Multilingual GEO

1. What multilingual GEO is

Multilingual GEO is the GEO discipline applied across a language boundary — the study of which steps of the answer loop are language-invariant and which are not, and what to do about the ones that are not.

Definition (GEO Wiki working definition): multilingual GEO is the practice of optimizing for AI engines that retrieve, ground, and cite from a per-language slice of the web, with the recognition that four mechanisms — source pool, cross-language entity binding, chunk citability, and trust corroboration — vary with language even though the underlying answer-loop shape does not.

The four mechanisms that vary, in one phrase each:

• Source pool — retrieval is per-language: the corpus an engine pulls from is selected by the query’s language, not unified across languages.
• Entity binding — your brand has a surface form in each language but a single canonical entity; whether the engine joins surface to canonical across languages is its own gate.
• Chunk shape — citability heuristics (paragraph density, sentence length, punctuation, structure) were calibrated on language-specific corpora; English-trained extraction does not transfer cleanly to Chinese, and vice versa.
• Trust pool — what counts as an authoritative corroborating source differs across language pools; E-E-A-T is a stable concept, the observable pattern that satisfies it is not.

Language is a routing dimension inside each step of the generative engine optimization loop, not a translation problem layered on top of an English-default workflow. The framework below is general; the worked examples (zh ↔ en) are the pair the field has the most live evidence for.

2. The four axes — at a glance

The table this entry’s mechanism stories hang off. Each row is a separate § below.

The rest of this entry walks each axis once, starting with the structural one.

3. Per-language source pools — and the engines that draw from them

An AI engine builds its answer from a retrieved sentinel set of sources, and that set is selected from the language-matched slice of the web — not from a global corpus that languages share. This is the headline structural fact: appearing in an en answer is not evidence you appear in a zh answer, and vice versa, because they are different retrieval events over different webs.

                  "best CRM for SMB"
                        │
            ┌───────────┴───────────┐
            ▼                       ▼
      en query path           zh query path
            │                       │
       retrieves from          retrieves from
       en web slice            zh web slice
       (Wikipedia-en,          (Wikipedia-zh, Baidu
       vendor docs,            Baike, Zhihu, vendor
       SE Land, Reddit,        cn docs, Weixin 公众号
       G2, Capterra…)          articles, 36kr, …)
            │                       │
            ▼                       ▼
       en answer               zh answer
       (different sources, often different conclusions)

The two pools are not a translation of each other. Wikipedia-zh runs at roughly a fifth of Wikipedia-en’s article count; Baidu Baike is significantly larger than Wikipedia-zh inside China but invisible to Western engines; Zhihu and WeChat 公众号 content are primary discovery surfaces with no exact Western analogue; vendor documentation translated into zh often lags or is missing. The same query against the same brand produces, in practice, two different SERPs and two different answers.

The engines that read each pool, with what crawlers they admit:

A few load-bearing details to read carefully. Yuanbao retrieves from WeChat 公众号 content — a corpus no Western engine can reach, and the cleanest single example of the source-pool asymmetry. Baiduspider self-documents at www.baidu.com/search/spider.html and respects robots.txt directives; the other Chinese vendors are less transparent. DeepSeek’s product is a chat surface backed by API-served models, not a web retriever — there is no DeepSeek crawler to admit, and being “cited by DeepSeek” usually means being in the model’s training data or attached via a third-party retrieval layer, not retrieved live.

The 2025 Chinese AI-search market is structurally different from the West — practitioner coverage in 36kr (Feb 2025) tracks 230M+ Chinese AI users and a three-way ecosystem split between traditional Baidu/Google search, social search dominated by Xiaohongshu (~600M daily searches inside the app), and AI-native assistants. The Xiaohongshu shape, in particular, has no exact Western analogue — discovery happens inside a social-content app, not at a search box.

Blocking GPTBot does not block Baiduspider; blocking Baiduspider does not block GPTBot. Each pool is gated by its own crawlers, and absence from one pool is invisible from the other pool’s tooling — your en-side rank tracker will not see your zh-side coverage gap.

4. Cross-language entity binding

A brand has a surface form in each language but a single canonical entity. Whether the engine joins surface to canonical across the boundary is its own gate — the harder version of the resolution problem entity recognition owns for the single-language case.

The mental model: one canonical entity, multiple surface forms across languages. The Wikidata Q-id is the language-agnostic spine; the language-specific Wikipedia pages (en-Wikipedia, zh-Wikipedia, etc.) are surface attestations of the same node. Wikidata centralizes all interlanguage links between Wikipedia articles about the same topic since 2013, with labels and descriptions stored in any number of languages and a fallback chain when a label is missing in the requested one (Wikidata Help:Multilingual).

The explicit join key for cross-language resolution is sameAs pointing at multiple language Wikipedias plus the Wikidata URI. Schema.org’s own definition is exactly this shape — sameAs is the “URL of a reference Web page that unambiguously indicates the item’s identity. E.g. the URL of the item’s Wikipedia page, Wikidata entry, or official website” (schema.org/sameAs). For the JSON-LD pattern, see Schema.org for AI; the cross-language case is the hardest test of the join — multiple surface forms must collapse onto one node without any of them contradicting the others.

The failure shapes, ordered by frequency:

The honest bound, the cross-language counterpart of entity recognition §6’s long-tail finding: resolvability rises with how widely an entity is attested in BOTH language pools. Cross-lingual entity linking research treats Wikidata as the spine and successfully links mentions in 100+ languages to a shared knowledge base of ~20M entities (Botha, Shan & Gillick, EMNLP 2020) — but its measured success is on Wikipedia abstracts, not long-tail brand resolution on the open web. A famous global brand resolves reliably across the boundary; a long-tail brand often does not, even with sameAs, because there is nothing on either side for the join to corroborate against.

5. Citability, attribution density, and trust signals are not language-invariant

Three concerns the field treats as language-invariant but is not: how a passage extracts, how attribution reads, and what counts as a trusted source. The shared mechanism is the same — the heuristics that calibrate each were trained or tuned on a specific language’s web. The concepts hold; the observable pattern that satisfies each in a given language does not transfer.

5.1 Chunk shape and citability

The chunk-extraction heuristics behind citability — paragraph length, sentence density, header cadence, scannability — were calibrated on English-dominant corpora. Chinese chunk shape differs structurally:

The fix is not “mimic English structure” — that produces stiff, AI-flavored zh that reads worse to humans without reliably reading better to machines. The fix is explicit structural scaffolding inside zh idiom: real H3 anchors, a one-sentence lead per paragraph, summary tables, key claims surfaced before the supporting prose. Western scaffolding inside native zh rhythm, not the other way around.

5.2 Attribution density and citation form

Western practitioner content typically marks attribution with numbered inline citations or footnote-style superscripts ([1], [2]). Chinese practitioner content frequently inlines the source name in prose — “据 IDC 报告” (“according to IDC’s report”), “Gartner 数据显示” (“Gartner data show”) — without any numbered cite at all. The argument citation vs mention makes about the link between named-source density and citability holds in both languages, but the form that satisfies it differs.

An engine calibrated on the numbered-citation pattern may under-count zh attributions even when they are present and well-formed. A zh content strategy that ports English’s footnote habit verbatim into zh tends to read as unnatural; one that uses idiomatic in-flow naming still satisfies the citability gate, just on the zh side.

5.3 Trust pool and corroboration

E-E-A-T’s pool of “what counts as an authoritative source” is per-language. A short comparison:

Authority cues differ. Government-source weight is higher inside the zh pool than the en pool. Platform-native weight (Zhihu, WeChat 公众号, Xiaohongshu posts) is much higher inside the zh pool because the public-square-via-platform pattern dominates Chinese discourse. KOL weight in a vertical can outweigh institutional weight inside the zh pool in ways less common inside the en pool. Trust corroboration is per-language, not because trust is a different concept across languages, but because the stock of trusted attestors is.

Citability and trust are stable concepts; the observable pattern that satisfies each in a given language is not.

6. The technical layer — hreflang, URL i18n, and regional crawlers

Hreflang is hygiene; per-language source pools are the mechanism. This § exists to keep that order honest.

A site with perfect hreflang and zero presence in the zh source pool will not be cited in zh. A site with broken hreflang but a strong zh-pool attestation chain will be. The technical layer matters because crawler access is its own gate; hreflang matters because traditional search still mediates a meaningful share of AI engines’ retrieval. Neither is the mechanism — the mechanism is which sources land in the retrieved set.

For the per-language file alongside robots.txt, see llms.txt; for the broader crawler-access discipline, see AI Crawlers.

7. What the evidence says — and what it does not

The mechanism direction (pools differ, binding fragments, extraction varies) is well-attested. The brand-level coefficient — how much a zh-engine cites differently from an en-engine, for what kinds of sources, by what factor — is not rigorously published.

The honest gap, stated plain: there is no published rigorous benchmark, as of 2026-05, of zh-engine vs en-engine citation-preference deltas at the brand level. No public study systematically compares ChatGPT, Perplexity, or Gemini answers against Doubao, Yuanbao, or Baidu AI Search on matched-intent queries for the same brand in each engine’s dominant language. Anyone who quotes a precise percentage for “zh engines cite X% more brand-named sources than en engines” is over-claiming. Read the direction (the four axes really do vary), build for it, and resist importing en coefficients into zh decisions.

8. Anti-patterns — multilingual misreads

The failure mode is rarely “we forgot to translate.” It is treating one of the four axes from §2 as language-invariant when it is not.

9. Why this matters for GEO + how to act

Multilingual GEO is not a separate discipline — it is GEO done once per language pool, with the four axes from §2 as the per-pool audit checklist.

The practical reading: run the four-axis audit for each language pool you actually serve. Most teams discover that they have one pool well-resourced and another pool either invisible or actively miscalibrated — usually because they assumed translation plus hreflang covered the rest.

References

Academic:

• Wu, S., Tang, S., Yang, J., Wang, S., Jia, R., Yu, S., Yao, S. & Su, J. (2024). Not All Languages Are Equal: Insights into Multilingual Retrieval-Augmented Generation. arXiv:2410.21970
• Botha, J. A., Shan, Z. & Gillick, D. (2020). Entity Linking in 100 Languages. EMNLP 2020. ACL Anthology
• Aggarwal, P. et al. (2024). GEO: Generative Engine Optimization. KDD ‘24. arXiv:2311.09735 · paper summary — boundary reference; English-only measurement

Official / standards:

• Schema.org — sameAs — “URL of a reference Web page that unambiguously indicates the item’s identity”
• Google Search Central — Tell Google about localized versions of your page — hreflang annotation methods and the explicit “Google does not use hreflang for language detection” note
• Wikidata — Help:Multilingual — multilingual labels, fallback chains, and centralized interlanguage links

Chinese AI engines (official surfaces):

• Baidu — chat.baidu.com (consumer AI search) · Qianfan Intelligent Search Generation API
• Alibaba — Qianwen (formerly Tongyi; Qwen model family)
• ByteDance — Doubao
• Tencent — Yuanbao (Hunyuan, integrates WeChat 公众号 content as citation source)
• DeepSeek — deepseek.com (API-served; no public web-crawler UA)

Industry:

• Hunt, M. (2026-01-21). International SEO in 2026: What still works, what no longer does, and why. Search Engine Land
• 36氪 (2025-02-11). 2025 搜索之战愈演愈乱：从新旧王朝到三’族’鼎立. 36kr.com