Start with a measured language inventory
Detect the language of every crawled page using its declared HTML language and visible text. Report page count, average readiness, high-risk pages, authorship, source-ready pages and material claims per language. Preserve the original text and language code. Do not translate findings silently or assign an English score to a Spanish page.
A supported language that has no crawled page should be labelled not observed. That is a capability statement, not a recommendation to enter the market. Target-market decisions require business strategy, demand and operational capacity beyond a crawl.
Audit the complete hreflang graph
Retain every declared alternate URL and language-region code. For each page, test self-reference, target availability, language agreement and the reciprocal return link. A multilingual alternate graph should normally have one genuine x-default route when a default or language selector exists. The target should return a public success response and declare an appropriate canonical.
Hreflang is descriptive metadata, not a ranking guarantee. A correct graph helps systems understand alternate relationships; it cannot guarantee indexing, retrieval or citation. A broken graph is still actionable because the published relationship can be tested and repaired.
Compare evidence parity inside explicit locale sets
Build locale sets from reciprocal alternate links rather than guessing from similar URL paths. Compare readiness score, word count, material claims, adjacent sources, authorship, visible dates and answer-passage coverage. A large score gap, a localized page less than roughly half the depth of its strongest sibling, or a source-free version of an evidence-rich page should trigger review.
Parity does not mean literal translation. A German tax explanation may need German law, a Spanish incentives page needs the current administering authority, and a French pricing page may require different currency and market dates. The correct target is equivalent decision usefulness and provenance, not identical sentences.
Join market evidence to measured outcomes
Map the latest Search Console or Bing rows to the language of their landing page. Show impressions and clicks without calling them AI prompt volume. Map explicitly identifiable AI-referral sessions, conversions and revenue to the same page-language inventory without redistributing direct traffic.
This exposes valuable decisions. A weak Spanish evidence set with high measured search demand and growing AI referrals can outrank a technically untidy locale with no demonstrated demand. A strong German page generating revenue may need protection during template work. Totals must reconcile to the imported page rows.
Keep buyer questions market-specific
Monitoring questions should remain in the audience's language and preserve local terminology. Translating an English question mechanically can miss how buyers describe regulation, procurement, risk or product categories. Version the question portfolio whenever a market-specific question is added, and compare provider observations only against unchanged language and provider sets.
Turn findings into owned work
- International web operations: repair inaccessible targets, reciprocal links, canonicals and language-code disagreements.
- Local editorial: close content-depth and answer-completeness gaps without copying irrelevant claims.
- Research and legal: validate jurisdiction-specific claims, sources, dates, currency and limitations.
- Analytics: preserve landing page, country, language and source dimensions in equivalent windows.
- GEO programme owner: prioritize markets using measured demand and value alongside evidence severity.
Define exact acceptance tests
A good international action names every affected URL and the condition that will clear it. Recrawl the full locale set, not one page. Require every intended target to return successfully, self-reference, reciprocate, match its declared language and remain canonical. For parity work, require the weaker locale to close the readiness gap, restore local evidence and clear its answer-passage threshold.
Keep the report honest
A multilingual audit observes public pages and declarations. It does not prove how a private model segments markets or guarantee that a correct alternate will be cited. Search demand, AI referrals and provider citations are separate evidence layers. Displaying them together supports a commercial decision; it does not make them causally interchangeable.