number of different technical sectors.
The Key Challenges of Bringing AI Into Patent Search
Alongside these requirements, the interviews also surfaced a clear set of key challenges facing anyone trying to bring AI into this space. Patent searching, the researchers concluded, is a highly interactive and genuinely complex process, one that typically requires multiple searches, a diversity of search strategies, and careful ongoing management of the whole search effort. From a linguistic and semantic standpoint, the core challenges are consistent with what the literature review had already suggested: dense legal wording, unusually long sentences, heavy use of acronyms, and the inherently technical nature of patent claims.
The researchers also grounded their thinking in a well established framework from the information retrieval field, which holds that the usability of any information retrieval system really comes down to three distinct aspects: its effectiveness, its efficiency, and the satisfaction of the people using it. For the purposes of this particular feasibility study, the team chose to focus primarily on effectiveness, in other words, how well the system manages to surface documents that genuinely meet specified relevance criteria. The usual gold standard for measuring this is manual judgement by human experts. But here the researchers flagged an important subtlety, backed by wider research in the field: human judges don't always agree with one another about what counts as "relevant." Interestingly, though, people tend to agree with each other far more readily when asked a comparative question, "which of these two documents is more relevant to the query?", than when asked to make an absolute, standalone judgement, such as simply "is this document relevant to the query?" This distinction matters a great deal for how any evaluation of a prior art search system should actually be designed and interpreted.
From Ranked Retrieval to Concrete Technical Requirements
One more foundational point from the information retrieval field shaped the study's overall approach: nearly all modern search technologies rely on ranked retrieval rather than simple Boolean matching, and it's widely accepted within the information retrieval research community that ranked retrieval is almost always more effective than a pure Boolean approach. This principle, the idea that a system should return documents ordered by likely relevance, rather than simply flagging documents that either do or don't match a query, underpins much of the technical design that followed in later stages of the project.
Building directly on these interviews and observations, the researchers translated the examiners' needs into five concrete technical requirements that would guide the actual system development. The first, TR1, was automated query expansion, the system suggesting synonyms, meronyms (part of relationships), hyponyms (more specific terms), and hypernyms (broader terms) to enrich a search. The second, TR2, was automated document classification, meaning the system would suggest additional classification codes for a given application. The third, TR3, was automated identification of similar documents, using semantic similarity measures rather than simple keyword matching. The fourth, TR4, was producing an automated ranked list of relevant documents, based on how similar each document is to the query. And the fifth, TR5, was visualising the distinguishing characteristics of retrieved documents through topic modelling, giving examiners an intuitive, high level view of what's actually in a large set of search results.
Measuring Search Quality: Precision, Recall, and Beyond
Finally, to actually measure how well any of this worked, the researchers turned to two classic metrics from information retrieval: precision and recall. Precision indicates how many irrelevant documents got swept up alongside the genuinely relevant ones in a search, essentially a measure of exactness or quality. Recall, by contrast, measures how many truly relevant documents were missed altogether, a measure of completeness. Mathematically, precision is calculated as the number of relevant items retrieved divided by the total number of items retrieved, while recall is calculated as the number of relevant items retrieved divided by the total number of relevant items that exist anywhere in the whole collection.
Both of these measures require documents to be manually labelled and assessed for relevance by human experts, which makes it genuinely impractical to evaluate every single document in a very large collection, in such cases, researchers typically fall back on precision alone. It's also worth noting that total recall isn't actually necessary in a real prior art search: an examiner only needs to find one reference that predates the filing of the application in question to establish that it isn't novel. In practice, most searchers do try to find more than just one such reference for thoroughness, but there's no formal requirement to achieve complete recall across the whole collection.
Given that the ranking of documents matters so much in practice, the feasibility study specifically chose to measure what's known as "precision at k," or precision@k, where k represents a cut off point within the ranked list of retrieved documents, for instance, looking only at the top 10, 20, or up to 100 results. The value of k wasn't fixed in advance, instead, a range of possible values was tested. Alongside this, the study also used the F measure, a weighted harmonic mean combining precision and recall, specifically to assess how accurate the classification algorithm was. Beyond these purely quantitative measures, the researchers also incorporated more human centred methods, including an agreement measure to assess how interpretable the topic modelling results actually were to real users, and a focus group discussion designed to explore the broader user experience of working with the system.
Manually recreating this level of analysis is exactly what makes prior art search so labor intensive for examiners and applicants alike. If you'd like to see how these principles translate into a practical tool, you can explore the WOIPS Novelty Search service and how it applies ranked, AI-assisted retrieval to your own invention.
Frequently Asked Questions
What is a prior art search in patent examination?
A prior art search is the process patent examiners use to find existing patents, applications, or publications that could affect whether a new application is novel and inventive.
How do patent examiners search for prior art?
Examiners analyze the application's claims, form a search statement, identify classification codes, and run queries in tools such as EpoqueNet, then manually sift and narrow the results using highlighting and clustering features. You can see a simplified version of this process yourself with a free novelty search.
What tool do patent examiners use to search prior art?
Most examiners at the European and UK patent offices rely on EpoqueNet, a professional search platform developed by the European Patent Office, alongside classification databases and internal repositories.
Why is defining the search statement the hardest part of a prior art search?
Because it requires the examiner to interpret the core novel aspect of an invention in a way that captures relevant prior art, even using words that never appear in the original claims, and it's often revised multiple times during a search.
Can AI improve the patent prior art search process?
Yes. AI can assist with query expansion, automated classification, semantic similarity matching, ranked retrieval, and visual clustering of results, reducing the manual sifting that examiners currently do by hand. Explore how this works in practice on the prior art search page.
This article is adapted from: UK Intellectual Property Office, AI-assisted patent prior art searching – feasibility study, April 2020, ISBN 978-1-910790-80-9, © Crown Copyright 2020, licensed under the Open Government Licence v3.0.