We ran 1,250 Bahasa Indonesia consumer questions through Google's latest Gemini Flash model with Search grounding and logged every citation, search query, and brand mention. What decides whether your site can even be cited — and how crowded the answer is once you get there.
The core
What we measured
We executed 1,250 consumer prompts in Bahasa Indonesia — 250 each across Electronics, Fashion, Beauty, Travel, and Food — against Google's latest Gemini Flash model with Google Search grounding enabled, and captured the full answer, every citation, and every search query the model issued. Only 427 of the 1,250 prompts (34.2%) triggered a live search at all. Search intent, not topic, decides who gets in: local queries ground at 88.6% and commercial product-research queries at 69.1%, while informational "what/why/how" questions ground at just 21.1%. When Gemini does search, it cites 7.2 sources per answer drawn from a pool of 1,089 domains — 71.5% of which were cited exactly once.
Our verdict
AI visibility in Indonesian consumer search is two separate games. Game one is getting retrieved: it only exists for local and commercial queries, and it rewards query-shaped content — price-bracketed, year-stamped, city-named — because that is literally what Gemini types into Google. Game two is being remembered: 47.7% of all answers name at least one brand, and in Travel and Electronics the model recommends brands just as densely without searching as with it. Content optimization wins game one. Only sustained brand presence wins game two — and most brands are playing neither deliberately.
The key numbers
1,250
Consumer prompts executed
250 × 5 categories, Bahasa Indonesia
34.2%
Triggered live Google Search
427 of 1,250 prompts
3,089
Citations captured
Across 1,089 unique domains
≤7%
Max citation share of any domain
No winner-take-all in any category
253
Brands & platforms named in answers
5,651 mentions · 47.7% of answers
998
Search queries Gemini issued
The exact strings typed into Google
Transparency
This is a primary-data study, not a synthesis. Every number below was computed from a single logged dataset, and the two analysis scripts that produced them are kept alongside the raw JSON so the results can be regenerated.
Data collection
On 22–23 July 2026 we executed 1,250 consumer prompts in Bahasa Indonesia against Google's latest Gemini Flash model with the Google Search grounding tool enabled, via the google-genai SDK — the same client, grounding tool, and generation config Intura's platform uses for its brand-visibility experiments. Each record logs the prompt, category, search-intent label, full answer text, every grounding citation (domain + redirect URI), and the exact web-search queries Gemini issued.
Model naming, transparently
The run used Google's rolling "latest Flash" alias, which Google resolves server-side to the newest production Gemini Flash model. We report it as such rather than pinning a version number we did not verify at runtime. Results are a July 2026 snapshot of that model's grounding behavior — they describe Gemini Flash with Search grounding, not AI search in general.
Scope
The study covers five consumer categories — Electronics, Fashion, Beauty, Travel, and Food — with 250 prompts each, n = 1,250. At 250 prompts per category, any within-category percentage carries a worst-case margin of error of ±6.2 points at 95% confidence; pooled figures carry ±2.6 points.
Statistics
Proportions are reported with Wilson 95% confidence intervals. Group differences use Kruskal–Wallis and Mann–Whitney U (the metrics are right-skewed; medians are reported). Categorical associations use Pearson chi-square with Cramér's V effect sizes and standardized residuals. Correlations are Spearman rank. Citation counts are zero-inflated — 65.8% of prompts produce zero citations — so grounded and ungrounded answers are modeled separately rather than deleting "outliers" from a count variable. Significance threshold α = 0.05; robustness checks (outlier removal, winsorization) left every conclusion unchanged.
Entity-mention analysis
Brand and platform mentions were counted inside the answer texts using a curated lexicon of 344 entities (219 consumer brands, 34 marketplaces/OTAs/social platforms plus aliases), built from capitalized-token frequency lists of the corpus itself plus known Indonesian consumer brands, then hand-filtered. Sub-brands roll up to the parent (Redmi → Xiaomi, Galaxy → Samsung, ThinkPad → Lenovo). Matching is case-sensitive with word boundaries, and names that collide with common Indonesian words or feature names (Mango, Coach, Aqua, Guardian) were excluded after context review — we chose undercounting over false positives. 253 of the 344 entities actually appear in the answers.
Limitations — read before quoting
The first gate
The single most important structural fact in the data: two-thirds of consumer questions are answered purely from the model's memory, with zero citations — meaning no website gets any visibility for those queries. And the gate swings very differently by category. Travel prompts trigger a live search 50.8% of the time, twice the rate of Fashion (25.6%) or Beauty (26.0%) — plausibly because travel answers depend on current prices, schedules, and visa rules, while fashion and beauty advice is mostly evergreen. These gaps are far too large to be a sampling fluke. One nuance: once a search fires, its depth is similar everywhere — 6.5 to 8.2 sources per answer.
| Category | Grounded | Grounding rate | Wilson 95% CI | Sources when grounded |
|---|---|---|---|---|
| Travel | 127/250 | 50.8% | 44.6–56.9% | 7.19 |
| Electronics | 93/250 | 37.2% | 31.4–43.3% | 7.07 |
| Food | 78/250 | 31.2% | 25.8–37.2% | 6.55 |
| Beauty | 65/250 | 26.0% | 21.0–31.8% | 8.17 |
| Fashion | 64/250 | 25.6% | 20.6–31.4% | 7.45 |
χ²(4) = 48.33 · p = 8.1×10⁻¹⁰ · Cramér's V = 0.20
What it means
In Fashion and Beauty, roughly 3 in 4 consumer questions never touch the live web. Content optimization only moves the needle on the grounded minority — the intent table below shows exactly which queries those are.
The real driver
Search intent predicts whether Gemini will search far better than the topic does — knowing why someone is asking tells you more than knowing what they are asking about. Local "near me / in this city" queries hit the live web 88.6% of the time and commercial product-research queries 69.1% — but informational "what/why/how" questions, which make up 75% of the prompt set, ground at just 21.1%. Fashion and Beauty's low grounding rates are largely their prompt mix skewing informational, not something special about the categories.
χ²(4) = 301.4 · p = 5.4×10⁻⁶⁴ · Cramér's V = 0.49
What it means
Local and commercial queries are where being citable matters — they almost always hit the live web. Traditional informational SEO content has limited AI-citation upside in this segment: Gemini answers those from memory 4 times out of 5.
The citation map
We classified all 3,089 citations into a two-level site taxonomy using ~250 hand-built domain rules. News and media outlets take 27.6% of citations and e-commerce 15.5% — but the biggest bucket is the long-tail: 36.4% of citations go to roughly 1,000 domains that each hold under 0.3% share. Which site type wins depends heavily on the category. Beauty leans on health portals like Halodoc and Alodokter far more than any other category — Gemini treats skincare questions as quasi-medical. Food skews to national news and UGC, Travel scatters across small destination blogs, and Fashion cites official brand sites strikingly rarely.
National news · lifestyle media · tech review media
Marketplaces · travel OTAs · specialty retail
YouTube · Lemon8 · forums, blogs & Q&A
Manufacturer sites · government · operators
Health portals · travel guides · encyclopedias
Each under 0.3% of citations
χ²(20) = 286.6 · p = 4.4×10⁻⁴⁹ · Cramér's V = 0.15
Even the #1 domain in any category captures at most about 7% of that category's citations. Concentration is real — the top 10% of domains take 58.4% of all citations — but 71.5% of the 1,089 cited domains appear exactly once, and the ten most-cited domains together hold only 23.8%. Grounding rewards query-specific relevance over raw domain authority: niche pages genuinely break through.
Travel
Electronics
Beauty
Fashion
Food
Concentration
Entity-density analysis
Citations only measure who gets linked. The answers themselves also name brands — and that is arguably the metric that moves buying decisions. Counting mentions against our 344-entity lexicon: 47.7% of all 1,250 answers name at least one brand or platform, 253 distinct entities appear 5,651 times, and an answer that names anything names 3.6 entities on average. The crowdedness varies wildly. Electronics is a brand brawl — 91.6% of answers name at least one brand and the average answer carries 4.2 unique entities (up to 14 in one answer). Food is nearly brand-empty: 27.2% of answers, 0.5 entities each, and the top "brands" are actually TikTok and Instagram.
| Category | Unique entities | Answers naming ≥1 | Avg per answer | Avg per mentioning answer | Max in one answer |
|---|---|---|---|---|---|
| Electronics | 93 | 91.6% | 4.18 | 4.56 | 14 |
| Travel | 67 | 49.6% | 1.63 | 3.29 | 12 |
| Beauty | 82 | 36.4% | 1.24 | 3.42 | 11 |
| Fashion | 62 | 33.6% | 0.98 | 2.92 | 9 |
| Food | 51 | 27.2% | 0.52 | 1.93 | 8 |
Curated 344-entity lexicon · 253 appear (219 brands, 34 platforms) · 5,651 mentions · Gini 0.72
Electronics
Travel
Beauty
Fashion
Food
Entity density by search intent
Does searching the web change how many brands Gemini names? Only sometimes — and this is the most strategically loaded finding in the report. In Beauty, an answer built from a live search names almost three times more brands than one written from memory (2.35 vs 0.85 per answer). Fashion and Food show a real but smaller lift. In Travel and Electronics the difference disappears: Gemini names Traveloka, Samsung, and Apple just as freely without ever touching the web.
Average number of distinct brands named per answer — with vs without web search
| Category | With search | Without search | Mann–Whitney |
|---|---|---|---|
| Beauty | 2.35 | 0.85 | p = 9.5×10⁻¹² · significant |
| Fashion | 1.41 | 0.83 | p = 0.001 · significant |
| Food | 0.72 | 0.44 | p = 0.027 · significant |
| Electronics | 4.53 | 3.97 | p = 0.065 · not significant |
| Travel | 1.49 | 1.78 | p = 0.60 · not significant |
| Overall (all 5 categories) | 2.13 | 1.49 | p = 9.2×10⁻¹² · significant |
How to read this
Read the table as a map of where content can still change the outcome. Where the "with search" column is clearly higher — Beauty, Fashion, Food — brand mentions flow through retrieved pages: the brands Gemini reads about are the brands it names, so citable content directly buys mentions. Where the two columns are level — Travel, Electronics — the recommendation list was decided long before your content loads. It lives in the model's memory, built from years of web presence, and no landing-page rewrite will change it this quarter.
Recommended actions
| Category | Interpretation | Action item |
|---|---|---|
| Beauty | Web search nearly triples the brands named (2.35 vs 0.85 per answer) — the strongest lift in the study. Gemini leans heavily on fresh web content for skincare and beauty recommendations. | Update content frequently (weekly to monthly), keep ingredient and review pages fresh on the health portals and communities Gemini cites, and monitor competitor content. |
| Fashion | Search lifts brand mentions noticeably (1.41 vs 0.83). Trends flow into answers through retrieved pages — what lifestyle media and marketplaces publish this season is what Gemini names. | Publish around seasonal collections, new products, and trend roundups ("rekomendasi" content). Refresh existing pages regularly instead of letting them age. |
| Food | Search improves brand presence, but the lift is smaller (0.72 vs 0.44) and mentions are rare overall — answers revolve around recipes and places more than brands. | Update menus, recipes, and seasonal content periodically — and move fast on viral moments, because "viral" is what triggers search in this category. |
| Electronics | No meaningful lift (4.53 vs 3.97). Gemini already knows the established brands and product lines from its training data — answers are brand-dense either way. | Prioritize evergreen product documentation, comparisons, FAQs, and schema markup. Update around major launches and spec changes rather than on a fixed schedule. |
| Travel | No meaningful difference (1.49 vs 1.78) — answers name the same platforms with or without search. The Traveloka-and-friends slate lives in the model's memory. | Focus on authoritative destination, price, and regulation resources. Update when prices, rules, or attractions change — and invest in long-term brand presence that future model versions absorb. |
| Overall | Across all 1,250 answers, search lifts brand mentions by roughly 40% (2.13 vs 1.49 per answer) — but that average hides two different games: search-driven categories and memory-driven ones. | Aim content at commercial and local intents (69–89% search rate), and run a recurring prompt panel measuring the share of answers that name you versus competitors — with and without search. That number tells you which game you are winning. |
✓ = brand mentions are search-driven (fresh content pays off directly) · ⚠ = memory-driven (build long-term presence).
What it means
For a challenger brand the math is concrete: appearing in a commercial-intent Electronics answer means sharing it with ~3.6 other brands, while a Beauty answer that names you shares you with ~2.4 others. And local brands do win here — in Fashion, Ventela, Aerostreet, and Patrobas out-mention Nike and Adidas; in Beauty, Somethinc and Azarine lead while global names trail.
The retrieval spec
When Gemini grounds, it rewrites the user's question into Google queries — 998 of them in this dataset, and we logged every string. The rewrite grammar is remarkably formulaic. 43.9% of queries carry a location name. Electronics queries stack price anchor + year + "terbaik" ("rekomendasi HP gaming 3 jutaan terbaik 2025 2026") — and Gemini appends both the current and the next year, a freshness heuristic worth mirroring in titles. Food over-indexes on viral/trending hooks (18.3% of its queries), matching its citation skew toward Lemon8 and YouTube. Beauty is the least templated and leans on review/comparison hooks — consistent with Gemini treating skincare as quasi-medical.
| Search pattern | All queries | Strongest in |
|---|---|---|
| Location (Jakarta, Bandung, Bali, "indonesia"…) | 43.9% | Electronics 55.6% · Food 52.2% · Travel 51.1% |
| "harga" (price) | 15.6% | Electronics 34.7% |
| Price anchor ("3 jutaan", "500 ribuan") | 13.9% | Electronics 28.9% |
| Year freshness (2024/2025/2026) | 10.8% | Electronics 17.8% · Fashion 15.0% |
| "rekomendasi" | 8.4% | Fashion 16.2% |
| "murah" (cheap) | 8.2% | Travel 11.2% · Food 10.2% |
| viral / trending / kekinian | 6.6% | Food 18.3% |
| "terbaik" (best) | 5.9% | Electronics 12.4% |
| review / vs / perbandingan | 3.8% | Beauty 6.0% |
What it means
If your content's title reads like the query Gemini types — price-bracketed, dual-year-stamped listicles for Electronics; "[thing] + [city] + murah" for Travel and Food; "rekomendasi" trend roundups for Fashion — you are aligned with the retrieval pattern. That is the practical core of GEO for this model.
What to act on
Two-thirds of consumer prompts produce zero citations
AI-visibility competition happens only inside the grounded 34.2% — and intent, not topic, decides whether a query is in that pool. Local (88.6%) and commercial (69.1%) queries are the battleground; informational queries mostly are not (21.1%).
No domain dominates — the long-tail is real
The top domain in any category holds ≤7% of its citations; 71.5% of the 1,089 cited domains appear exactly once. Gemini's grounding rewards query-specific relevance over domain authority — niche sites genuinely compete.
Each category has different gatekeepers
Health portals guard Beauty (Halodoc, Alodokter), tech media and marketplaces guard Electronics, OTAs and government sites guard Travel (Traveloka, imigrasi.go.id), news/UGC guard Food, and marketplaces plus lifestyle media guard Fashion. Pitch your content — or your PR — where your category's citations actually come from.
Brand mentions are a separate game from citations
47.7% of answers name at least one of 253 brands/platforms — and in Travel and Electronics, Gemini names them just as densely without searching. A large share of AI brand visibility is baked into the model's memory, where content tweaks cannot reach it. Beauty is the exception: grounding nearly triples brand density (2.35 vs 0.85 per answer).
Write titles in Gemini's query grammar
The 998 logged search queries are a retrieval spec: price-bracket + dual-year listicles for Electronics ("… 3 jutaan terbaik 2025 2026"), location + "murah" for Travel and Food, "rekomendasi" roundups for Fashion, clinical-adjacent authority content for Beauty. Match the grammar and you match the retrieval.
For the rigorous
| Test | Statistic | p-value | Effect size |
|---|---|---|---|
| Grounding × category (chi-square, df = 4) | χ² = 48.33 | 8.1×10⁻¹⁰ | V = 0.20 |
| Grounding × search intent (chi-square, df = 4) | χ² = 301.40 | 5.4×10⁻⁶⁴ | V = 0.49 |
| Category × cited site type (chi-square, df = 20) | χ² = 286.60 | 4.4×10⁻⁴⁹ | V = 0.15 |
| Sources per grounded answer across categories (Kruskal–Wallis) | H = 13.71 | 0.008 | — |
| Answer length, grounded vs ungrounded (Mann–Whitney) | medians 3,235 vs 3,652 chars | 4.0×10⁻³⁰ | — |
| Entity density, grounded vs ungrounded — Beauty (Mann–Whitney) | 2.35 vs 0.85 | 9.5×10⁻¹² | — |
| Entity density, grounded vs ungrounded — Travel (Mann–Whitney) | 1.49 vs 1.78 | 0.60 (n.s.) | — |
| Entity density, grounded vs ungrounded — Overall (Mann–Whitney) | 2.13 vs 1.49 | 9.2×10⁻¹² | — |
All within-category proportions carry a margin of error ≤ ±6.2 pp (n = 250, 95% confidence); pooled figures ±2.6 pp. Wilson intervals throughout. Removing Tukey outliers and winsorizing answer length at p99 left every conclusion unchanged.
The raw data itself
Five records from the dataset, one per category, quoted verbatim (answers truncated). They show the three behaviors the statistics describe: query rewriting on commercial prompts, deep multi-query research on local prompts, and fully-formed brand recommendations produced with zero web search.
Prompt (verbatim)
“Mau beli laptop baru buat anak SMA, budget maksimal 6 juta enaknya merek dan tipe apa yang speknya oke buat belajar?”
Search queries Gemini issued
laptop 5 6 juta terbaik 2024 2025 indonesialaptop harga 5-6 juta untuk sekolah SMA terbaikDomains cited
advandigital.com · youtube.com · els.id · tirto.id · jba.co.id · kanakomputer.com · carisinyal.com · kumparan.com
Answer (truncated)
Untuk budget maksimal Rp6 juta, Bapak/Ibu sudah berada di sweet spot (titik paling pas). Di rentang harga Rp5–6 juta, laptop sudah sangat nyaman untuk anak SMA: tidak lemot, responsif untuk multitasking, Zoom/Google Meet, mengetik tugas, presentasi, hingga editing foto/video ringan. […] Patokan Spesifikasi Minimum (Wajib Ada): 1. Prosesor: Minimal Intel Core i3 (Gen 12/13) atau AMD Ryzen 3 (Seri 7000). Hindari Intel Celeron/Pentium…
Why this record matters
Textbook commercial grounding: Gemini converts a conversational budget question into two price-bracketed, year-stamped Google queries — including both "2024 2025" — then assembles an 8-source answer that names Intel, AMD, and specific model lines.
Prompt (verbatim)
“Moisturizer Skintific yang 5x ceramide itu beneran bagus buat skin barrier yang rusak gak sih, tolong review jujur dong.”
Answer (truncated)
Jawab jujur: Ya, moisturizer Skintific 5x Ceramide ini memang BAGUS dan efektif untuk memperbaiki skin barrier yang rusak, terutama untuk kerusakan tingkat ringan hingga sedang. Formulasinya memang dirancang khusus untuk fungsi barrier repair. […] Berikut adalah review jujur dan objektif dari segi kandungan, kelebihan, hingga kekurangannya…
Why this record matters
The most consequential behavior in the dataset: a direct product-review request answered with zero web search. Gemini "reviews" Skintific's moisturizer — verdict included — entirely from parametric memory. No review site, no brand page, no publisher got a citation for this purchase-intent query.
Prompt (verbatim)
“Hotel bintang 4 di Bandung yang dekat stasiun kereta tapi ada kolam renang buat anak-anak ada rekomendasi, budgetnya maksimal 700 ribu per malam.”
Search queries Gemini issued
vue palace artotel curated kolam renang anakarion suites hotel bandung bintang 4 kolam renang anak hargahotel bintang 4 dekat stasiun bandung under 700 ribuaston braga hotel bandung kolam renang anak hargael hotel bandung kolam renang anak hargahotel bintang 4 dekat stasiun bandung ada kolam renang anakDomains cited
traveloka.com · skyscanner.co.id · agoda.com · hotels.com · trivago.co.id · booking.com · kompas.com · expedia.co.id · youtube.com · trip.com · disway.id · tiket.com · detik.com · el-hotels.com
Answer (truncated)
Berikut adalah beberapa rekomendasi hotel bintang 4 di Bandung yang lokasinya dekat dengan Stasiun Kereta Api Bandung, memiliki kolam renang anak, dan harganya berkisar di bawah/maksimal Rp 700.000 per malam… 1. Arion Suites Hotel Bandung — Jarak ke Stasiun: Sangat dekat (hanya ± 200–350 meter / 3–5 menit jalan kaki)…
Why this record matters
Deep-research behavior on a local query: 6 search queries, 14 sources — near the dataset maximum of 15. Gemini first researches specific hotel candidates by name, then verifies each one's kids-pool and price. OTAs (Traveloka, Agoda, Booking.com) dominate the citation slate.
Prompt (verbatim)
“Gimana ya cara memilih dan mengolah udang yang benar biar nanti pas digoreng jadi udang balon hasilnya bisa garing dan mengembang sempurna kayak yang viral itu?”
Search queries Gemini issued
resep udang balon viralcara membuat udang balon garing mengembangDomains cited
youtube.com · idntimes.com · liputan6.com · haibunda.com
Answer (truncated)
Membuat udang balon yang viral—dengan tampilan bulat melembung, kriuk di luar, dan juicy di dalam—memang membutuhkan kombinasi teknik pemilihan udang, racikan tepung, dan teknik menggoreng yang pas. 1. Cara Memilih Udang yang Tepat — Ukuran: Gunakan udang ukuran sedang hingga agak besar (seperti udang vaname atau udang jerbung)…
Why this record matters
One of the informational exceptions that does ground — because the trigger word is "viral". Gemini's own queries keep the hook ("resep udang balon viral") and the citations land on YouTube, IDN Times, and Liputan6: exactly the news/UGC skew the site-type analysis predicts for Food.
Prompt (verbatim)
“Aku lagi cari handmade perhiasan atau aksesoris lokal yang unik buat kado ulang tahun teman, coba kasih rekomendasi beberapa brand yang lagi hits di media sosial dong.”
Search queries Gemini issued
rekomendasi brand perhiasan aksesoris lokal unik handmadebrand aksesoris lokal handmade indonesia hits instagram tiktokDomains cited
femaledaily.com · lemon8-app.com · ekrut.com · loulourose.co · rukita.co · scmedia.id · cxomedia.id · asya-azalea.com
Answer (truncated)
Berikut adalah beberapa rekomendasi brand perhiasan dan aksesoris lokal handmade yang unik, berkualitas, serta sedang hits di media sosial (Instagram & TikTok)… 1. Gaya Playful, Quirky & Colorful (Trend Y2K / Pop Art) — Segi Jewelry (@segijewelry): Terkenal dengan kreasi perhiasan dari glass beads (manik-manik kaca) yang dipadukan dengan batu alam…
Why this record matters
The long-tail breakthrough in action: alongside Female Daily and Lemon8, Gemini cites two small brand-owned domains — loulourose.co and asya-azalea.com — that appear nowhere else in the 3,089 citations. Query-relevant niche pages beat domain authority; this is why 71.5% of cited domains appear exactly once.
Dataset, analysis scripts, and the entity lexicon are archived with this report. Statistics computed with Wilson intervals, chi-square/Cramér's V, Kruskal–Wallis, Mann–Whitney U, and Spearman rank correlations; α = 0.05.