Why Prior Art Searches Fail Before the First Database Query Is Run

The Mistake Happens Before the Search Begins

Ask most people what goes wrong in a patent prior art search, and they will say: incomplete database coverage, keyword selection, not searching enough sources. These are real problems. But the most consequential mistake in patent searching happens before the first database is opened — in how the claims are read and interpreted.

Claim interpretation determines search scope. Search scope determines which prior art is found. And wrong claim interpretation — reading claims too literally, too broadly, or without understanding the underlying technology — produces searches that are thorough by any conventional metric while completely missing the prior art that matters.

The Literal Reading Problem

Consider a patent claim: "A method for compressing sensor data in an IoT device comprising: sampling sensor measurements at a variable rate based on a rate of change of sensor values; applying predictive encoding to generate a residual signal; and transmitting only the residual signal when said residual signal falls below a threshold."

A non-engineer reads this and searches for: "IoT sensor data compression," "variable rate sampling," "predictive encoding IoT," "residual signal IoT." They search USPTO, EPO, and maybe Google Patents. They find some related IoT patents from 2015-2020 and report that the prior art landscape is relatively clear.

An engineer reads this and thinks: variable rate sampling based on signal derivative is essentially just signal activity detection — if the signal is changing rapidly, sample faster; if it is stable, sample slower. This is adaptive Nyquist sampling. Predictive encoding generating a residual is differential PCM (DPCM) — a standard technique from digital audio compression in the 1970s. The threshold-based transmission of residuals is essentially silence suppression from VoIP compression. The claim, stripped to its engineering fundamentals, is a combination of DPCM, adaptive sampling, and VAD (Voice Activity Detection) from audio signal processing — all developed decades before IoT existed.

The Terminology Translation Problem

Every technical domain has its own vocabulary for the same underlying concepts. Phrases that mean the same thing in different domains:

• "Adaptive threshold" in image processing = "soft decision boundary" in communications = "hysteresis control" in power electronics = "dead band" in process control — all describe the same fundamental concept of a threshold that adjusts to avoid oscillation.
• "Neural network classifier" in AI = "multi-class discriminant function" in statistics = "pattern recognition system" in older signal processing literature — the concept predates the terminology by decades.
• "Energy harvesting" in IoT = "parasitic power capture" in RFID = "thermoelectric generation" in industrial sensors — the underlying principle (converting environmental energy to electrical power for a low-power system) is identical across these domains.

If a prior art search uses only the terminology in the patent claim, it will miss every piece of prior art that describes the same concept using different words from a different technical domain. This is not a minor problem — it is the primary reason valuable prior art goes unfound.

The Novelty vs Non-Obviousness Conflation

Many patent search firms find themselves running patentability searches that are actually only § 102 novelty searches — looking for single references that describe all claim elements. § 103 non-obviousness is treated as an afterthought. This is backwards for the majority of modern patent claims, where perfect single-reference anticipation is rare while § 103 obviousness combinations are the primary invalidity strategy.

Effective prior art search for patentability purposes must simultaneously identify: (1) § 102 references that individually anticipate the claim; and (2) pairs or small combinations of § 103 references that render the claim obvious under the KSR flexible approach. The second category requires understanding which technical domains provide which elements, and whether a POSITA would have been motivated to combine them — a judgment that requires engineering domain knowledge.

The Database-First Problem

Most IP search firms are organised around database access — which databases they subscribe to, how to use them efficiently, which keyword strategies generate the most relevant results. This database-centric approach creates a structural bias: you search for what the databases can find, rather than identifying what needs to be found and then determining which sources contain it.

Non-patent literature (NPL) is the critical example. For any technology domain with an active academic research community — semiconductor devices, wireless communications, machine learning, biotechnology — the most technically detailed descriptions of innovations appear in journal papers and conference proceedings before they appear in patents. The IEEE Transactions journals, ACM proceedings, Nature/Science/Cell publications, and conference records (ISSCC, IEDM, NeurIPS, ICML, ICLR) contain prior art that is never found by patent-database-only searches. At Bullseye, NPL search in IEEE Xplore, ACM Digital Library, arXiv, and technical standards documents is standard practice, not an add-on.

The Engineer's Approach: Function First, Search Second

The correct methodology for any prior art search — patentability, invalidity, or FTO — begins with claim decomposition at the functional level:

• What technical problem does this claim solve?
• What are the underlying physical principles at work?
• What other technical domains solve the same problem using the same principles?
• What terminology does each of those domains use?
• What sources — patent and non-patent — contain the primary literature for each domain?

Only after this analysis is complete can you construct a search strategy that will actually find the relevant prior art. The search itself — the database queries, the keyword combinations, the citation chaining — is the final step, not the first.

"We never open a database until we have spent significant time understanding the claim at an engineering level — decomposing it into its functional elements, mapping those elements to all the technical domains where they have been implemented, and identifying the prior art universe that must be searched. Most search firms do this backwards."