USPTO vs AI Scoring: DuPont Factor 1 in 2026 Class 9

Class 9 software marks faced a Section 2(d) refusal rate during FY2024 first actions, yet filings pre-screened with AI embeddings cleared at a lower rejection threshold. This divergence reveals a structural blind spot: the USPTO’s legacy search interface prioritizes exact spelling matches while examiners apply DuPont Factor 1, which weighs appearance, sound, meaning, and commercial proximity.

Traditional keyword queries miss phonetic overlaps and visual similarities that trigger likelihood-of-confusion rejections. Neural-network models now parse those multidimensional signals simultaneously, surfacing conflicts that Vienna Classification codes routinely overlook. The result is a more predictable examination landscape where refusal patterns align with measurable similarity metrics rather than subjective examiner discretion.

Cost structures reflect this technological shift. Enterprise platforms operate between $10,000 and $100,000 annually, while mid-market alternatives sit comfortably between $50 and $500 per year. Firms leveraging these tools report 40 to 60 percent faster clearance cycles without sacrificing coverage across the 14 million USPTO records or global registries.

Inside DuPont Factor 1

Factor 1 similarity is where Class 9 applications live or die, and public search tools do not see what examiners see. Under Lanham Act Section 2(d), registration is barred where an applicant mark creates likelihood of confusion with a prior live mark. That determination runs through the 13-factor test from In re DuPont, but in practice Factors 1 and 2 control: similarity of the marks and relatedness of the goods. If those two align against you, the other eleven rarely save the filing.

Examiners apply that test under TMEP with an internal advantage applicants lack. The examining attorney runs X-Search across direct hits plus phonetic equivalents plus truncated stems, then maps that result across IC 009 goods that the USPTO treats as inherently related — downloadable software, electric batteries, smart eyewear, sensors, and adjacent hardware. That is why NOVA for battery management software can draw a 2(d) citation to NOVAQ for smart glasses even when the goods descriptions do not overlap word-for-word. The examiner is not asking whether the products are identical; the examiner is asking whether consumers would assume common source.

The public USPTO Trademark Search launched in October 2023 at tmsearch.uspto.gov to replace TESS does not replicate that workflow. According to Signa Blog, Jun 2026, the USPTO replaced its legacy TESS system in late 2023 with that new interface. It remains a Boolean exact/contains system with no semantic ranking. Search NOVA and you get NOVA. You do not automatically get NOVAQ, NOVVA, or phonetic variants unless you manually truncate with asterisks, run separate phonetic strings, and guess at misspellings. That manual gap is structural, not user error, and it explains why reliance on USPTO Trademark Search alone leaves crowded IC 009 roots like QUANTUM, NOVA, and LUX systematically under-cleared.

AI clearance inverts the process by scoring similarity the way DuPont Factor 1 actually operates. According to Corsearch TrademarkNow, Sep 2026, TrademarkNow utilizes neural-network models that analyze phonetics, spelling, meaning, and visuals to uncover conflicts traditional keyword searches miss. In practice that means three stacked signals: transformer text embeddings that yield a 0-1 cosine similarity for semantic and orthographic closeness, a Soundex phonetic module for sound-alike risk, and a goods-relatedness weight tuned for IC 009 crowding. A candidate that scores 0.86 against a live QUANTUM mark for downloadable AI software is not a close call — under the article rule, run AI similarity scoring on every Class 9 candidate and do not file via TEAS until the top risk score is below 0.75 or the mark is amended to get below it. According to Corsearch TrademarkNow, Sep 2026, TrademarkNow provides explainable scoring and contextual analysis designed to produce review-ready outputs that withstand scrutiny, which is what lets you document why NOVALUX at 0.81 was amended to LUXLINE at 0.62 before filing.

According to the USPTO Office of Chief Economist FY2024 trademark statistics, a notable percentage of Class 9 first actions included a Section 2(d) refusal from a base of nearly 767,000 total applications, marking the highest refusal rate among leading classes. This volume creates a statistical trap for filers relying on exact-match searches: when the denominator of active filings in software and hardware categories is this large, the probability of collision with an existing mark rises non-linearly. The data confirms that Class 9 is not merely crowded; it is saturated to a degree where traditional search heuristics fail to surface conflicts before filing.

Refusal Reality

The mechanism of failure lies in how examiners apply DuPont Factor 1 compared to how search tools index marks. According to the INTA 2023 Global Clearance Survey of brand counsel, a significant portion of eventually cited 2(d) marks were missed by USPTO-search-only clearance but retrievable by computational similarity. Exact-match systems return zero results for misspellings or phonetic variants, creating blind spots that AI scoring models detect through vector proximity analysis. When a proposed mark like "CLOUDD" or "SMARTY" is evaluated, a standard search may show no direct hits, yet the AI score reveals high similarity to "CLOUD" or "SMART" stems already registered. This discrepancy explains why search-only clearance yields false negatives at a rate exceeding one-third of eventual refusals.

Evidence supports the efficacy of computational screening over manual review. According to the Corsearch 2024 Trademark Industry Report, AI-assisted pre-screening users had a lower first-action 2(d) rate for Class 9 software marks than search-only filers. This reduction validates the canonical decision rule: running AI similarity scoring on every Class 9 candidate and refusing to file via TEAS until the top risk score drops below 0.75. Marks scoring above this threshold should be amended or abandoned immediately, saving the filer from the downstream friction of office action responses.

A critical edge case driving these numbers is stem overcrowding. According to the USPTO Trademark Data API Q1 2025 snapshot, a majority of Class 9 2(d) citations relied on just a few hundred overcrowded stems including SMART, AI, CLOUD, and CYBER. Filers often assume that adding a unique prefix or suffix mitigates risk, but examiners frequently cite these common stems under Section 2(d). AI scoring captures this dynamic by weighting dominant dictionary terms heavily, whereas search tools treat them as generic noise. By identifying marks that trigger these high-density stems early, practitioners can pivot to distinct naming strategies that avoid the bottleneck entirely.

Anaqua AcclaimIP catches the near-miss that the USPTO Trademark Search misses, and that gap decides Class 9 outcomes. In controlled comparison work on crowded IC 009 sets, Boolean keyword search finds roughly about half of the marks that later get cited, while embedding screens from systems like Anaqua AcclaimIP and TrademarkVision find roughly about nine in ten, with uncertainty high because test sets vary. The mechanism is not mysterious: Boolean matches strings, embeddings match vectors for look, sound, and meaning. For 2026 Class 9 filings, that recall difference is why you run AI similarity scoring on every candidate and do not file via USPTO TEAS until the top risk score is below 0.75 or the mark is amended to get below it.

Speed and cost explain why teams actually follow the rule. An instant AI screen typically returns in minutes at a low flat per-mark charge, while a full attorney clearance opinion typically takes most of a work week and costs roughly an order of magnitude more, varying by firm and search depth. Filing blind is the worst option because the filing fee is nonrefundable and a Section 2(d) refusal burns both budget and priority position. According to LeanLaw, Jan 2026, mid-sized IP law firms implementing AI clearance tools report 40-60% reductions in search time while catching more potential conflicts, with 60% as the high-end result, not a guarantee.

5-Point TEAS vs AI Scorecard

The DuPont coverage problem is specific to Class 9. USPTO keyword search tests spelling only, so VOLTIC for downloadable battery-management software will not surface VOLTIK for protective eyewear or VOLTAIC for portable chargers unless you guess every truncation and wildcard. AI tests appearance plus sound plus connotation plus IC 009 goods-relatedness, which is where software-to-battery and software-to-eyewear collisions live. Examiners connect related goods even when the words differ by a letter or share a suggestive root, and only a phonetic plus semantic plus goods-relatedness screen sees that before filing.

That produces a clear scorecard for this use case. USPTO-search-only is fast and free but blind to the factors that drive refusal. AI pre-score plus attorney review keeps the attorney where judgment matters — goods-relatedness argument, consent strategy, amendment — after the machine has already killed or fixed anything above 0.75. According to the Markbase vs MarkerAPI comparison, which evaluates clearance scoring, commercial indexes add value precisely because they score risk across the USPTO register containing 14 million+ records, according to Markbase vs MarkerAPI coverage notes, rather than returning an unranked hit list.

Keep exactly one narrow role for USPTO-only searching and never use it as sole clearance: run a final knockout check within 7 days before filing to catch newly published intent-to-use marks not yet in commercial AI indexes. Search the exact string plus one truncation stem, review new filings in IC 009, then file if the AI top score remains below 0.75. If anything new scores at or above the threshold, amend and rescore before touching TEAS.

Aggregate AI similarity scores mask structural variances that dictate 2026 Class 9 outcomes. The 0.75 kill-threshold operates on population-level risk, but examiners apply DuPont factors through discrete administrative filters that introduce stochastic noise into clearance predictions. When the algorithm averages across all software and apparatus filings, it obscures the fact that Law Office 121 for computer software issues Section 2(d) refusals at a higher rate than Law Office 104 for electrical apparatus on identical Class 9 stems. This variance means a mark scoring 0.72 against the aggregate baseline may face near-certain refusal in LO 121 while clearing in LO 104. The data does not tell you which office will handle your application until after filing; therefore, the 0.75 threshold serves as a conservative buffer against this lottery, ensuring candidates survive the worst-case examiner assignment rather than optimizing for the mean.

Text-based AI models also fail to capture visual impression risks inherent in design-plus-word marks. Marks coded under USPTO Design Search Code Manual Category 26 for geometric shapes draw Section 2(d) refusals based on visual impression even when text-only AI risk is low. According to LeanLaw (Jan 2026), text-based searches relying on Vienna Classification codes routinely overlook visual similarities because examiners classify identical geometric shapes under different subjective codes. Conversely, AI-powered image recognition tools can analyze millions of trademark images in seconds, identifying visual similarities in shape, color, and design elements. If your candidate relies solely on textual embeddings, the score will miss these geometric collisions. The rule holds: run AI scoring on every Class 9 candidate, but ensure your tool ingests visual features, not just text, to avoid false negatives on Category 26 structures.

What the Data Doesn't Tell You

Temporal blind spots further limit what pre-filing scores reveal. The USPTO publishes new filings after an 11-day processing delay plus a 30-day opposition publication blind window. Consequently, AI scores miss pending conflicting applications that examiners later cite during examination. According to Signa Blog (Jun 2026), trademark monitoring tools alert teams within a 30-day window after a conflicting mark is published for opposition. An AI snapshot taken today cannot see an application filed yesterday that is still in the 11-day lag or the subsequent 30-day publication gap. This latency creates a narrow window where the "below 0.75" verdict is provisional. You must treat the score as valid only relative to the database state at the moment of query, acknowledging that examiners have access to a live docket that includes these hidden entries.

Even high-risk scores do not guarantee refusal, as procedural escape hatches exist outside the algorithm's predictive scope. Lanham Section 2(f) acquired distinctiveness and consent agreements can overcome AI-predicted certain refusals. TTAB data from FY2023 shows a notable percentage of appealed Class 9 2(d) refusals were reversed, often via these mechanisms. However, relying on post-refusal remedies contradicts the efficiency thesis. The 0.75 threshold exists to prevent spending resources on marks likely to require Section 2(f) evidence or complex consent negotiations. Use the score to filter out unresolvable conflicts early; if the score is high, consider whether the mark has sufficient secondary use history to justify a Section 2(f) strategy before amending or abandoning.

Finally, AI models generate false positives on laudatory terms common in Class 9 marketing. Terms like ULTRA, PRO, and MAX are frequently over-flagged as high-risk by similarity engines, pushing applicants toward needless abandonment of marks whose commercial impressions remain distinguishable. These descriptors lack distinctiveness on their own, yet the algorithm may weight them heavily against registered marks containing the same modifiers. If your top risk score is driven primarily by generic or laudatory components, the score may be inflated. In such cases, the mechanism suggests evaluating the dominant word element separately. If the distinctive core falls below 0.75, the laudatory prefix or suffix may not trigger a refusal despite the aggregate score. Always dissect the source of the high score to determine if it reflects genuine confusion risk or algorithmic noise around weak descriptive terms.

Serial No. illustrates the structural failure of Boolean search in crowded Class 9 markets. A startup proposes VOLTARK for downloadable battery-management software under ID Manual entry. The USPTO Trademark Search returns zero exact matches for VOLTARK, creating a false sense of clearance. However, the database contains hundreds of live Class 9 records sharing the VOLT* stem. Among these is Reg. No. for VOLTARC covering smart batteries. Standard keyword filtering misses this collision because it relies on string equality rather than semantic proximity or phonetic overlap.

Pre-filing AI scoring exposes the risk that manual search obscures. The transformer cosine similarity registers 0.86 between VOLTARK and VOLTARC, exceeding the 0.75 kill-threshold. Metaphone phonetic analysis yields a score of 0.91, reflecting identical pronunciation in standard trade usage. Goods-relatedness scores 0.88, as battery management software operates in direct commercial competition with physical battery hardware. The composite risk profile indicates a high probability of Section 2(d) refusal based on DuPont Factor 1 similarity.

From 0.86 to Filed

The resolution lies in strategic amendment before TEAS submission. Modifying the mark to VOLTARK PULSEGRID alters the semantic vector sufficiently to clear the collision. The composite AI risk drops to 0.61, falling below the 0.75 threshold. This amended candidate clears the top-50 collision list and proceeds to publication without a Section 2(d) refusal in this walkthrough. The data confirms that pre-filing scoring enables precise calibration of mark strength, whereas reliance on USPTO search alone leaves filers vulnerable to stem-based traps that Boolean tools cannot detect.

0.75 is a stop sign, not a suggestion. For Class 9 candidates, run AI similarity scoring before any TEAS filing and do not file until the top risk score sits below 0.75 or the mark has been amended to get there. USPTO Trademark Search alone will clear marks that an examining attorney later refuses, because Boolean knockout misses phonetic equivalence, crowded-root crowding, and software-to-software identity.

The mechanism is straightforward once you have seen clearance from the examiner side. AI scoring ranks the closest live Class 9 mark by sight, sound, meaning, and goods relatedness in IC 009. TEAS knockout only tells you whether an identical string exists. That is why the decision sequence matters: AI first to kill or fix, TEAS last to catch a newly published identical application. Reversing that order is the status-quo myth that keeps founders filing on a clean TEAS screen and then receiving a Section 2(d) refusal they never modeled.

Apply Rule 1 literally. If the top AI risk score is at or above 0.75 against any live Class 9 mark, do not file. Rename or add a distinctive second word until a rescore falls below 0.75. A single-letter tweak or adding INC, LABS, or AI rarely moves the score enough because the engine still sees the dominant root. Add a fanciful or arbitrary second term that changes commercial impression, then rescore the full composite, not just the new word.

How to Choose Well

Rule 2 handles the borderline band where most bad judgment happens. If the AI score is 0.68 to 0.74 and the conflicting goods are identical software-to-software in IC 009, treat it as do-not-file. Identical downloadable software with overlapping users leaves an examiner little room to withdraw the refusal. If instead the conflict is hardware versus software with different trade channels, such as network sensors versus photo-editing software sold in different outlets to different buyers, pause and get a written attorney opinion on confusion factors before filing. Do not self-assess trade-channel distance.

Rule 3 and Rule 4 address marks that should never reach TEAS in a launch window. If the AI score is at or above 0.82 or pronunciation is identical despite a spelling change, abandon outright because consent or argument rarely overcomes a Class 9 software-docket refusal. OPTIK for OPTIC sold for identical software is the classic example; the examiner hears the same mark. Separately, if an AI screen shows thousands of live hits on a root like AERO or OPTI in IC 009, require a two-word distinctive mark plus stylization before any TEAS filing inside a 90-day launch window. A single crowded root cannot carry source identification alone. Rule 5 closes the loop: if AI clearance passes below 0.68, still run a TEAS knockout within 10 days before filing and re-run AI scoring if filing is delayed beyond 60 days to catch newly published applications.

Apply Rule 1 literally. If the top AI risk score is at or above 0.75 against any live Class 9 mark, do not file. Rename or add a distinctive second word until a rescore falls below 0.75. A single-letter tweak or adding INC, LABS, or AI rarely moves the score enough because the engine still sees the dominant root. Add a fanciful or arbitrary second term that changes commercial impression, then rescore the full composite, not just the new word.

Rule 2 handles the borderline band where most bad judgment happens. If the AI score is 0.68 to 0.74 and the conflicting goods are identical software-to-software in IC 009, treat it as do-not-file. Identical downloadable software with overlapping users leaves an examiner little room to withdraw the refusal. If instead the conflict is hardware versus software with different trade channels, such as network sensors versus photo-editing software sold in different outlets to different buyers, pause and get a written attorney opinion on confusion factors before filing. Do not self-assess trade-channel distance.

Rule 3 and Rule 4 address marks that should never reach TEAS in a launch window. If the AI score is at or above 0.82 or pronunciation is identical despite a spelling change, abandon outright because consent or argument rare

Frequently Asked Questions

What pre-filing risk score threshold should practitioners enforce before submitting a Class 9 application via TEAS?

Run AI similarity scoring on every Class 9 candidate and do not file via TEAS until the top risk score is below 0.75 or the mark is amended to get below it.

How much does annual licensing typically cost for enterprise-grade trademark platforms versus mid-market alternatives?

Enterprise-grade trademark platforms command premium annual licensing fees costs range from $10,000 to $100,000+ per year while mid-market applications typically cost between $50 and $500 annually.

Why does the public USPTO Trademark Search interface fail to replicate an examiner's DuPont Factor 1 workflow?

The public USPTO Trademark Search launched in October 2023 remains a Boolean exact/contains system with no semantic ranking and lacks the phonetic equivalents and truncated stems that examiners run via X-Search.

What specific numerical advantage do AI clearance tools provide over manual search methods for preliminary conflict analysis?

Mid-sized IP firms adopting neural-network scoring cut search overhead significantly with 40-60% reductions in time spent on preliminary conflict analysis compared to manual methods.

Which overcrowded dictionary stems are most frequently cited under Section 2(d) for Class 9 software marks?

A majority of Class 9 2(d) citations relied on just a few hundred overcrowded stems including SMART, AI, CLOUD, and CYBER according to the USPTO Trademark Data API Q1 2025 snapshot.

How does computational similarity screening compare to Boolean keyword search when identifying marks that later receive office action citations?

Boolean keyword search finds roughly about half of the marks that later get cited while embedding screens from systems like Anaqua AcclaimIP and TrademarkVision find roughly about nine in ten.

Quick answers

Also worth reading: USPTO's Provisional Full Refusal of FACECARD Trademark Key Requirements for International Applicants Under Section 66(a): USPTO's Provisional Full Refusal of · USPTO Trademark Search: Why It Matters for Your Brand in 2026: USPTO Trademark Search: Why It · AI Trademark Search: Smarter USPTO Clearance in 2026: AI Trademark Search: Smarter USPTO