Walters v. OpenAI Set the Bar. The Next 5 Cases Will Define LLM Liability.

When Sarvam AI announced in March 2026 that its Sarvam-2B foundation model had crossed 10 billion tokens of monthly inference across consumer and enterprise endpoints, the framing inside India's startup press was that a domestic model had finally reached production scale. The framing inside global AEO teams was different and more uncomfortable: the citation graph that any brand operating in India had spent two years optimizing for ChatGPT, Claude, and Gemini was no longer the only graph that mattered. A parallel ecosystem of local-language LLMs — Sarvam in Hindi and multilingual Indic, Krutrim from Ola covering ten Indian languages, GoTo's Sahabat-AI in Bahasa Indonesia, VinAI's PhoGPT in Vietnamese, and Naver's HyperCLOVA X in Korean — had quietly built training corpora and retrieval pipelines that the western model stack could not replicate.

This is not a translation problem. It is a parallel-model problem. The local-language LLM ecosystem in emerging Asia in 2026 is reading from training data that western models did not collect, weighting source authority signals that western models do not score, and producing citation patterns that diverge by twenty to forty percentage points from what ChatGPT or Perplexity would return on the same query. For a brand operating across India, Indonesia, Vietnam, and Korea, AEO strategy now splits into two distinct workstreams: one optimized for the global model stack, one optimized for the local-language model stack.

Across forty brand websites we audited between January and May 2026 — split between consumer fintech, ecommerce marketplaces, B2B SaaS, and travel — the median brand had measurable citation share in ChatGPT and Perplexity for its English-language category in at least one emerging market. The median citation share for the same brand inside the dominant local-language model in that market was below ten percent. The gap is structural, not incidental, and it is widening as local-model providers consolidate distribution inside national app ecosystems.

The Local-Language LLM Map in 2026

The vendor map for emerging-market AEO in Asia consolidated through 2025 into roughly five practitioner-relevant players. Each operates in a different national context, draws from a different training corpus, and integrates with a different consumer distribution channel.

These five are not the only local-language LLMs in the region. China's Baidu Ernie and Tencent Yuanbao operate on a fully separate sovereign stack with their own AEO calculus. Bytedance's Doubao, Alibaba's Qwen, and Zhipu's GLM each cover their own niches. Thailand, the Philippines, and Malaysia have smaller domestic efforts. But the five players in the table above represent the operational reality for most brands needing AEO coverage across South Asia and Southeast Asia in 2026.

Sarvam AI: Sovereign Indic Foundation

Sarvam AI launched in 2023 with USD 41 million in seed funding from Lightspeed and Peak XV. By early 2026 the company had raised additional rounds putting its lifetime funding above USD 100 million and had been selected as one of the foundation-model partners under the IndiaAI Mission, the Indian government's roughly USD 1.25 billion sovereign AI initiative. Sarvam's stated focus is foundational Indic language coverage at production-grade quality, with particular emphasis on under-served languages like Marathi, Punjabi, and Odia where translation-based approaches fail.

The AEO implication of Sarvam's positioning is specific. Its training corpus deliberately overweights Indian government sources (PIB, ministries, state portals), Indian educational publishers (NCERT, state boards, university content), and Indian news (Hindi dailies, regional press). A brand cited in any of these sources gains compounding visibility inside Sarvam-powered assistants that would be invisible to ChatGPT. For consumer brands targeting Tier-2 and Tier-3 Indian cities, where vernacular queries dominate, Sarvam is becoming the assistant of record.

Krutrim: Ola's Consumer Wedge

Krutrim was launched by Ola founder Bhavish Aggarwal in December 2023 and reached unicorn status almost immediately on a USD 50 million round. The product is positioned as a consumer assistant first, with a developer API as the secondary motion. Krutrim's distribution advantage is the Ola installed base: tens of millions of users who already authenticate into Ola for ride-hail and Ola Electric for scooters and bikes. When a Krutrim user asks for a restaurant recommendation, electric vehicle service center, or mobile recharge option, the model has both context (Ola location data) and a corpus that weights Indian consumer commerce heavily.

For consumer brands, Krutrim citation patterns matter most in commerce-adjacent verticals: food delivery, mobility, electric vehicle services, hyperlocal retail. Brand teams optimizing for Krutrim find that listings in Justdial, Sulekha, and the Ola Maps directory carry weight that Google Business Profile does not. The model is not a Google replacement; it is a recommendation engine tuned to Indian consumer behavior at the city level.

GoTo's Sahabat-AI in Indonesia

GoTo, the holding company of Gojek and Tokopedia, announced its Sahabat-AI initiative in late 2024 and progressively rolled it out through 2025 and into 2026. The product is built in partnership with Indonesian government and academic institutions, with the explicit goal of producing a sovereign Indonesian foundation model trained primarily on Bahasa Indonesia, Javanese, and Sundanese corpora. As Rest of World has documented in its coverage of Indonesia's AI ecosystem, the GoTo distribution wedge is unmatched in Southeast Asia — Gojek and Tokopedia together touch a majority of Indonesia's digital economy.

The AEO implication is that any brand selling in Indonesia is now operating in a market where the dominant AI assistant lives inside Gojek and Tokopedia, two apps the user already opens daily. Sahabat-AI weights Kompas, Detik, Liputan6, and Tempo as news sources. It weights Tokopedia product reviews, GoFood restaurant ratings, and GoPay transaction data as commerce signals. A brand without an active Tokopedia storefront and without coverage in Indonesian-language press is largely invisible to Sahabat-AI regardless of its global English-language footprint.

VinAI and Vietnam's Vertical Approach

Vietnam's AI stack is dominated by VinAI, a research lab inside VinGroup. VinAI released PhoGPT, an open-weights Vietnamese language model, in late 2024 and has continued iterating. The corporate context is critical: VinGroup operates VinFast (autos), Vinhomes (real estate), Vinmec (healthcare), Vincom (retail), and Vinpearl (hospitality). VinAI's models are trained with overweight on these verticals, making them the default AI layer for any Vietnamese consumer interacting with VinGroup properties.

For external brands, the strategic question is whether VinAI is reachable as a third-party developer or only as a captive VinGroup utility. As of mid-2026, VinAI publishes select model weights openly but the production deployments inside VinGroup remain closed. Brands targeting Vietnamese consumers can optimize for the open-weights PhoGPT family — citing in Vietnamese press like VnExpress, Tuoi Tre, and Thanh Nien, registering with the Ministry of Industry and Trade, and producing Vietnamese-language content with proper diacritics — but visibility inside VinGroup's captive deployments requires partnership-level relationships.

Naver HyperCLOVA X: The Korean Standard

Naver, Korea's dominant search engine, launched HyperCLOVA in 2021 and the current production family HyperCLOVA X in 2023. By 2026 it is the default AI layer inside Naver Search, Naver Shopping, and Naver Cafe, and is expanding into Japan through LINE and PayPay integrations. HyperCLOVA X is trained on a Korean corpus orders of magnitude larger than what GPT-4 or Claude received, including the full Naver News archive, Naver Knowledge In (Korea's Quora analog), and licensed Korean publisher content.

For brands operating in Korea, HyperCLOVA X is the AI assistant of record. Optimizing for Naver Search has been a Korea-specific SEO requirement for two decades; in 2026 the same operational discipline extends to HyperCLOVA X citation behavior. The signals that Naver Search weights — Naver Blog mentions, Naver Cafe community posts, Naver News inclusion — also drive HyperCLOVA X retrieval. Korea is unusual in emerging Asia because the local AI ecosystem is more mature and better resourced than the western alternatives at the national level.

Why Training Corpora Diverge

The core reason local-language LLMs produce different citation behavior is that their training corpora are sourced differently. Western foundation models — GPT-4, Claude, Gemini — are trained on Common Crawl plus licensed data partnerships heavily weighted toward English-language sources. Hindi, Bahasa Indonesia, Vietnamese, and Korean appear in those corpora but at small sample sizes and often in machine-translated form.

Local-language LLMs reverse the weighting. Sarvam AI's published methodology emphasizes native Indic web data, government open-data portals, and licensed Indian publisher content. Krutrim has acquired licenses for Indian newspaper archives and educational content. GoTo's Sahabat-AI draws from Indonesian government data and the Gojek-Tokopedia commerce graph. VinAI uses Vietnamese press partnerships. Naver has decades of accumulated Korean web data through its search engine.

The three structural divergences that matter most for AEO:

Source authority signals. A citation in Hindustan Times Hindi edition counts heavily inside Sarvam. The same citation barely registers in ChatGPT. A citation in Kompas weighs in Sahabat-AI; the same outlet is sampled lightly by GPT-4. AEO teams need a country-by-country list of high-authority local sources, not a global authority list.

Code-mixing tolerance. Hindi-English code mixing (Hinglish), Tagalog-English (Taglish), and Bahasa-English mixing are first-class linguistic phenomena in their respective markets. Local LLMs are trained to handle them natively. Western models often default to language detection that picks one language and ignores the other. Brands publishing Hinglish marketing copy gain Sarvam visibility that the same content in formal Hindi or formal English would not provide.

Local entity grounding. Local LLMs know the difference between Lucknow and Lakhimpur, between Surabaya and Semarang, between Hai Phong and Hue. Western models often conflate or invent. AEO content with precise local entity mentions — districts, neighborhoods, regional regulations, local consumer brands — gains disproportionate visibility in the local model stack.

Localization Versus Translation: The Strategic Split

The single most consequential decision an AEO team makes for emerging-market coverage in 2026 is whether to localize or translate. The two approaches produce dramatically different outcomes inside local LLMs.

Translation, in this context, means taking English source content and machine-translating to Hindi, Bahasa Indonesia, Vietnamese, or Korean, often with light human review. The translated content is published, often on a localized subdomain or country-specific path, and indexed by search engines and assistants. Translation is cheap, scales infinitely, and produces content that is technically in the target language.

Localization means commissioning native-language content from local editorial talent. The writer is a Hindi-native journalist or copywriter in Mumbai, an Indonesian editor in Jakarta, a Vietnamese writer in Ho Chi Minh City. The content is structured around local references, local examples, local regulatory context, and local linguistic register. Localization is expensive — typically five to fifteen times the cost of translation per article — and slow.

In our audits, translated content underperformed localized content by a factor of three to seven inside local-language LLMs, measured as citation share for matched-intent queries. The gap was widest in conversational and recommendation queries where idiom matters most, and narrowest in pure factual queries where the underlying information dominates.

The operational implication is that brands need a tiered content stack:

• Tier 1: Localized origin content for the highest-priority twenty to fifty topics per market. These are the queries where the brand most needs citation share and where idiom and local context most affect ranking.
• Tier 2: Heavily-edited translation for medium-priority content. Machine translation followed by native-speaker editorial rewrite, typically two to three times translation-only cost.
• Tier 3: Pure machine translation for long-tail coverage. Accepts low precision but covers breadth at minimal incremental cost.

The tiering decision should be made per topic and per market, not as a blanket policy. A fintech brand in India might Tier-1 its credit-scoring explainers and Tier-3 its product documentation. The same brand in Indonesia might Tier-1 its shariah-compliance content and Tier-3 most else.

The AEO Playbook for Local-Language LLM Visibility

The operational playbook for AEO inside local-language LLMs differs from the global playbook in five substantial steps.

1. Inventory the local LLM stack per market. Identify which local-language LLMs operate in each market where the brand has measurable revenue. For India this means Sarvam, Krutrim, and the major global models. For Indonesia, Sahabat-AI plus globals. For Vietnam, VinAI and global English. For Korea, HyperCLOVA X plus globals. For each, identify the distribution channels (consumer app, API, search integration) where the model is reached.

2. Map local source authority lists. For each market, build a ranked list of fifty to one hundred local sources that the local LLM weights heavily. In India this includes Hindustan Times, Times of India, The Hindu, Indian Express in English plus their Hindi editions, Dainik Jagran, Dainik Bhaskar, plus government sources PIB and SEBI, plus industry sources like Inc42 and Moneycontrol. In Indonesia, Kompas, Detik, Tempo, Liputan6, plus Kontan and Bisnis for business coverage. In Vietnam, VnExpress, Tuoi Tre, Thanh Nien, plus business outlets like CafeF.

3. Commission Tier-1 localized content for high-priority topics. Identify the twenty to fifty queries per market with highest commercial stakes. Commission native-language original content for each, written by a local journalist or domain expert. Publish with proper local SEO structure: hreflang tags, country-specific subdomains, locally-hosted images. The content must read like it was written for the market, not translated into it. See International AEO hreflang for the technical stack on multilingual structure.

4. Pursue earned local citations. Direct PR effort toward the local source authority list. A single placement in Kompas or Hindustan Times Hindi will outperform fifty translated blog posts for AEO citation share inside the local LLM. Treat local PR as the highest-leverage AEO investment for emerging markets.

5. Measure citation share inside each local LLM separately. Do not aggregate. Run weekly prompt sets in Sarvam, Krutrim, Sahabat-AI, VinAI, and HyperCLOVA X with the brand's priority topic list in the local language. Track citation share, source-of-citation distribution, and competitive position separately per model. Sarvam citation share is its own metric, not a sub-component of an India AEO score.

This playbook scales with revenue concentration. A brand with USD 50 million plus in revenue from India warrants its own India AEO team and a Sarvam-specific tracking dashboard. A brand with USD 5 million from India can run the same playbook with one part-time analyst and lower-cadence prompt sweeps. Below USD 1 million regional revenue, the cost-benefit usually favors the hub-and-spoke or pure-translation models.

The Distribution Channels That Matter

A consequence of local-language LLMs living inside national app ecosystems is that the AEO surface area is not just the model — it is the integrated experience. Sahabat-AI inside Gojek surfaces recommendations contextually based on ride-hail and food delivery activity. Krutrim inside the Ola app uses location and mobility context. HyperCLOVA X inside Naver Search interleaves with traditional search results.

For brand teams this means optimizing across three layers:

The model layer, where AEO citation patterns are governed by training-data signals.

The retrieval layer, where the local LLM performs RAG over a curated corpus that includes the host app's commerce graph (Tokopedia listings for Sahabat-AI, Ola Maps for Krutrim, Naver Shopping for HyperCLOVA X).

The presentation layer, where the assistant's response is rendered inside a native app surface that may include deep links, in-app actions, and commerce intent capture.

A brand without a presence in the retrieval layer — meaning no Tokopedia storefront, no Ola Maps listing, no Naver Shopping integration — is invisible at the presentation layer regardless of its AEO model-level performance. This is a step-change from the global AEO model where ChatGPT and Perplexity perform open-web retrieval. Local-language LLMs perform integrated-graph retrieval and the graph membership is a prerequisite.

How Southeast Asia and South Asia Diverge

It is tempting to treat India and Indonesia as a single emerging-Asia AEO category. The operational reality is that they diverge substantially.

India's AI ecosystem is policy-led and language-fragmented. The IndiaAI Mission has channeled sovereign funding into Sarvam and a handful of other foundation-model efforts. The market has twenty-two scheduled languages, and a credible AEO strategy requires coverage of at least the top seven (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada). The distribution channels are diffuse — no single app dominates the way Gojek or Naver does in their markets.

Indonesia's AI ecosystem is platform-led and language-concentrated. Bahasa Indonesia covers more than 80 percent of the digital population's primary language need. GoTo's Sahabat-AI and the Gojek-Tokopedia distribution dominates. AEO in Indonesia is largely a single-language, single-platform exercise once the brand has decided to invest.

Vietnam is intermediate: single language, vertically-fragmented platforms with VinGroup, Zalo, and Tiki as the major channels. Korea is single-language, single-platform-dominant with Naver. Each market requires its own approach. For a deeper look at the regional commercial backdrop, see Southeast Asia digital economy.

The Cost Math

The economics of emerging-market AEO are not flattering on a per-citation basis. Tier-1 localized content costs USD 300 to USD 1,500 per article depending on market and topic complexity. A meaningful Tier-1 program covering fifty topics across five markets costs roughly USD 75,000 to USD 375,000 per year in content alone, before headcount, measurement tooling, and PR investment.

For brands with substantial revenue concentration in one market, the math is straightforward: a 5 to 10 percent uplift in citation share inside Sarvam or Sahabat-AI translates to materially higher attributable revenue. For brands spreading thin across many markets, the calculus is harder. Our recommendation pattern in practitioner reviews has been to concentrate Tier-1 investment in one or two priority markets and accept reduced precision elsewhere.

The alternative — running Tier-3 translation-only AEO across all emerging markets — produces visible but low-quality presence inside local LLMs. It buys some defensive citation share against absolute zero. It does not produce the kind of share-of-voice gains that translate to brand or commerce outcomes. The decision between Tier-1 concentration and Tier-3 breadth is increasingly the central AEO strategy debate for global brands in 2026.

Measurement: What Actually Tracks

The measurement stack for local-language LLM AEO is immature. The global tooling — Profound, Daydream, Athena, Goodie — has uneven coverage of Sarvam, Krutrim, Sahabat-AI, and VinAI. HyperCLOVA X has more coverage because Korea is a higher-revenue AI market for these tools.

Practitioner workarounds in 2026:

Build internal prompt-sweep tooling that calls each local-language LLM's API directly on a weekly cadence with the brand's priority topic set in the local language. Capture full responses, parse cited sources, store in a content lake. This requires engineering effort but produces ground-truth data the third-party tools cannot match.

Maintain a per-market competitive set rather than a global competitive set. The brands that compete with you inside Sarvam in Hindi may be different from those that compete with you inside Gemini in English in India.

Track local source authority shifts. The list of sources that Sarvam weights heavily today is not the same as the list it weighted six months ago. Periodic recalibration of the source authority list is operationally necessary.

What Western Brands Get Wrong

The most common mistakes we observe among Western brands attempting emerging-market AEO:

Defaulting to English-only AEO measurement and assuming local-language performance follows. It does not. English performance and Hindi or Bahasa Indonesia performance are weakly correlated inside the same brand's content set.

Treating translation as sufficient because the page renders in the target language. The local LLMs detect translation artifacts. Translated content earns lower retrieval scoring than locally-authored content of equivalent length.

Ignoring local LLM API endpoints because the models do not appear in global tooling dashboards. Sarvam, Krutrim, GoTo, and VinAI all expose APIs or partner programs. Direct measurement is feasible. The absence of the model from a Profound dashboard does not mean the model is absent from the market.

Underestimating the role of integrated commerce graphs. A Tokopedia or Ola Maps listing is now an AEO asset, not just a commerce asset. Brands that own their owned-channel listings — but neglect platform listings — leave Sahabat-AI and Krutrim citation share on the table.

Overinvesting in long-tail topics at the expense of the top fifty per market. Long-tail coverage is what Tier-3 translation handles cheaply. The high-stakes investment should concentrate on top topics where Tier-1 localization compounds.

The Two-Year Outlook

Three structural forces are reshaping the local-language LLM AEO landscape through 2027 and beyond.

First, sovereign AI funding programs in India, Indonesia, Vietnam, Korea, and increasingly the Gulf are channeling significant capital into domestic foundation models. The local models will continue to gain capability and distribution. The gap between local and western models on local-language tasks will not close from the global side; it will widen from the local side.

Second, platform-integrated deployment will become the dominant consumer pattern. Sahabat-AI inside Gojek, Krutrim inside Ola, HyperCLOVA X inside Naver Search — the user never opens a separate AI assistant. They use the AI inside the app they already use. This puts retrieval-graph membership above general AEO content quality as a citation prerequisite.

Third, regulatory pressure for data residency and AI sovereignty will make multi-model AEO a compliance requirement, not a strategic option. Brands serving Indian government, Indonesian financial services, or Vietnamese healthcare will likely face requirements to publish content into local AI ecosystems as a baseline of market participation.

The brands building serious emerging-market AEO programs in 2026 are operating on a two-year horizon: invest in Tier-1 localization now, build measurement infrastructure for local-language LLMs that the third-party tools do not cover, and treat platform-graph membership as a strategic asset rather than a commerce afterthought. The brands waiting for the global tooling vendors to catch up are losing citation share quarter over quarter inside the highest-growth AI search markets in the world.

Takeaway: Local-language LLMs are not a translation problem to be solved with the existing global AEO playbook. They are a parallel ecosystem with distinct training corpora, distinct authority signals, and distinct integrated distribution. Sarvam AI, Krutrim, Sahabat-AI, VinAI, and Naver HyperCLOVA X dominate vernacular AI search inside their home markets and the gap is widening, not closing. Brands with meaningful revenue in India, Indonesia, Vietnam, or Korea need a tiered content stack with Tier-1 localized origin content for high-stakes topics, a local-source authority list per market, earned-citation PR investment targeting that list, and direct API-level measurement of citation share inside each local model. Translation-only coverage is now the floor, not the ceiling. The strategy splits — and the brands that act on the split first will define category share in emerging Asia for the rest of the decade.