GEO practice: how an official site gets cited when 700M users ask AI

The short version: China's CNNIC reported this September that generative-AI users have passed 700 million, with penetration above 50% — and about 76% of them use AI to answer questions. The battlefield of customer acquisition has moved forward: instead of competing for a position on a results page, companies now compete for whether the AI mentions them — and describes them accurately — when a buyer asks the question. The craft is called GEO (generative engine optimization). We rebuilt our own site to that standard; this post documents what we can verify, not folklore.

The same shift is visible abroad: ChatGPT now serves more than one million business search users, and Google's AI search converts 27% better than classic search. The first step of a B2B purchase is changing from "open a search engine, type keywords, read ten pages" to "ask AI one question, get a structured answer, contact the recommended vendor". That changes the role of an official site: it is no longer just a facade for humans — it is the material library an AI draws on when composing answers.

Old path · search-engine era Type keywords Read ten pages Open each site Piece facts together Contact a few, compare slowly New path · AI-search era Ask AI one question "line-software vendors near Suzhou?" AI answers with 3–5 named candidates who gets named depends on what can be trusted Contact the named vendor most comparison already happened in chat The paging era competed for rank; the AI era competes for a place inside the answer — that is the real split between GEO and SEO
The front of the buying path compresses from "ten pages of comparing" into "one conversation": the 3–5 vendors named in the answer are the new first screen

1 · GEO and SEO are not the same thing: earn trust, not clicks

DimensionSEOGEO
What you winRank and clicks on a results pageBeing mentioned and described correctly inside a generated answer
AudienceCrawlers and ranking algorithmsA model's semantic understanding and citation preferences
Main leversKeyword density, backlinks, technical fitSource authority, structured content, consistency across platforms
Content shapePages built around keywordsAnswers built around questions that can be fully resolved
Best fitBasic product lookupDeep decisions: vendor comparison, solution selection

The two are complements, not substitutes — classic search still exists while AI search grows fast. Keep both legs: SEO defends the legacy entrance, GEO takes the growing one. For budget-constrained SMEs we suggest building the GEO foundation first, because much of it (structured data, factual density, readable pages) also improves SEO quality.

2 · What kind of company content does an AI actually cite

  • High factual density. Concrete numbers, delivery lists, stacks, process steps — not "deep industry experience, one-stop solutions", which any company can write. AI adopts verifiable statements.
  • Clear structure. Question-and-answer sections, tables, lists. AI extracts along structure; an orderly FAQ beats a lyrical company profile.
  • Cross-source consistency. The site, trade media and knowledge communities describe the company the same way. Orphan claims are discounted; contradictions get skipped.
  • Traceable freshness. Dates and versions. Content that is stale and undated loses citation priority.
In one sentence: an AI is grading a reading-comprehension test on source credibility. You cannot bribe the grader — you can only write answers that survive checking.

3 · The six-point self-check — we applied it to this site

None of this requires an agency; a technical owner can audit and fix it in half a day to two days. This is the checklist we ran on our own site:

  1. llms.txt at the site root telling AI crawlers what the site contains and what each page covers — the robots.txt plus sitemap of the AI era.
  2. JSON-LD structured data: Organization, Service, FAQPage, Article each in place. Structured fields are far cheaper to extract than full text; better labels mean better odds of being cited correctly.
  3. A FAQ page of the questions buyers really ask, each answer giving the conclusion first and the explanation after — exactly the granularity AI assembles answers from.
  4. Case pages that write process, not adjectives: modules, flows, trade-offs, pitfalls — all verifiable statements. Adjectives do not earn citations.
  5. Canonical and hreflang across the bilingual pages, so AI does not mistake two versions of the same content for two companies.
  6. One clear answer per page: the title is the conclusion ("Modbus or OPC UA?"), the first paragraph gives the judgement. That is the paragraph an AI summary grabs.

4 · Three lines not to cross

  1. Do not fake sources or stuff entries. Citation preferences keep converging on authoritative sources; mass-produced filler content has near-zero marginal return — and regulatory risk under generative-AI rules.
  2. Do not believe "guaranteed citation rank". Generated answers differ from query to query; anyone promising "always first" is not credible. What can be promised: content rewritten to citable standards, complete structured data, monitored visibility.
  3. Do not make GEO the only channel. AI search earns you a place on the shortlist; converting the lead still depends on cases, transparent pricing and response speed — the things the rest of this site covers.

An honest expectation: even done right, you will not be cited every time — candidate slots are few and competition is dynamic. Done wrong, you are guaranteed absent. Build the foundation correctly and let long-term content quality do the rest.

If you want the same audit on your site, start with llms.txt and JSON-LD — the gap is visible the same day. We also run an AI visibility check: we ask your industry's real questions on the mainstream AI platforms, see whether you appear and how accurately, then hand you a remediation list.

Planning a similar production line or enterprise system?

Start with our pricing logic, then let's talk requirements — cost structure, technology comparison and acceptance criteria, all laid out in the open.