The AI Citation Question Nobody in PH Can Answer

A Filipino asks ChatGPT for the best payroll software for a Cebu retailer. Does the answer cite Philippine sources, or does it reach for US ones?

Nobody has published the answer. Not one agency in Manila. Not one in Singapore. No academic group, no analytics vendor.

That gap is not a curiosity, it is the most consequential unknown in Philippine AI search visibility, because the entire case for local content investment rests on an assumption about it. If engines already reach for Philippine sources when the query is Philippine, building local authority is a strong play. If they default to global English incumbents regardless, then a Philippine brand competes against the whole English-language web on every query. Very different problem. Much harder one.

Four independent research passes commissioned for this article reached the same conclusion separately. The answer is unavailable. What follows is what the adjacent evidence does show, why the extrapolations on offer do not hold here, where citations in this market actually seem to come from, and how to run the test on your own category rather than waiting for somebody else to publish it.

What the adjacent research establishes

Three findings sit near this question without answering it. Each is worth understanding precisely, because each gets quoted as though it did.

Engines shift toward local-language domains when the prompt language shifts. A study covering roughly 400 questions across four languages and four systems in five countries, published July 2026, found local-language prompting moved citations sharply toward local-language domains. In the Japanese portion, citations to .jp domains rose from around one percent under English prompting to twenty-six percent under Japanese prompting. It also found English queries pulled more sources than local-language queries in every market tested.

The markets tested were France, Spain, Argentina, Japan and the United States. Not the Philippines. Applying that mechanism here is inference, and there is a specific reason to be careful: if a large share of Philippine querying happens in English anyway, the local-language-unlocks-local-sources effect may barely fire in this market at all.

ChatGPT frequently searches in English regardless of prompt language. An analysis of over ten million prompts and twenty million background search queries, published February 2026, found 43 percent of ChatGPT's background searches running in English even when the user prompted in another language. 78 percent of non-English sessions included at least one English fan-out query. Filipino and Tagalog were not among the languages tested. Directionally interesting. Not a Philippine measurement, and it should never be presented as one.

Engines differ sharply in which source types they favour. The same July 2026 study found Gemini and Google AI Overviews leaning heavily on YouTube and Reddit, while ChatGPT and Claude favoured brand, editorial and institutional pages. Again, not tested in the Philippines. But it means "optimise for AI" is not one activity, and a programme built for one engine can produce very little in another.

Why the Philippines is a harder case than the studies suggest

There is a structural reason the local-language mechanism may not rescue Philippine brands the way it appears to rescue Japanese or French ones.

In Japan, a Japanese-language query enters a thinner competitive pool. The corpus of Japanese-language commercial content on any given topic is smaller than the English-language corpus, so a well-built Japanese page competes against fewer alternatives. The local-language effect is partly an authority effect and partly, maybe mostly, a scarcity effect.

Philippine business publishing runs largely in English. Proficiency is high and formally measured: the EF English Proficiency Index 2024 places the Philippines 22nd of 116 countries on a score of 570, second in Asia after Singapore. So a Philippine brand publishing in English is eligible for citation, which sounds like an advantage. It also means competing directly against the entire English-language web, including US and UK incumbents with far deeper link and mention profiles, rather than inside a thinner local-language pool.

Publishing in Filipino as a workaround has its own problem. FilBench found the strongest of 27 models scoring 72.23 percent aggregate on Filipino with a 46.48 percent text-generation subscore, and a separate 2025 study found retrieval-augmented performance highest in English, lower in Taglish, lowest in pure Tagalog. Filipino content sits in a thinner pool but gets handled less reliably by the models doing the retrieving.

So the Philippine problem is not "translate to win local citations". It is closer to "build enough English-language local entity authority to be selected over global English incumbents". Harder brief. And a different one from what most GEO advice written for other markets prescribes.

What gets misused as an answer

Claim you may be shown	What it actually measures	Why it does not answer the question
A screenshot of ChatGPT naming the brand	One answer, one session, one moment	No query record, no repeat runs, no engine or model date, and no way to separate a stable result from answer variance
"We rank first, so AI cites us"	Organic position	Ranking and citation are separate observations. A first-position page can go uncited and a third-page result can be cited repeatedly
Local-language citation percentages from Japan or Europe	Source selection in markets with thinner local-language corpora	The Philippines publishes commercially in English, so the scarcity effect driving those numbers may not apply here at all
AI Overview prevalence figures	How often the feature appears, in a given query corpus	Prevalence is not citation. One Philippine figure exists, Ahrefs placing the Philippines at 29.1 percent of keywords across 108 million AI Overview queries, disputed by other research and method-dependent. Global comparators run 15.69 to 60.32 percent
A brand mention inside an answer	That the model named you	Being named is not being cited. A citation attributes information to your source. Both worth tracking, in separate columns
An .ph domain appearing in a result	Domain extension	Many Philippine institutions publish on .com and many .ph domains are operated from abroad. Country of ownership is the variable, not the suffix

That last row quietly breaks most informal attempts at this question. Coding a citation as Philippine because the URL ends in .ph produces a clean-looking dataset that measures domain registration rather than local authority.

Where citations plausibly come from in this market

Even without the national study, the surface mix in the Philippines is unusual enough to shape a programme. Three numbers make the point.

Reddit's potential advertising reach in the Philippines grew by 10.8 million people between the end of 2024 and late 2025, an increase of roughly 250 percent, taking it to about 15.1 million users or 12.9 percent of the population. That is not a mature platform inching upward. That is a platform arriving.

YouTube reaches 85 percent of Philippine internet users aged 16 and over on a monthly basis. Facebook reaches 94.9 percent and accounts for 84.17 percent of social-referral web traffic to third-party sites in the country.

Now put those beside the engine-preference finding. Gemini and Google AI Overviews lean heavily on YouTube and Reddit, the two surfaces where Philippine presence is either enormous or growing fastest. ChatGPT and Claude favour brand, editorial and institutional pages, which is where owned content and digital PR live.

The implication is a dual-layer programme rather than a single content push. An owned layer of extractable, well-structured pages for the engines that prefer brand and editorial sources. An earned layer of genuine community participation and structured video with full transcripts for the engines that prefer community and video. Cover one and you have made a choice about half the market, whether or not you meant to.

One honest caveat, because this is where the argument gets stretched. No study proves that Facebook or TikTok content is cited by AI engines for Philippine queries. The reach figures establish where Filipinos are. They do not establish what gets retrieved. Crawlable, indexable surfaces come first for that reason, with social treated as distribution and community presence treated as a reasonable bet rather than a measured lever.

How to run the experiment on your own category

The national study does not exist and may not for a while. The version scoped to your category is entirely runnable, and more useful to you than the national one would be anyway.

Build a balanced prompt panel. Between forty and eighty prompts covering the intents that matter commercially: informational, comparison, vendor-selection, regulated-topic, local-service, B2B. Write them the way a buyer would actually type, not the way a keyword tool would phrase them.

Write three language variants of each. English, Filipino, naturally code-switched Taglish. Have a Filipino native speaker confirm the three variants actually mean the same thing, because a mechanical translation introduces a difference that will later show up in your results looking like a finding.

Run from Philippine IP addresses, in clean sessions. Personalisation and session history contaminate results in ways that are invisible afterwards. A logged-in account with months of history measures its own history as much as it measures the engine.

Repeat, and record the model date. A single run measures nothing, because generative answers vary between identical prompts. Repeat on a schedule, and log which engine and which model version produced each answer, so a result that changes after a model update reads as information rather than noise.

Classify sources by ownership, with the rule written down first. Decide before you start what makes a source Philippine. Publisher or institution ownership, editorial base, something else. Write the rule, then apply it. Deciding case by case as data arrives is how a coding scheme drifts toward the answer you expected to find.

Log the answer text, the cited URL and the position. Cropped screenshots are not evidence. Full answer capture lets you re-read context later, which matters a great deal when you are trying to work out why a citation disappeared.

Report per engine, never blended. Given how sharply engines differ in source preference, a blended figure hides exactly what you needed. Show ChatGPT, Google AI Overviews, Gemini, Perplexity, Copilot and Claude separately, then discuss the pattern across them.

Why nobody has run it yet

The study is not technically difficult, which is the genuinely interesting part. Philippine IP access, clean sessions, a validated multilingual prompt set, repeat runs, a pre-registered source-coding rule. None of that is exotic.

Two explanations are worth putting side by side, because they are not mutually exclusive.

The first is capability and maturity. Swarm's Philippine AI Report 2025, surveying 175 organisations between October and November 2025, found 92 percent had used AI in some capacity while 65 percent remained stuck at proof-of-concept stage, with 54 percent having used generative AI for more than twelve months. The barriers respondents named are the telling part: talent scarcity at 57 percent, and security and privacy concerns at 40 percent. Broad adoption, shallow production depth, and a shortage of people who can build measurement infrastructure. That is not a market with slack research capacity sitting idle.

The second explanation is incentive, and it is the less comfortable one. Almost everyone positioned to run this study sells a service whose pitch depends on the answer. A study that could contradict your own positioning is an uncomfortable thing to commission, and an easy thing to postpone indefinitely. That applies to us as much as to anyone else, which is why this article states the gap rather than filling it with a number.

What to do while the question stays open

Waiting for a national study is not a strategy. Neither is pretending one exists. Three things hold regardless of how the question eventually resolves.

Build English-language authority with heavy local entity density. Name the Philippine regulator that applies to your category, the specific cities you serve, at least two Philippine publications, the peso context. This is the version of localisation that works whether or not the local-language mechanism fires, because it makes content unambiguously about the Philippines while staying in the language the market publishes in.

Build the dual-layer source footprint described above. Owned extractable pages, plus community and video presence with real transcripts. The engine divergence is documented, even if not documented here, and covering only one side should be a considered choice rather than an accident.

Measure your own citation baseline now, before changing anything. The most common regret in this work is not a wrong tactic. It is having no baseline, which makes every subsequent result unfalsifiable in both directions at once.

Our GEO service for the Philippines starts with exactly that baseline, and pairs with our SEO service for the Philippines because engines still retrieve from crawlable, indexed content. Tessar Napitupulu covers the measurement discipline behind it in Cited or Silent, available as a free gated edition.

---

Frequently Asked Questions

Do AI engines cite Philippine sources or default to US ones?

Nobody has published an answer. No controlled study has run a fixed Philippine prompt set from Philippine IP addresses across the major engines while coding each cited source by country of ownership. Four independent research passes conducted for this article reached that conclusion separately. Any vendor answering this confidently is offering a hypothesis, not a measurement.

Does asking in Filipino make AI more likely to cite Philippine sources?

Untested here, and there is reason for caution. A July 2026 study covering France, Spain, Argentina, Japan and the United States found local-language prompting shifted citations sharply toward local-language domains, with Japanese-domain citations rising from around one percent under English prompting to twenty-six percent under Japanese prompting. The Philippines was not tested. Because Philippine business publishing runs largely in English, the scarcity effect driving those numbers may barely apply here.

Should we publish in Filipino to get cited locally?

Not as the foundation. Filipino content sits in a thinner competitive pool but gets handled less reliably by the models doing the retrieving: FilBench found the strongest of 27 models scoring 72.23 percent aggregate on Filipino with a 46.48 percent text-generation subscore, and a separate 2025 study found retrieval-augmented performance lowest in pure Tagalog. Build the authority layer in English with heavy local entity density, and treat Filipino and Taglish as targeted supplements written by native speakers.

Can we just check whether AI cites Philippine sources by looking for .ph domains?

No, and this is the most common methodological error in informal attempts. Many Philippine institutions and publishers use .com domains, and many .ph domains are operated from outside the country. Domain suffix measures registration, not local authority or editorial base. Define an ownership rule before collecting data and apply it consistently.

Does Reddit matter for AI citation in the Philippines?

Enough to plan around, though not enough to treat as proven. Reddit's potential advertising reach in the Philippines grew by 10.8 million people between the end of 2024 and late 2025, roughly 250 percent, reaching about 15.1 million users or 12.9 percent of the population. Separately, research covering five non-Philippine markets found Gemini and Google AI Overviews drawing heavily on Reddit and YouTube while ChatGPT and Claude favoured brand and editorial pages. Nobody has tested whether that engine preference holds for Philippine queries, so treat community presence as a reasonable bet on a documented pattern rather than a measured lever.

Is a brand mention in an AI answer the same as a citation?

No. A mention means the model named you. A citation means the model attributed information to your source, usually with a link. Both are worth tracking and they belong in separate columns, because a programme can move one substantially without moving the other, and reporting them together hides which happened.

How many prompts does a citation study need?

No standard exists for the Philippine market. As a working design, forty to eighty prompts spread across informational, comparison, vendor-selection, regulated-topic, local-service and B2B intents, each with English, Filipino and Taglish variants, run repeatedly rather than once. The repeat runs matter more than panel size, because generative answers vary between identical prompts and a single run cannot distinguish a real result from variance.

Why has nobody published this study?

Two reasons that reinforce each other. Capacity is one: Swarm's survey of 175 Philippine organisations found 92 percent using AI in some capacity but 65 percent still at proof-of-concept, with talent scarcity named as a barrier by 57 percent and security or privacy concerns by 40 percent. Incentive is the other, and the less comfortable one: almost everyone positioned to run the study sells a service whose pitch depends on the answer, and a study that could contradict your own positioning is easy to postpone indefinitely.

---