Thai AI-Citation Index — where AI actually gets its answers when Thai consumers ask what to buy

Three findings

1. The two engines reward completely different assets

Google AI Overview drew 39% of its citations from platforms and community sites — facebook.com alone accounted for 43 citations, with pantip.com (Thailand's largest forum) next at 11.

ChatGPT inverted that. Platforms were just 8%; 92% of its citations went to businesses', media outlets' and institutions' own websites — including small independent clinics, not only large brands.

2. Citations are roughly 80% volatile — the most important number here

We re-ran the identical query set about 13 minutes later, changing nothing.

About three in four citations changed inside an hour. Anyone showing you a single ChatGPT screenshot as proof that a brand "is winning AI search" is showing you one draw from a random process. We publish this number even though it weakens our own measurement, because withholding it would mislead the reader — and because any AI-visibility report that does not disclose its sampling should be treated as marketing.

3. Small businesses do get cited — on the right engine

ChatGPT's results included independent clinics and local operators cited directly by domain, not only national brands. On Google AI Overview, smaller businesses appeared mostly through their Facebook pages rather than their websites.

Most-cited domains — Google AI Overview

Most-cited domains — ChatGPT

What this means if you sell in Thailand

This dataset shows what AI cites. It does not show what causes AI to cite you — those are different claims and we are not going to convert correlation into causation. Stated carefully:

• If your buyers use Google, your Facebook page and the forum threads that mention you may carry as much weight as your website, in the picture we measured.
• If your buyers use ChatGPT, your own website is the asset — almost all of its citations went there.
• Do not trust single screenshots, ours included. Look at trends from repeated measurement using an identical method.
• Start with the free check. Open your robots.txt and look for GPTBot, ClaudeBot, PerplexityBot and Google-Extended. If any is disallowed, those systems cannot quote you at all — this is the one lever the AI vendors document themselves.

Method (repeatable)

• Query set: 48 Thai-language queries with commercial or pre-purchase intent — 8 industries × 6 queries (dental & aesthetics, hotels & accommodation, restaurants & cafés, automotive, real estate, professional services, health & fitness, home services).
• Geographic spread: Bangkok 8 · Pattaya 8 · Chiang Mai 8 · Phuket 7 · Khon Kaen 7 · national 10.
• Engines: Google AI Overview and ChatGPT search, collected via the Apify google-search-scraper with countryCode=th and languageCode=th.
• Runs: 2 per query per engine, about 13 minutes apart, on 22 August 2026.
• "Cited" means: the domain appeared as a source in the generated answer. Infrastructure domains (engine assets, map tiles, font hosts) are excluded. Results are reported split between platforms/communities (Facebook, Pantip, Instagram, Wongnai, OTAs and similar) and own websites.
• The query set was independently reviewed before collection to remove leading questions, questions that explicitly ask the AI to name websites (which mechanically inflates citation counts), and questions requiring a real user location that a headless run cannot resolve.

Limitations — read before citing these numbers

• High volatility. As shown above, run-to-run stability was 24% and 21%. Two runs are not enough to produce stable averages; treat every figure as indicative.
• Single point in time. One collection day. Vendors change models, triggering logic and citation UI continuously.
• Collection conditions differ from real users. Datacenter/proxy IPs, no account, no search history, no real location. Real users may see different answers.
• Thai language only. English-language demand from expats and international buyers — significant in hotels, Phuket and Pattaya property, and dental tourism — is not covered.
• Small sample. Six queries per industry is a sliver of the real query universe, chosen by editorial judgement against stated criteria rather than drawn from measured search-volume data.
• City weighting is editorial. Five provinces are weighted roughly equally, though Bangkok's real query volume is far larger, and the publisher is based in Pattaya.
• Some industries are aggregator-dominated by nature. Hotels skew to OTAs, restaurants to review sites. That is a property of those categories, not a bias in the measurement.
• Citation is not quality. It means the engine linked to a domain — nothing about whether that business is good, its information accurate, or its prices fair. It also says nothing about traffic or sales.