How the Index works

The Thailand AI Visibility Index records which brands ChatGPT and Google AI Mode name in response to unbranded category questions in Thai and English. Each category has a fixed prompt set and a fixed brand panel. Every prompt is run three times on each engine before a brand table is published.

A mention rate is the percentage of valid answers in which a brand appears. We show it separately for each engine and language, together with the overall rate, the range across the three runs, and the number of valid answers behind the edition. DLVX publishes the Index. DLVX is not ranked, and its own score is not published.

We collect answers from ChatGPT and Google AI Mode. The collection route, model or product label returned by the provider, timestamps, and raw responses are retained in the edition log. Where the provider supplies no model label, the log says so rather than guessing one.

ChatGPT is measured with web search enabled. Google AI Mode is measured as its own answer surface, not inferred from ordinary Google rankings. A cited link and a brand mention are different events: a link can appear without the brand being named in the answer.

Prompts are written as buyer questions about a category, not as requests to evaluate a particular brand. They do not name any panel brand and do not instruct the engine which brands to include or exclude. Each set covers comparison, selection, location, price or practical use where those intents fit the category.

The full prompt set is published with the category edition. Changes to wording, prompt count, language, location, engine, or collection route create a new edition. We do not compare a new edition with an older one as though the instrument were unchanged.

Every category uses the same question intent in natural Thai and English. The Thai prompts are written for Thai readers rather than translated word for word. Each prompt is issued with Thailand as the location setting and its language recorded.

Location and language are measurement settings, not proof that every answer is identical to what every person in Thailand would see. Product version, account state, personalisation, time, and the engine's own retrieval choices can still change an answer.

We run the complete set three times on separate collection runs. A run that fails, is blocked, returns an error page, or contains no usable answer is marked invalid and excluded from the denominator. Invalid and missing responses remain visible in the log.

A category table is published only after all three runs are saved and independently checked. If a brand's mention rate varies by more than 30 points across runs, we show that range. If more than one third of the panel crosses that threshold, we publish category-level findings only and do not rank the brands.

A mention counts only when the brand's official name or a reviewed alias appears in the visible answer text. We maintain English and Thai aliases for each brand. Matching ignores letter case where the language has case, and uses the complete name or approved alias rather than a generic word fragment.

Images, alt text, URLs, source lists, and citations do not count as mentions. A domain cited without the brand in the answer text is recorded separately as a citation, not upgraded to a mention.

Common-word names and ambiguous aliases receive human review. If a match could refer to another company, place, or ordinary phrase, it is not counted until a reviewer confirms the intended entity. Corrections to an alias are logged with the reason. Property and group matching run independently, so a property mention can count in both hotel tables.

A mention rate of 60 means the brand appeared in 60 percent of the valid answers measured for that engine, language, category, and edition. The overall rate uses all valid answers across the included engines and languages.

The table is ordered from higher to lower overall mention rate. Brands with the same rate share the same position in the table. Our ordering is not the order in which brands appeared inside an AI answer. Brands with zero in every run are grouped at the bottom as not named in any answer.

The score is not a brand quality rating, a recommendation, an endorsement, or a measure of customer satisfaction. It does not prove market share, revenue, enquiries, bookings, accuracy, sentiment, or the quality of a brand's website or marketing support.

A low score does not prove that a brand is weak. A high score does not prove that a brand is best. Both describe only the answers collected under the published conditions.

AI answers vary between runs. Research by SparkToro and Gumshoe found that recommendation lists and their ordering changed frequently, while repeated appearances across runs still provided a useful visibility measure. We therefore report rates and ranges, never a single answer as a verdict.

The panel is curated, not a census of every brand in Thailand. Results depend on the prompts, location, language, engines, and collection route. A different prompt set can produce a different table. We do not extrapolate the results to all AI products or to every search a customer might make.

Third-party collection can differ from a person's signed-in experience. We record that limitation in the edition log. We do not treat blocked pages, CAPTCHA screens, or provider errors as brand absence.

Categories are refreshed monthly where the measurement route remains available. Each edition carries its collection dates and stays in the archive with its prompt set and summary counts. Material changes to an engine, prompt set, or matching rule trigger a new edition rather than a silent rewrite.

A brand can request a correction by identifying the answer, brand name, and alias involved. We check the retained response before changing a score. If we made an error, we correct the entry, mark the correction date, and explain what changed. A proven scoring error pauses the next category until the cause is fixed.

Every run contains 40 prompts, 20 Thai and 20 English, across both engines.

These prompts were used in all three runs. No panel brand is named in a prompt.