Patent Citations Rank Search Leads, Not Inventions

Network centrality.

Field-normalised impact scores.

Assignee graphs where the little circles communicate their importance by glowing teal.

Somebody will split your population into technology classes, adjust for age, subtract the self-citations, and produce a decimal that looks very disappointed in a garage-door sensor.

Then there is the citation list.

It has document numbers, dates, and a modest ambition.

It can help you find another document.

It’s a filing clerk with a string board.

It doesn’t know whether the focal invention was brilliant, patentable, commercially successful, novel, or the last thing standing between humanity and a municipal pigeon problem.

It knows that a record points at another record.

And sometimes who put the point there.

That difference is not cosmetic.

The historical NBER file contains roughly 3 million US patents and more than 16 million citations.

Enough material to make a bar chart look like a personality test.

Scale gives citation analysis its appeal.

It also preserves the machinery that made the citations. Applications, examiners, search reports, prosecution records, changing rates of patenting, time.

The citation does not arrive by immaculate conception.

It arrives with paperwork.

A score, and several routes

Begin with your source.

In the US grant sample from 2001 to 2003, Alcácer, Gittelman and Sampat reported that examiners added 63% of citations on the average patent.

And that 40% of granted patents had all citations added by examiners.

Those are descriptive results for that sample.

Not universal rates.

They’re still a vigorous objection to a dashboard that treats every reference as an applicant-authored confession of technical ancestry.

An applicant-submitted reference and an examiner-listed reference can describe the same document while doing different work for you.

A compliant IDS records applicant submission and examiner consideration.

A PTO-892 records an examiner listing.

The examiner-listing asterisk in MPEP 1302.12 tells you the route.

It is not a tiny gold star for the reference that won the science fair.

Put the source in its own field.

Don’t let a column called citations eat the evidence of how those citations got there.

The pending file hands you the next correction.

The useful question isn’t whether a future front page has a long bibliography.

It’s whether the document was submitted, considered, raised in an Office action, used alone, combined with another reference, distinguished after amendment, or simply present in the room.

Under MPEP 707, a rejection notice identifies the rejected claims and explains a reference’s pertinence where it isn’t apparent.

An obviousness rejection explains the modification or combination of references.

So your bare list of 15 documents becomes claim by claim.

Cited.

Relied upon.

Distinguished.

Never used.

That’s where your citation list stops impersonating a verdict and becomes an index.

It takes you to the dated IDS, the PTO-892, the Office action, the response.

Then it stops.

It doesn’t do their jobs.

A reference has no obligation to have been dispositive merely because it appears on a list.

Which is fortunate, because lists would otherwise become terribly ambitious documents.

Direction is not decoration

Your next small embarrassment is direction.

A backward citation goes from your focal application or patent to an earlier document.

A forward citation comes from a later document to the focal record.

The first can disclose antecedents already in an earlier record. The second can surface later terminology, improvements, assignees, inventors.

Writing “citations” without direction is like writing “north” on a map after removing the coast.

Time needs the same care from you.

A PCT application is generally published around 18 months from priority. US ordinary publication is ordinarily after 18 months from the earliest benefit date.

So a publication date isn’t a decorative timestamp beside an application identifier.

It affects when the record can enter an observed citation population.

A later publication has had fewer opportunities to acquire visible forward citations by a fixed extract date.

The literature calls that incomplete observation truncation.

An older patent with 10 observed forward citations cannot be mechanically placed above a two-year-old patent with 4.

Not without a shared window, a cohort adjustment, or a model of expected accumulation.

A plain rule beats the haunted decimal you were offered.

Choose a fixed 5-year post-grant forward window.

State the cutoff.

Keep the focal publication date.

Your result is not timeless influence.

It’s an observed count under a declared clock.

For a defined cohort, write the count so it admits what it is.

received_citations = count(later records citing focal record)
                    where provenance = applicant or examiner
                    and publication date falls inside the stated window
                    and the stated cutoff has passed

That’s the whole machine.

It doesn’t need to explain invention quality.

It needs to say whose later records were counted, in which direction, over what interval, and whether applicant and examiner entries stayed separate.

The rest of the dashboard can wait outside with its gradients.

Search is the job

Your useful workflow starts with a seed document found through text or classification searching.

Follow backward links for earlier records.

Follow forward links for later records.

Keep the ordinary things that let somebody else repeat your excursion.

Focal application identifier.

Publication date.

Direction.

Source database.

A reason for retention.

Your core search log has four fields.

Focal document, direction, source database, reason for retention.

The identifier and the publication date stop the log becoming a set of confident adjectives.

None of this replaces a search strategy.

MPEP 904 calls for a field of search, appropriate tools, and a strategy.

Citation chasing can miss uncited art, documents in other languages, and records omitted from an accessible database.

The citation list is a corridor, not the building.

A sensible searcher opens the doors it offers, and also looks for the ones it forgot to mention.

The same restraint makes your ranked portfolio list useful.

A review of 200 patents can order candidates with separate columns.

Forward citations in a stated window.

Examiner-origin share.

References used in Office-action rejections.

Age of relevant cited art.

That queue allocates your reading time.

It doesn’t decide validity, enforceability, acquisition value, ownership, maintenance status, claim scope, or product fit.

Those questions have their own documents.

Inconsiderate, but legally tidy.

For a map or a study, preserve your reporting elements.

Unit of analysis.

Jurisdiction.

Grant or application cohort.

Source-attribution rule.

Citation window.

Self-citation treatment.

Field or time adjustment.

Together they tell a reader what your metric measures.

Otherwise “citation impact” wanders the streets unsupervised.

When your focal record is an application, preserve its procedural address as well as its bibliographic face.

Your filing verification checklist has five fields.

Application number.

Filing date.

Applicant and inventor data.

Document list.

Fee status.

The later publication can be found by application number, publication number, or applicant name.

Your working list should keep the identifier you actually used to retrieve it.

A citation table that loses the application identifier is a treasure map that has removed the island in the interests of clean design.

Association has a paperwork problem

The literature does find associations.

Hegde and Sampat linked patents issued in 1992, 1996 and 2000 to renewal outcomes, and found examiner citations to a patent stronger predictors of renewal than applicant citations.

Renewal is a cost-bearing decision, and a private-value proxy.

Nothing more.

It is not market value, social welfare, technical merit, or a coronation conducted by the Patent Office.

The qualification matters, because source can be a feature of the production process rather than a magic ingredient.

Examiner searches may find legally important art more consistently. The source field may also stand in for characteristics of applications and fields that predict renewal.

The observed reference comes first.

The metric you chose comes next.

An association may come after that.

Causation is still somewhere else, reviewing its correspondence.

Backward citations have different roles again.

A 2022 study separating blocking backward citations, meaning references identified as invalidating claims for novelty or obviousness, from non-blocking applicant and examiner citations, finds different associations with grant.

That doesn’t let you translate “more citations” into “less grantable.”

Count and source don’t tell you whether a reference constrained the claim.

The claim language, the amendments, the statutory requirements and the examiner’s reasoning are all still employed.

So the mundane object to build is a list.

Not an invention league table.

focal application | publication date | direction | cited document
provenance         | source database  | window    | cutoff
retention reason   | prosecution-record locator | limitation

A label on the box

Use the list to decide what to open next.

Preserve your direction, your application identifier, your publication date, your source, your provenance, your window and your reason.

Then read the record that gives your connection its procedural role.

Call that modest column the citation-provenance field.

It is not an oracle.

It’s an honest label on the box.

Which is about as much personality as a citation deserves.