Local vs Global Brands in AI Recommendations

Six of the ten best-selling skincare brands in Indonesia by e-commerce GMV are homegrown: Glad2Glow, Wardah, Skintific, Facetology, Hanasui and Scora, outselling global majors like Garnier, Cetaphil and La Roche-Posay in actual transactions. A separate, rigorous audit of AI beauty recommendations found the exact opposite pattern, an AI system naming exclusively global and Western brands, e.l.f., Maybelline, Charlotte Tilbury, NYX, L'Oréal Paris, and never a single homegrown name. Both findings are real. Neither one is wrong. They simply were not measuring the same market, and the gap between them is the single most useful strategic question in Indonesian FMCG GEO right now, because almost nobody has actually tested it directly.

Two Real Findings That Do Not Agree, and Why That's the Point

The AI beauty recommendation audit, run by Foundation Agency and reported by Cosmetics Business in July 2026, analysed 1,335 brand mentions and 690 source citations across ChatGPT, Gemini, Perplexity and Claude. It drew 44% of citations from editorial earned media and just 7% from brand-owned content, with Reddit as the single most-cited individual source. But the audit ran in English, in a UK context. It never tested a Bahasa Indonesia prompt, and it was never designed to.

The Indonesian e-commerce data, from Magpie IQ, covering March to May 2026, measures something entirely different: actual purchases, not AI-generated recommendations. Local brands leading real Indonesian sales and global brands dominating an English-language AI audit are not contradictory facts. They are two separate measurements of two separate behaviours, and treating one as evidence about the other is exactly the kind of inference that sounds authoritative and is not actually supported by either study.

What the Academic Research Actually Shows

The most rigorous evidence available on this question comes from brand bias research published at EMNLP 2024 by Kamruzzaman and colleagues, examining how large language models associate brands with positive or negative attributes. The finding is nuanced rather than absolute: models trained predominantly on English-language data do show a systematic bias toward global brands in undifferentiated queries, but a documented country-of-origin effect changes the outcome when a prompt specifies a country or cultural context. A query phrased as "rekomendasi produk di Indonesia" measurably increases a model's willingness to surface local brands compared with a generic, context-free version of the same question.

That single finding reframes the entire local-versus-global question from a fixed disadvantage into a solvable content and prompting problem. Global brands retain a default advantage in context-free, English-language queries, largely because they have more entity authority accumulated in the English-language training data these models learn from. Local brands have a real, evidence-backed lever available specifically in Indonesian-context, Indonesian-language queries, one that has nothing to do with advertising budget and everything to do with how much structured, native-language content and entity data exists for a model to draw on.

What that native-language content actually needs to sound like matters as much as the fact that it exists. Real Indonesian search queries frequently code-switch between formal Bahasa Indonesia, English loanwords and casual abbreviations within a single sentence, something like "moisturizer ceramide lokal yang BPOM approved, budget di bawah 150rb." Content written in clean, formal Bahasa Indonesia throughout, or content that is simply an English page run through automated translation, matches neither register and reads unnaturally against how people actually type. Content that mirrors this genuine code-switched pattern is a closer match to the country-of-origin effect's actual mechanism than generic "write in Bahasa Indonesia" advice captures on its own.

Clinical Positioning Beats Advertising Budget, Repeatedly

A second, separate pattern in the global beauty data is worth understanding on its own terms, because it offers local brands a second lever beyond language and context. The eMarketer AI Visibility Index for the first quarter of 2026 analysed more than 5,200 ChatGPT responses across nine personal care and beauty categories and found La Roche-Posay appearing in 81% of facial skincare queries, the single most frequently recommended brand across any category the index tested, with CeraVe appearing in 71%. A related index, the AI Beauty Authority Index from 5W and Everything-PR, found three dermatologist-positioned brands, Drunk Elephant, La Roche-Posay and SkinCeuticals, topping AI search results despite spending considerably less on traditional advertising than the legacy luxury brands they outperformed there.

The mechanism behind this is consistent across both indices: AI systems appear to reward source corroboration, dermatologist content, peer-reviewed research, independent editorial recommendation, more heavily than paid placement or brand scale. A separate 2026 academic study on skincare recommendations found the same underlying dynamic in a more extreme form: well-known brands were recommended 100% of the time when competing products were functionally equivalent on paper, a pattern the researchers termed a "Conditional Monopoly," but that dominance collapsed with as little as a 0.1-star rating advantage for a competitor. Brand size is a weak, brittle defence in this environment. A specific, verifiable quality or clinical claim is not.

Why Perplexity Deserves More Weight in an Indonesian Prompt Set

Any brand testing its own AI visibility needs to weight engines according to actual local usage, not global market share alone. Telkomsel, Indonesia's largest mobile operator, partnered with Perplexity in May 2025 to bundle Perplexity Pro into both consumer and enterprise mobile plans, the first such distribution deal in the country. That single partnership plausibly lifts Perplexity's real Indonesian user base well above what its global market share would predict, and Perplexity's own citation behaviour, leaning more heavily on niche expertise and customer reviews than Gemini or ChatGPT do, makes it a meaningfully different surface to monitor, not an interchangeable stand-in for "AI search" generally.

Indonesia is also building its own large language model, Sahabat-AI, trained specifically on Bahasa Indonesia, Javanese and Sundanese, a joint effort between Indosat Ooredoo Hutchison and GoTo. Its existence is itself a signal worth reading carefully: a country investing in a locally trained model is implicitly acknowledging that global, English-trained models under-serve the language, which is the same structural gap the country-of-origin research identifies from the opposite direction, through model behaviour rather than through infrastructure investment.

Platform Differences Compound the Local-Global Question

Layered on top of the language and context question is a genuine difference in how individual AI engines source their answers. One analysis of 6.8 million AI citations found Gemini drawing roughly 52% of its citations from brand-owned websites, while ChatGPT drew closer to 49% from third-party sites, a meaningfully different trust model between the two. A local brand investing only in its own website content, hoping that alone will carry it across every engine, is optimising for Gemini's apparent preference while leaving ChatGPT's third-party-weighted surface largely uncontested. The practical implication is that "getting cited by AI" is not one strategy. It is closer to three or four related but distinct strategies, one per major engine's actual sourcing behaviour.

How to Actually Test This for Your Own Brand, Instead of Assuming an Answer

Given how thin the direct evidence is for Indonesia specifically, the most useful thing a brand can do is not wait for someone else to publish the definitive study. It is to run a small, disciplined version of that test internally. The methodology does not need to be academic-grade to be useful; it needs to be consistent enough to produce a comparable reading over time.

Build a fixed set of 20 to 30 prompts covering the brand's actual category, phrased the way a real Indonesian consumer would type them, not the way a marketing brief would phrase them. Run each prompt in both Bahasa Indonesia and English, on the same day, across ChatGPT, Perplexity, Gemini and Claude. Record which brands appear, in what order, and whether the specific brand being tracked shows up at all. Repeat the exact same prompt set on a fixed schedule, monthly is a reasonable cadence, because AI answers are documented as inconsistent from one run to the next, and a single test session measures a coincidence as easily as it measures a real pattern.

Two comparisons matter more than the raw presence-or-absence result. First, compare the Bahasa Indonesia results against the English results for the same prompt, since that gap, if one exists, is the country-of-origin effect showing up directly in a brand's own category rather than in someone else's published study. Second, compare results across engines for the same prompt, since a brand appearing reliably in Perplexity's answers but never in Gemini's is learning something specific and actionable about where its structured data and third-party corroboration are and are not working, rather than a single undifferentiated "AI visibility" score that hides which surface is actually the problem.

This is close to the same baseline audit methodology used at the start of a structured GEO engagement, just run informally and repeatedly rather than as a one-time diagnostic. A brand that builds this habit before hiring an agency, or continues it independently between formal audits, arrives at every subsequent strategic conversation with real, brand-specific evidence instead of a general industry statistic that may or may not describe its own category accurately.

The honest summary is less dramatic than either extreme framing would suggest. Global brands do not have an insurmountable structural advantage in Indonesia, because the country-of-origin effect and the clinical-positioning pattern both offer real, evidence-backed counter-levers unrelated to advertising spend. Local brands are not automatically favoured either, because that default advantage only activates when a prompt specifies Indonesian context, and a global competitor writing genuinely native Bahasa Indonesia content can capture the same lever a local brand assumed was exclusively its own. The market has not been tested directly enough, in Bahasa Indonesia, at scale, for anyone to state a confident final answer either way, and any agency claiming certainty here is describing an inference, not a measurement.

What is measurable, and worth doing regardless of which way that open question eventually resolves, is building the specific things both bodies of research point to: native-language structured content, halal and BPOM data as machine-readable schema rather than a badge image, and third-party corroboration from credible, named sources. These are covered in more operational depth in our overview of GEO for FMCG, alongside the broader Citation Share measurement approach needed to track whether any of this is actually moving the needle for a specific brand over time. None of it works without the underlying product and category content foundation described in SEO for FMCG, which both local and global brands need regardless of how the local-versus-global citation question eventually settles.

Frequently Asked Questions

Is there direct proof that AI engines favour global brands over local Indonesian brands?

Not for Indonesia specifically, and not in Bahasa Indonesia. The clearest evidence of a global-brand bias comes from a UK, English-language audit. No study we are aware of has directly tested Bahasa Indonesia prompts for local-versus-global brand bias in Indonesia, which remains a genuine, unanswered question rather than a settled one.

Should a local Indonesian brand write content in English to compete internationally, or focus on Bahasa Indonesia?

For the Indonesian domestic market specifically, native Bahasa Indonesia content is the stronger lever, since it is what activates the country-of-origin effect and reaches the primary local search population. English content matters for international or diaspora audiences, but should be treated as a separate audience and a separate content track, not a substitute for native-language domestic content.

Does halal certification actually help a brand get cited by AI, or is that just a compliance requirement?

It is genuinely both, though the AI citation benefit specifically remains unproven by direct measurement. Halal certification is a hard, verified legal requirement regardless. Structuring the certification data as machine-readable schema, rather than only a packaging badge, is a low-risk step worth taking either way, but should not be oversold internally as a guaranteed citation lever.

Why does the same brand seem to get recommended by one AI engine and not another?

Different engines source their answers differently. One large-scale citation analysis found Gemini favouring brand-owned content more than ChatGPT, which leans more heavily on third-party sites. A brand strong on one surface and weak on the other will see genuinely different recommendation patterns across engines, which is a sourcing difference, not a bug or an inconsistency to be alarmed by.

How should a brand decide how much weight to give Perplexity versus ChatGPT or Gemini in Indonesia?

Weight it higher than its global market share alone would suggest. Perplexity's distribution partnership with Telkomsel gives it a larger real Indonesian user base than global usage statistics capture, which is a market-specific factor most brands do not account for when building their AI monitoring prompt set.

The full evidence base and practical framework for building Indonesian FMCG AI visibility, including the country-of-origin and platform-sourcing dynamics covered in this piece, is explored in more depth in Cited or Silent. The free gated edition is available now, alongside Kindle, Google Play and Apple Books editions live internationally.

Get the Free Ebook

Written by Tessar Napitupulu, Founder and CEO of PT Arfadia Digital Indonesia, Forbes Agency Council member, and author of Found Before They Search and Cited or Silent. Arfadia has positioned itself as Indonesia's GEO pioneer since 2023.