# Codebook: pricing-page corpus, 9 August 2026

Every field in `corpus.csv` is machine-derived from a rendered page that is archived. Two rows per
vendor, one per language arm. `frame.md` is the vendor list, fixed before any page was opened.

## The definition

> **The pricing page a visitor reaches from the vendor's own homepage navigation**: English under
> `Accept-Language: en-US`, German under `de-DE`.

Not "a pricing page exists somewhere". A page nobody can navigate to is not a disclosure. This is
**stricter** than what other published counts appear to measure, which is part of why the rate here
is lower: a fact about definitions, not a claim that anyone else is wrong.

## Fields

| Field | Meaning |
|---|---|
| `arm` | `en` or `de`: the language the request asked for |
| `status` | `coded`, `no_pricing_link_in_nav`, or `UNMEASURED_*` |
| `valid` | false when the fetch never returned a real page. **Missing data, never a negative** |
| `content_lang` | the language the page actually came back in, detected from the text |
| `shows_price` | ≥1 currency amount rendered on the page |
| `contact_wall_near_price` | a contact-sales prompt **within 1,500 characters of a price**: the strict, tier-level reading |
| `contact_wall_anywhere` | a contact-sales prompt anywhere on the page: the loose reading |
| `n_price_points` | count of distinct currency amounts. A **proxy** for priced tiers, not a tier count: add-ons and monthly/annual pairs inflate it |
| `dom_sha256` | first 16 hex of the rendered DOM hash |

## Why the contact-wall figure is two fields and not one

The first version of this corpus reported the loose code as a single number. It should not have:
**16 of the 27 pages it counted carry the same phrase on the vendor's homepage**, where there are no
tiers at all, so the loose code cannot separate a tier-level contact wall from site-wide chrome.
Both codes are published and the honest statement is the range between them.

## Two collection defects, recorded rather than buried

1. **A first collection scored 53 CDN challenge pages and consent interstitials as vendors with no
   pricing link.** The cause was an unfingerprinted headless browser and no validity screen on a
   render. Every observation now passes an anchor-count and byte floor with one retry, and anything
   that still fails is `UNMEASURED_*`.
2. **A same-domain filter rejected vendor subdomains**, discarding DocuSign's `Plans & Pricing`
   link because it points at `ecom.docusign.com`.

Residual: **21 of 128 observations are unmeasured**, across 13 vendors, and they are in the file
with that status rather than counted as opacity.

## Standing limitations

- Collection ran from a German IP, so geo-routing and `Accept-Language` can disagree. This is
  measured rather than assumed: `content_lang` records what actually came back, and 19 of 34
  German-arm pages came back in English.
- One observation per vendor per arm, on one day. Page churn is unmeasured.
- Contact-CTA detection is an English/German phrase list; unusual wording reads as no wall.
- The frame is US-listed public cloud companies. It is not "SaaS" and generalises to nothing wider.

Full archives of every rendered page, the collector as run, and the audit that caught defect 1 are
retained in the source library alongside this file.