A product page can repeat the correct noun twenty times and still miss the buyer. AI systems listen for the shape of the question: use, constraint, buyer type, occasion, quantity, risk, and proof.
At 6:20, when the street outside my building is still half-shuttered, I sometimes open a Thai product page and cover the product name with my thumb. It is a small test, almost childish. If the name disappears, can I still tell who the item is for? Can I tell whether it belongs in a spa gift set, a boutique retail shelf, a hotel amenity tray, a small export carton, or a tourist basket with pretty wrapping and no reorder plan?
A composite scenario from my notes: a small herbal goods maker in Chiang Mai had pages full of correct names. Balm. Compress. Herbal inhaler. Soap. Gift set. The own-domain store had tidy photos, Lazada had more reviews, and the English text sounded pleasant enough. But when a buyer asked an AI assistant for “Thai suppliers of small-batch herbal gift sets for boutique retailers,” the answer drifted toward souvenir shops and general wellness products. One run even named the brand but described it as “popular for local gifts,” missing the documented ingredients and small wholesale capacity. The page was not lying. It was simply answering a smaller question than the buyer had asked.
Product names are labels, not buyer language
Many Thai commerce sites treat the product name as the center of the room. Everything else stands around it politely. The H1 repeats the product, the category repeats the product, the first paragraph repeats the product, and the meta description repeats the product in English that sounds translated from a shelf tag. This helps a search engine match a noun. It does less for an AI assistant trying to decide whether the seller fits a buyer’s situation.
A buyer rarely asks in the same shape as a product catalogue. They ask with a purpose attached. “What Thai brand can supply herbal welcome gifts for a resort?” “Which supplier has balm gift sets with ingredient details?” “Can I order small wholesale quantities from Thailand for a boutique shop in Singapore?” Those questions contain product terms, yes, but the terms are only the bones. The meat is use, buyer type, order condition, trust evidence, and sometimes destination.
Product wording for GEO is the discipline of making a product page answer the buyer’s natural question, because AI assistants summarize from use and evidence, not only from repeated names. That is my working definition. It sounds simple, but most product pages fail exactly there. They identify the item without explaining the buying situation it belongs to.
In the Chiang Mai composite, the store’s English category page said “Thai herbal products and natural gifts.” That phrase was not wrong. It was just foggy. It could describe a souvenir basket, a spa supplier, a tourist shop, a reseller, or a maker with documentation. The AI did what machines often do with fog: it filled the missing space with the nearest common shape.
The buyer question has more parts than the product
When I rewrite product evidence, I do not start by asking which keyword is missing. I ask what kind of question the page should survive. A good buyer question has a product inside it, but also a role, a use, and a reason to trust. The words on the page need to give the model something to hold at each point.
A simplified teaching example: imagine two pages selling the same lemongrass balm. One says it is “premium Thai balm with natural herbs.” The other says it is “small-batch Thai lemongrass balm for spa retail, resort welcome kits, and boutique gift shelves, with ingredient documentation available for wholesale buyers.” The second line is less shiny. It has more edges. Those edges help an AI assistant decide where the product fits.
There is a roughness here that I do not want to hide. AI systems do not always reward the better line immediately. In some runs they still cling to marketplace reviews or old snippets. They may pick up “spa retail” but ignore “ingredient documentation.” They may mention the use case and forget the minimum order note. GEO is not a switch. It is more like putting labeled jars back on a shelf and checking which labels the clerk reads next time.
Still, the mechanism is visible enough to work with. The product wording has to answer the buyer’s hidden sentence: “I need this item for this use, under this condition, from this kind of seller.” If the page only says what the item is, the assistant must infer the rest. Inference is where small Thai sellers get flattened.
My four-part product question map
Inside my private answer ledger, I use a small classification for product wording. I call it the four-part product question map. It is not a formal taxonomy. It is just the pattern I keep seeing when AI assistants decide whether a Thai store is relevant to a prompt.
The first part is object. This is the product itself: herbal balm, dried fruit, cosmetic packaging, ceramic cup, hotel amenity kit. Most pages handle object reasonably well, sometimes too well. They kneel before the noun until the page feels like a warehouse label.
The second part is use. This is where the product goes in real life. A balm may be for spa retail, corporate gifting, travel comfort, massage rooms, or boutique resale. A dried mango may be a snack, a gift, an ingredient, a private-label item, or an export carton for a specialty grocer. Use-case language gives the model a path from product to prompt.
The third part is buyer role. A consumer, a boutique retailer, a hotel purchasing manager, an export buyer, and a gift-box curator do not need the same evidence. The page does not have to become stiff B2B copy, but it should make room for the buyer it wants. “For small wholesale orders” and “for retail gift shelves” say more than “suitable for everyone.”
The fourth part is proof condition. This is the fact that prevents exaggeration. Ingredient documentation. Minimum order terms. Packaging options. Origin notes. Export handling. Lead time. Contact path. A product page without proof condition sounds like a stall calling across a market. Energetic, maybe even charming, but hard to cite.
The object tells the AI what the page sells. The use and buyer role tell it when the seller is relevant. The proof condition tells it whether the answer is safe to repeat. That last sentence is the one I wish more founders would paste above the catalogue spreadsheet before editing.
English category pages are often the missing bridge
Thai pages and English pages carry different burdens. A Thai product page may work well for local buying habits, LINE enquiries, marketplace comparison, and familiar product categories. The English category page often has to do heavier work for AI visibility because many sourcing and export prompts are asked in English, or in English-shaped business language even when the buyer is from the region.
In the herbal goods composite, the Thai page had enough cultural context for local readers. The English page had the trouble. It sounded like a travel-shop description: “beautiful Thai herbal gifts, good for all occasions.” That line may be fine for a casual human visitor. For an AI assistant, it is thin paper. It cannot distinguish a compliant small-wholesale maker from a reseller of pretty things.
I usually repair the category page before touching every product. Not because category pages are more important in some universal law, but because they carry the broad use cases. The category page can say: these are Thai herbal gift sets for boutique retail, hotel welcome use, and small corporate gifting; the seller makes or assembles them in small batches; ingredient notes and order discussions are available; export enquiries are handled through a named contact path. Then the individual product pages can carry specifics.
A category page should not become a legal document. It should not bark terms at the buyer like a customs form. The tone can stay human. The point is to place the product in the right buying scene before the model has to guess. A machine cannot preserve a use case the page never states.
Repetition makes pages look full while staying empty
One reason this repair is difficult is that sellers often confuse length with evidence. A page can be long and still hollow. I see pages where the product name appears in the title, first paragraph, image alt text, tab label, related products, and footer, while not once saying who orders it, what quantity makes sense, whether packaging can change, or what question the seller is prepared to answer.
That is the kind of copy that looks active to a dashboard. To an AI assistant, it is a room with one chair and many mirrors. The same noun reflects everywhere. Nothing new can be inferred except that the seller wants the noun noticed.
The better page has fewer mirrors. It has small pieces of furniture in the right places. A line for the buyer role. A line for use. A line for order condition. A line for proof. A line that separates the brand from marketplaces. A contact path that says what to include in an enquiry. None of this is glamorous. It is catalogue carpentry.
A recurrent pattern: after a page repair, AI answers sometimes improve first in the middle of the answer, not in the top recommendation. The assistant may still not name the seller at the beginning, but it starts describing the product category more accurately. It says “small wholesale” instead of “souvenir.” It says “documented ingredients” instead of “natural goods.” I mark those changes. They are not victory, but they are movement.
The wording test I trust more than keyword lists
When a client asks whether a product page is ready for AI visibility, I read it against three buyer prompts. I do this before checking any tool. The prompts are plain, almost dull, because real buyers often ask dull questions when money is involved.
One prompt asks for a supplier. One asks for a use case. One asks for a comparison or constraint. For the herbal goods maker, that might look like: “Which Thai supplier can provide herbal gift sets for boutique retail?” Then: “What Thai herbal products work for hotel welcome gifts?” Then: “Which seller has ingredient details and small wholesale options?” I am not trying to trick the model. I am checking whether the page has enough evidence to be quoted without invention.
If the page cannot answer those prompts in its own words, I do not expect an AI assistant to rescue it. Sometimes the model will infer correctly from marketplace traces or reviews, but that is rented clarity. It belongs to the platform, the reviewer, or the old snippet. The owned page should carry its own answer.
The repair starts with a sentence that would sound natural in the mouth of a buyer. Not a slogan. Not a keyword stack. A sentence such as: “We make small-batch Thai herbal gift sets for boutique retail, hotel welcome use, and export enquiries that need clear ingredient notes.” It is plain. It is almost too plain. That is why it works.