# `sample-domains-10k.csv` — schema and sampling method

## Sampling method

Stratified systematic sample by Tranco rank, drawn from `run_id=top1m-2026-08-25-full-v2`
(2026-08-25). Every row in the Top-1,000 band is included; the other three bands are drawn at a
fixed rank-stride so the sample stays usable for reproducing the rank-band chart without needing
the full 1,000,000-row dataset — a plain top-10,000-by-rank draw would put roughly 0% of its
rows past rank 10,000, which is 99.9% of the actual universe.

| Band | Rank range | Rows in sample | Method |
| --- | --- | --- | --- |
| Top 1,000 | 1–1,000 | 1,000 | All domains in range (both reachable and unreachable) |
| 1K–10K | 1,001–10,000 | 2,250 | Every domain where `rank % 4 == 0` |
| 10K–100K | 10,001–100,000 | 3,000 | Every domain where `rank % 30 == 0` |
| 100K–1M | 100,001–1,000,000 | 3,600 | 40 equal-width rank slices across the range; first 90 domains by rank in each slice |
| **Total** | | **9,850** | |

Both reachable and unreachable domains are included in every band, at their natural rate (~16.8%
unreachable in this sample vs. 15.3% in the full census) — this is deliberate: it demonstrates
the `reachable=false` semantics described in `METHODOLOGY.md` directly in the sample data itself,
rather than only asserting it in prose.

## Fields included

| Column | Meaning |
| --- | --- |
| `domain` | The domain fingerprinted |
| `tld` | Top-level domain |
| `rank` | Tranco rank at scan time |
| `reachable` | Whether the homepage fetch succeeded — see `METHODOLOGY.md` |
| `status` | HTTP status code (empty if the fetch failed before a status was returned) |
| `scheme` | `http` or `https` |
| `final_url` | The URL that produced the adopted body, after redirect/`www.` fallback resolution |
| `technology_names` | Pipe-separated (`\|`) list of every detected technology name |
| `categories` | Pipe-separated list of every detected technology's category |
| `tech_count` | Count of detected technologies |
| `cms`, `ecommerce`, `cdn`, `web_server`, `server_language` | Single-value rollup fields (primary detected vendor in that category, if any) |
| `analytics` | Pipe-separated list of detected analytics technologies |
| `has_captcha` | Whether a CAPTCHA technology was detected |
| `render_tier` | `http` for every row in this release (no browser-render pass was run) |
| `seed_source` | `tranco` for every row (the `internal-harvest` corpus is out of scope) |
| `detector_version`, `run_id`, `run_date`, `probed_at` | Provenance / reproducibility |

## Fields deliberately NOT included, and why

| Field (on the source record) | Why excluded |
| --- | --- |
| `technologies[]` (nested: per-technology `confidence`, `version`, `evidence`) | The flattened `technology_names`/`categories`/`tech_count` columns already answer "what stack does this site run." The nested detail — confidence tier, detected version string, and the literal evidence snippet — is queryable-API-tier granularity, not aggregate-release granularity. |
| `failure_reason`, `probe_error`, `block_matched` | Scan-quality diagnostics, not tech-stack content. `block_matched` specifically lists literal block/challenge marker strings — out of scope for a technology-stack release. |
| `is_infrastructure` | A best-effort heuristic the source system's own documentation says cannot be trusted on a false negative. Publishing it as a column risks it being read as verified fact. |
| `secondary_pages[]`, `secondary_only_technology_names` | Would expose specific internal page URLs (e.g. `/careers`, `/login`) crawled per domain — crawl-target detail, not stack-composition detail. |
| `source_url`, `method_version`, `schema_version` | Internal pipeline plumbing with no reader-facing value; current values are stated once in `METHODOLOGY.md` rather than repeated per row. |
| `category` (the seed-time field, distinct from `categories`) | Usually empty; a leftover seed-stage field, not detector output. |

## Not included in this release at all

The full 1,000,000-row dataset is not published in any form — this sample is capped at ~1% of
the census. For the complete, continuously-rescanned dataset, query the
[live API](https://crawlora.net/datasets/techstack).