A few days ago something happened: I asked an AI for “recommended IT operations companies in Beijing.” It listed seven or eight, and we weren’t among them.
My first reaction was almost instinctive — the content isn’t good enough, write a few more articles.
But I didn’t rush into it. I ran three sets of controlled tests first. The results showed this was not a content problem at all, and that if I had followed my first instinct and produced more content, all that effort would have been wasted.
This article is about that three-layer diagnosis: when an AI doesn’t recommend you, the cause may lie at one of three entirely different levels, and the fixes do not overlap. Get the level wrong and no amount of work will help.
The conclusion first: three layers — diagnose before you act
The same symptom — “the AI doesn’t recommend me” — can mean:
① It doesn’t know you at all (missing entity) — what you need to fix is your basic information sources
② It knows you, but didn’t retrieve you this time (insufficient recall) — what you need to add is the type of page that can be retrieved
③ It retrieved you, but didn’t choose you (ranking behind) — what you need to add is credentials, endorsements, and comparable information
Most people assume they are case ① and write content frantically; in reality many companies are case ② or ③, and no amount of writing will get them in.
Layer One: Does it actually know you?
How to test: open a new conversation (don’t ask inside the earlier context) and ask directly, “What does XX company do?” — using the full, complete company name.
It can’t answer / answers with a different company → a Layer ① problem: missing entity. What you need to do is complete your basic information sources: business registration platforms, the company profile page on your own website, and any public descriptions that can be retrieved. Until this is done, everything else is a castle in the air.
It answers accurately and completely → congratulations, Layer ① is fine. You can rule out the “it doesn’t know me” hypothesis and move straight to Layer ②.
In our case the result was the latter — it got our founding year, our district, our three main business lines and even our in-house products right. So the whole idea of “write more content so the AI gets to know us” was wrong from the start.
Layer Two: Did it actually go online this time?
This is the most easily overlooked point, and the one that can save you the most money: the default mode of many AIs does not go online at all.
It isn’t that it searched and failed to find you — it didn’t search at all, and answered straight from what the model “remembers.” And what the model remembers are the large companies that appeared again and again in the public corpus it was trained on. In this mode, you can optimize your website to perfection and nothing will change.
How to tell whether it went online — two signals:
- Whether there is a notice such as “searched N keywords / referenced N sources”;
- Whether the answer carries source markers or links. A perfectly clean answer with not a single citation almost certainly means no internet access.
Ask the same question again with one line added manually — “Please search the web before answering and cite your sources” — and compare. If the two results differ greatly, what you tested before was not the same thing at all.
When we ran this comparison, the offline round returned only the big names in the industry; once we required web search, it explicitly noted that it had “referenced twenty sources,” and the list changed at once — but we still weren’t on it. So the problem fell to Layer Three.
Layer Three: It retrieved you — so why didn’t it pick you?
There is only one thing to do at this layer: look at whom it chose and on what basis. In web-search mode it usually marks its sources, and those sources are the answer.
We went through the cited sources one by one and found a pattern that runs against intuition:
When an AI needs to assemble a “here are a few recommendations” list, it pulls the “Company Profile / About Us” pages of individual companies, along with their profile pages on business information platforms — not the carefully written selection guide you produced.
The reason is plain enough: what it needs are company entities it can extract one by one. However well written your “How to Choose XX” article is, if it talks only about your own company, the AI cannot extract a list from it.
Another obvious feature: the selected companies generally carry verifiable credentials and third-party endorsements in their descriptions — system certifications, industry rankings, “specialized and sophisticated enterprise” designations and the like. And the first piece of selection advice an AI gives is often “verify the credentials first.” That is the real threshold at Layer ③: not how well you write, but whether you have anything others can verify.
The Most Surprising Finding: Someone Is Speaking for You with Information from Ten Years Ago
Following the “sources” trail further, we found the thing that really stung.
A B2B directory site carries a page about our company, and the AI is citing it as an authoritative source. That page says:
- The person in charge is wrong — it is not our company’s legal representative;
- The address is our old premises from years ago, and the phone number was discontinued long ago;
- The most damaging one: the industry category is listed as “office equipment” rather than IT services;
- The main business field reads “Mainly deals in related products. Unknown.”
Another platform is even worse: its title describes the company outright as an “office equipment rental company,” and its content dates back nine years.
What does this mean? When someone asks for “recommended IT operations companies,” some of the material about us that the AI retrieves states plainly that we are in office equipment. Of course it won’t put us on an IT operations list — and that is not its fault.
The defining trait of such pages is that you never look at them yourself, so you never know they exist. It may have come from a registration years ago, filled in offhand by a salesperson, and sat there for a decade. In the search era it carried little weight; in the AI era, it has become one of the “materials about you” in the machine’s eyes.
How to Find Out Who Is Speaking for You
- Use an AI with web search to look up your full company name, then open every source it lists. This is far faster than paging through a search engine yourself, and what you see is exactly the set of pages the machine trusts.
- Focus on the non-official sites: directory sites, business listings, industry platforms, rental and procurement sites. This is where the errors almost always are.
- Check four items one by one: company name, legal representative, address and phone, and industry category. The last is the most error-prone and the most damaging.
- For anything wrong, use the platform’s complaint or correction channel. Most platforms offer a report form that requires no login, or a support email; explain the situation and provide your business license to request a correction or removal.
Find your row: which action applies to you
| What the test shows | Which layer | What to do |
|---|---|---|
| You ask for the company’s full name; it cannot answer, or names a different company | Layer ① | Build up basic information sources: business registration platforms, your website’s company profile page, publicly searchable descriptions |
| It answers accurately, but cites no sources at all | Layer ② (it is not online) |
Switch to a web-enabled mode and test again. In this case, optimizing your website has no effect, so do not waste your effort |
| It is online, but the sources name other companies and not you | Layer ② | Create “company profile / vendor directory” style pages instead of writing more selection guides |
| You appear in the sources, but not in the recommendation list | Layer ③ | Add verifiable qualifications and third-party endorsements |
| The sources include a third-party page that describes you incorrectly | Handle first | Use the complaint / correction channels to have it fixed or taken down, especially where your industry category is wrong |
A few common questions
Q: Of the three layers, which one should I tackle first?
A: Spend ten minutes completing the diagnosis, then decide — that is exactly the point of this article. But if the diagnosis turns up a third-party page that gets your industry or your legal representative wrong, that jumps the queue and takes priority, because it keeps doing damage: unless you fix it, everything else you do above it is being cancelled out.
Q: How long after making changes will I see a difference?
A: Months. And bear in mind that AI answers are inherently non-deterministic — the same question may give different results at different times. Look at the trend over several months; do not draw conclusions from a single result.
Q: Do small companies have no chance on these “recommend a few vendors” questions?
A: On broad, generic terms, you genuinely cannot outcompete large vendors with market share and analyst reports. But the more specific the question, the better the chance for a small company — building your content and information around specific scenarios (how many employees, which industry, what requirement) is far more realistic than fighting for the broadest term of all.
Q: Can I check all this myself?
A: Yes, and it costs nothing. Just ask the mainstream AI web apps directly; one pass takes about ten minutes. The genuinely hard part is sticking with it once a month, and patiently working through each platform’s appeal process once you find incorrect information.
Want to know which category AI currently puts you in — and who is speaking on your behalf?
Siwenlide provides website and SEO/GEO content-led customer acquisition for small and medium-sized businesses in Beijing and the surrounding area: we use the three-layer diagnosis above to pinpoint which layer your problem sits in, find every third-party page that describes you one by one, route the incorrect ones through the proper correction process, then fill the gaps with targeted profile pages and credential endorsements on your website, and issue a comparison report every month. We ran this diagnosis on ourselves first — the incorrect information in this article is exactly what it turned up. Visit siwenlide.com (or take a look at our company profile first), or call 400-686-2011 to talk it over.
