1 in 8 patents has a flaw
you only find when you try to enforce it

A patent’s independent claims are the stand-alone sentences that define its broadest protection. We checked them for two families of flags.

Narrowness flags. Under the all-elements rule, a product infringes a claim only if it satisfies every requirement in the claim. Each added word is one more thing a competitor can decline to do. Claim length is the standard measure of scope in patent economics (Marco, Sarnoff & De Grazia 2019; Kuhn & Thompson validated word count against patent attorneys’ own scope judgments). We flag a patent only when even its broadest claim is in the longest tenth of litigated claims from its own technology area — a nine-way segmentation by the patent’s primary CPC classification, with the flag line running from 265 words in semiconductors to 462 in fintech (the full table is in Section 2). A long claim is lawful, and claims run longer in some fields than others: claims tend to narrow in crowded art, and a decade of shifting §101 eligibility law has pushed software and fintech drafters to trade breadth for allowance, line by line, in negotiation with an examiner. Segmenting the benchmark absorbs that field effect — each patent is measured against what enforceable coverage looks like in its own field — so what the flag tracks is how much coverage survived prosecution, relative to the patents from the same art that get enforced.

Defect flags. Six defects, each tied to controlling authority. Four can be measured from claim text across the whole corpus. The other two, divided infringement and unsupported claim terms, require reading a claim against its specification or its actors, so they are measured on samples and reported alongside the census headline:

We also tested four other defect categories that did not survive false-positive review; the methodology section describes them, and their samples and verdicts ship with the study’s data files.

What the flags do not measure. Everything in this study is read off the claim text. The test that decides what a patent is worth happens somewhere else: against a competitor’s product, element by element. A clean claim that no shipping product practices protects nothing, and a flagged claim that a rival practices anyway can carry a licensing program. We did not run that test; coverage is a claim-by-product mapping, and it is a different study. The flags say whether a claim is ready for enforcement, not whether there is anything to enforce it against.

The corpus. Google Patents Public Data on BigQuery, the standard public dataset of USPTO records. Every U.S. utility patent granted June 1, 2025 through April 21, 2026 whose primary CPC classification is in the electrical, computing, physics, mechanical, or medical-device arts: 235,577 patents. Chemistry, pharmaceuticals, and biotech are excluded by classification code, because constructions we flag elsewhere (closed transitions, exclusions) are standard intentional practice there and would inflate the counts. The claim parser was validated blind: a second model re-read 200 claims with no labels shown and agreed on 199. The single miss errs in the direction that makes our length numbers conservative.

The benchmark. The narrowness flag needs a yardstick built from claims that demonstrably matter. We assembled 9,873 U.S. utility patents asserted in district-court litigation from 2020 through 2025, from public court records, and ran their 29,115 independent claims through the same parser. The median asserted claim runs 159 words. The 90th percentile is 318. The 95th is 392. (There is a selection story here, and it works in the benchmark’s favor: owners rarely assert patents they expect to lose, which is what makes asserted claims a reasonable picture of what enforceable coverage looks like.)

The litigated population behind this benchmark is profiled in depth in our companion study, Shaping Your Portfolio from the Patent Up, which analyzes what separates asserted patents from the ones that never get enforced.

Median asserted-claim length is nearly the same in every technology area — but the tail is not, and the flag lives in the tail. So the yardstick is segmented: each asserted patent is assigned a technology area from its own primary CPC classification, the same rule later applied to the corpus, and each area contributes its own 90th-percentile flag line:

The medians span only 140 to 213 words, but the 90th percentiles span 265 to 462 — a 1.7x range. A single all-fields line (an earlier release of this study used one, at 311 words) sits at the 93rd percentile of litigated semiconductor claims and below the 80th of litigated fintech claims, which overstates narrowness in long-claim fields and understates it in short-claim ones. Segmenting removes that distortion; the sensitivity analysis comparing the two regimes ships with the study. A cross-check that segments the benchmark by the asserting party’s industry instead of the patent’s CPC gives the same picture: medians 152 to 235, 90th percentiles 280 to 552.

1. Of 235,577 patents examined, 12.6% carried at least one verified flag, each patent measured against the litigated claims of its own technology area.

2. Some companies had 28 times the flags of others. Cleanest large portfolio: a 1.0% flag rate. Highest: 26.9%.

3. Narrowness accounts for nearly all of it (10.9 of the 12.6 points). The median new patent’s broadest claim runs 174 words; the median asserted claim runs 159. For 55% of new patents, even the broadest claim exceeds the asserted median of their own technology area.

4. In a 1,500-claim sample, 1.3% of independent method claims were clearly divided (95% CI 0.8 to 2.0%): as written, no single actor performs every step.

5. Two defects require reading the specification, and both are rare. 368 patents (0.16% of the corpus) claim an exclusion the specification gives no basis for. And 7 of 1,160 individually reviewed patents (0.6%) claim a term the specification never describes, a higher rate than two of the defects measured across the full corpus.

Among companies with at least 140 grants in the window, all filing through the same office under the same rules, flag rates ran from 1.0% to 26.9%. The ten cleanest portfolios include Qualcomm, Nvidia, and Intel; the ten highest include two major banks and two payment networks. These figures already give every company the benefit of its field: each patent is flagged only against the litigated claims of its own technology area, so a fintech patent has to be long by fintech standards (past 462 words) to count. The banks and payment networks top the list anyway.

The named incumbents we track did hold an edge in aggregate, with a flag rate of 7.6% against 14.2% for the rest of the market. But the range within the tracked group is far wider than the gap between the groups.

The clean end is dominated by semiconductor firms (Qualcomm, SK Hynix, Nvidia, Intel, Micron, TSMC), whose portfolios pair short claims with near-zero defect flags. The high end mixes two very different failure modes, mapped in Section 6. In the rankings, we name only the companies that score well; poor scorers appear anonymized ("a payments network," "one major U.S. bank"). Patent examples are the one exception, since each links to a public record, and the record carries the owner’s name.

Narrowness: 10.9% of patents flagged; 55% above their field’s asserted median

A typical new patent’s broadest claim, at 174 words, is already past the 159-word asserted median. In 55% of new patents, not one independent claim comes in under the asserted median of the patent’s own technology area. The flag is reserved for patents whose best claim is longer than 90% of everything that gets enforced in their field. Those patents are 10.9% of the corpus. In absolute terms, more than 25,000 of them issued in under a year. One percent have no independent claim under 602 words.

By industry, banking and payments patents still stand apart — and now the comparison is fair, because each field is measured against its own litigated tail. Even with fintech patents needing 462 words to trip the flag, banking and payments run 15.1%; semiconductors run 3.9% against their 265-word line, and medical devices 3.5%. Under the earlier single 311-word line, banking and payments read 23.4% — segmentation cuts the excess roughly in half, which says half of the old gap was the field’s claim-length regime rather than the drafting. What remains is the finding: these portfolios are long even by the standard of the fintech claims that get enforced. One qualification survives: the benchmark selects — short, broad patents are the ones worth asserting — so every flag line is generous, and the rates here are floors.

Median claim-1 length rose from 138 words in 2005 to 181 in 2015, and it has stayed there since, sitting at 188 in 2025.

Divided infringement: 1.3% of method claims (sampled)

Half of all patents (49.6%) contain at least one independent method claim, for a total of 147,530 in the window. Under BMC/Akamai, a method claim is infringed only when one party performs or controls every step. The question "how many actors does this claim require?" has to be read claim by claim, so we drew 1,500 method claims at random and adjudicated them conservatively. 1.3% carried a clear divided-infringement defect (19 of 1,500; 95% CI 0.8 to 2.0%). Another 2.2% were borderline. Applied to the 147,530 method claims in the window, the clear-case rate works out to about 1,900. The defect is avoidable by construction, since a claim that anchors every step to a single actor cannot be divided. The per-claim verdicts are in the data appendix. Because the rate is sample-measured, it is kept out of the 12.6% headline and reported on its own.

"Adapted to": 1.1% of patents

When a claim says a part is "adapted to" do something, courts have read the phrase two ways: as "designed to" (Giannelli) or as merely "capable of." Those are incompatible scopes, and the question gets resolved only at claim construction, long after the drafting decisions were made. The flag concentrates in medical devices, where three of the best-known makers run 12 to 14% on it, because describing an instrument by what its parts are adapted to do is the field’s natural idiom.

No-weight "whereby": 0.6% of patents

A "whereby" clause that merely restates the result of the recited steps gets no patentable weight. A judge reads the claim as if the clause were not there.

One telecom-equipment maker writes claims through a house-style "whereby the device is operative to: …" construction more often than any other large filer. But a whereby clause that introduces the operative steps gets full weight under Texas Instruments, and adjudication confirms it: 93% of that company’s reviewed whereby clauses are genuine limitations, and its verified rate is closer to 1%. Every company figure in this report reflects that kind of correction; the methodology explains how.

Unsupported claim terms: 0.6% of patents (sampled)

Claims are routinely amended during examination, years after the specification was written. Sometimes the amendment introduces a term the specification never describes. In the strict sense this is a §112(a) written-description defect, and the only reliable way to find it is to read each claim against its own specification. We reviewed 1,160 randomly drawn patents, extracting each broadest claim’s 3 to 5 most substantive limiting terms and checking every one against the specification. Synonyms and paraphrase count as support, and every candidate hit was adversarially re-audited before being counted. 7 of 1,160 patents carried a substantive claim term with no specification anchor (0.6%; 95% CI 0.3 to 1.2%). It ties the whereby flag and exceeds both verified negative limitations and closed transitions. Like divided infringement, the figure comes from a sample and sits outside the headline.

Unsupported negative limitations: 0.16% of patents (verified by reading the specs)

A claim may exclude something ("free of adhesive") if the specification shows the inventors contemplated the exclusion. Under Novartis v. Accord, silence is fatal. Whether any given exclusion survives depends on its specification, so we read the specifications behind all 1,854 flagged patents. 45% (836) have explicit exclusion support in their own text. Of the 621 that mention the excluded term without explicit exclusion language, individual adjudication found 250 supported and 368 unsupported: the term appears only as a required component, giving no basis for excluding it. Of the 397 that never mention the term verbatim, a full-record audit found 94% are wording misses and only 6% genuine silence. All three tiers together leave 368 individually verified unsupported exclusions, 0.16% of the corpus, plus an estimated 60 more from the sampled tiers.

Closed transitions: 0.02% of patents

A claim that opens with "consisting of" covers the listed elements and nothing more; "comprising" would have left it open to additions. About one patent in five thousand does this, making it the rarest flag in the study.

Across large portfolios, narrowness and defect rates are nearly uncorrelated. Companies that fail one measure almost never fail the other.

The far right of the chart is the narrowness mode. One major U.S. bank has no claim under 260 words in its median patent, with a defect-flag rate near zero. The drafting is careful, and the coverage still sits in the longest decile of the asserted benchmark. The top is the defect mode: a telecom-equipment maker at 11% verified defect flags with claims of ordinary length (a figure that overstates even so, since the company’s house style trips the whereby flag at well above the average false-positive rate; see the whereby section above). The lower left is where both numbers are small: Qualcomm at 0.16% verified defect flags on a 132-word median, along with Nvidia, Meta, Intel, and SK Hynix, across thousands of patents each.

The danger zone is where both problems meet: a median claim over 200 words, beyond the asserted median of every technology area but fintech, and a defect rate more than double the market’s 1.9%. Exactly one large portfolio lands there. It is a payments network with a 248-word median claim and a 3.9% verified defect rate.

A portfolio review that tests for only one of them will overlook the other. The remedies live in different places, too. Undoing narrowness means different filing and prosecution choices, made years before grant. The defect flags are cheaper, because four of the six are pattern-level constructions (a result-only whereby, a bare "adapted to," a closed transition) that a pre-filing checklist catches in a single read of claim 1.

Population. US B1/B2 utility grants, 2025-06-01 to 2026-04-21 (the latest grant in the public dataset at extraction), primary CPC classification in sections B, F, G, H or subclasses A61B/A61F/A61M; 235,577 patents; source: Google Patents Public Data on BigQuery. An English-claims-text guard drops zero patents in this window. "Broadest claim" means the independent claim with the fewest words. Parser validation: two-way structural check across the census plus a 200-claim blind semantic re-read by a second model (199/200; the single miss documented, error direction conservative).

Benchmark. 9,873 U.S. utility patents asserted in district-court litigation 2020 to 2025, assembled from public court records (CourtListener); 29,115 independent claims parsed with the same pipeline. Grant-to-assertion gap across the benchmark: median 6.3 years, interquartile range 2.6 to 10.8, with 29% asserted ten or more years after grant. Overall p50/p90/p95 = 159/318/392 words. Each asserted patent is assigned one of nine technology areas from its primary CPC (the Section 2 table); the same deterministic mapping is applied to every corpus patent, and every corpus patent maps to exactly one area. Per-area p90s (the flag lines) run 265 to 462 words; per-area medians 140 to 213. A cross-segmentation by asserting party’s industry (assigned per asserter from the court records) is published as a consistency check: medians 152 to 235, p90s 280 to 552.

Revision note. The initial release of this study flagged narrowness against a single all-fields line of 311 words (the overall asserted p90 as then computed) and reported a 13.6% headline, a 12.0% narrowness rate, and a 25x company spread. Because the per-area p90s span 265–462, a single line over-flags long-claim fields (fintech read 38.0% at its worst) and under-flags short-claim ones. This revision segments the threshold by technology area. The bottom-10 company list is essentially unchanged (8 of 10 members identical); the four fintech portfolios still occupy the top four slots, at roughly half their previously reported rates; the spread widens to 28x. The full flat-vs-segmented comparison, including a p95 variant, ships with the study.

Flags and numbers. A patent is flagged when its broadest claim exceeds the asserted p90 of its technology area (265–462 words; Section 2 table) or an independent claim carries a defect flag: verified-unsupported negative limitation, no-weight whereby, adapted-to, or closed transition (chemistry-clean subset). Every pattern screen was adjudicated against the controlling cases on about 100 hash-drawn hits in two same-model passes (Claude Opus 4.8; a test-retest check rather than an independent replication), with per-claim verdicts published. All rates in this report are single corrected figures: each screen’s raw hit rate multiplied by its audited precision, taking the harsher of the two passes (whereby 0.59, adapted-to 0.49 including construction-dependent hits, closed 0.16). The negative-limitation figure counts only individually adjudicated unsupported patents. The whereby precision was subsequently replicated at scale: a second model (Claude Sonnet 5) adjudicated a non-overlapping 1,000 whereby-flagged claims and measured 62.2% true no-weight clauses, slightly above the 59% used, so the published correction is a conservative floor. Before correction, the raw screen union is 14.5%; the corrected headline is 12.6% (10.9 narrowness points, exact, plus corrected defect points, overlap deduplicated patent by patent). The four retired screens are described below.

Retired screens. Four screens did not survive false-positive review and were not kept, all published with their audit samples. Mixed apparatus-and-method claims (IPXL Holdings): 98 of 100 audited hits proved to be lawful capability language under MasterMine v. Microsoft. Exemplary claim language ("such as," "preferably"), a genuine MPEP 2173.05(d) category: audited at around 35% precision, too noisy to publish. Trademarks as claim terms: almost all of the audited hits were definite standards marks (Bluetooth®, Wi-Fi®). And result-shaped "wherein" clauses ("wherein … is achieved," "wherein … results in"): the same no-weight risk can in principle hide behind wherein, so we screened for it and found it in 1.0% of patents. But two adjudication passes (5/12/83 and 6/12/82 true/defensible/false-positive on the same 100 hash-drawn hits) agreed that at least 94% of those hits are false positives, because wherein clauses are presumptively limiting (Griffin v. Bertina; Allergan v. Sandoz) and the flagged constructions overwhelmingly turn out to be genuine limitations. The whereby figure therefore stands as the floor for the no-weight-clause property.

Negative limitations. All 1,854 presence-flagged patents joined to their full specifications; excluded terms extracted from claims and searched with context windows. Tier 1 (explicit exclusion language, n=836) audited at 90% precision (n=80). Tier 2 (term present, no explicit language, n=621) fully adjudicated per patent: 250 supported / 368 unsupported / 3 unclear. Tier 3 (term absent verbatim, n=397) audited by full-record review (n=80): 6.25% genuine silence, remainder wording misses. The headline counts only the 368 individually adjudicated unsupported patents; a presence-only screen would have overstated the defect by roughly a factor of four.

Unsupported claim terms. Census text screen (stemmed content words of the broadest claim vs. full specification, corpus-frequency stoplist) retired after a 100-hit audit measured 2% precision; published with the retired screens. The property was instead measured on 1,160 randomly drawn patents in two rounds: 160 (Claude Opus 4.8: 1 unsupported / 157 supported / 2 unclear) and a non-overlapping 1,000 (Claude Sonnet 5 first pass, then adversarial re-audit of all 19 candidate hits by Claude Fable 5 under the same support standard: 6 confirmed / 13 refuted). Both rounds independently measured 0.6%. Method: 3 to 5 substantive limiting terms per broadest claim, each checked against the specification with synonyms and paraphrase counting as support. Pooled: 7 of 1,158 decided (0.6%; Wilson 95% CI 0.3 to 1.2%); per-patent verdicts and audit rationales published. Sample-measured, and therefore not part of the headline.

Divided infringement. 1,500 independent method claims drawn by deterministic hash from 147,530, in two rounds: 500 (Claude Opus 4.8: 8 clear / 12 borderline) and a non-overlapping 1,000 (Claude Sonnet 5: 11 clear / 21 borderline), a cross-model replication with consistent rates (1.6% vs 1.1% clear). Adjudicated against BMC/Akamai, clear cases only; pooled 19 clear / 33 borderline; Wilson 95% CI; per-claim verdicts published.

Company figures. All company figures are precision-corrected, on the same basis as the headline. The correction matters most where a house drafting style concentrates hits: one telecom-equipment maker’s "whereby the device is operative to: …" construction fires the raw whereby screen on 14.5% of its patents, more than any other large filer, yet 93% of its flagged claims that landed in the 1,000-claim review proved to be false positives, leaving a verified whereby rate closer to 1%. Flag rate = expected verified flags per 100 patents: narrowness flags (each patent against its own technology area’s line) and individually verified negative limitations count exactly; whereby, adapted-to, and closed-transition screen hits are weighted by their audited precisions (0.59 / 0.49 / 0.16). The scatter’s defect axis is the corrected probability that a patent carries at least one true defect flag. Because a single patent can carry more than one flag, a flag rate can slightly exceed the share of patents flagged. Companies aggregate subsidiaries under published name-mapping rules; separately listed affiliates excluded; 140+ grants required for ranking. Era-drift figures use claim 1 on a 2% per-year sample. All pipeline SQL, audit samples, per-claim verdicts, benchmark export, and figure code are published with the study. The full replication package is public, under an MIT license, at github.com/arcprime-ip/patent-defect-audit.