Year: 2026

Our “About Us” page lists every one of them — ISO 9001, ISO 20000, and 21 software copyright registrations. Yet one AI said it “could not find any qualification certificate information for this company.” After searching around, we found the pattern: every company that gets surfaced for “having certifications” relies on a standalone article titled something like “Company X Passes Certification Y” — not a single one relies on a list of qualifications on its own website.

We ran a controlled experiment: give the AI only a company name and the official website ranks at the very top; add “how is it” and the website disappears entirely; ask “which one is best” and not a single company website shows up in the top 20 results. The reason is simple — when it comes to whether you are good or not, your own word does not count. This article shares a reproducible testing method and a four-step approach ranked by cost-effectiveness.

Previous posts covered what to write for human readers; this one covers what to write for machines: structured data — restating “who our company is” in a format machines can read directly. It addresses two failure modes: getting it wrong and leaving it out. We have seen AI misspell our company name with a homophone, and AI report “no qualification information found” when that information was right there on our website. This article lists which fields enterprise identity information should include, one important line not to cross (include aliases, but never include a wrong name — that amounts to admitting you go by it), three pitfalls we have all hit (written but not taking effect / inconsistent with the page copy / multiple conflicting blocks on one page), a two-minute verification method, and the right way to see it: not a master key, but a way to reduce error at the moment AI is already about to mention you.

A hidden problem: each page looks well written on its own, yet two pages are competing for the same keyword. Search engines usually pick only one page per site to compete for a given keyword, so you put in three times the effort and none of it ranks. In our own audit we caught a textbook case: the homepage and a service landing page had almost identical titles, and even their descriptions were two versions of the same paragraph. This article gives you a five-minute self-check (line up your titles and spot the look-alikes, search the keyword and see which page the engine picked), a role-split table for four page types, and how we handled it: we left the homepage alone (it is the main driver, and volatility is expensive) and only changed the service page from “which provider is best” to “what exactly we do”, so each page serves a different search intent. Three cautions included (do not touch pages that already rank, split the content as well, and do not try to fix it by deleting pages).

Earlier posts covered how to check indexing, rankings, and crawler activity. The most common question after that audit is: “I know I’m on page two—so what do I change?” The answer is to start with page titles. They take only minutes to fix, affect both search rankings and whether AI systems cite you, and most companies get them wrong. This article compares how the top dozen or so competitors write their titles (core business keywords up front, company name at the end), lists four common mistakes, and discloses two typical problems we found in our own audit: (1) an overly long site-name suffix that pushes business keywords into truncation; (2) service and contact pages using universal phrases like “Products & Services” and “Contact Us” that would fit any company. Includes four principles for writing titles, plus why descriptions should read as factual summaries rather than marketing slogans—because AI summaries often quote them directly.

Being indexed and being usable by AI are two different things. When AI goes online, it only looks at the top batch of results, so ranking 3rd and ranking 300th both count as indexed. We paged through the results to check where we stood, and the outcome was unexpected: we were not missing from search at all, we were ranked 3rd on page 2, just a little short of page one. That finding directly changed what we needed to do. This article covers a plain, low-tech way to page through and check your ranking, four solutions for four ranking positions (pages 2 to 3 offer the best value), and one problem we found along the way: the title and description shown in the search results were the old versions from several years ago, which means the index was never updated. The real value of proactive submission is not only getting new pages discovered, but refreshing outdated old records.

“Can our company self-host an AI? We don’t want our data going outside” — the question we’ve been asked most often lately. Yes, you can — but before you spend money on hardware, think through four things: (1) when self-hosting is genuinely worth it (data that cannot leave the internal network is the only reason that needs no cost-benefit math); (2) hardware is not just about buying a good GPU — VRAM capacity and memory bandwidth decide success or failure, and the golden rule for choosing a model is not to look at total parameter count alone; (3) the most valuable point: across three “it’s slow” investigations, not one turned out to be a hardware problem — thinking mode left on meant a long wait for an empty reply, output filed into the wrong field, and the real bottleneck was time to first token (with the right caching, from over ten seconds to under two); (4) factor operations manpower into the budget. Plus a suggested three-step path: spend a few hundred RMB validating the use case with an online service first, then decide whether to spend tens of thousands.

The most common GEO mistake is treating all AI assistants as a single thing to optimize. In practice, each AI assistant queries its own search engine when it goes online, so “an AI doesn’t recommend you” really means “you aren’t indexed, or don’t rank, in that AI’s search.” We ran that check ourselves: one major vendor’s search returned only 3 results for site:, even though the site has 326 pages — and those 3 were old titles and stale articles from years ago. This article lays out two five-minute checks (site: for indexed volume, business keywords for ranking), three steps to rule out problems on your own side (did the crawler arrive / is it blocked by your protection / verify content completeness with a crawler user agent), and three reasons pages can be crawled but not indexed, with the matching fix for each.

Most claims about “how AI sees you” are indirect guesswork. Only server logs are direct evidence: who came, how often, which pages they fetched, and whether you shut them out at the door. This article gives a three-step check (count each crawler’s fetch volume, see whether status codes show rejections, see whether the pages fetched are your key pages), plus two pitfalls we found when we ran it on our own site: (1) all six sitemap addresses listed in the robots file were dead; after redirects they returned web pages, not XML, and the one that actually worked was missing entirely, so every crawler entering the site received an invalid directory. (2) One search engine crawler was hammering 404s against an address structure that disappeared long ago, burning crawl quota for nothing. A ten-minute self-check checklist is included.

A client exported three days of firewall threat logs — 1,080 entries — and asked whether any of them needed handling. The answer after analysis: hold off. These logs have three problems that make them nearly unusable. ① 88% of the alerts point to the company’s own internal gateway (upstream address translation erased the real source), so blocking IPs from the log means blocking yourself and cutting off the whole company network. ② 34.2% of the actions were “alert only” with no block, including 6 high-risk events that were allowed through — among them two requests fired within ten seconds at publicly disclosed vulnerabilities. ③ The largest category, 29.3% of the total, was the company’s own systems being misidentified during normal communication — and counted twice. Includes the correct order for reading logs (check the action column first, not the threat level) and a five-point self-audit checklist.