Patents are one of the richest open sources of technical intelligence in the world. A patent has to describe an invention in enough detail for a skilled person to reproduce it. It names the applicant and the inventors, it is classified by technology, and it is usually published 18 months after filing, often years before a product reaches the market. Read in bulk, patents show who is investing in which technologies, where, with whom and how fast.
This guide covers how patent analytics works, the indicators analysts use and the mistakes that make patent studies misleading.
What a patent document tells you
| Field | Intelligence value |
|---|---|
| Applicant / assignee | Which organisations invest in a technology. Assignment records reveal acquisitions and licensing. |
| Inventors | Key people, team size, mobility between companies, collaborations |
| Priority date and country | When and where the invention started |
| Family members | Where protection is sought: a signal of which markets matter |
| Classification codes (IPC/CPC) | Which technologies the invention belongs to |
| Claims | What is legally protected, as opposed to what is merely described |
| Citations | Which earlier technology it builds on, and later, who builds on it |
| Legal status | Granted, pending, lapsed, opposed or litigated |
The concept that matters most is the patent family: the set of applications in different countries that protect the same invention. Counting individual documents instead of families inflates the numbers for applicants that file in many countries. This is one of the most common errors in published studies.
Building a patent landscape
A patent landscape maps the patents in a field to answer a question, for example: who leads in solid-state batteries, and is the lead shifting towards China? WIPO has published dozens of landscape reports that serve as good models. The workflow is usually:
- Define the question and scope. Technology boundaries, time window, jurisdictions.
- Build the search. Combine keywords (with synonyms and translations) and classification codes from the IPC or the more detailed CPC. Keywords alone miss documents that use unusual wording. Classification codes alone pull in noise.
- Test and refine. Read a sample of results, measure how many are relevant (precision), check that known key patents are included (recall), and adjust the search.
- Clean the data. Group families, and harmonise applicant names: “Toyota Jidosha KK”, “Toyota Motor Corp” and their subsidiaries need to become one entity.
- Analyse. Trends, leading applicants, geographic spread, sub-technology clusters, collaboration networks and citation impact.
- Interpret. Explain what the numbers mean for the decision, with caveats.
Free databases such as Espacenet, Google Patents, PATENTSCOPE and The Lens are enough for many landscapes. Commercial platforms add cleaned applicant names, legal status data and analytics.
Indicators and what they really measure
- Filing counts and growth. Activity and investment. Not quality.
- Family size and international filing. How much an applicant values the invention. Filing in many countries is expensive.
- Forward citations. How often later patents cite this one. Widely used as a quality proxy, but it depends on field and age. Compare like with like.
- Technology share and specialisation. An applicant’s or country’s share of a field relative to its share of all patents. This shows focus.
- Co-applicant and co-inventor networks. Partnerships, university–industry links and knowledge flows.
- Claims analysis. The only way to judge what is actually protected. It requires reading, and AI now makes reading at scale feasible.
Pitfalls that ruin patent studies
- Counting documents instead of families.
- Ignoring the publication lag. The last 18 months always look like a decline because many applications are not yet published.
- Treating Chinese filing volumes as directly comparable. China files more patent applications than any other country. A large share are utility models or domestic-only filings, whose average value differs from that of international families. Separate the types and look at international filings too.
- Unharmonised applicant names, which split one company into many.
- Confusing patents with products. Many patents are never commercialised. Many products rely on trade secrets rather than patents.
- Search bias. A query built in English with English keywords will under-count documents filed only in Chinese, Japanese or Korean.
Patents and innovation policy
Patent data is central to innovation policy. Governments use it to evaluate R&D programmes, identify strategic emerging industries and measure technology transfer from universities. Treat it with care in this context too. Subsidies and evaluation systems that reward filing counts produce more filings, not necessarily more innovation. Several countries, China among them, have changed their policies to reward patent quality rather than quantity for this reason.
The NPE debate
Non-practising entities (NPEs), also called patent assertion entities, hold patents without making products and earn money from licensing and litigation. Critics call the aggressive ones “patent trolls”. Supporters say some NPEs make markets for technology more efficient: they let small inventors get paid, aggregate patents and lower transaction costs.
The evidence suggests both are true for different business models. A 2016 study by the US Federal Trade Commission found that most demands came from “portfolio” entities with negotiated licences, while a large volume of low-value “litigation” entities filed many suits and settled for less than the cost of defending them. For analysts, the practical lesson is to check litigation databases and assignment records when you map a field. NPE activity is a risk indicator for anyone launching products in that space.
Patent analytics in the age of AI
Machine learning is changing patent analytics in three ways. Semantic search finds relevant documents by meaning rather than keywords, which reduces language bias. Automated classification and clustering map sub-technologies without hand-coding. Language models can read claims and summarise differences between thousands of documents. Search and summaries still need validation by someone who understands the technology and the law, but the cost of a good landscape has fallen sharply. That makes patent intelligence accessible to smaller companies, universities and journalists.