Methodology white-paper

Designed by a licensed patent attorney with 15+ years of EU and US trademark practice. This page describes how a Demarka clearance is run end to end — what we look for, how we score it, where we use AI and where we deliberately do not. We publish it so brand owners, in-house counsel and external attorneys can see exactly what they are getting, and where its limits are.

Two design choices shape everything below:

Our search is predictable and explainable. Given the same query and the same data, it returns the same results — every time. The scoring is transparent on every report, and every match shows you exactly why it was returned. It is not a black box.
AI is used only in two narrow, well-defined places — both fully bounded, both producing structured output that can be checked, retried and audited.

1. Why a dedicated trademark engine

Generic text search is built for documents. Trademark examiners and opposing parties do not read your mark as a string of letters — they read it as a sound, a shape, a meaning, and a position relative to other marks in the market. The conflicts that block registrations rarely look anything like the applied-for mark on paper:

Cyrillic СОВА cited against Latin COBA — visually identical, yet every character is different.
7UP ↔ Seven Up — examiners treat them as the same mark despite zero character overlap.
Gugl ↔ Гугл ↔ グーグル all read as GOOGLE across three writing systems.

Generic search returns near-zero recall against any of these. Demarka was built to catch them by default.

2. Why we are better than the search tools registries offer

The free tools published by national registries — USPTO TESS, EUIPO eSearch plus, WIPO Global Brand Database, TMView — are excellent for confirming what is on the register. They were never built for clearance. They offer one engine: a generic fuzzy text match with an optional phonetic toggle that was state of the art in the 1990s. Two consequences follow.

High false-positive rate. Generic fuzzy search rewards any shared substring. Examiners do not. A clearance run on a generic engine forces you (or your paralegal) to manually filter hundreds of irrelevant returns before the real conflicts surface.
Low recall on the patterns that actually block filings. Cross-script lookalikes, leetspeak, condensed phonetic spellings, number-word interchange and word-order rearrangements are systematically missed.

We invert that trade-off. Multiple narrow, well-targeted detection layers run in parallel against every record we index. A candidate that hits one strong layer or several weaker ones surfaces; pure surface coincidences are filtered out. The result is higher recall on the patterns examiners actually cite, with a fraction of the noise.

3. Coverage

100M+ trademark records across the registries we index.
90+ jurisdictions in scope for similarity and absolute-grounds checks.
One unified search across every registry we index. No need to pick a database. One query, one ranked result list.
Every official language of each jurisdiction read independently when we evaluate absolute grounds.
Continuous synchronization with source registries — every active record is re-evaluated as new filings arrive.

4. What our search catches — parallel detection layers

Each layer below targets a specific way trademark conflicts surface in practice. Every clearance runs every layer at once, against every record we index. We describe each layer by what it catches and why it matters.

Identity & spelling

Catches marks identical to yours, or differing only in trivial spelling variations — accents (CAFE ↔ CAFÉ ↔ MUNCHEN ↔ MÜNCHEN), punctuation and symbols (S&P ↔ S P ↔ S.P. ↔ SP), plurals and verb forms (CAT ↔ CATS, STREAM ↔ STREAMING ↔ STREAMER), and word-order changes in multi-word marks (Sky Blue ↔ Blue Sky). Examiners treat these as one and the same mark.

Phonetic

Catches marks that sound the same as yours when spoken aloud — even when they are spelled completely differently. This includes condensed phonetic spellings (GOOGLE ↔ Gugl ↔ Гугл) and substitute spellings (EXPRESS ↔ Xpress ↔ eXpres). Phonetic similarity is one of the most common grounds for refusal and opposition worldwide.

Cross-script

Catches the same name written in a different alphabet — Latin ↔ Cyrillic ↔ Greek ↔ Hangul ↔ Kanji ↔ Arabic. A trademark protects a name, not a script; examiners and opponents follow it across writing systems.

Visual look-alike

Catches marks that look identical or confusingly similar to yours at a glance — characters from different alphabets that share a shape (Latin COBA ↔ Cyrillic СОВА), Unicode lookalikes, and letter combinations that confuse the eye (rn ↔ m, cl ↔ d, vv ↔ w). Examiners read marks as visual objects, not just text.

Leetspeak & informal substitution

Catches modern brand-naming substitutions — numbers and symbols standing in for letters or whole words (2Fast4U ↔ Too Fast For You; u ↔ you; 4 ↔ for; 2 ↔ to). The underlying mark is the same; consumers and examiners see through the styling.

Structural & partial

Catches marks that share a core element with yours — partial overlaps, compound and domain-style variants (GoFinance, GetWallet, FinanceGo), and rearranged components. Examiners ask whether the dominant element of one mark appears in the other; they are not constrained by exact alignment.

Fuzzy / typo

Catches near-spellings, transposed letters and small slips that examiners read as conflicts (STARBUCKS ↔ Stabrucks ↔ Starbukcs). Small typing distances do not separate one mark from another in examination practice.

Number-word equivalence

Catches marks that swap digits for words and back (7UP ↔ Seven Up ↔ 7-Up), as well as Roman numerals (II ↔ 2). Examiners treat the number and its spelled-out form as the same element.

Semantic near-duplicate

A final AI-assisted layer that catches paraphrases and reformulations the other layers might miss. We treat it as recall insurance — additional candidates that the rest of the engine then scores and ranks alongside everything else.

5. How results are presented

Every candidate that the detection layers surface is ranked into a single relevancy figure that appears on your report.

We deliberately tune for high recall. It costs less for a clearance to return one borderline hit that is ruled out in 30 seconds than to miss a citation that ends a project six months in.

Every match in your report is transparent: you and your counsel can see why it was returned, which examination ground it touches, and how it compares on visual, phonetic and conceptual grounds. There is no opaque score we cannot explain — but the result is not noise either: results that do not reflect a real conflict are kept out, so the report stays useful.

What the search step is not: it is not an AI guessing. It does not invent hits. It does not re-weight itself depending on which office you are filing in — office-specific differences are handled in the absolute-grounds analysis (Section 7), not by changing how the search itself works.

6. Where AI is used — and where it isn’t

We separate the two cleanly.

Step	How it works
Searching for similar marks and scoring them	Predictable parallel detection layers; reproducible result list; transparent on every report. No generative AI.
Filtering obvious non-matches	A language model is used as a recall-biased filter on large candidate sets. When in doubt, it keeps the candidate. It never adds candidates the search did not find.
Absolute-grounds analysis	A language model reads the mark, the goods/services and the curated per-jurisdiction statutory reference, then returns a structured assessment for each of the eight absolute-ground categories.
Aggregating the final registration outlook	Predictable. Combines per-class scores and overall risk into the final registration-probability envelope and outlook tier (Clear / Fixable / At Risk / Critical).

Both AI-touching steps are bounded: they return their findings in a predictable, validated structure — never free-form opinion. They cannot fabricate citations — every similar-mark hit in your report comes from the search engine, never from a language model.

7. Absolute grounds — eight categories, every official language

Beyond similar marks, every jurisdiction has its own list of refusal grounds. Most search tools skip this layer entirely; the few that touch it run a static English-only keyword check. Demarka evaluates eight absolute-ground categories for every selected jurisdiction:

Inherent distinctiveness — is the mark descriptive, generic, or non-distinctive for the claimed goods and services?
Deceptive or misleading character — does it suggest qualities, origin, or composition the product does not have?
Public order and morality — vulgarity, hate references, content contrary to local public policy.
Prohibited or state-protected elements — flags, emblems, official seals and designations protected under Article 6ter and analogous national rules.
Geographic indications and protected designations — PDO, PGI and other protected names.
Linguistic and cultural sensitivities — local slang, taboo terms, religious connotations, historically charged terms.
Personality and name-image rights — collisions with known persons, surnames, public figures.
Well-known and reputed marks — collisions with marks whose reputation extends beyond their registered classes.

Each category is evaluated:

In every official language of the selected jurisdiction, plus English as a cross-check. A mark that is arbitrary in one official language and descriptive in another is flagged accordingly.
Against the actual statutory grounds of that office, anchored to a curated per-jurisdiction reference (primary statute, official languages, examination practice, locally protected designations, religious and culturally sensitive terms, special considerations). The reference is maintained for every jurisdiction we cover.
With cultural and linguistic context — slang, historical usage, regional sensitivities.

The result is a structured risk rating per ground (LOW / MEDIUM / HIGH or N/A) with a plain-language reason — not a single global “absolute-grounds score” you have to take on faith.

8. The assessment rubric — what every report contains

Every Demarka assessment is produced against a fixed, structured rubric. The rubric is part of the public report, so every score is explainable.

Mark analysis — type of mark (word, figurative, composite, …) and distinctiveness on a five-tier scale: fanciful · arbitrary · suggestive · descriptive · generic.
Absolute grounds — risk per category (the eight above) plus an overall rating.
Relative grounds — for each cited mark returned by the search, the analysis scores it on three axes — visual · phonetic · conceptual — and rates the goods/services overlap (identical · related · unrelated) and the combined conflict risk.
Composite-mark and dominant-element analysis — for multi-word and figurative marks, the assessment identifies which element is doing the legal work and weights the comparison accordingly. This matters because the cited mark and the applied-for mark are not always conflicting on their dominant component.
Per-class assessment — for each Nice class you applied in: initial registration probability, post-response probability, difficulty tier (easy → very hard), the specific hurdle, and a plain-language path to overcome it.
Overall assessment — combined risk, registration-probability percentage (with a best/worst-class range), and a short summary.
Registration outlook — a tier on a fixed ladder: Clear · Fixable · At Risk · Critical, with an actionable checklist and indicative timeline.

Every field of the rubric uses a fixed category list. The AI cannot return free-form text in place of a registration outlook, a risk level or a distinctiveness tier.

9. Per-jurisdiction context

Office-specific differences are applied at the assessment layer, not the search layer. For each jurisdiction we maintain a curated reference covering:

The primary trademark statute and the office that examines it.
The official languages the office reads.
Whether the office examines relative grounds ex officio or only on opposition — this materially changes what “risk” means for the same citation in different offices.
Locally protected designations and regulated terminology.
Religious and culturally sensitive terms.
Special considerations specific to that office’s examination practice.

When the absolute-grounds analysis runs, this jurisdiction reference is part of the input. The output reflects how that office is likely to treat the mark — including bilingual offices that read the application independently in each official language (e.g. CH in DE / FR / IT / RM, CA in FR + EN, BE in NL / FR / DE).

10. Human in the loop

Software handles what software is good at: high-recall pattern matching, statutory enumeration, structured per-jurisdiction scoring at a scale no human team could match. People handle the judgment calls.

Paid plans include human review and consultation. The exact scope — written QA, a video consultation with a qualified trademark attorney, attorney-led action on specific items — depends on the plan you are on. The current scope of human review per plan is listed on our pricing page.

Whether you review a report yourself, hand it to your own counsel, or use the included human review, every match is auditable: which detection layers fired, which examination ground was triggered, which jurisdiction-specific rule applied. The reasoning is fully transparent.

11. What we deliberately do not claim

We are not an oracle. Trademark examination involves discretion. We surface what a reasonable examiner is likely to cite; the final call is legal, not statistical.
No engine has 100% recall. New filings appear daily. Common-law marks, unregistered usage and pending applications below our data-source horizon can still surface in opposition.
The AI components do not invent citations. Every cited prior right in a report comes from the deterministic search. AI is used for recall-biased prescreening and for absolute-grounds reasoning — never to manufacture a similar-mark hit.
We do not replace counsel. We make counsel an order of magnitude faster — we do not remove the need for them.

12. Continuous improvement

Every flagged match that turns out to be a false positive — or a missed citation surfaced by human review — feeds back into evaluation. The patterns above describe what we cover today; new ones are added when examination practice or naming conventions shift. Recent additions reflect AI-generated brand names, emoji and Unicode-symbol marks, and new generative-script transliterations. The jurisdiction reference is refreshed whenever a covered office updates its examination practice.

← Back to clearance