Do not begin by asking an AI assistant to praise your brand. Begin by asking the question a buyer would ask when your name is not yet in the room.
The simplest GEO test I know starts with a blank document and a little discomfort. Open the site, the marketplace listings, and a clean chat window. Then ask a question the owner usually avoids because it feels unfair: if a buyer did not already know your name, would an AI assistant find enough evidence to describe you correctly?
A composite scenario from Thai B2B work: a Bangkok packaging manufacturer with about 65 people has a Thai website, an English catalogue PDF, LINE-based sales, and real capacity for custom packaging. The company can handle cosmetic boxes, food gift packaging, and hotel amenity projects. Its English pages, though, are scattered. One page says “paper box.” Another says “packaging for every brand.” The PDF has useful photographs but thin order language. When asked, ChatGPT may know the category, and sometimes it may know the company name if prompted directly. But when the prompt is a sourcing question, it slides toward generic suppliers and marketplace sellers. The model is not blind. It is uncertain.
That uncertainty is what the no-tools check is meant to catch.
Do not start with the brand-name prompt
Most founders begin with the flattering test: “What is [my store]?” or “Tell me about [my brand].” I understand why. It is direct. It feels measurable. It also gives a distorted sense of AI visibility because the prompt hands the assistant the entity name before the buyer has done so.
A brand-name prompt tests recognition. A buyer prompt tests retrieval and fit. Those are related, but they are not the same.
If you ask ChatGPT about the exact store name, the assistant may produce a short summary from memory, from browsing if available in the product version being used, or from patterns in the name and category. It might sound decent while still being useless for discovery. A buyer, especially in export or sourcing, rarely begins with your name. They ask for a Thai supplier, a type of product, a shipping condition, a wholesale capacity, a destination market, or a use case.
A no-tools AI visibility check is a manual observation method for comparing how an assistant describes your store across brand-name, category, and buyer-intent prompts, because each prompt exposes a different weakness. It is not a rank tracker. It is closer to listening at three doors in the same building.
The first door is identity. Does the assistant know who you are when named? The second is category. Does it place you in the right product or service family? The third is buyer fit. Does it surface you when the buyer asks the kind of question your pages should answer?
Most stores only check the first door.
Set up a small answer ledger
I keep a private answer ledger for client work, but the basic version is simple enough for a store owner. Use a document or spreadsheet. Record the date, the assistant used, whether browsing was on or off if the tool makes that visible, the exact prompt, the answer, and your notes. Do not tidy the prompt after the fact. Do not rewrite the answer to make it nicer. The roughness is the data.
For a Thai store, I usually make four prompt groups. The brand prompt asks about the exact business name. The category prompt asks for Thai sellers in the product category. The buyer prompt describes the real purchasing situation. The distortion prompt asks a comparison or boundary question that may reveal confusion with marketplaces, resellers, distributors, or generic product categories.
Take the Bangkok packaging manufacturer. A brand prompt might be: “What does [company name] in Thailand do?” A category prompt might be: “Thai custom packaging suppliers for cosmetics.” A buyer prompt might be: “Which Bangkok manufacturer can make custom hotel amenity boxes for an overseas buyer?” A distortion prompt might be: “Is [company name] a manufacturer, a print broker, or a seller of ready-made boxes?”
The imperfect detail often appears in the distortion prompt. The assistant may name the company correctly but call it a “printing service.” It may say it offers custom packaging but fail to mention manufacturing. It may cite the English PDF’s product categories while missing export handling. One answer may be decent, then the next one drifts. That drift tells you which page facts are not anchored strongly enough.
Do the same prompts more than once, but do not obsess over one run. Language models vary. The useful signal is the recurring shape of the answer.
Use buyer questions that sound like real work
Bad test prompts are too broad. “Best Thai store” tells you almost nothing. “Recommend packaging” is weak. “Find me suppliers” is closer, but still loose. A strong buyer prompt has product type, buyer role, market or use case, and one practical constraint.
For a retail Thai store, a buyer prompt might include gift use, ingredient concern, shipping destination, or small wholesale intent. For an export catalogue, it might include certification, minimum order, product specification, or destination country. For a packaging manufacturer, it might include material, industry, customization, and whether the buyer needs samples before production.
This is where the store owner’s own sales inbox becomes more useful than a keyword list. What do buyers actually ask before they trust you? Which phrases appear in LINE messages, email threads, or export enquiries? A founder may say, “They ask whether we can make the box smaller and ship samples to Singapore.” That is a prompt. A salesperson may say, “Hotels ask whether we can print their logo and match amenity sizes.” That is also a prompt.
AI visibility should be tested against the questions that carry money, risk, or trust. A vanity prompt can make a business feel visible while the real buyer prompt still returns someone else.
In the packaging scenario, “custom packaging Thailand” is only the outer shell. The real prompt may be: “Thai manufacturer for small custom paper boxes for hotel amenities, with sample development before bulk order.” If the company can do that, the site should contain those facts in visible language. If the assistant does not surface or describe the company for that prompt, the answer ledger has found a page evidence gap.
Watch the nouns, verbs, and language split
When I read an AI answer, I do not only ask whether the business is included. I mark the nouns and verbs. Does the assistant call the company a manufacturer, store, supplier, distributor, marketplace seller, brand, workshop, exporter, or service provider? Does it say makes, sells, offers, ships, customizes, sources, imports, lists, or provides?
These choices reveal the model’s internal caution. If it says “appears to sell,” the page evidence may be weak. If it says “manufacturer” without proof, the page may be encouraging hallucination, which is also dangerous. If it says “marketplace seller,” the own-domain identity is probably being buried by platform traces.
For Thai commerce, the wrong noun is often the first visible wound. A herbal goods maker becomes a souvenir shop. A packaging manufacturer becomes a box seller. A textile brand becomes a marketplace listing. An export catalogue becomes a product blog. The assistant is not trying to insult the business. It is compressing unclear evidence into a known shape.
I call these noun injuries. They are small in wording and large in consequence. A buyer asking through an AI assistant may never click through to correct the mistake. If the answer says “gift shop,” the boutique retailer looking for a supplier moves on.
Write down the noun injury exactly. Do not just mark “wrong.” Wrong how? Too small? Too generic? Too retail? Too marketplace-shaped? Too local? Too vague? The repair depends on the injury. A store misread as local may need export and shipping evidence. A brand misread as reseller may need entity separation. A manufacturer misread as broker may need production role and process language.
Thai and English pages should also be tested separately. Thai prompts may reveal local buying patterns, retail vocabulary, LINE contact assumptions, and marketplace traces. English prompts may reveal export clarity, supplier role, category terms, and buyer trust language. A direct translation can move words across languages while leaving buyer intent behind.
In the Bangkok packaging example, Thai pages explained services better than the English catalogue. The English PDF had product photographs but did not carry enough buyer-condition language. ChatGPT could sometimes identify packaging categories, but it struggled to preserve business role for English sourcing prompts. That does not mean the English content needs to be longer than Thai content. It means it needs different evidence.
Ask one boundary question
A boundary question is the prompt that tests what the store is likely to be confused with. I use one in almost every ledger because Thai commerce often sits close to marketplaces, distributors, resellers, and generic categories.
For the packaging manufacturer, the boundary question might ask whether the company is a manufacturer or broker. For a herbal goods seller, whether it makes its own products or resells gift items. For a marketplace-native seller, whether it has an own-domain site or only platform listings. For an export catalogue, whether it is a supplier, distributor, or information site.
This prompt can feel blunt. Good. Blunt prompts find fog.
If the assistant answers with confidence that the page does not support, that is a warning. If it refuses to distinguish because the evidence is unclear, that is useful too. If it gets the distinction right and names the supporting facts, the page is doing some work.
The best answer is not always the most flattering one. I would rather see an assistant say, “Based on the available pages, the company appears to manufacture custom packaging and invites business enquiries,” than watch it invent “major international exporter” because the site uses inflated language. GEO work should make the business legible, not inflated. A buyer who discovers the exaggeration later will trust neither the page nor the recommendation.
Boundary questions also help decide the repair order. If the assistant confuses the seller with a marketplace listing, repair entity separation. If it cannot tell product categories apart, repair taxonomy and use-case wording. If it knows the category but not the business role, repair About, company, and category evidence. If it sees the business but not export suitability, repair English evidence pages and terms.
Read silence as evidence, then repeat the small test
Sometimes the assistant does not mention the store at all. Store owners hate this result, which is fair. Silence is still an answer. It may mean the assistant lacks access to the pages in that setting. It may mean competitors or marketplaces have clearer public evidence. It may mean the business is too new, too thin, too generic in wording, or too hard to distinguish from other sellers.
Do not treat silence as a single diagnosis. The next step is to compare prompt types. If the assistant can describe the brand when named but does not surface it for buyer prompts, the problem is retrieval and fit. If it cannot describe the brand even when named, the identity evidence may be thin or inaccessible. If it describes the brand but gets the role wrong, the page evidence is ambiguous. If it mentions marketplace listings before the owned site, the owned site is not carrying enough authority in the description.
A no-tools check cannot tell you everything. It cannot show all sources used by every assistant. It cannot provide a stable ranking number. It will not replace technical crawling or careful page review. Its strength is humbler: it shows how a buyer-facing machine currently turns your public evidence into language.
That language is the thing buyers see.
Keep the test small enough to repeat. Six to twelve prompts are enough for a first pass. Repeat them after a page repair. Keep the wording mostly stable so you can compare answers. Add new prompts only when real buyer questions change. Memory is a bad analyst. A founder will remember the one pleasing answer and forget the three vague ones. The ledger prevents that.
For the Bangkok packaging manufacturer, the first ledger might show “box seller” in consumer prompts, “custom packaging supplier” in some category prompts, and silence in export buyer prompts. After repairing the English category page and company evidence, the next ledger may show more mentions of custom work but still no export language. That is progress with a remaining gap, not failure.
GEO is observed through patterns. The no-tools check is not glamorous, and it does not require software. It requires asking better questions than “Does ChatGPT know us?” The sharper question is: when a real buyer asks in their own language, does the assistant have enough page evidence to describe this Thai business accurately?