An AI assistant does not recommend a Thai store because the page sounds confident. It looks for small public facts that make the recommendation less risky: who sells, what they sell, where the proof sits, and what the buyer can do next.
The first thing I look for on a Thai commerce site is usually not the hero line. It is the dull sentence near the bottom of the page, the one that says whether the company manufactures, distributes, packs, exports, or only sells retail. On one Bangkok packaging site, a composite scenario from several observations, that sentence did not exist. The homepage showed boxes for cosmetics, food gifts, and hotel amenities. The English catalogue PDF had dozens of product photographs. The sales team could handle B2B orders and export paperwork. Yet an AI assistant described the company as “a packaging supplier marketplace” in one run and “a Thai box printing service” in another.
Both answers were half-right in the annoying way that creates commercial damage. The business did make custom packaging. It was not a marketplace. It was not only a printer. It had staff, production coordination, order handling, and export experience. But the own-domain pages made the model assemble the company from scattered nouns: box, cosmetic, food, hotel, catalogue, print. The evidence had no spine.
Recommendation starts with risk, not enthusiasm
When a human recommends a store, they often draw on memory, taste, trust, and social cues. An AI assistant has a thinner job. It generates an answer from patterns in available information, and when asked to recommend a business, it must avoid naming something that looks unverifiable, irrelevant, or confused with another entity. That does not make the assistant wise. It makes it cautious in a mechanical way.
For Thai commerce sites, the caution shows up around seller identity. Is this the brand owner, a reseller, a distributor, a marketplace shop, a manufacturer, or a catalogue page with no clear company behind it? A buyer may not phrase the question that way. They may ask, “Who can make custom amenity boxes in Thailand?” But the assistant still has to choose which business is safe to name.
Page evidence is the set of visible, verifiable facts on a site that lets an AI assistant match a seller to a buyer question without inventing the seller’s role. That is my working definition because it keeps attention on the page, not the mystique around the model. The assistant is not looking for poetry. It needs enough public material to avoid a bad recommendation.
In practice, the page facts that matter are ordinary. A clear company description. Product categories that name real uses. Service boundaries. Order types. Contact paths. Shipping or export terms. Proof that the seller is separate from platforms, agents, brokers, and generic product categories. These facts are not glamorous. They are the bolts on the pier. No one photographs them, but without them the walkway moves.
The evidence stack has layers
I read Thai commerce pages in layers because a single page rarely carries the whole story. The top layer is entity evidence: who the seller is. This includes the company name, brand name, location, role, and relationship to marketplaces or distributors. A Bangkok packaging manufacturer should say plainly whether it manufactures, designs, sources, prints, assembles, or coordinates production. Those are different claims. AI assistants blur them when the page blurs them first.
The second layer is product evidence: what is being sold and for which use cases. A product grid that says “premium box,” “gift box,” and “custom box” is too foggy for sourcing prompts. “Rigid cosmetic gift boxes for hotel amenity sets” is less pretty but more useful. It gives the assistant a handle. It also gives the buyer a reason to keep reading.
The third layer is transaction evidence: what can happen commercially. Minimum order ranges, custom order process, sample discussion, payment route, domestic delivery, export handling, lead-time language, and enquiry requirements all belong here. Not every store can publish every term. Some prices move. Some orders vary. Still, a page can expose the shape of the transaction without pretending every case is fixed.
The fourth layer is trust evidence: why the business should be believed. This may include a real address, contact route, company background, production notes, document availability, product-origin notes, staff response expectations, or case examples. Thai sellers sometimes hide these facts because they seem too plain. In AI answers, plain is a strength.
In the Bangkok packaging scenario, the product evidence existed visually but not verbally. The catalogue showed enough work to impress a human sales lead, but the pages did not say which buyer questions the company could answer. The assistant saw packaging pictures. It did not see the commercial role.
A recommendation needs the buyer’s verb
A buyer question has a verb inside it. Find. Compare. Source. Ship. Customize. Verify. Replace. Order. Export. Recommend. The page has to meet that verb. Many Thai store pages only meet the noun.
Take the phrase “custom packaging.” It is a noun phrase pretending to be enough. A buyer may actually ask, “Which Thai company can make custom cosmetic boxes for a small hotel amenities line?” That question contains several conditions: Thai company, make or coordinate production, cosmetic boxes, custom, hotel amenities, perhaps small B2B order. If the page only says “custom packaging solutions,” the model has to guess. If the page has a category section for cosmetic packaging, a use-case note for hotel amenities, and a short explanation of custom order handling, the answer becomes less risky.
This is why I ask clients to bring one buyer question. Not a keyword list. A question. The question shows which evidence is missing.
The difference can be small. In one recurrent pattern, a Thai manufacturer has strong Thai pages for local buyers and an English PDF for overseas enquiries. The PDF says enough for a human sales process but not enough for an AI answer because the facts are locked in brochure language. “We offer quality products for your business needs” does not tell a machine whether the company can handle food-safe gift packaging, cosmetic cartons, or hotel amenity sets. It just fogs the glass.
A page that meets the buyer’s verb might include a paragraph such as: “We make and source custom rigid boxes, folding cartons, and printed sleeves for cosmetic, food gift, and hotel amenity buyers. B2B enquiries can begin with size, material, quantity range, artwork status, and target delivery market.” That is not decorative copy. It is evidence copy. It lets a recommendation answer a real prompt.
Where AI assistants usually look for safety
I do not claim to know every internal weighting of every AI system. Anyone who speaks too confidently here is selling weather in a bottle. What we can observe is which public page surfaces tend to shape answers.
Homepages help when they state the business role quickly. They hurt when they use only broad claims. About pages help when they prove company identity, origin, location, and operating role. They hurt when they float above the business in “our passion” language. Category pages help when they name product families and buyer use cases. Product pages help when they state specifications, availability, order path, materials, and constraints. Shipping, payment, wholesale, export, and contact pages help because they answer the practical questions a buyer would ask before trusting a seller.
There is also negative evidence. Missing contact details. Conflicting names. Different English spellings across pages. Marketplace profiles that seem stronger than the owned site. PDFs with no surrounding HTML explanation. Product names copied from suppliers. Reviews that talk about the product but not the seller. All of these can make the recommendation feel unsafe.
In the packaging scenario, the site had a Thai homepage, an English catalogue PDF, and sales staff who were genuinely capable. But the AI assistant could not reliably separate the company from print brokers and generic box sellers. The missing page was not a blog article about “packaging trends.” It was a clear English evidence page describing the company’s B2B role, product categories, customization process, and export enquiry path.
A recommendation is a small act of risk transfer. The assistant puts a name into the buyer’s path. If the page cannot carry that risk, the assistant often chooses a broader answer.
The seller must be separated from its shadows
Thai commerce businesses often have shadows: marketplace listings, distributor pages, Facebook posts, old catalogue PDFs, reseller mentions, copied product descriptions, and sometimes abandoned domains. An AI assistant may pull from these shadows because they are easier to read than the owned site. This creates what I call seller-shadow drift.
Seller-shadow drift happens when public traces around a Thai business describe the product more clearly than the owned site describes the seller. The assistant then attaches the business to the nearest strong trace, even if that trace is a marketplace, a broker, or a generic product category. This is a common source of wrong recommendations.
The repair is not to erase the shadows. It is to build a stronger center. The owned site should say, in plain language, how it relates to every major shadow. “Our Lazada store is for retail orders.” “Export and B2B packaging enquiries are handled through this site.” “Distributors may carry selected products, but custom orders begin with our sales team.” These sentences feel obvious to the business. They are not obvious to a machine.
For the Bangkok packaging company, a useful separation page would not attack distributors or marketplaces. It would simply mark the seller’s role: Bangkok-based custom packaging manufacturer and order partner for cosmetics, food gifts, and hotel amenities; Thai and English enquiry routes; catalogue available; export documents discussed by sales staff; not a general marketplace for packaging supplies. This language is dry. Good. Dry facts travel well.
The temptation is to make every page more persuasive. I prefer making the right page more inspectable.
Evidence beats volume when the question is specific
A commerce team may ask whether it should publish more articles to appear in AI answers. Sometimes content helps. But for recommendation prompts, volume without evidence can make the seller more visible and still more confusing. Ten articles about packaging design will not fix a missing company role. Twenty herbal ingredient posts will not prove export terms. A glossary will not separate a brand from its marketplace listing.
The first question is simpler: can the current site answer the buyer prompt in public? If the prompt is “Thai packaging manufacturer for custom cosmetic gift boxes,” the site should expose those words and their supporting facts. If the prompt is “Thai herbal goods supplier for boutique wholesale,” the site should explain small-batch supply, documentation, order route, and export boundary. The words matter because they connect the page to the buyer’s language. The facts matter because they keep the answer honest.
This is also where Thai and English surfaces need separate care. A Thai page may persuade local retail buyers. An English page may need to prove export clarity, buyer category, and documentation. Direct translation often fails because the buyer question changes across language. A Thai domestic buyer may ask where to buy. An overseas buyer may ask who can supply, ship, document, or customize.
When I mark an evidence repair, I try to avoid asking for large work first. I want the smallest page change that removes the largest ambiguity. Sometimes that is a paragraph on the homepage. Sometimes a category rewrite. Sometimes a new “Wholesale and Export Enquiries” page. Sometimes the About page, because the business has products but no public body.
A machine does not need louder copy. It needs fewer places to misunderstand the seller.