Arabic Keyword Normalization: One Query, Many Spellings

Arabic writes the same word in several legitimate-looking ways. A searcher who means أسعار مكافحة الحشرات in Riyadh may type اسعار مكافحه الحشرات في الرياض — the hamza on alif dropped from أسعار, the ta marbuta of مكافحة rendered as bare ه. Keyboards and typing habits make this normal behavior rather than an edge case, and every spelling is a different string to a literal tool.

String-level tooling treats each spelling as a different keyword, and the damage compounds quietly: demand for one service looks split across five rows, clustering proposes near-duplicate topics that differ only by orthography, and anchors written against one spelling stop matching pages written in another. Stemming and matching describe how engines cope. A research process that owns its vocabulary does better: it canonicalizes first.

Normalization changes keys, not sentences. Three real pairs show the pattern: a hamza that appears and disappears, a ta marbuta typed as its bare-haa twin, and a phrase carrying optional tashkil. Alif maqsura pairs — مستشفي against مستشفى — resolve by the identical rule. Each surface form collapses onto the same canonical string, and each row keeps a human-readable meaning beside it.

What normalization never does is rewrite visible content. The rendered page keeps human spelling; the canonical form lives in the keys underneath — cluster comparisons, demand rows, anchor pools. An auditor reading the page reads Arabic; an auditor reading the keyword records reads canonical strings. That separation is what lets matching stay deterministic while the prose stays human, which is the entire trade.

Per node, the map carries exactly one canonical form. Every orthographic variant folds into it before demand scoring, so a topic is approved once against its full demand instead of three times against fragments. Unnormalized variants surface later as suspicious sibling topics — near-identical question frames, overlapping H2s — and someone burns a merge cycle discovering they were always one keyword.

Vernacular spellings are a different question, and a disciplined research process separates them on purpose. Register — عزل فوم, تسليك مجاري — is mined as data through dialect mining across Gulf, KSA, Egypt and Levant lexicons, and a vernacular term earns its own node when demand justifies one. Orthographic variants, by contrast, never become map nodes; they stay a content-layer concern, resolved by normalization.

Because normalization happens once at intake, every downstream artifact inherits it unchanged. The EAV matrix keys its rows to canonical strings, so a price band is stored once rather than scattered across spellings. The topical map scores demand against the canonical keyword, and SERP-overlap checks compare like with like — no reviewer has to guess which spelling is being judged.

Anchor policy and audit close the chain. Anchors are drawn from canonical terms, which lets the anchor-diversity ceiling count identical anchors correctly — two spellings of one anchor would look like diversity to the counter and like spam to a reader. The audit re-checks density and headings against the same form. One spelling from intake to publication; no layer re-litigates orthography.

Brand spellings and proper nouns stay as written — always. A client whose name is stylized with a particular spelling owns that spelling; normalizing it would be mistyping their name on their own site. Place names such as الدرعية and product names such as iPhone are identities, not keywords, and a canonical-form discipline is built to recognize them without ever editing them.

Deliberate vernacular in H2s and extractive answers is voice, not error. An Arabic heading that asks the question in the trade’s own register is doing its job; rewriting it into flattest MSA sands off exactly the human-voice quality the audit’s Arabic check exists to protect. Normalization is a discipline for matching and counting. The moment it starts editing prose, it has left its lane.