Technology deep dives

Patent analytics: patent counts mislead, citation networks predict

GetFocus

·

November 20, 2026

·

7

min read·Updated

Patent analytics: patent counts mislead, citation networks predict

Stay ahead of the curve

Get exclusive insights delivered directly to your inbox. Join thousands of tech leaders.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

TL;DR

  • A filing count measures effort and consensus. In 2012 3D television drew six times the filings of 4K and HDR, two years after 4K and HDR read 67.8% a year against 27.6% for 3D.
  • MIT research on 28 technology domains found no reliable relationship between the number of patents in a domain and how fast it improves. Citation measures tracked improvement strongly.
  • The citation network gives two signals, cycle time and knowledge flow, which together yield an improvement rate per technology. GetFocus reports R² ≈ 0.7 against measured rates across 150+ domains.
  • The rate is directional and comparative. Check any patent analytics by asking for the unit, the question the chart answers, a rate comparison between rivals, and a replay on past data.

Patent analytics turns patent data into evidence for a decision, and for the decision "which technology do we fund?" the filings-per-year chart has a poor record. In the technology races GetFocus has replayed, the technology with the most filings was repeatedly the one that lost, while the citation network, the record of how later patents build on earlier ones, pointed at the winner years earlier.

We build GetFocus, which measures how fast technologies improve from the way patents cite each other. This article explains why the two readings disagree, what the MIT research behind the method found, and how to check any patent analytics you are shown. The figures come from the public backtest library, where each reading uses only the patents that were public at that date.

What is patent analytics, and which question is it answering?

Patent analytics is the analysis of patent documents and their data (filing dates, applicants, classifications, citations) to answer a business or technical question. The work falls into three groups, and each needs a different measurement.

  • What has been patented? Patent search, for prior art and IP work.
  • Who holds what, and how strong is it? Landscape and portfolio analytics, which map filings per applicant, country and year.
  • Which competing technology will win, and when? Forecasting, which needs a measure of how fast each technology improves.

Counts suit the second question. The trouble starts when a count chart made for it is used to answer the third, and a rising filing curve is read as momentum. Our page on AI patent analytics sets the three groups of tools side by side.

What did patent counts say in a race we can check?

Take data storage in 2001. Hard drives out-filed solid-state storage six to one that year: 3,316 new patent families against 580. A reader of the filing chart saw an industry committed to magnetic disks. The improvement rate, computed from the patents public that year, read 49.6% a year for solid-state against 29.2% for hard drives, according to the storage backtest.

The filing chart took nine more years to agree. Solid-state out-filed hard drives for the first time in 2010, and in October 2015 Western Digital announced it would buy SanDisk for about $19 billion. In 2026 the filing picture is inverted, with 3,172 families for solid-state against 383 for hard drives.

An improvement rate works like an interest rate on performance. Solid-state at 49.6% against hard drives at 29.2% compounds to a relative gain of more than three times over nine years (our arithmetic on the 2001 readings, as an illustration; the backtest shows the gap widening after 2001).

The same split shows up in other cases in the library:

  • Powertrains. Hydrogen fuel cells still out-file batteries in 2026, at 2,841 patent families against 504 for NMC lithium batteries. The first reading, in 2001, had NMC improving at 54.2% a year against hydrogen's 21.4% (powertrain backtest).
  • Displays. 3D television filings peaked in 2012 at 2,017 families, six times the count for 4K and HDR. In 2010, 4K and HDR read 67.8% a year against 27.6% for 3D (display backtest).
  • Imaging. Silver-halide film ran 2,611 families in 1997, more than every digital imaging technology combined. In 1980, CCD imagers read 89.8% a year against film's 17.3% (Kodak backtest).

In lighting, the filament led LED on filings until 2002, twelve years after the first reading favoured LED (lighting backtest). In payments, cash infrastructure led online payments on filings until 2015. The seventh case, the rotary engine, tests the method on a loser. Dozens of manufacturers bought Wankel licences between 1961 and 1970, and at the first reading in 1972 the rotary improved at 4.8% a year against the piston engine's 6.6% (rotary backtest).

Why do patent counts mislead on the funding question?

A filing count measures inventive effort: how many inventions companies paid to protect in a year. Effort follows budgets, and budgets follow the consensus of the day. Hydrogen's filing curve recorded what carmakers, oil majors and governments had agreed to fund.

MIT researchers tested the assumption behind the count chart. Benson and Magee took 28 technology domains with known improvement rates and asked whether the domains with more patents improve faster. Their 2015 paper in PLOS ONE reports no reliable relationship between the number of patents in a domain and its rate of improvement. The measures that tracked improvement strongly came from citations and dates: how often a domain's patents are cited, how recent the patents are, and how soon after publication later patents cite them.

Counts also carry two mechanical distortions, both described on the GetFocus method page. One invention filed in fifteen countries appears fifteen times unless the analysis works on patent families. And when patents are read one by one, the breakthroughs show up late, while the citation pattern underneath them forms years earlier.

How does a citation network predict what a count misses?

Patents cite the earlier inventions they build on. Linked together, those citations form a network that records how knowledge moves through a field. GetFocus reads two signals from that network:

  • Cycle time: the gap, in years, between an invention and the earlier one it improves on. Shorter cycles mean a technology is advancing faster.
  • Knowledge flow: how much later inventions build on a given one, read from forward citations. Higher flow means bigger steps.

Together the two signals give each technology its improvement rate. Think of a building site. The filing count is the headcount at the gate, and the citation network shows how quickly each new floor goes up on the one below.

Storage shows why the headcount misled. Hard drives and solid-state iterated at a similar pace, with cycle times between four and six years for both. The difference sat in knowledge flow: solid-state's roughly doubled over the period, from 0.8 toward 1.7, while magnetic storage stayed between 0.3 and 0.8. Each generation of solid-state inventions carried more of the field forward.

The research behind this runs through 2020. Triulzi, Alstott and Magee tested patent-based predictors on 30 technologies and found that the centrality of a domain's inventions in the overall US patent citation network is an accurate and highly reliable predictor of its improvement rate (Technological Forecasting and Social Change, 2020). GetFocus automated the measurement and reports that its patent-based rate matches measured improvement rates at R² ≈ 0.7 across more than 150 domains, so it explains about 70% of the variation. Why technology improvement is predictable covers the research in full.

How do you check the patent analytics you are shown?

Four questions sort a count chart from a forecast.

  1. What is the unit? Ask whether the chart shows documents or patent families. Family-level analysis keeps a company that files widely from reading as a company that invents more.
  2. Which question does this chart answer? A filing curve answers who is active and where the money goes. Treat it as a map of the consensus.
  3. Are competing technologies compared on rate? A rate means most beside the rate of a rival doing the same job.
  4. Can the method be replayed? Every patent carries a date, so any method can be run on the patents public in a past year and scored against what happened. Our article on how to backtest a technology forecasting method gives the rules for a fair test.

For the count-based landscape itself, why patent landscaping is hard walks through the points where it breaks and how to fix each one.

What are the limits of reading the citation network?

The improvement rate is directional: an R² of about 0.7 leaves roughly 30% of the variation in measured rates unexplained, and the year a market tips also depends on cost, manufacturing scale and regulation.

The rate is also comparative: cash-payment infrastructure improved at a healthy 26% to 32% a year throughout the payments backtest and still lost, because online payments read between 200% and 390% a year in each listed reading from 2004 onwards.

The rate also needs data: GetFocus sets the floor for a reliable reading at 50 patent families per technology, and the signal is weaker where progress runs on secrecy or tacit knowledge.

Seven cases chosen by the company that sells the method prove little on their own, which is why the aggregate figure and a replay on your own industry's history matter more. The MIT results are correlations across 28 and 30 domains, and Benson and Magee caution in their paper that past results are an imperfect guide to future ones.

Filing counts keep their job, and they do catch up. In 2007 LED filed 7,899 patent families against 2,625 for incandescent lighting, five years after the filings crossed and seventeen years after the rate first favoured LED. The count confirmed the outcome late.

FAQ

What is patent analytics?

Patent analytics is the analysis of patent documents and their data, such as filing dates, applicants, classifications and citations, to answer a business or technical question. It covers three jobs: searching what has been patented, mapping who holds what, and forecasting which competing technology will win.

Are patent counts useless?

No. Filing counts show who is active, where and since when, which is what an IP or competitor question needs. They mislead when a rising filing curve is read as proof that a technology is winning. In 2012, 3D television filings ran at six times the count for 4K and HDR, the technology with roughly twice the improvement rate.

What is a patent citation network?

Patents cite the earlier inventions they build on. Linked together, those citations form a network that records how knowledge moves through a field. GetFocus reads two signals from it: cycle time, the gap in years between an invention and the earlier one it improves on, and knowledge flow, how much later inventions build on it.

How early does the citation signal lead the filing count?

In three GetFocus backtests the lead ran from nine to seventeen years. Solid-state storage led on improvement rate in 2001 and on filings in 2010. LED led incandescent lighting on rate in 1990 and on filings in 2002. Online payments passed cash infrastructure on filings in 2015, seventeen years after the 1998 reading put them at 99.9% a year against 25.6%.

Which chart is your next funding decision resting on?

If the case for a technology in your portfolio leans on a filing curve, a second reading is cheap. Send us the decision. Get a free 48-hour Decision Brief: a first-pass radar of the competing options and the improvement rate of each, read from the citation network.

Ready to see it in action?

Bring us one technology decision you're weighing. We'll send back the competing routes mapped by improvement rate and maturity.

Talk to us