Classify any domain into IAB categories —
more accurate than a frontier LLM, ~50× faster
ZeroGPU's Domain Classification model reads nothing but the domain — e.g. espn.com — and returns standard IAB content categories. On a held-out, independently-labelled benchmark it is more accurate than GPT-5.4 Nano while running about ~50× faster.
01What the model does
A purpose-built classifier for high-volume adtech: it turns a raw domain into structured IAB content categories, topics, keywords, and intent signals — from the domain alone.
Input is just the domain (no page fetch, no crawl) — up to ~10× smaller payload than page-level classification. Ideal for dead, parked, or un-crawlable domains where there is no page to read.
Low-latency, high-volume by design — bidstream enrichment, contextual targeting, brand-safety screening, and domain-level intelligence at ad-auction speed.
Structured IAB 1.0 and IAB 2.2 outputs through ZeroGPU's OpenAI-compatible API. Supports batch processing; additional IAB tags available on request.
02How this benchmark was measured
A like-for-like, bias-free comparison: identical inputs, a stronger independent referee for ground truth, and the same scoring for both systems.
Real web addresses the model never saw in training, spanning the full breadth of IAB tier-1 categories (technology, finance, shopping, sports, health, travel, education, and more).
Every URL's correct labels were set independently by GPT-5.5 — a larger, stronger model than either system under test. It did not produce the model's training labels, so it cannot favour it.
The baseline, GPT-5.4 Nano, was prompted to do the identical URL→IAB task from the same sparse input. Both systems are scored on content micro-F1 against the gold labels.
03Accuracy — and the shape of the win
On the 784-URL gold benchmark the model reaches 0.385 F1 vs GPT-5.4 Nano's 0.353. The win is precision-driven: when it assigns a category it is right far more often (0.425 vs 0.333), at essentially the same recall — so downstream targeting and brand-safety decisions carry less noise.
Bars scaled to 0.50. ZeroGPU leads on precision and F1; recall is a near-tie (0.351 vs 0.374). Net: +0.032 F1 (+9%) over the frontier baseline.
04Where it wins, by content area
Broken down by IAB tier-1 category, the model beats GPT-5.4 Nano in 18 of 26 well-populated categories — and the wins are largest exactly where contextual targeting spends the most.
High-value contextual segments — shopping, technology, business & finance, sports, healthy living, and movies.
Small, ambiguous, or naming-driven categories where a bare domain under-determines the topic and the LLM's broad world knowledge helps — the natural next targets for additional training data.
05Why it's state-of-the-art for this task
Built on ModernBERT, the strongest current compact text encoder — newer data, longer context, more efficient attention. On short, keyword-like inputs (exactly what a URL is) it extracts far more signal than older small encoders.
Trained on ~1 million domains labelled by a capable LLM that could see the live page at labelling time. It distils that page-aware knowledge into the URL string — so it reproduces page-level judgments from the domain alone.
A tuned per-decision confidence threshold means it commits only to categories it is sure of. That single step lifted precision from ~0.26 to ~0.42 — turning a near-tie into a clear win.
07Speed, reliability & input
A bidstream classifier must be fast, well-formed, and tiny to call. The model answers in tens of milliseconds, never invents a label, and takes only the domain.
~50× faster — tens of milliseconds vs ~2 seconds per URL.
No page fetch, no system prompt, no taxonomy in the request — send a domain, get structured IAB categories back. Batch processing supported for high-volume pipelines.
Run details
| Task | URL-only IAB content classification (domain in → IAB content categories out) |
| Test set | 784 held-out URLs, never seen in training, across IAB tier-1 categories |
| Gold standard | GPT-5.5 — independent referee; did not produce the model's training labels |
| Baseline | GPT-5.4 Nano, prompted on the identical URL→IAB task |
| Metric | content micro-F1 (precision / recall balance) — same scoring for both |
| Result | F1 0.385 vs 0.353 (+0.032, +9%) · precision 0.425 vs 0.333 · recall 0.351 vs 0.374 |
| Latency | ~36 ms vs ~1,900 ms median per URL (~50× faster) |
| Model | ZeroGPU Domain Classification (zlm-v1-iab-domain-classifier) — ModernBERT backbone, ~1 million-domain page-aware distillation, calibrated |