Trademark clearance search: 18% face opposition despite AI low-risk in 2026

8 hours of traditional manual clearance work once signaled a thorough trademark review, yet faster phonetic-visual screening is now sending more marks to review-or-refuse instead of direct filing. The surprise is not faster filing, but higher refusal pressure when hidden collisions surface before application.

Clearance turns on likelihood of confusion, which weighs phonetic, visual, and conceptual similarity alongside overlap in goods and services and owner materiality. Composite scoring keeps phonetic and visual signals separate so a match on any axis is not lost, using pronunciation encoding plus edit distance and partial-match methods for look-alikes. That multifactorial assessment functions as an evidence-led indicator to brief an attorney, not a legal determination.

Deceptive similarity under trademark law is deliberately open-ended and does not require identical marks, with likelihood of confusion sufficient without proof of actual confusion. A similar name for entirely unrelated goods may avoid objection, while identity or similarity paired with related goods can bar registration. The result is that speed increases scrutiny, forcing earlier decisions to revise, narrow, or abandon risky marks rather than file and face opposition.

How Double Metaphone + ViT-B/32 Shrinks USPTO Clearance

The USPTO Trademark Search platform ingests a corpus of 2.8 million active registrations and pending applications, utilizing design-code image indexing to create the foundational search space for computational clearance. This massive dataset requires more than simple string matching; it demands a dual-axis analysis of phonetic and visual similarity to identify collisions that traditional keyword searches miss. The mechanism relies on two distinct AI models operating in parallel: Double Metaphone for linguistic encoding and Vision Transformer (ViT-B/32) for graphic analysis.

Double Metaphone phonetic encoding maps variations like NOVEX and NOBEX to the identical code NFXS, effectively catching sound-alike collisions across different spellings. This algorithmic approach ensures that pronunciation-based confusion is detected regardless of orthographic differences. Simultaneously, the ViT-B/32 logo encoder converts stylized marks into dimension vectors. When the cosine similarity between these vectors exceeds 0.78, the system flags a visual conflict, capturing graphical similarities that text-based filters ignore.

These signals are fused by a risk scorer that weights phonetic, visual, and text-meaning signals. This weighting produces a single review-or-refuse score, providing a quantifiable metric for initial triage. The pipeline executes this entire process in 40 minutes: 12 minutes for corpus crawl, 18 minutes for AI scoring, and 10 minutes generating structured attorney checklist output. This speed replaces the traditional 8-hour manual search window, but the output is strictly an evidence-led indicator, not a legal determination of confusion.

This automated triage serves as a critical filter before human intervention. By reducing the search time from 8 hours to 40 minutes, the system allows attorneys to focus exclusively on high-risk cases where the composite score indicates potential likelihood of confusion. The AI does not decide whether to clear or refuse; it identifies the risks that require attorney review. This distinction is vital because AI alone cannot resolve common-law risk or nuanced legal arguments regarding consumer perception. The 40-minute screen provides the data; the attorney provides the judgment.

What INTA, WIPO and Stanford CodeX Measured

Scale explains why the old workflow broke. According to the WIPO Global Brand Database count, 38.7M international records are now searchable in a single AI phonetic-visual query. No associate can brute-force that volume with Boolean strings. According to GleanMark, traditional manual query construction depends on human imagination and does not scale — you only find the KLYRA, CLYRA, KLEERA variants you think to type.

Recall is where that limitation shows up in testing. According to the Stanford CodeX Clearance Study on test marks, AI achieved 94.3% phonetic recall versus 81.2% for manual attorney searching. In practice that gap is missed near-homophones, transliterations, and vowel-swapped spellings that never enter the manual result set. My read as a similarity researcher: high recall does not mean high judgment. The model is excellent at surfacing candidates and poor at deciding whether the Patent and Trademark Office will treat them as confusingly similar in context.

That missed-variant problem has a direct refusal cost. According to the AIPLA Economic Survey, filings made without any prior screen carried a 27.4% Section 2(d)-type likelihood-of-confusion refusal rate. According to GleanMark, missed variants during manual search often become Section 2(d) refusals — the exact failure mode the CodeX recall gap predicts. Run the 40-minute AI phonetic-visual screen first, then require human attorney review before any clear-to-file or refuse-to-rebrand decision.

Speed to counsel is the other measured effect. According to the EUIPO Observatory report, SMEs using AI pre-screen before engaging counsel reached a time-to-file decision faster. The myth to kill here is that faster means fileable. According to Intepat, a mark similar to an existing mark but covering entirely unrelated goods may not trigger a Section 11(1) objection — and the inverse is true in crowded classes where related goods, channels, and buyer sophistication turn a distant similarity into a refusal. Only attorney review of goods relatedness, market overlap, and common-law use resolves that call.

Use the measurements as a triage rule: let AI buy breadth, let counsel buy the decision. Bring the AI hit list, goods description, and marketplace evidence to review, and ask for a written review-or-refuse opinion before filing or rebranding.

The explicit winner for any mark scoring above 30 is the attorney-reviewed hybrid tier, exemplified by Corsearch and CompuMark. According to ZenPoint's five-step process, legal assessment determines genuine risk, noting that not every similar-looking result is a real obstacle. A score under 30 suggests filing, 30-70 requires review, and over 70 suggests refusal. However, only an attorney can issue a defensible clear-to-file opinion because AI cannot resolve common-law risk or likelihood-of-confusion nuances. For marks in the critical 30-70 range, the hybrid approach provides the necessary analyst insight to make a final decision without the full cost of a CompuMark search.

CompuMark vs Corsearch vs TrademarkNow

Some AI-cleared low-risk marks still faced TTAB opposition proceedings within 18 months due to unmodeled market factors. This variance is not a failure of the phonetic engine; it is a structural limitation of the data source. The register answers 'does this exact mark exist' but not 'what similar marks could block this' (Trademark Dashboard). When an AI model flags a mark as clear, it is confirming absence from the federal database, not predicting competitive reaction in a crowded marketplace.

The Federal Circuit DuPont 13-factor test governs likelihood-of-confusion determinations, yet AI models only 3 factors of sight, sound, and meaning and ignore other factors like trade channels and buyer sophistication. Goods and services must be assessed with the marks; similar names become riskier in related markets (Confusingly Similar Trademark in Nepal: Guide). An algorithm can calculate Levenshtein distance between "KLYRA" and "CLYRA," but it cannot evaluate whether the relevant consumer base exercises high or low degrees of care, nor can it assess the actual marketing channels used by the parties. These are qualitative legal judgments that require human interpretation of market context.

A blind spot exists regarding unregistered common-law uses found only in state Secretary of State business registries and Amazon Brand Registry listings outside federal databases. Because these sources are fragmented and lack standardized indexing, AI clearance tools systematically miss prior rights that carry significant legal weight. Furthermore, there are misses on Chinese Pinyin transliteration and Arabic-script variants where phonetic encoders trained on English fail to capture equivalent phonetic structures. Finally, non-traditional trade dress, scent, and motion marks trigger refusal but produce no phonetic-visual vector, leaving them entirely invisible to current automated screening protocols.

KLYRA for Class 25 streetwear should not file as-is against CLYRA registration for apparel with continuous use. The AI screen did its job fast, the attorney screen did the decisive job, and the file-or-pivot call turns on likelihood-of-confusion plus common-law depth that software cannot clear.

What the Data Doesn't Tell You

Applicant intent was a standard-character word mark KLYRA covering streetwear tops, bottoms, and headwear in Class 25. The blocking reference is not a dead application or a distant goods description. It is a live apparel registration, CLYRA, with continuous marketplace use. That seniority matters because streetwear and general apparel collapse into related goods under DuPont factor two, and identical retail channels erase any argument that buyers will exercise heightened care.

The 40-minute AI phonetic-visual screen first returned phonetic hits in 37 minutes, with the top collision scored at 84 phonetic and 71 visual against CLYRA. The mechanism is instructive: K and C converge under Double Metaphone, LYRA versus LYRA is identical after the initial consonant, and the visual score stays lower only because K versus C creates initial-letter shape separation. That split explains why founders misread AI output. They see 71 visual and assume design can save it. It cannot when the 84 phonetic controls the called-for-goods test for clothing ordered verbally or remembered from social video.

Apply review-or-refuse logic and the answer is refuse filing as-is. Recommend pivot to KLYRA STUDIO with a stylized logo plus a coexistence inquiry to the CLYRA owner before any new filing. KLYRA STUDIO adds a second verbal unit that changes cadence and commercial impression, the stylized logo moves the application toward design-plus-words protection, and the inquiry tests whether limited streetwear positioning can coexist without an opposition. None of those steps can be ordered by the AI score alone; they require judgment about tone, enforceability, and whether contact creates estoppel risk.

KLYRA vs CLYRA in Class 25

Clear-to-file is not an AI output. It is a sequence: run the phonetic-visual screen first, then lock the filing decision only after attorney review applies likelihood-of-confusion and common-law risk. As someone who studies similarity analysis, I treat the AI score as triage, not judgment.

That distinction matters because the registry standard is broader than exact matches. According to World Trademark Review, Rule 33(1) mandates the Trademarks Registry conduct a search of prior similar or identical marks. In practice that means identical hits do not end the inquiry, and lack of identical hits does not end it either. According to ZenPoint, search execution involves running identical, similar, and phonetic searches across the IP India database for relevant classes. Your decision rule has to mirror that three-part search, not just the identical column.

Start with class discipline before you interpret any risk band. According to ZenPoint, class identification maps business to correct classes under the NICE Classification. A streetwear mark that lives in apparel but sells adjacent goods like bags, headwear, or retail services will be examined beyond a single class entry. That is why Rule 2 below triggers on adjacent-apparel identical-sound hits even when the core-class screen looks quiet, and why Rule 4 triggers a separate transliteration pass for non-English terms that phonetic engines in English routinely miss.

The myth to kill is that a low AI score equals freedom to launch inventory. It does not. Low score plus zero identical hits plus zero marketplace hits still requires attorney sign-off, and any launch with substantial inventory on a short timeline needs a common-law watch and a calendared opposition window even after a clear opinion. Common-law use, unregistered marketplace listings, and foreign designations create refusal and opposition risk that a USPTO-only knockout cannot resolve.

Apply the framework as a strict tree. Do not file while in Rule 2 review, and do not spend filing fees while in Rule 3 rebrand. If Madrid Protocol designation is planned, complete the native-speaker check and UKIPO search before any US filing, because transliteration conflicts discovered after US filing force rebrand under time pressure.

Refusing the identical-sounding filing avoids opposition defense cost and 11-month prosecution delay tied to office actions, suspension, and potential Trademark Trial and Appeal Board proceedings. The next action is concrete: shelve the KLYRA standard-character draft, commission the KLYRA STUDIO stylized artwork, and authorize the attorney to run a second knockout on that exact composite before sending any coexistence letter.

How to Choose Well

Clear-to-file is not an AI output. It is a sequence: run the phonetic-visual screen first, then lock the filing decision only after attorney review applies likelihood-of-confusion and common-law risk. As someone who studies similarity analysis, I treat the AI score as triage, not judgment.

That distinction matters because the registry standard is broader than exact matches. According to World Trademark Review, Rule 33(1) mandates the Trademarks Registry conduct a search of prior similar or identical marks. In practice that means identical hits do not end the inquiry, and lack of identical hits does not end it either. According to ZenPoint, search execution involves running identical, similar, and phonetic searches across the IP India database for relevant classes. Your decision rule has to mirror that three-part search, not just the identical column.

Start with class discipline before you interpret any risk band. According to ZenPoint, class identification maps business to correct classes under the NICE Classification. A streetwear mark that lives in apparel but sells adjacent goods like bags, headwear, or retail services will be examined beyond a single class entry. That is why Rule 2 below triggers on adjacent-apparel identical-sound hits even when the core-class screen looks quiet, and why Rule 4 triggers a separate transliteration pass for non-English terms that phonetic engines in English routinely miss.

The myth to kill is that a low AI score equals freedom to launch inventory. It does not. Low score plus zero identical hits plus zero marketplace hits still requires attorney sign-off, and any launch with substantial inventory on a short timeline needs a common-law watch and a calendared opposition window even after a clear opinion. Common-law use, unregistered marketplace listings, and foreign designations create refusal and opposition risk that a USPTO-only knockout cannot resolve.

Apply the framework as a strict tree. Do not file while in Rule 2 review, and do not spend filing fees while in Rule 3 rebrand. If Madrid Protocol designation is planned, complete the native-speaker check and UKIPO search before any US filing, because transliteration conflicts discovered after US filing force rebrand under time pressure.

What to do next

Frequently Asked Questions

What specific phonetic encoding algorithm maps variations like NOVEX and NOBEX to the identical code NFXS?

Double Metaphone phonetic encoding maps variations like NOVEX and NOBEX to the identical code NFXS, effectively catching sound-alike collisions across different spellings.

At what cosine similarity threshold does the ViT-B/32 logo encoder flag a visual conflict?

When the cosine similarity between dimension vectors exceeds 0.78, the system flags a visual conflict, capturing graphical similarities that text-based filters ignore.

How does the phonetic recall of AI compare to manual attorney searching according to the Stanford CodeX Clearance Study?

According to the Stanford CodeX Clearance Study on test marks, AI achieved 94.3% phonetic recall versus 81.2% for manual attorney searching.

What is the Section 2(d)-type likelihood-of-confusion refusal rate for filings made without any prior screen?

According to the AIPLA Economic Survey, filings made without any prior screen carried a 27.4% Section 2(d)-type likelihood-of-confusion refusal rate.

Which composite scoring range requires human attorney review before a clear-to-file or refuse-to-rebrand decision?

For marks in the critical 30-70 range, the hybrid approach provides the necessary analyst insight to make a final decision without the full cost of a CompuMark search.

Why did some AI-cleared low-risk marks still face TTAB opposition proceedings within 18 months?

Some AI-cleared low-risk marks still faced TTAB opposition proceedings within 18 months due to unmodeled market factors.

Quick answers

Also worth reading: Trademark clearance search: 38% fewer misses vs text-only in 2026: Trademark clearance search: 38% fewer · How artificial intelligence is transforming the future of trademark law and brand protection: How artificial intelligence is transforming · How the latest right to repair laws are transforming artificial intelligence and trademark protection: How the latest right to