Standard taxonomies describe the average customer's world. Yours is not average. Our LLM-based classification pipeline maps websites and text into any customer-defined taxonomy — your compliance risk buckets, your media-buying suitability tiers, your internal verticals — starting from nothing more than category names and one-line definitions. Delivered as a dedicated real-time endpoint, as batch classification of your URL lists, or as the entire 102M-domain database pre-classified into your scheme.
IAB Content Taxonomy, web-filtering categories, Google Shopping verticals — standard schemes are excellent when your workflow was designed around them. Most enterprise workflows were not. Compliance teams reason in internal risk classes, media buyers in client-specific suitability tiers, marketplaces in their own vertical trees, and CRM systems in industry codes that predate every public standard. When the classification layer speaks a different language than the systems downstream of it, every result needs a mapping table, every mapping table needs an owner, and every edge case becomes a meeting. The categories are supposed to remove ambiguity; borrowed categories add it back.
Your compliance categories, media-buying whitelists, internal verticals and CRM industry codes evolved around your business rules. Forcing them through IAB tiers or filtering labels loses exactly the distinctions your process depends on.
A classical supervised classifier needs an ML team, thousands of labeled documents per category, annotation guidelines, evaluation infrastructure and permanent model ownership — a multi-quarter project before the first useful prediction.
Businesses reorganize categories constantly. In a trained-model world, splitting one label into three means new annotation, new training runs and new regression testing — so teams quietly stop improving the taxonomy at all.
Most categorization products ship one fixed scheme, take it or leave it. Your "leave it" option is a brittle crosswalk spreadsheet that silently mislabels everything that falls between two of their categories.
Our platform already operates multiple taxonomy families in production — IAB Content Taxonomy v2 and v3, web filtering categories, Google Shopping and Shopify product taxonomies, IPTC news codes, plus per-client custom deployments. That breadth is possible because the underlying pipeline is LLM-based: the model reads the page content and reasons about which of your labels applies, guided by the label names and short definitions you supply. Standing up a new taxonomy is therefore a configuration and calibration exercise measured in days, not a machine-learning project measured in quarters. Here is the four-step process.
You provide your category names with a one-line definition for each — optionally a few example sites or texts per label where boundaries are subtle. Flat lists and hierarchical trees are both supported. That document, typically a single spreadsheet, is the entire “training set.”
We classify a sample of your real URLs into the draft taxonomy and return the results with a confidence score per label. Your domain experts review, we refine the definitions where the model and your experts disagree, and repeat until the sample matches your expectations.
Choose your delivery mode — or combine them:
Add a category, rename one, split one in two: that is a definition change, not a retraining
project. We version the taxonomy (customer_taxonomy_v1,
_v2, …) so your downstream systems migrate on your schedule, and historical
results stay reproducible.
Three representative scenarios — not client testimonials, but the shape of projects this capability is built for.
A payments compliance team maintains its own merchant risk classes — prohibited, restricted-with-enhanced-diligence, elevated-monitoring, standard — defined by its regulators and card-network agreements, not by any public taxonomy. Their onboarding flow calls a dedicated endpoint with the merchant's website; the response is one of their risk buckets with a confidence score, so low-confidence cases route straight to a human analyst while clear cases auto-clear. The nightly batch re-screens the existing portfolio against the same definitions.
A media agency's largest client hands down a bespoke brand-suitability framework — its own tiers, its own exclusion themes, stricter than GARM in some places and looser in others. Instead of approximating it with a generic taxonomy plus a crosswalk, the agency deploys the client's framework as a custom taxonomy and batch-classifies its entire publisher universe into it. Inclusion lists come out already speaking the client's language, and the quarterly framework revision is handled as a definition update.
A B2B marketplace organizes supply into a proprietary two-tier vertical tree that its search, pricing and sales-territory logic all depend on. It licenses the 102M-domain database pre-classified into that tree as an offline delivery, enriching every company record with the marketplace's own vertical — and uses the real-time endpoint to slot newly registered sellers into the same tree at signup.
A custom taxonomy is served from a dedicated endpoint provisioned for your account. The request shape is identical to our standard classifiers — a URL or raw text plus your API key — and the response carries your label vocabulary and the taxonomy version, so integration is a drop-in for anything already consuming our standard API.
# Dedicated endpoint provisioned for your taxonomy
curl -X POST "https://www.websitecategorizationapi.com/api/custom/acme_risk/classify.php" \
-d "query=https://example-merchant.com" \
-d "api_key=YOUR_API_KEY"
{
"classification": [
{
"category": "Restricted - Enhanced Due Diligence",
"confidence": 0.91
},
{
"category": "Elevated Monitoring",
"confidence": 0.06
}
],
"taxonomy": "customer_taxonomy_v1",
"status": 200
}
The category values are your labels,
verbatim — no mapping layer between the API and your downstream systems. Confidence scores
let you set your own automation threshold: auto-accept above it, human review below it.
import requests, csv # 1. Classify a URL list against your taxonomy with open("merchant_urls.csv") as f: urls = [row[0] for row in csv.reader(f)] results = [] for url in urls: resp = requests.post( "https://www.websitecategorizationapi.com/api/custom/acme_risk/classify.php", data={"query": url, "api_key": API_KEY}, timeout=120, ) results.append(resp.json()) # 2. Route by confidence against your own threshold for r in results: top = r["classification"][0] if top["confidence"] >= 0.85: auto_apply(top["category"]) else: review_queue.add(r)
For large lists you do not need to loop over an API at all: send us the file and receive it back classified, or license the pre-classified 102M-domain database and join on domain locally — the right mode depends on volume, latency needs and data-residency constraints, and mixing modes on one taxonomy is normal.
Custom does not replace the standards — IAB and web-filtering schemes remain the right choice when your counterparties expect them. Custom is for when the taxonomy itself is part of your competitive or regulatory logic.
| Standard taxonomies (IAB, filtering, ecommerce) | Your custom taxonomy | |
|---|---|---|
| Time to deploy | Immediate — live endpoints today | Days: define, calibrate on a sample, go live |
| Label ownership | Industry bodies (IAB, IPTC, Google) own and change the labels | You own the labels, the definitions and the versioning |
| Fit to internal systems | Requires a mapping table to your compliance / CRM / buying categories | Native — the API speaks your vocabulary directly |
| Retraining needs | None (fixed scheme) | None — category changes are definition edits, not model retraining |
| Delivery formats | Shared API endpoints, database exports | Dedicated per-customer endpoint, batch runs, full 102M-domain database in your scheme |
| Best for | Interoperability with the wider ecosystem | Compliance buckets, suitability schemes, vertical trees, industry codes |
There is no meaningful minimum — we have run schemes from a handful of risk buckets to hundreds of labels. Flat lists and hierarchical trees are both supported; for hierarchies we classify tier by tier, the same way our multi-tier IAB deployment works, so a two- or three-level tree behaves naturally. What matters more than size is that each label has a definition a colleague could apply consistently — if a human can adjudicate it from the definition, the pipeline can too.
Accuracy is established empirically on your data during calibration, not quoted as a generic percentage. We classify a sample set of your URLs, your experts review the output, and definitions are refined where model and expert disagree — usually the disagreements reveal genuinely ambiguous label boundaries, and tightening one sentence in a definition resolves them. Every prediction ships with a per-label confidence score, so after go-live you can hold a high-precision automated lane and route only low-confidence cases to humans.
As a project setup fee covering taxonomy onboarding and calibration, plus usage: per-call pricing on your dedicated endpoint, per-URL pricing for batch runs, or a database license if you take the full 102M-domain delivery. The mix depends on volumes and delivery modes, so pricing is quoted per project — contact us with your taxonomy outline and expected volumes for a concrete proposal.
Send your category list with one-line definitions and a sample of URLs — we will return the sample classified into your scheme so you can judge the fit on your own data before committing to anything.
Start a Custom Taxonomy Project Read the API Docs