Buyers Never Search Your Company Name in AI

A hardware manufacturer asked me recently: when I search our company name in Doubao, the founding year, the main products and both certifications are all correct, not a word wrong. And yet in the past three months we have had zero inquiries from AI.

I asked him one question in return: do your buyers know what your company is called?

The problem is not whether AI knows you. It is that buyers never ask for your name. When they are ready to buy they ask "who can do small-batch stainless steel stamping", "what actually separates 304 from 316", "which tests do I need to export to the EU" — sentences, not names.

This is the most commonly inverted thing in GEO: companies spend all their effort making AI know them, while the real entry point is whether AI thinks of you when a buyer asks about a category, a specification, a problem or a situation. Those four word types are the entry point; the company name is the result.

The boundary first. This article is about how a buyer runs into you inside AI, and it does not apply to companies that already have steady branded search volume — for them the bottleneck is conversion, not exposure. Nor does every business work this way: for custom equipment priced in the hundreds of thousands, buyers usually ask for a solution before a product, and problem keywords carry more weight than category words. Finally, everything below only raises the probability of being mentioned and cited. It cannot guarantee you appear in any answer. Generative engine optimization has no recognised certification of results to this day, and anyone who guarantees you a ranking has ended the conversation.

One question in, several queries out

Google's own documentation calls this query fan-out: AI search does not retrieve with your sentence as written. It first splits the question into a set of smaller sub-queries, searches each of them, then assembles an answer.

The buyer's sentence might be: "I need a factory that can do small-batch stainless steel stamping, ideally in Guangdong, and it has to pass salt-spray testing." Inside the system that becomes something like "small-batch stainless steel stamping supplier", "stainless steel stamping minimum order quantity", "salt spray test standard stamped parts", "Guangdong stamping factory".

To be assembled into that answer, you first need a position in one of those sub-queries.

Nowhere along that chain does your company name need to appear once.

This is not speculation. A Semrush study of more than 540,000 AI answers contains a figure that stings: 62% of AI citations never produce a brand mention. AI is citing pages constantly and organising its answers with what those pages say, without ever naming the brand.

For a company doing AI search optimization this points to a measurement error: watch only whether AI mentions your company and you miss most of the real exposure, then spend more in the wrong place. Measure the wrong indicator in a GEO programme's first phase and the next several months of accounting follow it.

Only four kinds of words ever get asked

The first step in GEO is not writing. It is finding words. Lay out what gets asked in your industry and there are four kinds. How each is phrased, the content shape each demands and the competition on each are entirely different; look at them as one lump and you draw the wrong conclusion.

Category words: what people inside the industry call the thing. "Precision stamping parts", "custom hardware parts", "CNC machining". The most contested, and where the large factories are thickest.

Specification words: asked when a buyer is comparing against a drawing. "How much salt-spray resistance separates 304 from 316", "can you hold ±0.05 tolerance consistently", "what is the thinnest wall you can run". Fewer people ask, but whoever asks is already choosing a supplier.

Problem words: asked when a buyer is in trouble. "How do you fix distortion after bending stamped parts", "is plating blistering a pre-treatment issue or a current issue". Whoever phrases it that way usually has a batch of rejects sitting in the plant.

Situation words: asked when a buyer describes their own circumstances. "Which factory takes small-batch trial production", "which certifications do I need to export to the EU", "how do I negotiate price on a small order".

A factory usually has real substance in one or two of these. Standard-part makers are strong on category and specification; custom makers are strong on problem and situation. Covering all four is rare, and where it happens someone on the team has genuinely been writing technical content for years.

So invert the order of decision: look at what real material you hold, then pick the two types to attack — rather than setting a "cover everything" target and writing all four thin.

Twenty years of focus, and not one category AI can read

This is the most common misalignment in AI search optimization. The first line of the homepage reads "twenty years focused on precision manufacturing, serving customers worldwide", and nowhere on the page is there a specific category word or a single parameter. That sentence works on a human being. To a retrieval system it carries zero information — nothing on which to file you as a precision stamping parts supplier.

There is a harder layer. Even with category words all over your site, results are not assured. AirOps' "State of AI Search" 2026 states that 85% of brand mentions come from third-party pages. That explains something many owners cannot reconcile — what you say about yourself carries less weight than what others say. GEO and traditional SEO part company here: in the SEO era the website could be the destination; now it is only the starting point.

A University of Toronto preprint measures it more finely: when assembling answers, AI's preference for third-party sources runs between 53% and 95%, with the actual point depending on query type. So the preference is not an absolute rule, but the direction is consistent.

Two more academic reference points. A UC Berkeley preprint found the correlation between content structure and citation to be in the 0.63 to 0.68 range (preprint, not peer-reviewed). A Yonsei University preprint measured a 22% lift in retrieval hits from structured treatment — a relative change under laboratory conditions, not an online result.

Put together, they point to one thing: content has to be written not only with the right words but in a shape the system can cut. A three-hundred-word paragraph with no breaks and a four-row specification table are a long way apart in the odds of being taken. Generative engine optimization has never been about keyword frequency; it is about structure that can be taken apart.

How to check: ask the four types, not your name

The self-check is not complicated. One person can finish the first round in half a day.

1. List 20 words, five from each type. Use the buyer's own phrasing, not industry jargon. "Small-batch stamping" is closer to a real question than "multi-variety small-batch flexible stamping".
2. Ask once each in Doubao, DeepSeek, ERNIE Bot, Tencent Yuanbao and Qwen, copying the same sentence verbatim with no rewriting along the way.
3. Record three things per answer: whether your name appeared, what it said when it did, and which competitors it named.
4. Ask again two weeks later. A single reading is led astray by chance.

The point of this step is not to produce a score. It is to establish two things: which category you are currently filed under, and which one is a complete blank.

Why five engines? A SIGIR 2026 paper tested exactly this: hand the same question to different systems and the overlap between the pages they cite — Jaccard — is only 0.18. Test one and you have given up on the other four. In AI search optimization there is no such thing as one engine standing in for all.

One aside: those 20 words need not be complete. The first batch is for orientation, not coverage. Get the orientation right and inferring from actual inquiries later takes far less effort.

A word list is not one page; it is one landing point per type

The most common mistake is piling the words onto one "Products" page with a title like "Products, Precision Stamping, Custom Hardware, CNC Machining". A page that wants everything gets nothing. That is not what a GEO word list is for.

What works is one standalone page per word type, with the page heading being the question itself.

Category words go on the homepage and main product pages: what you make, for whom, to what standard. Specification words go on a selection and comparison page that sets out the differences between two grades in a table. Problem words go on a troubleshooting page written in three parts — symptom, cause, remedy. Situation words go on an application page covering small batches, rush orders and heavy certification demands.

Each page answers one type of question only. There is a direct benefit: when AI splits out the sub-query "salt spray test standard stamped parts", your site has a page devoted to it, rather than a passing sentence on some other page.

One detail that is easily missed. Meltwater analysed 9.5 million AI citations and found that about 75% of LinkedIn citations come from personal profiles and only 25% from company pages. This is an overseas-platform figure and cannot be carried straight across to Chinese platforms, but the principle holds generally: content carrying a named author and specific experience is easier to cite than a release written in an institutional voice.

Branded search volume is a result, not a method

A common form of reporting: three months of work, and searches for the company name went up. That may be true, but it does not prove GEO works. It proves someone remembered you.

The order runs like this: a buyer searches a category word, sees you in the answer, remembers the name, and only then may search you directly. A rise in branded volume is the result at the far end of the chain. Treat it as the only indicator and you have taken the last link for the whole chain.

There is a trap on the other side. If almost everything on the site is About Us, company culture and milestones, your content density sits where nobody asks. Those pages can exist, but they should not outweigh product and technical content. This is what makes AI search optimization awkward on many factory sites: what has to move is not the copy but the content structure.

Three common questions

Our product is too technical. Do buyers really ask at that level?

The more technical the product, the more specific the questions — people who do not understand it never come to you at all. Buyers who actually place orders ask about parameters and exception handling. The total volume of such queries is small, but each one sits very close to a purchase.

Do overseas buyers always ask in English?

It depends on the target market. English-speaking regions ask in English; the Middle East, South America and Southeast Asia each have their own main language. One point easily missed: Google states plainly that it determines page language from the visible body text, not from the lang attribute and not from the URL path. Renaming a path to /en/ while the body stays Chinese does nothing for language matching.

How many words do we need to start?

A list of 20 is enough for the first round. After that, work backwards from real inquiries — the sentences buyers use in their emails are the words most worth covering. That is far more accurate than brainstorming in an office. Get the words right and the rest of the GEO work means something.

Where this leaves you

Back to the manufacturer at the start. His problem was not that AI does not know him. It is that he took "our company introduction is well written" for "buyers can find us". There is no necessary connection between the two.

A more useful self-check: cover the headings on your homepage, leave only the body text, and give it to someone who has never heard of your company. Ask them what the business does. If they cannot answer, the odds are AI cannot either.

Yihang Network's own position is similar. The company worked in cross-border B2B and Alibaba.com-related services before making GEO its main line in 2024. Our first move with a client is not to ask about budget; it is to lay out the four word types they hold and compare them, and whichever is missing gets filled first.

Data and effect boundaries: the two statements cited from Google Search Central (the query fan-out mechanism, and page language being judged from visible body text) are platform rules whose effect is limited to the Google ecosystem. The Semrush figure of 62%, from more than 540,000 AI answers, and the 85% in AirOps' "State of AI Search" 2026 are vendor-published observational statistics, not independently verified, and describe differences between groups rather than causation. The University of Toronto range of 53–95%, the UC Berkeley correlation of 0.63–0.68, the Yonsei University 22% and the SIGIR 2026 Jaccard of 0.18 are preprint or paper findings under laboratory conditions — relative changes, either not peer-reviewed or bounded by the query set tested. Meltwater's 75% is an overseas-platform figure and does not apply to Chinese platforms. None of the above describes a ranking mechanism or represents the results of any specific company. The methods described raise the probability of being retrieved and cited; they do not guarantee that a company will appear in any AI answer, and they promise no timeline and no business outcome.

Sources

1. Google Search Central: "AI features and your website" (query fan-out; no special schema or markup required for AI features), accessed 2026
2. Google Search Central: "Managing Multi-Regional and Multilingual Sites" (page language determined by visible body text, not the lang attribute or URL)
3. Semrush: "ghost citations" study across more than 540,000 AI answers (62% of AI citations never produce a brand mention; vendor-published, observational)
4. AirOps: "State of AI Search" 2026 (85% of brand mentions from third-party pages; vendor-published, observational)
5. University of Toronto: arXiv:2509.08919 (AI answers prefer third-party sources 53–95%; preprint, not peer-reviewed)
6. UC Berkeley: arXiv:2509.10762 (correlation between content structure and citation 0.63–0.68; preprint, not peer-reviewed)
7. Yonsei University: arXiv:2602.12187 (structured treatment lifts retrieval hits by 22%, laboratory conditions; preprint, not peer-reviewed)
8. SIGIR '26: study of citation overlap across systems (Jaccard 0.18; peer-reviewed)
9. Meltwater: analysis of 9.5 million AI citations (about 75% of LinkedIn citations from personal profiles, 25% from company pages; overseas-platform figures, vendor-published)