How ChatGPT generates brand recommendations
When a user asks ChatGPT something like "What's a good digital agency in Jakarta?" or "Recommend an AI marketing agency in Indonesia," ChatGPT does not pull from a single ranked list. It synthesizes an answer from two sources:
- Training data — the snapshot of the web ChatGPT was trained on. This is months to years out of date and weights heavily toward sources that appeared often and authoritatively in the training corpus.
- Live retrieval — when ChatGPT's search is enabled, it queries Bing's index in real time and reads candidate sources via GPTBot. This is the surface where current GEO work has the most leverage.
The brand it names is the brand that appears credibly in both layers. That's why GEO is not a single tactic — it's a coordinated push across entity, content, and citation work.
The four levers that move ChatGPT citation rates
1. Entity authority
ChatGPT needs to be confident the brand exists as a coherent entity. The fastest signals: a Wikidata entry, a claimed Google Business Profile with a Knowledge Panel, a Crunchbase profile, and a complete LinkedIn Company page. Indonesian brands that skip this layer are invisible to ChatGPT for entity-level recommendations.
2. Third-party citations
ChatGPT pattern-matches authoritative sources. For Indonesia, that means appearing in DailySocial, Marketeers.com, SWA, Tech in Asia, Detik, Kompas, and category-specific publications. Even more important: "Top X in Indonesia" listicles. ChatGPT treats these as ready-made recommendation lists. Getting included in one Indonesian listicle for your category often moves the needle more than ten of your own blog posts.
3. On-site structured content
ChatGPT extracts content most easily from:
- FAQ schema — Q&A pairs marked with FAQPage JSON-LD get pulled almost verbatim.
- Definition lists —
<dl><dt>What is X?</dt><dd>X is...</dd></dl>is one of the highest-extractability formats in 2026. - llms.txt and llms-full.txt — emerging standards that let you serve AI-readable summaries of your site directly to language models.
- Comparison tables — when AI is asked "X vs Y", a clear comparison table on either party's site is often quoted directly.
4. Bahasa Indonesia coverage
ChatGPT returns different answers when prompted in Bahasa Indonesia vs English. Indonesian brands that exist only in English-language sources are invisible for prompts like "agensi digital terbaik di Jakarta". Bilingual coverage — your About page, schema descriptions, and third-party mentions in both languages — doubles your prompt surface.
What an Indonesian ChatGPT optimization sprint looks like
A realistic 90-day sprint for a mid-market Indonesian brand:
- Weeks 1–2: Audit. Run 60–150 representative prompts through ChatGPT (in both English and Bahasa) from Indonesian IPs. Log every mention, citation, and competitor reference.
- Weeks 3–4: Entity foundation. Wikidata entry, Knowledge Panel claim, Crunchbase, LinkedIn Specialties, schema cleanup, llms.txt deployment.
- Weeks 5–8: Content. FAQ schema across key pages, definition lists for category terms, comparison tables for known competitor queries, in-language answer blocks.
- Weeks 9–12: Outreach. Pitch inclusion in 5–10 Indonesian listicles for your category. Guest posts on DailySocial / Marketeers / category publications.
Re-audit at week 12. Most brands see measurable lift in citation rate and sentiment by then; the compounding gains build over the following 6–12 months.