| name | loki-yan-seo |
|---|---|
| description | Use when diagnosing or executing Google SEO, international/cross-border sites, AIO, or GEO. Unofficial engine distilled from Loki Yan (@loki_yan_seo) public tweets. Not a generic SEO encyclopedia. Does not imitate his voice. Not affiliated or endorsed. |
Loki Yan SEO
Executable engine. Knowledge SSOT is methodology.md in this directory. Internalize the method, then work the site. Do not lecture about Loki unless the user asks for source or evidence.
Bounds
- Corpus now includes posts through 2026-09-03 UTC. Last full public-post recount remains 979 posts (367 originals, 588 replies, 24 reposts). Loki-authored +23 since last cutoff (was 854; now 877). This run only ingested 23 verified doctrine posts, not a full dump recount. Full originals/replies/reposts recount was not rerun. Do not treat this as the whole account. Last folded doctrine: fake authors / fake profiles (id=2095347481655840877).
- Conversation-partner tweets in the corpus are not his claims. Empty / quote-only RTs and thanks-only replies are ignored.
- The law-heavy window is almost all 2026-03-15 → 2026-09-03. One 2025-10-15 reply already states the Quality-vs-EEAT split (id=1978391845844426780).
- The cleanest before/afters (fix render → traffic jumped; stop $20K/mo links → no impact; traditional SEO only → ChatGPT rose) are all [asserted]: no site name, no GSC, no chart in the tweets.
- He said, exactly: AEO / GEO “I currently have no way to answer you” (id=2091727382726205949). Do not invent a how-to.
- “Do not mechanically execute SOP” (id=2091761728241766791). Below is a scan order, not a list you can run with your brain off.
Core stance
Ten laws from methodology §1. Cite (id=…). Do not invent ids.
- SEO does three things: crawl, index, serve. Writing blogs is not SEO. (id=2091364706703581191) (id=2092059738062348681) (id=2081519248011370590) 铺词 / 铺内容 are fake propositions (id=2090954259680694666) (id=2089165558516437461). On a serious long-term project, every page has one unique, searched, speakable query (id=2090938323284185332). Doing SEO (and future GEO) is catching demand (id=2091668405367791619). SEO captures demand, does not create it (id=2092387406242787628). Many businesses should not start with SEO — no search demand / no 种草. Paid can scale when ROAS works (id=2092380170426765423). SEO is maximizing a site and a brand’s exposure (id=2091463375003983897). Two kinds
[once]: growth (rank, traffic, convert) vs hold-the-hill at Canva / Zillow / Booking — “stop other people from messing around” (id=2091321150127366357). - You do SEO to make money, not for love. Optimize the Money Page and Trust. High-ticket buyers will not become paying users because of a Reddit thread or a ChatGPT blog (id=2090983234146574708). Conversion and money in the pocket have value; traffic that does not convert is charity (id=2033807753547043235). All-blog sites still have to answer “do you convert? I think you should do a money page” (id=2091775475635073373). Blog-only sites lose traffic as AI Mode answers; users like summaries; money page still required (id=2093475631803826251).
- Quality is a site-wide ranking factor. Its opposite is Spam. EEAT itself is not. “EEAT is not a ranking factor (Google said). EEAT determines quality, and Google said Quality is a ranking factor” (id=1978391845844426780). Quality should consider the whole site (id=2092027807719444899). Quality’s definition is Effort and Original (id=2091362717907251353). Later four-part of high-quality content: First-hand experience + Originality + Non-commodity + Effort (handwritten, not AI) (id=2092977176501228007). Do not drop Effort + Original. Authority cannot offset AI Spam / AI Slope / junk blogs (id=2089165558516437461). Keep the March tension
[once]: “It’s not about quality. It’s about how it was created” (id=2033808549445570669) — same human-input insistence; do not quote one without the other. - The EEAT door is Effort + Originality. Generic automation cannot cross it. Tokens are not Effort. (id=2092095823731462302) Your own process / agent / prompts are fine; do not sell a universal Agent (id=2092073561171398727). Dunk viral “Claude did SEO / 1M traffic” claims: “重点是，人家用claude的是大佬。。而不是claude” (id=2097125101875675367) (id=2097152234756989288). Keep tension: Claude-hype dunk vs own-agent/process is fine. Any batch work needs increment (original text / images / video / insights) or it is judged low effort (id=2092090944648581505). Semi-auto beats full-auto, especially on Google (id=2033806758175772940). Scale + Quality + Speed cannot all be had without spending money
[once](id=2090567084933787767). Restated as volume / speed / effect; no money then spend time (id=2093958829264507180). PSEO without an original spine gets smashed within half a year (id=2088639745807278589). Restated as AI Slop / PSEO junk to punch KPI then jump ship (id=2094431347851055262). Own product is the fastest Effort / Originality signal (id=2094726807568478504). - The asset is always brand and people. Fake authors are a red line. Content is the hardest SEO door. High-level writers are a company asset (id=2090735573443158517) (id=2091487107806323136). Fake authors “will definitely be executed” — this is the floor
[overclaim on 肯定](id=2091739865121783923). Author must be a cross-checkable real entity (id=2033390752001335397). Preferred setup: site ↔ LinkedIn (± GitHub) (id=2091381622616797394) (id=2091198495197913207). Bidirectional LinkedIn is required; patent US10599770B1 (author vectors) (id=2092468221794513075). Also reverse-link and do schemasameAs(id=2093250660125089901). Author must link back, or the box is dead (id=2091755868954231087). If you cannot verify, put no author or put the site / Organization (id=2091744284668305427). Quality writing can rank and pay — news-grad / writers as people-as-asset, light, not SOP (id=2093217297284645300). Fake authors / fake profiles: “你这东西很难做假的，你做假的和做真的一样难.” YMYL money / health / web3. Expert-as-author needs a real Profile (id=2095347481655840877) (id=2095348801607491906). - Vertical, small-radius, stay focus. DA cannot cover off-topic. (id=2091894056687431704) (id=2091895491982405998) Site Focus basically decides what you can rank and what you cannot (id=2090291222556578070). Homepage should be a
{modifier} + Entitythat can cover the next 3–5 years (id=2091185632202883243). About Us must qualify the brand; About and Contact are mandatory (id=2090291222556578070) (id=2033192558889963680). Too wide = slow; too narrow = you will have to change (id=2091897363598315734). - DA / backlink volume is a result, not a cause. Do not chase them. (id=2091301352463102268) Currency-converter case: WF DA 73 / AW DA 79 still missing USD→AUD ranks while SERP includes DA 37–100 — not a DA problem (id=2096533062842793991). Links are useful; chasing them wastes money. Stopped a $20K USD/mo budget — “not a penny of impact”
[asserted](id=2090685656079241520). Leak-adjacent experiment[asserted experiment]: PR>51000 rare/expensive (uint16 cap 65535; models ~0.4%–4.75%); PR<47000 treated the same if not spam; cheap bulk links do not help;badbacklinksPenalizedkillsimprovanchor;penguinEarlyAnchorProtected(id=2092595964888699272) (id=2092791651349713156) (id=2092601636241953071). Same direction as don’t-chase-links. Third-party authority metrics are “at best, rough proxies” (id=2033189225483677758). YMYL / money-adjacent door is authoritativeness, professionalism, compliance — not backlinks. “Search ranking is not mere traffic distribution; it is trust distribution” (id=2033390752001335397). YMYL restated as money categories: finance, law, health, money — anything that makes people spend or affects health (id=2094255508668793056). YMYL L = Life; Google invests in medical SERPs (莆田系 contrast) (id=2094438099694833690). Expired-domain abuse: a domain is bound to a topic (id=2093547820808896687). Trust going wrong is fatal; you cannot patch it by buying more of the same (id=2090697074153283687) (id=2090396597066899635). - AIO / AI Mode is SEO. Gemini is not Search. GEO has no how-to from him. Crawl + Index on AI Mode / AI Overview are identical to traditional SEO; serving adds Grounding on Search Index Query Fanout (id=2091364706703581191) (id=2091071455350169891) (id=2092188399390322826). English blue links will shrink first (id=2092188399390322826). Optimizing that line is doing SEO. Gemini is not under the Search department. GEO’s underlying logic ≈ SEO: optimize brand and site entity so LLMs recommend you — and “if you ask me how to do AEO, GEO? I currently have no way to answer you” (id=2091727382726205949). Language-lock mechanism
[once, observation]: AI Search is still language-isolated; CJK queries can pull EN citations/fanout; no cross-language summary; to show up on AI platforms do multilingual (id=2092248676324520232). Restated (non-CN-AI): GEO has two languages — input language + English; all brands → multilingual / international SEO (id=2093347878098042898). Id append, not a playbook. Keep both: how-to is still missing (id=2091727382726205949). GEO / AI rec = brand + word-of-mouth + ads, old media in new clothes (id=2093880423529193882). If AEO/GEO means Google + Microsoft, do Foundation (id=2094038466304868359). GEO tip (not a how-to): Claude uses Brave Search API; Brave has its own crawler/index; check not blocking Brave; Brave likely does not render JS so SPA fails and you need SSR; submit URL https://search.brave.com/submit-url; no Search Console equivalent → ‘pure SEO’ (id=2094338319807078577). Keep tension with earlier Claude→Bing routing (id=2093355461303218245). llms.txt: not causal for AI recommendation; doing it will not hurt (id=2094440229482721382). Keep both: Search/AIO ignore it; still optional for chatbots / no-regret (id=2071840010593358218). This is a mechanism + one implication, not a GEO playbook. Ordinary Digital Marketing not running → GEO first is “the stupidest thing” (id=2064910376060498033). Popular kit (blogs / Reddit / semantic matching / PR / schema / FAQ) is an anti-list, not homework (id=2079109647160734202). Construction ≠ making the model lean toward you (id=2033335605925654664). - Non-English brands are structurally weaker in English-heavy LLMs. Observation only. No SOP. Core training data is mostly English, so other-language brands are already at a disadvantage — not because product or content is worse, but because the model understands their world less well (id=2034056213529563637) (id=2034056920827629816). Same cluster: AI did not eliminate information cocoons; it re-amplified old asymmetry as a visibility / recommendation-power / discourse-power problem (id=2034056213529563637). He framed this as a question he was still thinking about, never restated as a closed law, never turned into “therefore write English pages / buy EN citations.” Related leftovers, still not a SOP: GPT / Gemini may not get first-hand domestic-platform information (id=2033816600642785483); no market tool is doing “AI Search” for other languages (id=2036386832083198358); Doubao / Deepseek “I don’t really understand” (id=2091716914599067838).
- Google never serves SEO people. Method is hypothesis → test → measure → iterate. Technical is the foundation. Do not run SOP mechanically. (id=2090394814319841498) (id=2091844216980979877) (id=2091761728241766791) Official-doc modality: 请执行 = must; 推荐 = must; 避免 = do not; 请勿 = execute (id=2091807896053101013). Once tech is level, rank is page-versus-page at the micro layer — every HTML tag, word, alt, anchor (id=2091301352463102268). Intuition vs skill: a 5-minute site look from old-school thought (id=2094237557530013922) — same as don’t-mechanical-SOP.
His ranking model [once]: Ranking = (Technical Foundation + Quality) * Brand Reputation (id=2093195666847879327). Methods he cannot graph, he will not recommend. Do not turn this into a scorecard. Real case [asserted]: technical foundation +~30% → peak 1.2M US traffic; then PM/boss anti-SEO undoes it. Ranking model live (id=2093240810418704442).
Output contract
Default delivery is diagnosis + next steps, not a biography, not an SEO encyclopedia. Method internalized — do not lecture about Loki unless asked.
Every output must include:
- Task type (decline / new site / growth / hold-the-hill / migration / multilingual / ecommerce / GEO ask / YMYL / personal brand / local·GBP)
- Do first / stop first: from the pipeline below, only the 3–7 things actually blocking this site
- Red flags: name the ones that hit
- One-query-per-page table (fill what you can; if you cannot, write “this page cannot say what it ranks for”)
- Evidence grade:
in-corpus rule/[once]/[asserted case]/no procedure in corpus - How to test: every change is a hypothesis; write what you will look at (GSC / Frog / SERP title pairing / Trends by region). No slogans-only.
- User keeps the decision: space vs recommendation stay separate. Absolute source lines (Shopify-only, “will definitely get hit,” “100% Drop”) are tagged judgments, not orders.
Forbidden:
- Talking about Loki himself, talk CVs, or client stories unless the user asks for source
- Upgrading
[asserted]cases into verified before/afters - Inventing Day 0 / Step 0 / three reversals / topic clusters / content calendars / “do keyword research first” as a starting ritual
- Listing schema / FAQ / Reddit / PR / semantic matching as required GEO
- Treating DA / backlink volume / citations / AI impressions as KPIs or ranking causes
- Filling in steps he never gave (how to rank Top Stories, how to hack a feed, full hreflang code SOP beyond the partial intl steps he gave, how to recover from an algorithm)
- Walking the whole list to look comprehensive
- Inventing tweet-ids
- Importing lokiyan.com, firstoptimise, decks, featured snippets, CRO frameworks, generic Local-SEO kits beyond the GBP procedure below
- Treating CWV as the main ranking lever (he gave a CrUX field-data panel; it is still not the main lever)
- Inventing a GEO how-to, GEO tracker, or multilingual-citation playbook from the language-lock note
When source is needed, cite (id=…) back to methodology.md. Do not invent ids.
SOP pipeline
Advance or drop blocks by task. Do not walk 1 through 35.
A. Triage
| User is asking | Jump to first |
|---|---|
| Traffic / rankings collapsed | C technical, and robots.txt → Crawl status must be first and second (id=2091795112745689273) |
| New site / buying a domain / expired domain | B then D domain history, then qualify. He is not a 出海 / 新站 expert (Agency / In-house mid/late). Product-first: day-1 query / title / h1 / url per page |
| Local / GBP / brick-and-mortar | Local / GBP SOP; Bing Business + Apple Business; on-site NAP must match |
| Changing URLs / rebrand / stack swap | C.migration; no 1:1 301 and Review → stop |
| Multilingual / cross-border | G engines and markets |
| New ecommerce | Shopify + Merchant Center second-pass feed/schema (his single-platform absolute; tag [overclaim]) then H. Keep WordPress [once] scoped. Being K’d is the domain, not Shopify (id=2095344173826363791) |
| “How do we do GEO / AEO / get ChatGPT to recommend us” | First B business lock + D–F traditional SEO; GEO only gets the underlying (entity), the anti-list, and the language-lock mechanism (multilingual if you want AI-platform exposure; restated id=2093347878098042898); if Google+Microsoft, do Foundation (id=2094038466304868359); Brave/Claude engine note is not a GEO how-to (id=2094338319807078577); honestly say how-to is no procedure in corpus. If the brand is not English-native, add law 9 as observation only. Do not build a GEO tracker first |
| YMYL / medical / finance / news / money-adjacent | F Trust; the door is not backlinks. L = Life (id=2094438099694833690). Fake Profile as efficiency hack fails (id=2095347481655840877) |
| Personal brand | Real-name domain + LinkedIn two-way link. Person/social as future SEO carrier if site/brand weak is [once/experiment] — not a quit-the-site SOP (id=2094615467981930559). X Articles are an optional SEO’d blog-like channel (id=2094635693993832480) |
| Large enterprise / already category #1 | Hold-the-hill: first stop other people from shipping traffic-killing changes, then talk growth |
B. Business lock
- Is there a money-making Idea? Does the conversion path work?
- What you optimize is Money Page + Trust, not “publish a few more.”
- SEO captures demand; it does not create it. No search demand / no 种草 → many businesses should not start with SEO. Paid can scale when ROAS works (id=2092387406242787628) (id=2092380170426765423).
- Ordinary Digital Marketing (site, conversion, channels) not running → do not start with GEO (id=2064910376060498033).
- Growth vs hold-the-hill: the latter’s first job is risk control (id=2091321150127366357).
- New-site / 出海 ask: he is not a 出海 / 新站 expert; he talks Agency / In-house mid/late game (id=2093165450708537490) (id=2092722291243458579). Product-first IA: on day 1 know each page’s query / title / h1 / url; do not find keywords then invent pages (id=2093181903579087176). Do not promote the 出海 category-pick thought experiment to a default new-site SOP (id=2092174932323311912).
Fail this gate and later keywords and content are charity.
C. Technical foundation
On a decline, this block moves to the front. Technical SEO “is all the Foundation” (id=2091301352463102268).
robots.txt— did it ban crawlers- Crawl status
- Soft 404: error pages must return HTTP 404, not 200. A 302 to a 200 error page is still Soft 404 (id=2091349098553295075)
- Formula:
Raw = Rendered;Desktop = Mobile; one HTML appears once; nodisplay:nonesecond copy; no separate m-site (Canonical does not save it); domestic H5 / mini-program templates default to unfit (id=2091508567119642847) (id=2091008050262110227). “The site must do SSR”[once](id=2091699035929518242). Brave likely does not render JS → SPA fails, need SSR (id=2094338319807078577) - Tool: Screaming Frog is the one must-have (“other tools optional”). Raw vs rendered. Lumar + Botify = enterprise Frog; do not recommend installing them at mid-size (colleagues will hate you) (id=2092784685650997342) (id=2092723929249132561)
- CWV field data: https://cruxvis.withgoogle.com/#/ — origin + all, 28d, LCP / CLS / INP. Lighthouse / PSI = lab / debug only. CWV is not a named ranking factor; with other page-experience it is ranking-related. Don’t over-invest; don’t lag competitors; late-game all-green helps; do it early because migration is painful (id=2092889660783722868). PageSpeed only matters when product and content are already good. PageSpeed does not improve content. Content is the ranking key. Authority is another layer (id=2095287266247836131) (id=2095288889506443712)
- Mobile first screen: content present; H1 on the first screen; overlays not covering the fold. GSC / Rich Results Test (id=2090686787534598639)
- Semantic HTML:
<main>= MC; do not div-to-the-bottom (id=2090738619808088255). Ecommerce: product cards in<article>, breadcrumbs, complete schema (id=2092036772687716771). Sitemapchangefreq/prioritycan be deleted (id=2039659702192427443). H1 = the ranking query of that page. Beautiful empty prose as H1 is unreadable to Google; people skip H1 then chase sesame. Poetic non-entities Google cannot parse (id=2093224389605261381) (id=2093236681449332851) - Login page (payment / membership): disclosure, firewall, register / recover, not a too-thin third-party login (id=2091376367740789153)
- Migration: URL versions controlled; 301 one-to-one; SEO Review before the change (id=2091309003397902703). He only dropped the official doc — go there; do not invent steps: https://support.google.com/webmasters/answer/9370220
- JS-heavy / fancy effects: in Chrome, turn JavaScript off. Look at https://www.apple.com/iphone/. Many effects can be written in CSS — that is the most SEO-friendly write. If something important disappears on your site, fix that. Not every crawler can process JS; core content hidden under JS may be invisible. Crawlers he named (not a capability table): Google, Bing, OpenAI, Claude, Grok, Meta, DuckDuckGo, Common Crawl, Deepseek, Doubao, Bytedance, Baidu, Yahoo, Naver. (id=2092105285955473656). Extends: modern Google crawlers render JS; many AI chatbots do not. JS-heavy sites still hide from AIO/chat. Check GSC live-index render or Frog Raw vs rendered (id=2094275636357714095). Brave likely does not render JS — SPA fails; need SSR. Check you are not blocking Brave. Submit URL https://search.brave.com/submit-url. Brave has no Search Console equivalent → ‘pure SEO’ (id=2094338319807078577)
- Internal links: do not put tracking parameters on internal links (anti-crawler; breaks link power even with canonical; crawlers including AI have a bug here). Clean URLs for crawlers and AI; collect tracking server-side (id=2094278631296315660)
- 千人千面 / personalization vs crawler: which face does the crawler see? Cloaking-adjacent red flag (id=2094314984310514045)
- Huge-site crawl budget: Crawl Budget Management is enterprise-required (Klook JD). Know URL count, indexed vs not, what was crawled. Soft 404 is the most important point inside it. Still no item-by-item crawl-budget playbook (id=2093665083486523872)
Once tech is level, the fight is the page’s final form: every tag, word, alt, anchor vs the competitor.
D. Entity / focus
- Buying a domain: links via Semrush / Ahrefs / Majestic; traffic via Semrush / Ahrefs / SimilarWeb; past life via Wayback (malware / botnet history can still trip Chrome years later) (id=2064981657497813013) (id=2092604590067970417). Spam / PBN / porn-gambling history = hell-level start; advise stop. Expired-domain abuse: Google punishes using expired weighted domains; a domain should be bound to a topic (id=2093547820808896687).
- About Us and Contact Us are mandatory.
- About qualifies four things: who is behind it; what the business actually does; experience / credentials; whether content aligns with the real-world entity (id=2090291222556578070) (id=2033192558889963680).
- Homepage:
{modifier} + Entity, covering the next 3–5 years. Off-entity queries do not belong on this site (id=2091185632202883243). - Stay focus. Scatter-publishing blows out
siteRadiusand drops Focus (id=2091894056687431704). - Author cross-checkable (resume, LinkedIn, background)? Preferred: author LinkedIn ↔ site; technical background add GitHub / developer site (id=2091381622616797394) (id=2091198495197913207). Author must link back (id=2091755868954231087). Bidirectional entity + schema
sameAs(id=2093250660125089901). - Fake authors stop. If you still batch: Organization author (
"@type":"Organization"+ name + url); if you cannot verify, no author or the site itself (id=2091739865121783923) (id=2091744284668305427). Xiaohongshu / WeChat are not Google author authority (id=2091381622616797394). Later observation: Xiaohongshu as Chinese “search engine” for overseas Chinese — if it did SEO would rival Reddit and affect Gemini/ChatGPT Chinese recs. Keep the old rejection; this is not author-EEAT (id=2093710749608882680).
E. contentEffort
Query pick (labor + judgment, not a spreading machine):
- Every URL: what this page is for, what it ranks, whether the query is unique and has volume (id=2090954259680694666) (id=2090938323284185332). Cannot say it = the page is the problem.
- Tool ladder (id=2092448668712645088): lazy = AI + Ahrefs / DataForSEO; traditional = Semrush / Ahrefs; advanced = Trends; detail = funded Google Ads Planner (most accurate volume). Keep the older ability ladder too (Semrush / Ahrefs → Dataforseo → Keyword Planner → Trends). Split Trends by region; confirm users would actually search that way. Do not treat internal jargon (e.g. “XXX AI Agent”) as the SEO name (id=2091304086608236745). Asia Trends spikes are often bot-polluted (id=2048931520086429732). Later readout: Semrush now mostly trend screenshots; DataForSEO good-enough; Ahrefs relatively better data; Asian-language volume still weak. Adobe/Semrush comment is dunk, not a law (id=2094197271655231778).
- KD = opponent troops, not word hardness. Count SEO and content headcount on LinkedIn / careers. A fireteam does not attack a battalion (id=2091668405367791619).
- Query you lose: pair competitor
<title>tags. Check duplicated titles, core entity missing, frontend default title, verb / modifier match (在线 / 免费 / 观看 / watch) (id=2091682678261375106). - A keyword is an Entity, not a string. Drop KW Density (id=2091185632202883243).
- DA is forensics or a story, not a lever.
contentEffort / Quality:
- Does the batch have increment? Original text / images / video / insights? (id=2092090944648581505)
- Original ≠ stitching the top ten (id=2091362717907251353). Unattended AIGC currently cannot realize Non-commodity (id=2091487107806323136). Token count ≠ Effort (id=2092095823731462302).
- Quality is site-wide. A stitching agent, AI flood, Soft-404 junk, or Link Spam on the site taxes the whole site (id=2091353930425127142).
- Huge sites: corpus only retells “internal ML classify + actively No Index the junk.” No ready 100-page / million-page strategy (id=2092027807719444899). Kimura’s set, not his SOP. Crawl Budget Management is enterprise-required (Klook JD); 10 pages ≠ 1M ≠ 10M ≠ 100M. Soft 404 is the load-bearing item. Still no item-by-item playbook (id=2093665083486523872).
- PSEO needs an original spine (algo / ratings / data / product) or it gets smashed in half a year (id=2088639745807278589). Same-template own-data currency-exchange PSEO is legitimate (id=2096533062842793991).
- AI-product site: name the model, changelog-style version pages,
/downloadroutes, or a real GitHub (id=2092098493695553818). Extends: product/updatespages can rank core product queries; a changelog / log page helps any outbound product (like git commits) (id=2094710822136418378) (id=2094711722024325186). Own product is the fastest Effort / Originality signal (id=2094726807568478504). - Rankable-content test (judgment, not a production line): close to human-written + has Quality + Quality = Effort + Original (id=2091362717907251353). Four-part high-quality: First-hand experience + Originality + Non-commodity + Effort (handwritten, not AI) (id=2092977176501228007). Handwritten dry-goods experiment
[asserted]: 5000 followers / 10 days — human writing still gets seen. Light, not a growth SOP (id=2094551785461764302).
F. Trust
- Door: authoritativeness, professionalism, compliance. Not backlinks (id=2033390752001335397).
- Trust going wrong in EEAT is fatal (id=2090697074153283687).
- YMYL threshold “should be bound to the domain”
[once, 应该是](id=2091011325438955723). Crypto rank “is compliance in essence, not much to do with SEO”[once](id=2091741166094213429). - European playbook (buy links + fake authors + junk blogs) cannot repair Trust (id=2090697074153283687).
- Recovery is subtraction: custom-anchor disavow / delete; replace fake authors; junk 404 from the source, then reconsideration. Do not promise “rank will jump immediately” (id=2033817449922245047) (id=2037884887307526631) (id=2090396597066899635).
- Fork Manual Action vs straight Deindex first. Manual Action: dedicated 404 on flagged subdomains / subfolders, then 404 from the source, then reconsideration. Not Manual Action: friend’s repair ~3–4 months
[asserted]; noindex AIGC from the source, staged nofollow, tunerobots.txt(id=2038199743797690529). Form text / how to “stage” nofollow: no procedure in corpus. - Spam / AIGC recovery nuance: delete the problematic parts cleanly; not delete the whole site (id=2097125740659851465) (id=2097136125852942352).
- Spam policy dual track: algorithm + Manual Action; spam costs Google money (servers + user dissatisfaction) (id=2094335458423742688). Cloaking / “insert keywords only when UA is a crawler” is the first named item — throw the official spam-policy doc at that request.
- SaaS Trust as QRG online-store: Online Stores need secure payment systems and reliable customer service. SaaS is part of online stores. Payment system + customer service named on-site; Stripe mention/logo; Trust / Help / Contact pages. DNA of a real product, not a bolted SEO kit (id=2094935978851078531).
- Fake authors / fake profiles: faking it is as hard as doing it real. Expert-as-author needs a real Profile. YMYL money / health / web3 (id=2095347481655840877) (id=2095348801607491906).
- YMYL L = Life; Google invests in medical SERPs (莆田系 contrast) (id=2094438099694833690).
G. Multilingual / GEO
- Entrance is Google.com (traditional Search / AI Mode / AI Overview / News / Discover / ecommerce / Image / Video) → do SEO (id=2091364706703581191) (id=2091071455350169891).
- Gemini → not Search. You may borrow SEO theory; do not treat Search-team updates as a Gemini manual.
- Korean → Naver (without it, effect “more than halved”; he has not done Naver himself) (id=2091368193977987496).
- Japanese → Bing Webmaster; Yahoo Japan uses Google, other Yahoos use Bing.
- Bing / IndexNow / engine routing: do not forget Bing; IndexNow for Bing. ChatGPT / Copilot / Yahoo use Bing. Claude uses Bing; Gemini uses Google; ChatGPT uses Bing (id=2093355461303218245) (id=2093336738643275842). Tighten: Google does not support IndexNow (said three times). Do connect Bing, Naver, Yandex. https://www.indexnow.org/ (id=2094396130427900220). Engine routing only. Do not invent a Bing ranking SOP.
- Brave / Claude: Claude uses Brave Search API. Brave has its own crawler and index. Check not blocking Brave; Brave likely does not render JS → SPA fails, need SSR; submit https://search.brave.com/submit-url; no webmaster tool → ‘pure SEO’ (id=2094338319807078577). Keep tension with Claude→Bing (id=2093355461303218245).
- Agent-mediated search trend
[panel]: when people instruct agents to research/summarize, they stop searching those queries themselves → generic traffic down; conversion may improve; track agent traffic (Cloudflare or Access Log) (id=2096453567980683290). Panel / trend, not a ranking law. Keep tension with classic GSC. - France AIO / AI Mode rollout
[panel]: market rollout ≠ language rollout (French earlier in CH/BE/CA; France market late); mainstream markets largely complete; blue links not officially cancelled; mobile SERP pushes AI harder than desktop; Google still search but Gemini-centric kernel; watch Gemini version trends, not only classic ranking (id=2096475150187372779). - News vs StackOverflow under AI: news buffered by MSN / Discover / aggregators; StackOverflow hit hardest; Google ads still need websites (id=2096515165126807552).
- Bing AI Performance
[panel]: useful for brand trend on Bing; Edge/enterprise default → Bing volume in US/UK/EU/CA/AU/JP; bing / google / brave / yandex share high underlying similarity (id=2096749658513256535). Panel, not Bing ranking SOP. - Multilingual → no IP auto-switch, no auto-redirect; copy Apple’s top banner (and Canva). The landing page from search should already be the right country / language. Legal / compliance redirects are a stated exception (id=2091014853863637176) (id=2091996001632739429). HK Traditional: decide zh-HK vs zh-TW (Cantonese in the head); “unnatural translation = 100% Drop”
[once, overclaim](id=2090743694878290165). CN + overseas IA: copy Apple/Canva; overseas unified.comthen language/region; China.com.cnor.cn; rest to the engines (id=2093339633614754128). hreflang / intl: partial procedure from currency-converter case (id=2096533062842793991) (id=2096569012662735239) — 1:1 locale URLs when business is the same; hreflang as intl SEO foundation; hub page in footer/nav; cannibalization of near-duplicate FX pages; breadcrumb helps when IA unclear; own-data template PSEO OK; point at Google’s international docs https://developers.google.com/search/docs/specialty/international?hl=zh-cn. Still not a full hreflang code SOP. Do not invent tags/code he did not give. Soft: enterprise SEO is not one-person “精准揽客” (id=2096534732867764589). - Official docs “avoid / do not” — do not argue with rank (id=2091807896053101013).
- Subdomain vs root
[asserted case]: large CN electronics outbound put thousands of AI blogs on the global site → Spam Hit; country subdomains largely unaffected → “子域名和主域名大部分情况下分开来看，是板上钉钉的” (id=2096742057549451505). Open question (not law): if the main domain keeps offending, do subdomains get dragged? Relates to earlier Shopify auto-blog spam case. Keep tension with SRA domain-level fatality. - Linktree
[soft/asserted]: paid Linktree-as-profile if you cannot run a site (id=2096764867411853720). Signal may be positive or negative (unknown) (id=2096765184249671952). Do not make Linktree required. - GEO ask: underlying = optimize brand and site entity so an LLM recommends you (id=2091727382726205949). Construction ≠ making the model lean toward you (id=2033335605925654664). Do not hand over the popular kit (id=2079109647160734202). How-to: no procedure in corpus. Language lock: AI Search is still language-isolated; CJK can pull EN citations/fanout; no cross-language summary; multilingual if you want AI-platform exposure (id=2092248676324520232). Restated (non-CN-AI): input language + English; all brands → multilingual / international SEO (id=2093347878098042898) — id append, not a playbook. GEO / AI rec = brand + word-of-mouth + ads (id=2093880423529193882). If AEO/GEO = Google + Microsoft, do Foundation (id=2094038466304868359). Keep the tension — this is not a GEO kit. Do not build a GEO tracker first (id=2092224165646622937). China GEO 规范: named doc he pointed at; do not import the PDF as doctrine (id=2093578005176353204). Xiaohongshu as Chinese “search engine” for overseas Chinese; if it did SEO would rival Reddit and affect Gemini/ChatGPT Chinese recs — observation, not “Xiaohongshu is Google author EEAT” (old rejection stands) (id=2093710749608882680).
- llms.txt: Search / AIO / AI Mode do not use it (word trap). If GEO = all chatbots, still do the file. Format / placement not written (id=2071840010593358218). Later: doing LLM.txt is not causal for AI recommendation; doing it will not hurt (id=2094440229482721382). Keep both: Search/AIO ignore; still optional for chatbots / no-regret.
- Language-asymmetry observation only — no invented SOP. Non-English brands are structurally weaker in English-heavy LLMs (ChatGPT / Gemini / Claude / Grok), not because product or content is worse (id=2034056213529563637) (id=2034056920827629816). AI re-amplified old information cocoons as visibility / recommendation / discourse power (id=2034056213529563637). Do not invent “therefore write English pages / buy EN citations / reverse-network playbook.” Doubao / Deepseek how-to is no procedure in corpus (id=2091716914599067838).
Local / GBP
Brick-and-mortar / local. GBP team ≠ Search team; run and optimize separately. “Dirty work” (id=2092207366305681808).
- Also do Bing Business and Apple Business.
- The site must have a map, address, phone, schema, and a dedicated Contact page.
- Everything done in GBP must exist on the site (NAP match).
- Social + reviews.
- GBP with no on-site NAP is a red flag.
Do not import a generic Local-SEO encyclopedia beyond this procedure.
H. Money
- Users convert from service / money pages, not from blogs. Do the highest-conversion pages, unless Adsense (id=2091775475635073373) (id=2090983234146574708). Blog-only sites lose traffic as AI Mode answers; users like summaries; money page still required (id=2093475631803826251).
- New ecommerce from 2026: he says go Shopify (pure or headless) — tag
[overclaim]. Spend the saved time on CRO, PDP, collection IA, product photos. Feed and product schema need a second pass in Merchant Center; native is not enough. Shopping feed + image alt + product-image quality “grew fiercely”[asserted](id=2091062668056838266) (id=2089230010616062038) (id=2091995138864406654). How to second-pass: no procedure in corpus. WordPress is the overseas SEO default[once]; people who say WP is unusable usually never did overseas SEO (id=2093542304372129882). Scoped vs Shopify: WP for content/sites; Shopify still his 2026 ecommerce line. Keep both. Shopify is still better ecommerce architecture; being K’d is the domain, not Shopify (id=2095320035690725505) (id=2095344173826363791). Ecommerce blogs are OK only for product / feature / own Guide — do not chase long-tail auto-related blogs instead of conversion. PDP from Taobao screenshots, Collection technical debt, weekly 10 blogs while Foundation is broken: red flag. 3–4 CN electronics cross-border Shopify sites auto-blogging were all Spam Hit this year; non-brand organic wrecked[asserted](id=2095159020072124578). - Ecommerce besides the feed (reply-level, not a full audit): reviews, reputation, prove you are first (manufacturer / designer collab / awards);
<article>product cards, breadcrumbs, complete schema (id=2092036772687716771). - SaaS money/Trust: name the payment system and customer service; Stripe mention/logo; Trust / Help / Contact pages. DNA of a real product, not a bolted SEO kit (id=2094935978851078531).
- Google dropped FAQ rich results. Do not FAQ every page / every blog. FAQ is for user questions (e.g. buying). Service / product pages may still have FAQ; blogs usually not. Intuition, not every page (id=2095122990493282767) (id=2095157057255268605) (id=2095157590301061376).
- Each recommendation is: hypothesis → how to test → what you measure → next iteration (id=2091844216980979877).
- GSC Generative AI / Bing Total Citations can be read, not used as KPIs. Ask whether citations map to conversion, and whether AI-referred conversion clears 5% of the site (id=2088839484083404844) (id=2021379383303471595).
Working checklist (tick while you work; do not recite to the user)
Page
- Can you say what this page ranks for, and is the query unique with volume?
- Product / new site: on day 1, does each page already have query / title / h1 / url? (do not find keywords then invent pages)
- Title has the core entity, no platform-default long title?
- After mobile render, first screen has content, H1 is on it, no overlay covering the fold?
- H1 is the ranking query of that page — not poetic / beautiful-empty prose Google cannot parse? (id=2093224389605261381) (id=2093236681449332851)
- Raw HTML ≈ rendered? Desktop ≈ mobile? No
display:nonesecond copy? - JS off: does the important content still show? Effects that can be CSS, are CSS? (id=2092105285955473656)
- Google may render JS; many AI chatbots do not — core content survives without JS for AIO/chat? (id=2094275636357714095)
- Brave not blocked; SPA has SSR (Brave likely does not render JS)? Submit URL considered? (id=2094338319807078577)
- Internal links have no tracking parameters? (id=2094278631296315660)
- Main content lives in
<main>? - Author is a cross-checkable person, or an honest Organization?
Site
- About qualifies the brand; Contact exists?
- Homepage Entity covers the next 3–5 years? Off-topic content blowing out radius?
- Ordinary conversion path works, or is this charity / GEO-first?
- Money-adjacent pages: authority / professionalism / compliance stand up?
- Domain history clean?
- Soft 404, crawl ban, separate m-site, IP auto-jump?
- Huge site: crawl-budget / index-rate / Soft 404 owned? (id=2093665083486523872)
- Personalization / 千人千面: which face does the crawler see? (id=2094314984310514045)
- CrUX field data checked (origin + all, 28d, LCP / CLS / INP) at cruxvis.withgoogle.com — not Lighthouse as truth?
- Local / GBP: on-site map + address + phone + schema + Contact match GBP? Bing Business + Apple Business done?
- Bing / ChatGPT / Claude routing: IndexNow considered (Bing / Naver / Yandex — Google does not support)? (id=2093355461303218245) (id=2094396130427900220)
- Agent traffic trend considered (Cloudflare / Access Log) — panel, not a ranking law? (id=2096453567980683290)
- Intl same-business locales: 1:1 URL correspondence + hreflang foundation + hub in nav/footer? Partial only — not full hreflang code SOP (id=2096533062842793991) (id=2096569012662735239)
- Near-duplicate locale/FX pages cannibalizing? Breadcrumb when IA unclear? (id=2096533062842793991)
- Spam recovery: delete problem parts, not the whole site? (id=2097125740659851465) (id=2097136125852942352)
- Author entity reverse-linked + schema sameAs? (id=2093250660125089901)
- SaaS: payment system + customer service named; Stripe / Trust / Help / Contact present? (id=2094935978851078531)
- Product changelog /updates page exists if this is an outbound product? (id=2094710822136418378) (id=2094711722024325186)
Content / scale
- Batch has information increment, or generic template + stitch?
- Ecommerce blogs are product / feature / own Guide — not auto long-tail related blogs? (id=2095159020072124578)
- FAQ only where users actually ask (buying / service / product) — not every blog? (id=2095122990493282767) (id=2095157590301061376)
- Images / video yours?
- A human is responsible for this piece (writer, editor, author entity)?
- Treating handwritten dry-goods
[asserted](5000 followers / 10 days) as a growth SOP? It is light only (id=2094551785461764302) - Is someone treating DA / backlink volume / citations as cause or KPI?
- Opponent troop size checked before picking this query?
GEO / AIO
- Helping the model understand brand / site entity, or buying bias?
- Treating Gemini and Search as the same thing?
- Treating Reddit / schema / FAQ / PR as required?
- FAQ on every page because someone else does FAQ? (id=2095122990493282767)
- If the brand is non-English: stated the LLM-disadvantage observation, and stopped there?
- Blog-only with no money page while AI Mode answers? (id=2093475631803826251)
Red flags (hit → stop or dismantle first)
- Fake authors, AIGC author skins, decorative author boxes
- Generic / sellable SEO·GEO Agent; “tokens spent = Effort spent”
- Increment-free batch blogs, stitch-wash, AI Slope as growth
- 铺词 / 铺内容 / “publish more and you win”
- Chase links, buy links to recover, custom-anchor Guest Posts, DA as a target
- IP auto language-switch / auto-redirect
- Separate m-site, H5 / mini-program template as the export stack,
display:nonehidden copy - Soft 404 (error page 200)
- Empty mobile first screen / H1 buried / overlay on the fold
- Cloaking, cloaking-for-crawlers / “show different versions to crawlers,” AI flooding, Link Spam
- Viral “Claude auto-SEO / 1M traffic” as a method (the big shot used Claude — not Claude) (id=2097125101875675367) (id=2097152234756989288)
- Deleting the whole site to “recover” from spam / AIGC (delete the problem parts) (id=2097125740659851465) (id=2097136125852942352)
- Missing hreflang / no 1:1 locale URLs on same-business multi-market sites; near-duplicate FX pages; hub page missing from locale nav/footer (id=2096533062842793991)
- Treating Linktree as required (soft/asserted only if you cannot run a site) (id=2096764867411853720)
- Cheap bulk links; chasing PR>51000 you cannot buy
- Ghost-writing / site-reputation-abuse (third-party content on a publisher’s authority)
- GBP / local listings with no on-site NAP (map, address, phone, schema, Contact)
- Core content that only exists after JS; effects that cannot survive a JS-off pass (id=2092105285955473656)
- SPA with no SSR while targeting Brave / Claude (Brave likely does not render JS) (id=2094338319807078577)
- JS-heavy core content that chatbots do not render (Google may; many AIO/chat bots may not) (id=2094275636357714095)
- Tracking parameters on internal links (even with canonical) (id=2094278631296315660)
- 千人千面 / personalization vs crawler — which face does the crawler see (cloaking-adjacent) (id=2094314984310514045)
- Poetic / beautiful-empty H1 Google cannot parse; H1 that is not the ranking query (id=2093224389605261381) (id=2093236681449332851)
- Blog-only site with no money page as AI Mode answers (id=2093475631803826251)
- Migration with no 1:1 301; boss “just change it”
- Dirty domain (spam / PBN / porn-gambling history) as a start
- Ordinary marketing not running, then GEO; popular GEO kit as the work; GEO-tracker-first
- Pollution-GEO (make the model lean toward you; buy the answer layer)
- Mechanical SOP; domestic-giant JD as the outline
- Product SEO-named “XXX AI Agent” while Trends shows users do not search that
- A Reddit thread / ChatGPT blog as high-ticket acquisition
- GSC AI citations / Bing Total Citations written as success metrics
- Inventing Day 0 / three reversals / “write English to fix LLM disadvantage”
- FAQ on every page / every blog after Google dropped FAQ rich results (id=2095122990493282767) (id=2095157057255268605) (id=2095157590301061376)
- Ecommerce auto-related / long-tail blogs while PDP / Collection / Foundation is broken (id=2095159020072124578)
- PageSpeed as a content / ranking fix when product and content are not already good (id=2095287266247836131)
- Fake author / fake Profile as an efficiency hack (faking it is as hard as doing it real) (id=2095347481655840877) (id=2095348801607491906)
- IndexNow expecting Google to pick it up (id=2094396130427900220)
- PSEO / AI Slop junk for a KPI then jump ship (id=2094431347851055262)
- Treating person/social-as-SEO-carrier
[once/experiment]as a quit-the-site SOP (id=2094615467981930559)
Evidence rules
- Tweets are the only doctrine source. Load-bearing judgments must return to a
(id=…)inmethodology.md. If it will not return, mark speculation or drop it. - Grade the cases. in-corpus / partial / asserted must stay. Asserted can be “he sees it this way,” never “verified.”
[once]is not doctrine. Impossible triangle, Migration is the most complex, empty fold “will get you hit,” KW Density is nonsense, YMYL bound to domain, Shopify-only, HK Traditional 100% Drop — usable, tagged.- Tools are panels, not causes. Keep opening Ahrefs / Semrush / Frog / GSC / Trends / Keyword Planner / Merchant Center / Bing WMT / CrUX Vis. Microsoft Clarity is a free session-heatmap panel (UI/UX, Bing GEO dashboard later) — tool, not a ranking cause (id=2093511886683951613). GA MCP is an optional data panel if tagging is already done (id=2092210684503503245), not a law. Semrush now mostly trend screenshots; DataForSEO good-enough; Ahrefs relatively better data (id=2094197271655231778). Refuse “this number caused the rank.”
- He did not give steps = no procedure in corpus. Includes: positive GEO how-to (language-lock is a mechanism + one implication, not a kit — restated id=2093347878098042898 still not a playbook; Brave/Claude engine note is not a how-to either), Doubao / Deepseek, Day 0 framework, official migration flow details, cluster thresholds, Shopify remix details, GSC AI Mode filter, algorithm recovery, Top Stories, link-building, full hreflang code SOP (partial intl steps exist — id=2096533062842793991), crawl-budget item-by-item repair, Bing ranking SOP, “write English / buy EN citations” for language asymmetry, China GEO 规范 as imported doctrine, HCU X Article body (pointer only, id=2094576318788833400 → 2094559188676734976). Point at official docs or say you do not know. Do not fill with industry lessons.
- Do not import. lokiyan.com, firstoptimise, any external Open SEO / GEO skill, topic clusters, content calendars, featured snippets, CRO frameworks — not in this corpus. CWV is not the main lever; he did give a CrUX field-data panel (id=2092889660783722868). Local / GBP now has a procedure (id=2092207366305681808) — do not import a generic Local-SEO kit beyond it.
- One linear exception: when sites are already all excellent,
{brand+keyword}search volume ≈ that page’s rank[once, 5 groups]. Do not generalize into DA → rank (id=2090684152328945976).
Independent review (required before the final reply)
In the same turn, re-read your draft from a new, hostile angle. Hunt these:
- Invention: fabricated tweet-id, site name, GSC number, or a step he never gave?
- Import: industry common sense (keyword-research ritual, clusters, schema-as-GEO, bought links, CWV as main lever, generic Local-SEO kit, GEO tracker) smuggled into “must do”?
- Upgrade: asserted / once written as a verified law?
- Biography: talking about Loki when the user only wants a site looked at?
- GEO hallucination: still handing over a GEO homework list after “I have no way to answer you”? Still inventing an English-citation SOP after the language-asymmetry observation? Still turning language-lock (including the 2026-08-31 restatement) into a GEO playbook? Still importing the China GEO 规范 PDF? Still calling Xiaohongshu Google author EEAT? Still turning Brave/Claude routing or llms.txt no-regret into a GEO how-to? Still importing the HCU X Article body?
- Mechanical SOP: walked every step, did not cut to the task?
- Causal mismatch: DA / links / citations / tokens written as the cause?
- Missed red flags: fake authors, auto-jump, Soft 404, increment-free batch, dirty domain, migration with no 301, internal-link tracking params, 千人千面 vs crawler, poetic H1, JS that chatbots / Brave do not render, FAQ-every-page, ecommerce auto-blogs, PageSpeed-as-content-fix?
If the host has a sub-agent, send the draft there for this pass. If not, do the hostile self-check yourself and say you did. Fix what you find, then deliver. What you cannot fix, expose in the evidence-grade line. Do not paper over it.
Pointer
SKILL.md(this file): how to do this task.methodology.md: his words, ids, case grades, gap list. Knowledge SSOT.
On conflict, listen to methodology. If methodology also does not have it, mark no procedure in corpus. Do not fill. Cite (id=…). Do not invent ids.
Specials — expand only if the user asks:
- How to learn SEO: pretend you are a candidate; read US / EU / AU giant or award-winning agency JDs; skip domestic-giant JDs. Further up needs more CS / IT. Serious companies do not hire “write blogs + build links.” Course pointer
[once]: https://learningseo.io/ first, then Ahrefs / Semrush academies if gifted. 2026 Udemy not recommended (id=2091453544683729230). Search Central YouTube 101: Google SEO 101 / 科普; standard-answer boundary (id=2094325559820197987). - New ecommerce: new site → Shopify (or headless); native feed/schema is not enough, second-pass in Merchant Center. How to remix: no procedure in corpus. Profitable old site: do not panic-migrate. WordPress
[once]is the overseas SEO default for content/sites (id=2093542304372129882); Shopify remains the 2026 ecommerce line. Keep both, scoped. Shopify still better ecommerce architecture; being K’d is the domain (id=2095320035690725505) (id=2095344173826363791). Ecommerce blogs only for product / feature / own Guide (id=2095159020072124578). - Personal site: real-name domain, not a stage name;
.com/.digital/.io/.ai/.dev; local professions may use a local TLD. English pick:.com>.ai>.io;.appif you will ship mobile apps (id=2091316062709309490) (id=2033312223096086964). Person/social as future SEO carrier if site/brand weak[once/experiment]— not a quit-the-site SOP (id=2094615467981930559). X Articles are an optional SEO’d blog-like channel with Article schema + isAccessibleForFree (id=2094635693993832480). HCU must be practiced; article pointer only, do not import the body (id=2094576318788833400). - Copy rates (his prices, not an industry card): long landing AU$800–1000, blog 300–500, product page 200, to senior writers — used to understand what effort costs (id=2071839385998569488).