Patent citations: Tracing spillovers or chasing shadows?
Measuring how ideas move across firms and regions
Few things matter more for innovation and economic growth than the way ideas spread. Economists have long emphasized that when new ideas spill across organizational or regional boundaries, they fuel the progress of others and amplify the returns to discovery. But while the importance of these “spillovers” is widely accepted, measuring them is far from straightforward.
What are knowledge spillovers?
In an earlier post, we explained that knowledge is a special kind of good: it is non-rival (many can use it simultaneously) and often non-excludable (making it difficult to prevent others from using it). Unlike a barrel of oil, knowledge doesn’t run out when someone uses it. Once created, it can spread widely—across universities, firms, researchers, and inventors.
Economists generally refer to “knowledge flows” as the movement of ideas from one actor to another. These flows can be paid for (through a licensing contract, for instance) or not. Knowledge spillovers are the subset of flows for which the originator does not capture the full value. Spillovers can happen deliberately, as when scientists share results at a conference, or indirectly, for instance, when competitors learn from a patent disclosure. In practice, it is nearly impossible for innovators to appropriate all the value of what they create; part of the value “spills” beyond the original creator.
Why do we care about knowledge spillovers?
Knowledge spillovers matter for two main reasons. First, they imply that the social benefits of innovation exceed the private ones. A company may invent a new tool, material, or method, but suppliers, rivals, or even customers can benefit far beyond what the firm itself earns. Take semiconductors: when Intel develops a new chip design or manufacturing process, competitors study the patents, reverse-engineer the products, and adapt some of the underlying techniques. Even if Intel fiercely guards its IP, part of the knowledge inevitably spills over to the rest of the industry. The result is faster diffusion of better technology into laptops, smartphones, and data centers—benefits to society that Intel cannot fully capture.
Second, spillovers don’t just spread ideas more widely; they make innovation cumulative. Each generation of ideas becomes a platform for the next. The smartphone, for example, combined breakthroughs in semiconductors, wireless communication, touchscreens, and software—each of which built upon earlier advances that had spread across firms, universities, and industries. This cumulative process is what economists mean when they say innovation is “standing on the shoulders of giants.” But the metaphor can be sharpened: every time knowledge spills over, it does not just let others climb higher—it makes the giants taller, raising the platform for all future innovators.
In search of knowledge spillovers
Paul Krugman remarked in his 1992 book that knowledge flows “are invisible; they leave no paper trail by which they may be measured and tracked.” A frustrating observation, given their central role in economic growth. Just a year later, Adam Jaffe, Manuel Trajtenberg, and Rebecca Henderson countered: “knowledge flows do sometimes leave a paper trail, in the form of citations in patents.”
The central question of their paper was simple: are knowledge flows more likely to occur locally? To find out, they compared the location of citing and cited patents against a neutral baseline. Each citation was matched with a “control patent” from the same technology class and year but without a citation link. If citing inventors were unusually likely to be in the same place, it would suggest that geography matters for knowledge flows.
The results were striking. Citing patents were far more likely to be located near the cited ones than were the controls—three times more likely at the state level, and up to ten times more likely within a metropolitan area. The effect was strongest soon after the original patent was filed and diminished over time, consistent with the idea that knowledge starts local before spreading more widely.
Because spillovers are a subset of knowledge flows, much of the literature has focused on testing whether citations track flows in general. Only then can we return to the sharper question: do they capture the special case of spillovers?
Is the trail reliable?
The finding that patent citations capture knowledge flows became a cornerstone of the literature, sparking decades of debate, replication, and refinement. Criticism, however, has been twofold: methodological and interpretative.
On the methodology, the sharpest critique came from Peter Thompson and Melanie Fox-Kean, who argued that Jaffe and colleagues’ matching procedure was too coarse. Using finer technology subclass controls, the localization effect largely disappeared. Jaffe and colleagues countered that such fine matching reduces sample size and risks excluding real spillovers that cross technological boundaries. Later work by Yasusada Murata and co-authors, using actual distance measures, again found evidence of localization even under strict controls.
In other work, Peter Thompson sidestepped the matching problem entirely by exploiting newly released USPTO data that distinguish citations added by applicants from those added by examiners. He showed that applicant-added citations are far more localized than examiner-added ones—a result that Juan Alcácer and Michelle Gittelman somewhat tempered, but that was also found by Paola Criscuolo and Bart Verspagen at the European Patent Office.
Overall, these studies strengthened the case that citations can capture knowledge flows, but they also revealed a key weakness of citation data: because examiners (and patent attorneys) add so many references, the overall citation record dilutes the true signal of knowledge flows.
Errors of commission: citations without knowledge flows
The results above suggest the presence of “false positives,” namely citations that do not reflect knowledge flows. The prevalence of these “errors of commission” was quantified by Adam Jaffe, Manuel Trajtenberg, and Michael Fogarty using a survey of inventors. Each inventor was asked about two patents their own patent had cited, plus a placebo patent from the same year and technological class that had not been cited. A parallel survey asked cited inventors whether they had communicated with the citing inventor and whether the later invention appeared to build on their work.
The results painted a mixed picture. On the positive side, inventors were far more familiar with the patents they had cited than with the placebos, and around 40 percent reported learning from the cited invention during their work. On the negative side, roughly one-third of inventors were entirely unaware of the patents they had cited, and perhaps half of all citations did not correspond to any real knowledge flow. Many of these were likely inserted by attorneys or examiners. (The survey was done at a time when the USPTO did not release examiner vs. application citations.) The conclusion: citations can indicate knowledge flows, but many also reflect something else.
Errors of omission: knowledge flows without citations
If some citations occur without knowledge flows, the opposite is also true: many genuine flows leave no trace in the citation record. A striking example comes from Andrew Nelson’s study of the Cohen–Boyer recombinant DNA patents. Stanford University’s licensing office signed 464 agreements for the technology, yet only 135 organizations held patents citing the original rDNA patents, and just 55 overlapped with the licensees. In other words, citations missed 88 percent of the organizations that were actively building on the technology.
Michael Roach and Wes Cohen reached a similar conclusion by comparing citation data with survey reports from R&D managers. They found that citations systematically miss knowledge flows occurring through contract-based channels (consulting, sponsored research, collaborative projects) that rely on tacit, face-to-face communication. Citations also understate the role of public science in basic research, where outputs may not be patentable in the first place. The conclusion: knowledge flows can be captured by citations, but many also leave no citation trail.
Spillovers… or knowledge flows?
Up to this point, we have seen that citations can capture knowledge flows—but the original ambition of the literature went further. From the start, Jaffe, Trajtenberg, and Henderson interpreted local citation patterns not just as flows, but as evidence of knowledge spillovers. Their logic was straightforward: by excluding self-citations, they focused solely on links between independent organizations. The excess local clustering of cross-firm citations was taken as proof that citations do not just record knowledge flows, but specifically reflect spillovers—knowledge benefits that spill beyond the boundaries of the firm.
However, these localized citations may have been driven by market transactions. As a case in point, David Mowery and Arvids Ziedonis show that technology licenses—clear market-mediated transfers—are more geographically localized than patent citations. This suggests that proximity may matter just as much, if not more, for contracting and collaboration than for unpriced spillovers. Thus, localization in citations alone does not prove the existence of spillovers.
Other scholars argued that localized citations might simply reflect shared local environments: inventors clustered in the same region naturally work on similar problems and therefore cite one another. Ashish Arora, Sharon Belenzon, and Honggi Lee formalized this critique using “citation reversals”—cases where a citing patent has an earlier priority date than the cited patent, making sequential knowledge transmission impossible. If citations really captured spillovers, then standard citations (where transmission is possible) should be more geographically concentrated than reversals (where it is not). Yet both types of citations were similarly localized, suggesting that proximity in the citation record may reflect the geography of related inventive activity rather than knowledge spillovers.
So, do localized citations capture spillovers or just flows?
On the surface, these critiques suggest that localization in citations may have little to do with true spillovers. Yet some research points in the opposite direction: mobility, social ties, and personal presence show that genuine spillovers do exist—and that citations, read carefully, can reveal them.
Paul Almeida and Bruce Kogut showed that regions with high intra-regional labor mobility, like Silicon Valley, exhibit stronger localized knowledge flows, suggesting that ideas literally “move” with people. Ajay Agrawal, Iain Cockburn, and John McHale demonstrated that mobile inventors continue to generate citations back to their prior location, evidence that enduring ties facilitate diffusion even after departure. Jasjit Singh showed that inventors with direct collaborative ties are four times more likely to cite each other than unconnected inventors, and Stefano Breschi and Francesco Lissoni confirmed that once co-inventor networks and mobility are controlled for, the residual role of geography shrinks dramatically. Finally, Benjamin Balsmeier, Lee Fleming, and Sonja Lück provided quasi-experimental evidence on the importance of inventors as vectors of spillovers. When an inventor of a co-invented patent died between application and grant, local citations to that patent fell by about 25 percent within 20 miles of the deceased inventor’s location relative to the surviving co-inventors.
Together, these studies show that geography matters not because knowledge is “in the air,” but because it travels through people, relationships, and tacit interactions. That is exactly what we should expect if localized citations capture genuine spillovers rather than just parallel invention and shared local environment.
Concluding remarks
In the end, patent citations are neither a perfect lens on knowledge flows (and spillovers) nor an empty signal. They capture part of the picture, but with both noise and blind spots: many citations do not reflect true knowledge transfer, while many genuine flows leave no trace. What the literature now makes clear is that spillovers exist, they are localized, and they travel through people, networks, and institutions. Citations provide a valuable—if imperfect—paper trail, one that must be read critically and complemented with other evidence.
Please cite this post as follows:
de Rassenfosse, G. (2025). Patent citations: Tracing spillovers or chasing shadows? The Patentist Living Literature Review 8: 1–6. DOI: TBC.