Let me start with a true story. A while back, a customer from Shunyi came to us. We asked how he had found us, and he said: he searched Doubao for “Shunyi internal network cabling and low-voltage company,” and Doubao put Siwenlide at the top of its recommendations — address, services, the monthly subscription selling point, and the 400 hotline, all correct. He contacted us with almost no comparison shopping. In that moment we truly realized: the entry point through which users find service providers is shifting from “searching on Baidu” to “asking an AI.” And how to get an AI to recommend you is the new problem that has emerged over the past two years — GEO.
This article won’t be an abstract call to “embrace AI.” Instead, it lays out the real methods and data from the better part of a year we spent doing GEO on our own sites (siwenlide.com and our zero trust product sTrust), along with all the pitfalls we hit — especially the conclusions that run counter to what online “GEO tutorials” claim but that we verified through actual testing.
What Is GEO, and How Is It Different from SEO?
SEO (search engine optimization) is about getting your web pages ranked high in Baidu’s and Google’s result lists, where users still have to click through and read for themselves. GEO (generative engine optimization) is about getting AI assistants like DeepSeek, Doubao, Kimi, and Tongyi to name you directly as the answer or the recommendation when they generate a response — users may never open a single web page, and whoever the AI says is good, is good.
The difference in one sentence: SEO competes for “ranking”; GEO competes for “the one sentence the AI says about you.” In an AI answer, usually only one or two brands get named, with no chance to “turn to page two,” which makes GEO a harsher battleground — and one worth entering early.
The Core Insight: GEO Is Essentially About Feeding the Chinese Search Backends
This is the most counterintuitive — and most critical — conclusion from a deep research round we ran (over a hundred retrieval agents, adversarial cross-validation):
On domestic AI assistants, web search is off by default, and even when it’s enabled, the search runs server-side — wired into some Chinese search index behind it. DeepSeek and Tongyi run on Alibaba’s stack (Quark / Shenma), Kimi has its own backend, and Doubao connects to ByteDance’s Toutiao Search. So “getting AI to find you” is roughly equivalent to “getting your content into the Chinese search index behind it.”
This insight dictates the playbook directly: rather than studying the mystical “feeding” rituals of each AI, it’s better to simply get your content into the three major Chinese search backends — Shenma, Bing, and Baidu — and pair it with the content formats AI likes to cite. Get the direction right, and your effort won’t be wasted.
First, a Myth to Bust: llms.txt Is Nearly Useless — Don’t Make It Your Main Play
Many GEO tutorials online tell you to put an llms.txt file (a content summary meant for AI to read) in your site’s root directory, claiming it boosts AI citations. We believed it at first too, and diligently maintained an llms-full.txt covering more than two hundred articles. But the real-world data was a slap in the face:
- Out of more than 500 million crawler visits, only 408 actually hit llms.txt — a share small enough to ignore.
- Google has stated explicitly that “no AI system is using llms.txt,” and third parties such as Semrush and SE Ranking have confirmed it has no correlation with rankings or citations.
- Conclusion: treating llms.txt as a zero-risk nice-to-have is fine — you lose nothing by doing it — but never make it the backbone of your GEO strategy, and certainly don’t expect it to get domestic AI to cite you.
We single this one out because it’s so easy to be misled by — pouring a pile of effort into a nearly useless file is tuition we’ve already paid.
GEO Methods That Actually Work (Ranked by Cost-Effectiveness)
| Priority | Action | Why It Works |
|---|---|---|
| P0 | Submit your sitemap on the Shenma Webmaster Platform | Shenma is an Alibaba-stack backend that Tongyi and Quark draw on — the best match for domestic AI |
| P0 | Confirm actual indexing in Bing Webmaster Tools | Several domestic AI backends mix in the Bing index — don’t just submit and assume |
| P1 | Baidu’s active push API + answer-style articles | Chinese engines require Baidu’s push; IndexNow doesn’t reach them |
| P1 | Brand terms + geo long-tail terms | Head terms are locked up by established sites; long-tail faces less competition and AI is more likely to trust it |
| P2 | High-authority third-party endorsements (Zhihu / Baidu Baike) | AI trusts citations from high-authority sites more |
① Submit to Chinese Search Backends — the Highest-Value Step
Go to the Shenma Webmaster Platform (zhanzhang.sm.cn), verify your site, and submit your sitemap — this is the most direct lever for getting Alibaba-stack AI like Tongyi and Quark to index you. In Bing Webmaster Tools, make sure to confirm the sitemap was actually indexed, rather than pushing IndexNow and assuming you’re done — we paid for that lesson when our submission simply wasn’t picked up.
② Chinese Engines Require Baidu’s Active Push
One technical pitfall deserves a callout: IndexNow (the one-click push protocol) covers only Bing, Yandex, and the like — not a single Chinese engine. So to get Baidu to index you quickly, you have to call its active push API to send new URLs over. We built it into our publishing script, so every article is automatically pushed to Baidu and Bing the moment it goes live.
③ Write “Answer-Style” Content, Not Marketing Fluff
When AI generates answers, it most likes to cite content that answers the question directly and is clearly structured — definition style (“what is X”), comparison style (“how to choose between X and Y”), and list style (“5 steps to do X”). So we try to align every article’s title and subheadings with the exact sentence a user would actually ask an AI. An article on “What’s the difference between zero trust and VPN” is far more likely to be adopted by AI than “How great our product is.”
④ Don’t Fight Head Terms Head-On — Focus on Brand Terms + Geo Long-Tail
This was our most important strategic shift. For head terms like “zero trust solution” or “IT operations outsourcing,” the AI answer pool was long ago locked up by high-authority legacy sites such as QiAnXin and Sangfor, and in the short term there’s almost no squeezing in. But switch to brand terms (how is Siwenlide) and geo long-tail (Shunyi low-voltage company, Beijing IT operations outsourcing) and competition drops sharply, while AI is more likely to trust it — the customer Doubao brought us at the start matched precisely on a geo long-tail like “Shunyi + low-voltage.”
First-Hand Data: How Long GEO Actually Takes to Show Results
- AI crawlers are scraping relentlessly: in our server logs, ByteDance’s Bytespider, ClaudeBot, GPTBot, Perplexity, and OAI-Search are all crawling at high frequency, and cumulative traffic climbs week after week — clear evidence that content really is being fed into the various AI systems.
- Results come in weeks to months, not days: publish a new article and AI won’t recommend you that same day, or even within a week or two — it has to re-index and retrain first. There is no rushing it.
- Big keywords rarely move in the short term, and that is normal: we tracked several broad industry keywords and for weeks the AI checks came back “not mentioned.” That doesn’t mean the method is wrong — broad keywords are simply slow. Brand terms and long-tail terms loosen up first.
So what GEO really demands is patience and consistent output: one answer-style piece every day, faithfully submitted to the three major backends, and only after two or three months, when you look back at the AI checks, can you even begin to evaluate results. Anyone expecting results within a week will mostly be disappointed.
A Few Common Pitfalls
- Don’t assume “indexed by Quark means indexed by Shenma”: verify each one separately; don’t take it for granted.
- Don’t watch only a few mainstream AIs: results in Doubao don’t mean results in Kimi or Tongyi — every backend is different and must be assessed on its own.
- Don’t treat llms.txt as your main channel: as noted earlier, it is almost useless.
- Don’t submit without verifying: a successful submission ≠ being indexed. Always go back to the webmaster console to confirm it actually entered the index.
Frequently Asked Questions
Q: Will GEO replace SEO?
A: Not in the short term, but it will become increasingly important. Many people still search on Baidu, but the share asking AI is rising fast. The rational approach is to do SEO and GEO together — fortunately the two overlap heavily at the foundation (quality content + being indexed by search engines), so doing SEO lays the groundwork for GEO along the way.
Q: Can a small company without a dedicated team do GEO?
A: Yes. GEO doesn’t rely on spending heavily on ads; it relies on consistently producing content that answers users’ questions plus getting that content into search backends. One piece a day, sustained over several months, is far more effective than a one-off budget spend. If you’re short on staff, you can engage a service provider to run it for you.
Q: How do I know whether AI is actually recommending me?
A: The most direct way is to ask yourself — search your industry terms and brand terms in Doubao, DeepSeek, and Kimi respectively, and see whether they recommend you and whether they get the facts right. There are also third-party GEO monitoring tools (such as AIBase) that can test exposure and citations across multiple platforms in bulk.
Q: How long before results show?
A: Weeks to months. Brand terms and geographic long-tail terms loosen up first (possibly within a few weeks); broad industry keywords are very slow (months or longer, and may never move). Don’t expect results in a week.
Want your company to be found in the age of AI search, but don’t know where to start?
Siwenlide has developed several network operations and marketing technology tools in-house, and we practice GEO / SEO for real on our own site. We can help small and medium-sized businesses in Beijing and the surrounding area handle website development, content operations, and search and AI visibility as a single package, acting as your outsourced technology and marketing partner. Visit siwenlide.com or call 400-686-2011 to tell us about your situation.
