Ranking Trademark Records by Similarity to a Query Mark

A technical overview of the Logo Similarity Engine

Logo Similarity Engine, Inc.
Irvine, California · info@logosimilarityengine.com

Abstract

We compare a logo or mark against tens of millions of trademark records and rank the closest matches. A query may be an image, the text of a mark, or a serial number. Each record is represented on visual, textual, and phonetic axes, and the record metadata (owner, filing date, serial, Nice class, goods-and-services text, status, prosecution history) is treated as a first-class similarity dimension rather than a filter over image results. The system has been in continuous production since 2019, and the corpus is refreshed weekly. The closest public reference point is USPTO Image Search powered by Clarivate^[1]; this engine is the trademark analog on the same axis.

0Motivation

This project started with a failure.

Seven years ago I was starting a business. I paid for a logo, then business cards, printing, and an initial inventory of products with the logo on them. Then a cease-and-desist letter arrived. My logo was almost identical to a trademarked logo. Everything went into the trash, and I abandoned the business.

I was sure there was a faster, easier and better way to see a conflict before falling in love with a company name and a logo. So I started building one. The engine described below is that work, and the goal has not changed: help other people find the conflict before they spend money on it, not after.

Table 1. Incident log. Each step added cost before the conflict was known.
Step	Event	Outcome
	Logo designed and paid for	Sunk cost
	Business cards, printing, initial product inventory with the logo	Sunk cost
	Cease-and-desist received: logo almost identical to a trademarked logo	Conflict found too late
	All branded material discarded	Business abandoned
	Work begins on a pre-launch similarity check	This engine

1Scope

We operate a trademark and logo similarity engine. It ingests both the visual mark and the full trademark record: owner, filing date, serial, Nice class^[2], goods-and-services text, status, and prosecution history^[3]. Each dimension is embedded with a stack of vision and language models. For a query mark, the engine returns a ranked list of records from the corpus by similarity on visual, textual, and phonetic axes, with class-overlap and status filters applied at query time.

Queries may be made by uploaded image, by mark text, or by serial number. The system has been in continuous production for seven years and is refreshed weekly. A legal-IP reader will recognize the closest reference point as USPTO Image Search powered by Clarivate^[1]. This engine sits on the same axis, with one difference in design: record metadata is a similarity dimension in its own right, not a filter applied after image retrieval.

2Definitions

Definition 1 (Mark). A mark is a registered or filed trademark taken together with its record: the visual mark, its text, owner, filing date, serial number, Nice classes , goods-and-services text, status, and prosecution history.

Definition 2 (Corpus). The corpus is the set of all marks held by the engine. Its size is in the tens of millions, and it is rebuilt weekly from source records.

Definition 3 (Query). A query is an uploaded image, the text of a mark, or a serial number that resolves to a mark in .

Definition 4 (Similarity). For each axis , an embedding maps a mark to a vector, and the axis similarity compares the two vectors. The engine combines the axes into one score , reported as a percentage.

3Method

On each axis, similarity is measured by the angle between embedding vectors (cosine similarity^[4]):

(1)

The axis scores are combined into a single score . The combination function and model stack are shared under NDA. Given a set of Nice classes and permitted statuses chosen at query time, the engine returns the top marks

(2)

The verdict for the query is read from its highest score :

(3)

Every mark in is already registered or filed. The labels in (3) describe the query, not the marks it is compared with.

Figure 1 · Interactive

How similar is your logo to existing trademarked logos?

Every registered USPTO mark sits somewhere on this map, and logos that look alike sit close together. When you upload a logo, the engine drops it onto the map and measures how close its nearest neighbors are.

Try a logo

Each logo is turned into a fingerprint, a long list of numbers from an image model, and distance on the map stands for how different two fingerprints are. The rings are drawn where similarity crosses 75% and 95%. The sample logos and marks here are illustrations, not real registrations.

Universe of Trademarked USPTO Logos and Wordmarks

What the verdict means for your logo

OriginalBelow 75%. No registered mark falls inside the outer ring.

Not Original75% to under 95%. At least one mark sits between the two rings.

Trademarked95% or higher. A registered mark sits inside the inner ring.

Figure 1. A two-dimensional projection of the corpus with sample logos and marks. Each dot is a registered mark; marks that look alike sit close together. The query logo is dropped onto the map, and the rings are drawn where its similarity crosses 0.75 and 0.95; on this projection, distance from the query grows as . The verdict follows (3). Names and serial numbers are illustrative.

4Limitations

Similarity is not likelihood of confusion. The legal test weighs factors a similarity score does not see, including the relatedness of goods and services and conditions of sale^[5]. A score is an input to a legal review, not a substitute for one, and it is not legal advice.
Freshness. The corpus is rebuilt weekly. A mark filed after the most recent refresh will not appear until the next one.
Thresholds are operating points. The 0.75 and 0.95 boundaries in (3) are chosen for triage. They are not legal boundaries, and a score just under a threshold deserves the same attention as one just over it.
Figure 1 is a projection. Embeddings live in many dimensions. The two-dimensional map is an illustration with sample logos and marks, and distances on it only approximate distances in the full space.
No public benchmark. We do not publish accuracy figures. Product and corpus mechanics are shared over NDA.

5Team and posture

Three principals. A seven-year build. A clean corporate structure. The intellectual property is framed as trade secret and compilation copyright; there are no patents, by design.

We do not run a self-serve trial, a public benchmark, or a marketing funnel. Product and corpus mechanics are shared over NDA.

6Company facts

Table 2. Reference facts.
Entity	Logo Similarity Engine, Inc.
Jurisdiction	California, United States
Headquarters	Irvine, California
In production since	2019
Corpus	Tens of millions of trademark records, refreshed weekly
Contact	info@logosimilarityengine.com

AGlossary

Serial number: The eight-digit number the USPTO assigns to a trademark application when it is filed.
Nice class: One of 45 classes of goods and services in the international Nice Classification^[2].
Goods-and-services text: The wording in a record that says what the mark is used for.
Status: Where a record stands, for example live, pending, registered, abandoned, or cancelled.
Prosecution history: The record of events between filing and registration: office actions, responses, and decisions.
Embedding: A list of numbers that a model produces from an image or text, arranged so that similar inputs give nearby lists.
Cosine similarity: The cosine of the angle between two embeddings, as in (1). It is 1 when they point the same way.

References

USPTO Image Search powered by Clarivate. United States Patent and Trademark Office. uspto.gov/trademarks/search.
World Intellectual Property Organization. Nice Classification. wipo.int/classifications/nice.
United States Patent and Trademark Office. Trademark Status and Document Retrieval (TSDR). tsdr.uspto.gov.
G. Salton, A. Wong, C. S. Yang. A vector space model for automatic indexing. Communications of the ACM 18(11):613–620, 1975.
United States Patent and Trademark Office. Trademark Manual of Examining Procedure, §1207.01, Likelihood of Confusion. tmep.uspto.gov.

How to cite

@misc{lse2026overview,
  title        = {Ranking Trademark Records by Similarity to a Query Mark:
                  A Technical Overview of the Logo Similarity Engine},
  author       = {{Logo Similarity Engine, Inc.}},
  year         = {2026},
  address      = {Irvine, California},
  howpublished = {\url{https://logosimilarityengine.com}}
}