One Topical Map, Two Languages: Building Multilingual Content Networks

The drift is not dramatic; it is administrative. One team approves a node the other never hears about. A merge lands in the English map and the Arabic twin keeps both topics. Anchors rotate differently, so the two link graphs slowly describe two different sites. Within a quarter, the audit is grading two corpora — and neither report describes the actual site.

One ontology closes the gap structurally. The map is decided once — demand scoring, SERP-overlap checks and merge decisions all run against the shared node graph, and each decision is logged — and bilingual scope becomes a build-time property: every node produces two documents, one per language. Coverage asymmetry between languages then becomes structurally impossible; any deliberate asymmetry is a logged decision, not an accident of translation.

One rule decides the answer: whatever is shared is shared exactly, not approximately — shared means byte-identical, not roughly equivalent. This is the quiet advantage of building both editions from one ontology instead of translating — the two editions are artifacts of one plan, like two outputs of the same program, so the shared skeleton stays consistent. Five layers in particular never fork between the EN and AR editions:

• Slugs and URL structure — English by policy; stable addresses, no 301 churn
• The node graph — approvals, merges and demotions decided once for both editions
• The internal-link DAG — same edges, same equity direction in each language
• Section count and order — the AR H2s answer the same question sequence
• Publication dates — one editorial timeline, two renderings of it

The reasoning is mechanical. A slug is an address; translating addresses breaks recognition and invites redirect churn. An internal link must point at one URL or the graph splits its own equity. A date is evidence — the same publication event cannot have happened twice on two calendars. Localizing any of these would trade structural truth for cosmetic familiarity, which is a bad trade. Hreflang-aware publishing completes the picture: paired hreflang annotations declare each URL’s language twin, so engines consolidate signals for one network instead of splitting them across two rival sites.

Language enters exactly where a reader touches it. Prose, obviously — but also the H2s, re-asked as questions in the reader’s language rather than translated word-for-word, the date label, the author role line, and the register of the whole document: formal MSA with Arabic punctuation, not English syntax wearing Arabic words. The AR edition is a native document, not a mirrored template.

Terminology lockstep is the discipline that keeps it coherent. Every technical term has one agreed rendering per locale — topical map is always خريطة الموضوعات, ontology stays الأنطولوجيا, EAV stays EAV — decided once and applied everywhere. Without lockstep, each document improvises its own vocabulary and the bilingual corpus reads like it was written by a committee that never met.

Brand and entity names are addresses in the reader’s memory, and addresses do not translate. iPhone, Google, JSON-LD, Search Console stay Latin in the Arabic edition — exactly as a searcher who typed the brand name expects to find it. Translating them does not localize the page; it orphans the reader’s own query and merges the glossary into the brand. Recognition is the whole asset.

Numerals and structural ids follow the same logic. Numbers stay Western digits, the convention most Arabic web publications follow for statistics. Section ids, slugs and schema names are functional machinery — a translated id breaks the table-of-contents anchor that points at it. If a string exists to be matched by a machine or found by a URL, it is infrastructure, and infrastructure keeps its spelling.

The audit re-reads each edition as its own artifact, through per-locale checks. The Arabic human-voice and normalization checks run on the AR edition; a dedicated bilingual/Arabic syntax check examines native-voice signals the way the English checks examine theirs. Snippet budgets are language-specific: 30–45 words for an English extractive answer, 25–35 for Arabic — because the languages differ, not because one edition is graded easier.

Separate grading does not mean separate truth. Both editions draw their claims from the same set of verified facts, so a price band observed once is stated consistently in either language, and a claim that cannot trace to evidence fails in both editions at once. The fact record and its evidence rows are language-independent; only the prose that reports them localizes.