web-checks

The shared evaluation of two auditors that check the same thing on the same page: auditmysite via a live page in Chrome, astro-post-audit via a built dist/. The collection differs, the evaluation is the same.

Status

0.5.0, families robots, meta, social, structured_data and hreflang. More follow individually; render-blocking was examined and stays out (shared, it would be about five lines, the rest are different checks).

What the package does not do

• It fetches nothing. No HTTP, no file system, no browser.
• It words nothing. Finding identifiers, severities and texts stay with the host; they differ there and are multilingual. What lives here is data and predicates.

The second is a decision against the first draft: shared Finding return values would have forced unifying the identifiers of both tools. That is a separate, visible decision — not the side effect of a move.

Structure

What the merge uncovered

The two tools contradicted each other on the same bots — on the same page they would have reported opposite things:

Bot	auditmysite	astro-post-audit	now
GPTBot	Training	Citation	Training
ClaudeBot	Mixed	Citation	Mixed
anthropic-ai	Mixed	Citation	Mixed

What counts is what the vendors themselves write: OpenAI separates GPTBot (training) from OAI-SearchBot (search), Anthropic ClaudeBot from Claude-SearchBot and Claude-User. The search variants were missing from both registries and are now included.

That is the yield of this slice: not fewer lines, but one answer instead of two contradictory ones. For astro-post-audit the findings change accordingly — “AI citation bot blocked” becomes, for GPTBot, “AI training bot blocked”, which is the usual and unremarkable situation.

meta and social (0.3.0)

Both tools checked title and description length with the same limits (60 / 160) — and both in bytes: an umlaut counted twice, “Überschrift” was too long sooner than “Ueberschrift”. meta::length counts characters after browser-style whitespace collapsing. The minimum lengths (30 / 120) existed only in auditmysite; the package carries both limits, the host decides which it reports.

For OpenGraph, a tag counted as present in auditmysite as soon as the element existed, in astro-post-audit only with content. Now uniform: empty content counts as missing. The value checks from astro-post-audit (twitter:card only summary, summary_large_image, app, player; og:image absolute) are now available to both.

structured_data (0.4.0)

Both tools evaluate JSON-LD, and their tables contradicted each other on almost every shared type. The basis is now auditmysite’s table, because it is documented: every assessment references the Google Search Central page on the feature and carries the table’s version. The decisions in detail:

Point	astro-post-audit	auditmysite	now
Presence	key exists	not empty	not empty ("", null, [], {} count as missing)
@context	substring schema.org	exact, lists, @vocab, inheritance in @graph	exact like auditmysite; missing and foreign kept separate; under a foreign context only full schema.org IRIs count as a type (JSON-LD 1.1)
Root list, @type list	not resolved, only first type	resolved, all types	resolved, all types
Organization, Person, WebSite	name (and url) required	recommendations only	recommendations only (as Google documents it)
Article	headline required	recommendations only	recommendations only
NewsArticle publisher	required	—	recommended
WebSite potentialAction	low	—	recommended; the sitelinks search box no longer exists since 2024, the host decides the weight
LocalBusiness	name required	name, address required	name, address required
FAQPage	acceptedAnswer per question	only mainEntity	both, plus name per question and text per answer (Google FAQ)
BreadcrumbList	name also via item.name	≥ 2 entries, item except for the last, name directly	≥ 2 entries, item except for the last, name also via item.name
Duplicate @type across blocks	yes	no	duplicate_types, the host switches it
Event, Recipe, VideoObject, JobPosting, SoftwareApplication, ProfilePage, CollectionPage, ItemList, Merchant Listing	—	yes	yes

Two things that used to be sentences became data in the move: the hints for manual review (ManualReview instead of English sentences) and the structural problems (StructuralIssue instead of message text). Labels of the features and status texts stay with the host.

Comparison with visible content (needs the rendered page), consistency across all pages of a build, Microdata and RDFa stay out.

hreflang (0.5.0)

Both tools check x-default and the reference to the page itself, only auditmysite the language code. Comparison case by case:

Point	astro-post-audit	auditmysite	now
x-default present	exactly x-default	exactly x-default	regardless of capitalization
Reference to itself	every entry, also x-default; query discarded	every entry, also x-default; query kept	excluding x-default (Google: each language version names itself); query kept, fragment and trailing slash do not count
Language code	—	xx, xxx, xx-YY with uppercase region	language, optional script and region, any capitalization (Google’s subset of BCP 47)
Return links between pages	yes	—	stays with the host
Target exists in the build	yes	—	stays with the host

What changes: the German home page of heise.de names itself only as x-default and counted as complete in both; now it lacks the reference to itself. Versions that differ only in the query (?lang=de) counted as references to each other in astro-post-audit. auditmysite considered de-de and zh-Hant invalid. The 23 codes on nine of the 48 real pages examined are valid under both the old and the new check.

Public surface

Entry	Purpose
robots::parse	text → RobotsTxt with groups and sitemaps
RobotsTxt::wildcard_disallows_all, blocks, crawl_delays	the questions both hosts ask
Group::disallows_path	does a block apply to this path
classify_bot, BotClass	classification as data, without judgment
path_is_disallowed	longest-rule evaluation with * and $
meta::length, LengthRange::classify, meta::TITLE, meta::DESCRIPTION	length in characters, classification too short / ok / too long
social::is_present, is_complete, completeness	presence (empty = missing) and completeness via *_REQUIRED / *_FIELDS
social::is_valid_twitter_card, is_absolute_url	value checks
structured_data::parse_blocks, parse_block → Block { nodes, issues, json_error }	script text → normalized nodes and StructuralIssue
structured_data::has_schema_org_context, extract_types, normalize_schema_type	the building blocks of normalization
structured_data::assess_node, SchemaRuleAssessment, ProductRuleContext	assessment per node and type: missing required and recommended paths, feature, availability, source
structured_data::inventory_fields	fields per type for an inventory
structured_data::duplicate_types	@type in more than one block
structured_data::RULESET_VERSION	version of the rule table
hreflang::Alternate	a link rel=alternate hreflang with an href resolved by the host
hreflang::is_valid_code, invalid_codes, is_x_default, has_x_default	language code and x-default
hreflang::has_self_reference, same_page	reference to the page itself, URL comparison
third_party::classify_host, classify_url	host or URL → Match { provider, category }; the URL form applies path rules (facebook.net serves both the Like plugin and the ad pixel)
third_party::classify_storage_name	cookie or web-storage name → provider (exact or prefix, case-sensitive)
third_party::host_of	host of an absolute or scheme-relative URL; credentials, port and trailing dot stripped
third_party::PROVIDERS, RULESET_VERSION	the provider list, every entry with a source; version of the list

Limits

The provider list is a documented snapshot, not a census. An unknown host is None; absence from the list says nothing about a host. The seat is the provider’s own statement and says nothing about where data actually flows. Whether a provider may load before consent is a legal question the package does not answer.

The bot registries are a snapshot: vendors rename crawlers and add new ones. An unknown name is Unknown, not “harmless”.

Path evaluation follows Google’s interpretation (longer rule wins, on a tie Allow). Other search engines interpret the specification differently in places; anyone who needs that precisely checks against the target, not against this package.