Landscape reports arrange what is already sitting in the spreadsheet. The questions worth asking are not in any column.
A patent landscape is a study of all the patents in one area of technology. It shows who is inventing what, where, and how fast. Companies use it to decide where to spend on research. Law firms use it to advise clients. Universities use it to find who might license their work.
People have always wanted the same thing from a landscape: something specific enough to act on. Where is the opening? What should we do next? For a long time, that was out of reach. So a landscape settled for being a bird’s-eye view. You collected the patents, charted them, and looked at the shape of the field from above.
The reason was time. Every chart in a landscape rests on one slow step: an analyst reading the patents and sorting them into categories by hand. Turning that into actionable insight takes far longer. Imagine mapping ~5,000 patents, one by one, against the problems they solve and the solutions they use. The charts can only be as deep as the sorting, and the sorting could only go so deep.
Data analysis everywhere else has been through the same shift. Ten years ago, most business reports showed what had already happened. As the tools improved, the questions got sharper. A sales team today does not just see last quarter’s numbers. It sees which customers are likely to leave next.
Most landscape reports have not made that shift. They still show the same set of charts. Patents per year. The top companies. The countries where patents were filed. A breakdown by classification code (such as IPC, CPC, Japan’s FI codes or Derwent’s classes), the labels used to sort patents into technology categories. More and more often, an AI tool sits on top and writes a summary of what the charts already show.
The result is a slide deck. It stays with the IP team and a few people in R&D, who read it a couple of times. Then it stops, because it has no “what’s next.” Even when it reaches the people who decide budgets, it rarely moves them. Why this area? How sure are we? What should we do next? The charts were not built to answer that. Counting can raise awareness. It cannot make a decision.
The tools have changed, even if the reports have not. Patents are among the most structured public data there is. Every patent has a date, an owner, a country and patent claims: precise sentences that define exactly what the invention protects. Data this structured can, and should, be measured, not just counted. Until recently, measuring every patent against every other one was too slow and too costly. Now it is not.
Yet almost nobody does it. Nobody asks patent data a mathematical question.
Semantyr was built to ask those questions. It starts where the search ends. An analyst brings the patents they have already found with their usual search tools. Semantyr measures them, sorts them, and finds the relationships and meaning inside the set. The analyst stays in control at every step.
To show what that looks like, this article follows an illustrative landscape we ran on real patent data: ~2,500 patents on vehicle lighting. At each stage, it shows the usual way a landscape is done, and the measured way.
Every landscape starts with a conversation. Someone in the business needs an answer. Where should we invest? Who are we up against? Is there room for a new product? A landscape can serve eight or ten different purposes, from finding gaps in a field to weighing strengths and weaknesses against a competitor. The analyst’s first job is to find out which one this is.
Next, the analyst finds the patents. They search patent databases, often using classification codes to narrow the results, and download the matches into a spreadsheet.
Neither step is about software. Both are judgment: knowing what the business needs, and which patents answer it. That part stays with the analyst, and it does not change with Semantyr. Semantyr starts once the patent set is ready.
With the patent set ready, the slow part of a landscape begins.
The analyst reads the patents and sorts them into categories, usually as a tree: broad branches at the top, narrower ones beneath. For a few thousand patents, this can take weeks. Every chart that follows depends on it.
To save time, many analysts lean on classification codes. Every patent already has them, so it is tempting to sort by code instead of by reading. But codes were made for a different job.
A classification code is a label from a fixed list. Patent office staff choose it so the patent can be routed to the right examiner and found again in later searches. It describes the general subject, like a shelf in a library. The list is also revised a few times a year, so the same patent can move shelves over time.
Shelves are good for finding things. They are not built for comparing them. Two patents that solve the same problem in opposite ways can sit on the same shelf. So can two that have little in common. A landscape sorted by code inherits those shelves, chosen by someone who was not doing your analysis.
A code tells you which shelf a patent sits on. It does not tell you how close two patents are.
AI offers a newer shortcut. A language model, the kind of AI behind tools like ChatGPT, Claude or Copilot, can read a patent list, sort it into categories, and even flag the patents that look off-topic. It is quick, and the result reads well. Run it again on the same list, though, and the categories can come back different. The patents did not change. The answer did.
Semantyr takes a different route. At the core of its reasoning engine is a mathematical model. The model reads each patent’s claim, the sentence that defines what the invention protects, and measures how close every patent is to every other one. That closeness is a number. Two patents that protect nearly the same thing come out close. Two that only share a few words do not. Run the same patent set again and you get the same groups, every time.
Because it measures meaning, not words, language doesn’t matter. A claim filed in Japanese and a claim filed in German that protect the same mechanism end up side by side.
Think of ten patents in a spreadsheet. They are ten rows, and nothing about the rows tells you which patents are alike. Measured by closeness, the same ten patents form a map. Some sit together, some sit apart, and the distance between them shows how technically similar or different they are.
In a spreadsheet, patents are rows. Measured, they become a map.
Patents that sit close together form technology groups (called Classifications in Semantyr). Each group lists the patents it was built from, so the analyst can see why it exists before deciding whether to keep it. In the lighting study, each of the ~2,500 patents ends up in a group with the ones that take the same approach.
This is where the analyst takes over. The result is a whiteboard, not a final answer. The analyst can merge two groups, split one, or delete one that doesn’t serve the question. Even data cleaning works this way. If one company appears under five spellings, merging them can move it from fourteenth place to third in the rankings. That is a conclusion, not housekeeping, so the analyst sees the change and decides.
The math proposes the groups. The analyst decides which ones stand.
Sorted patents are still just a spreadsheet. The next step is to read them, and turn the groups into findings.
In a typical landscape, reading means charting. If you have read a landscape report, try to remember one finding from it. Not a use. One specific sentence about a technology that someone acted on.
Look inside the report and the reason becomes clear. Patents per year. The top twenty companies. The countries where they filed. A breakdown by classification code. Every one of those numbers was already in the spreadsheet before the analysis began. Sorting a column and drawing it as a bar chart is not a finding about the technology. It is the file, arranged.
The questions that lead to a decision are different. They come down to five:
Has most of the field agreed on one way of solving the problem, or is everyone still trying different things?
Which approaches were abandoned, and what replaced them?
Which companies are filing more, which are filing less, and who has just arrived?
Does one company quietly lead across several parts of the field?
Which earlier patents does much of the later work closely resemble, and who owns them?
You cannot answer any of these by sorting a column. Each one asks how patents relate to each other, and that has to be measured. This is the mathematical question from the opening.
Take the first question: is this field settled, or still up for grabs? The lighting study has five technology groups. We asked the question of one of them, made up of four sub-groups. The mathematical model compared every patent in each sub-group with every other one, and measured how much of it had come together on one shared approach. Each sub-group falls into one of three bands:
The smallest sub-group, ~50 patents on headlight power control, is settled: roughly two-thirds of it shares one approach. Integrated light modules, ~150 patents, is still forming, at about a quarter. The other two, ~500 patents between them, are wide open. Nothing in them has come together at all.
Four sub-groups, one question, four different answers. None of them could be counted from the spreadsheet.
A wide-open area of ~500 patents is something a business can act on. The field is still open, entry is still possible, and whoever finds the approach that takes hold may set the direction for everyone after. No standard chart shows this. It isn’t in the data until something measures it.
Measuring shows how each group behaves. The most valuable findings, though, are often smaller than any group.
A standard landscape is built for the view from above. It tells you how big a field is, how fast it is growing, and who files the most patents. That view is useful. But it works with totals and averages, and an average has a blind spot.
Think of a country’s average temperature in July. It can look perfectly mild while one state is in a heatwave and another is unusually cold. The average is correct. It just hides the places that matter.
Patent data works the same way. A company that has just arrived. An approach that quietly stopped. A new idea in a corner of the field. They are small, local and easy to miss from the top. An average hides the exact thing you were looking for.
Here is one from the lighting study. Inside one technology group sits a pocket: a small sub-group of ~15 patents that all tackle the same narrow problem.
For about four years, almost all of them came from one company. Patent after patent, it refined the same design: a separate circuit for each light source. One company, quietly owning a small technical problem.
Then new companies arrive. They had never filed in this pocket before, and their patents are not refinements of anything. They solve the same problem in a completely different way: one control loop driving whole clusters of tiny LEDs together.
If you work in that field, this is the most useful thing in the whole study. Someone has just walked into a room that one company had to itself for years, and they did not bring the same tools.
Now look for those patents in a standard report. On a patents-per-year chart, they barely register. In a ranking of top companies, the newcomers don’t appear. In a breakdown by classification code, they sit on a shelf with hundreds of others.
Every standard view of the study erases them.
Semantyr is built to go down to them. In a deep dive, the analyst can select any part of a technology group, down to a pocket of a few patents, and ask the same questions there. Is it settled or open? Who is arriving? What stopped? At that scale, a handful of patents is the whole world, and a couple of them can change everything.
This is what an actionable insight looks like. From above, the field looks steady. Up close, a new competitor has arrived with a different idea. One view tells you what happened. The other tells you where to look next.
Say a finding like that makes it out of the IP team. It goes onto a slide. The slide goes to someone senior who has to make a call, perhaps whether to spend next year’s research budget responding to the new companies. Somewhere in that meeting, a person who had nothing to do with the analysis asks the question that decides everything. Where did this come from?
There are only two kinds of answer. You can point at the patents that produced it. Or you can say the tool produced it.
The first answer keeps the conversation going. The second ends it. It also ends the report. Nobody carries a number into a decision they cannot defend.
This is where language models fall short. They are built to write. Ask one what is happening in a technology area and it will answer fluently. Ask again and it will answer a little differently. It is not measuring anything. It is producing likely sentences. That makes it a good writer and a poor analyst. A landscape is a measurement problem, and measurement is a job for math.
Think of a company’s annual report. Accountants produce the numbers. A writer explains them. Nobody lets the writer change a number to make a sentence read better.
Semantyr’s reasoning engine works the same way. It has two parts: the mathematical model, and a language model that only writes.
The math decides what is true. The language model only decides how to say it.
The mathematical model forms every technology group, counts every patent and computes every percentage. Only then does the language model see the results, and its job is to put them into plain sentences. It never chooses which patents belong together. It never produces a number.
That split is what lets every finding answer the three questions a meeting will ask.
Where did it come from? Every number points back at the patents that produced it, listed by patent number. That includes the awkward ones at the edges, the ones that fit least well. They can be opened and read in the meeting.
What does it actually say? The report speaks in plain terms, with no internal scores or codes that only the software understands. “Most of this area is refining one circuit design” can be checked and argued with. “Cluster 3, score 0.62” cannot.
Will it say the same thing next quarter? Yes. The numbers were measured, not written, so the same patent set gives the same answer. Present it to the board this year, run it again next year, and compare the two.
Technical papers call these provenance, explainability and repeatability. In a meeting, they are just three questions, and a report either answers them or it does not.
A report built this way invites the hard question instead of dying on it. That is the difference between a study that dies and one that gets used.
Even a finding that survives the meeting raises the next question: what now?
A typical landscape ends as a document. Someone presents it, a few colleagues read it, and it is filed. Without a “what’s next,” there is little it can lead to. Often the best outcome is a second study, commissioned to dig into one broad insight from the first. Weeks later, the cycle starts again.
The best outcome of a typical landscape is often another landscape.
With Semantyr, that second study happens inside the first. The open areas and the pocket from the lighting study are not where the report ends. They are where the next deep dive starts.
Take a problem the field has not solved. In a deep dive, the reasoning engine looks for places in the same patent set where a similar problem was solved by other means, and puts the two together as a possible direction. Then it checks that direction against the patent set and shows the closest thing that already exists. That is the closest within this study, never a claim that the idea is new.
An insight tells you where the opening is. An actionable insight tells you what to try there.
The study still ends. It just ends with your next move.
People always wanted landscapes to tell them what to do next. What held them back was time and accuracy. Sorting patents by hand is slow, and every chart is only as good as the sorting. That is why, for so long, a landscape stayed a view from above.
Measuring addresses both. The mathematical model sorts the patents by meaning, the same way every time. It is not instant, because measuring every patent against every other one takes real computation. But it takes a fraction of the time a manual sort does. The analyst spends their time where it counts: deciding what the results mean.
The data was always there. Nobody was asking it anything.