# Chapter 5 · Writing pages (4): 4 unverified page types, and checks before you publish > GEO Playbook · General Edition v1.0 · Canlah AI · CC BY 4.0 · Web page: https://canlah.ai/playbook/page-types-checklist/ > Markdown edition for AI assistants, same content as the web page. Figures are code blocks (wireframe / mermaid / bars / steps / split); "→" links open the matching section on the web, and the same URL with .md is its Markdown edition. ## 5.30 The four tier-C types: unverified page types **What you'll do in this section**: recognise that these four types have been seen only in an external teardown and that we have measured none of them ourselves. When you're done, you'll have a blueprint for the Buyer's selection framework (pt43), which is the strictly regulated side's lawful substitute for getting onto the list shelf, plus a one-line scheduling call for each of the other three types: all four types come after tier A and tier B. > **⚠️ None of the four types has been measured by us**: not one of the citations we measured supports them, and the teardown target's own pages of these types don't even rank organically. The teardown measures organic rankings and we measure AI citations; numbers from the two rulers cannot be read as one. We still write them up below so that the catalogue is complete and the next round can take them straight into testing. Evidence: see （→ 通用版 A.4 页型实测（二）：档 B / C 样本与 AI Mode 待补测清单）. ### pt43 Buyer's selection framework ```wireframe id=pt43 Figure: Block order for the Buyer's selection framework (not measured as a whole; spec taken from an external teardown; no block is marked cite); on the strict side and when switch E is on, the two blocks marked law are removed entirely title | How to Choose a for | req Quick answer | 2–3 sentences answering the title question directly, followed by a Disclosure statement | req H2 What does measure? | Opens with one punchy sentence, then 5–6 bold labels | opt H2 What a useful should show | Numbered, 5–7 items, one sentence each | opt H2 Research snapshot | Evidence paragraph: what the status quo is missing | opt H2 Evaluation dimensions | 5 "Label:" paragraphs, no bullets | req H2 Provider comparison matrix | Column headers: Provider / Dimension 1 / Dimension 2 / Best for | opt,law Per-provider card | One card per provider; write each limitation as a buyer task | opt,law H2 How to interpret common results | Short sub-sections without bullets, each opening with a label | opt H2 How to run a defensible baseline | Numbered, 5–10 imperative steps | opt H2 What a free diagnosis cannot prove | 100–160 words, spelling out what the diagnosis cannot prove | req H2 Questions to put in the RFP | 8 items, each a complete question ending in a question mark | req H2 Final recommendation | When to switch from X to Y, 2 paragraphs | opt FAQ + Method and sources | 5–6 questions, at most 2 naming your own brand | req ``` **When to use**: buyers ask "how do I choose a ?" It looks like （→ 通用版 5.18 资格流程页（pt07）） but the intent differs: pt07 answers "can I do this, what does it take, how do I apply"; this type answers "how do I choose a provider". Which leg cites it has not been measured. Going by the two-leg pattern, we infer it leans toward AI Mode, so only a follow-up measurement on the AI Mode leg can give the first piece of evidence: any citation moves it up from tier C; 0 citations keeps it at n = 0. **Why this is the only tier-C type we draw a blueprint for**: the strict side cannot self-build a list (see （→ 通用版 5.16 榜单页（pt05））), and a "how to choose X" criteria-based selection guide (it lists criteria, does not rank and names no one) is its lawful substitute for getting onto that shelf. **Not measured, 65% confidence**; we recommend scheduling one per industry as a control. **Compliance points specific to this page type**: how comparisons and lists are written on each of the three sides is covered only in （→ 通用版 5.3 三侧速查（二）：证言与评价、比较、榜单、头衔、外链、FAQ 与图注）. Applied to this type: on the strict side, and whenever switch E (peer comparison banned) is on, delete the two blocks marked law in the diagram (the comparison matrix and the per-provider cards), and keep only three blocks in the main body: evaluation dimensions, the limits of the diagnosis and the RFP questions. The lightly regulated side may include how fees are charged and a verification checklist, and its result numbers must pass the six checks for numbers. The unregulated side uses the blueprint as it is, per-provider comparison included. **Length and sample sentences**: the external teardown gives 2,100–3,700 words, 11–19 H2s and 1–2 tables. Treat these as a reference only: set the length by the page type and stop when the page is complete (see （→ 通用版 5.7 所有页型都成立的七条）). Sample sentence (not measured): write the own sentence as "a firm fact about us on + the check date". **Most common mistake** (inferred, not measured): writing it as "here's how to choose, and by the way we're the best". Restraint is the precondition for this type: the FAQ names your own brand in at most 2 items, and one item states "always confirm current scope with the vendor directly". ### The other three types: one row each | Page type | The measured neighbour to do first | Why it comes later | How to schedule it | |---|---|---|---| | pt44 Concept pillar page | pt12 Definition page | pt12 has measured citations; pt44 has 0 | Do pt12 first; pt44 is still to be verified. For how to split it from the Comparison page, see the figure below | | pt45 News and policy explainer | pt34 Change notice / old-vs-new page (about changes to your own business or your own jurisdiction) | Citations for rule-type questions land almost entirely on the official original page; commercial pages get 0 citations. AI is more likely to cite the original page directly than your explainer | Schedule it for defence and to piggyback on the rule-question leg, not as a traffic page. On the regulated sides, the strongest case is an explainer of a regulation amendment (for wording by side, see the page-type row in （→ 通用版 5.3 三侧速查（二）：证言与评价、比较、榜单、头衔、外链、FAQ 与图注）) | | pt46 Market observation page | None | The weakest evidence of all, and there is counter-evidence: the teardown target's own articles of this type are dead ends in the site's internal linking and do not even appear in the blog index | Without first-party data, don't write it and don't schedule it. A service provider may run it on its own blog for its own use, but there it works as a brand asset, not as a page that gets cited | pt12's card is in （→ 通用版 5.23 问题与规则类档 B（一）：监管义务、定义词条）, pt34's is in （→ 通用版 5.25 问题与规则类档 B（三）：日程、集合 FAQ、政策 hub、变更公告、误解纠正）. The blueprints for these three types are not in the main text; each is labelled in full "not measured, spec taken from an external teardown" and placed in （→ 通用版 B.5 未验证页型的施工图（供行业版引用））. When you test the four types, test 43, 44 and 45 first; 46 comes last. ```mermaid Figure: The title wording decides whether a page is a Comparison page or a Concept pillar page; only one page may take a given term, or the two pages cannibalise each other flowchart LR t{"Title contains"} -->|which is better / which to buy| p04["pt04 Comparison page"] t -->|what's the difference / what is| n{"How many concepts?"} n -->|one| p12["pt12 Definition page, measured"]:::hl n -->|two, long-form pillar| p44["pt44 Concept pillar page, to be verified"] ``` The same applies to English titles: "which is better" belongs to pt04, "what's the difference" belongs to pt44. If the two can't be cleanly split, merge them into one page. > **Example** (medical, not measured): the title "Dental implants vs bridges: differences, indications and cost difference" contains "differences", so it belongs to pt44, not to the Comparison page. The strict side can build it too, but it compares procedures only, not clinics. For pt45 to piggyback on the rule-question leg, there is only one way to write it: quote the rule word for word, date it, link back to the original page, then add one sentence on "what it specifically means for your line of business". In our measurements, what got copied was always these four things: "who / from which date / what must be done / the penalty for not doing it"; not one sentence of opinion was ever copied (inferred, with indirect support from measurements). Sample sentence (not measured): "As of , requires to do ; the direct impact on is ." ## 5.31 Handing over: five steps per page, the red line on lying, and self-checks **What you'll do in this section**: take every page through the same five steps before it goes live (brief → draft → non-author fact check → licence-holder compliance sign-off → launch acceptance, about 7 hours for an English page), and stop right there if any step gets stuck. Before launch, grep the whole site once against the banned-word list; before handing over, go through the self-check list for every page, and any question answered "no" sends the page back. ```steps Figure: Five steps per page; the next step does not start until the one before it is done 1 · 90 minutes | Write brief | Fewer than 5 exclusive facts: stop on the spot, change cluster 2 · 3 hours | Write draft | Missing a source: write a placeholder sentence, never make one up 3 · 60–90 minutes | Fact check | Checked by a non-author; if any item can't be found, it doesn't go live 4 · 30–45 minutes | Compliance sign-off | Licence holder checks the draft item by item and signs; unsigned means not live 5 · 45 minutes | Launch acceptance | If a non-cached load doesn't show the new H1, or the page source has no body text: roll back ``` What each step delivers, and who accepts it: | Step | Who does it | What it delivers | Who accepts it | |---|---|---|---| | 1 Brief | The brief writer drafts it; the person running probes produces the gap table | `brief-.md`: the cluster's 3–6 intents in the buyers' own words, H1/H2 questions, 5 or more exclusive facts + fact-table row numbers, 3 or more outbound-link candidates, a gap table of cited sources, and a list of banned phrases picked out by side | The owner's signature | | 2 Draft | The page writer | `draft-.md`, with frontmatter carrying intent_ids, page type, language and side; wherever a source is missing, write `[Source needed]` | — | | 3 Fact check | Non-author (a different person) | `factcheck-.md`: an item-by-item verdict + where the evidence is + the checker's signature and date | The checker signs and is personally responsible | | 4 Compliance | The owner + the licensed signatory | The compliance memo with every item ticked + the scan result against the Chinese–English banned-word list (hits must be zero) + the written sign-off | The signatory | | 5 Launch | The person who edits the site | The site change record + the `accept-.md` acceptance sheet + the page-register entry (intent cluster, buyer type, URL, launch date = split day, version) + one off-site target | The brief writer | A Chinese page counts as about 0.8 of a page of work and is scheduled on a separate day, through the same five steps. Only step 3 differs: it does not look for new evidence; it only checks that the fact-table row numbers and figures match the English page. None of the other four steps is cut: rewriting the question phrasing and writing independent answer sentences are done in step 2, and the three Chinese-page hard gates (the H1 contains the country name; the amount in the first paragraph carries its currency and says whether tax is included; the body names at least one real station or area, see （→ 通用版 5.9 内容形态、FAQ、中文页与外语页）) are checked in step 5. At step 5, when you accept the page, also log one off-site target, "which source should cite this page", and hand it to the off-site queue. If you write the page but push nothing off-site, the output never leaves your own site (for the other constraints on assigning tasks, see （→ 通用版 B.4 工单、台账与月报表头）). ### Self-check list Before handing over, check every page in this order: "positive gate → form → numbers → compliance → launch". Any row answered "no" is handled as the rightmost column says. ```split Figure: With only the ban gate, the editor's easiest move is to delete everything; the positive gate forces you to prove first that the page kept something Ban gate only (no positive gate) || Positive gate added The more you delete, the safer || Count exclusive facts first; go on only once you have 5 The page you hand over is identical to a competitor's site || Fewer than 5: sent back on the spot to change cluster, no last-minute patching AI has no reason to cite a page with no exclusive facts || Everything kept is a fact buyers will use to decide ``` | Gate | What to check | What to do if the answer is "no" | |---|---|---| | Positive gate | First run the ban scan; sentences it removes don't count. Then list, one by one, the checkable facts on this page that can't be found anywhere outside your own site, and check that there are at least 5, each of which would change a buyer's choice between candidates (generic information every peer has doesn't count). You can name which independent source (an official register, a government page, a third-party review, a ranking body) confirms what you are a fit for or your credibility, with its name and URL written out | Fewer than 5: send it back to step 1 to change the intent cluster; no patching facts in during review, and no keeping a sentence the ban gate hit just to make up the count. Can't name an independent source: fill in the off-site source plan before launch | | Form | The conclusion sits right under the H1 or in the first sentence of the first H2, and both sentences are in the first 30% of the body (on the strict side, and whenever switch E (peer comparison banned) is on, check that the second sentence is an official-basis sentence, see （→ 通用版 5.6 两条腿，两种句子）); the first sentence of every paragraph makes sense out of context, and none opens with "it depends" or "it's complicated"; one URL fully answers only one intent and one buyer scenario; length and how H2 questions are written follow （→ 通用版 5.7 所有页型都成立的七条）, and the conclusion block's spec follows （→ 通用版 5.8 结论块、密度闸、出处闸） | Fix it on the spot. A missing conclusion block, or one that doesn't match the body, is not an on-the-spot fix: send it back to step 2 to be completed | | Numbers | Every number carries its unit, threshold and how it is counted; the numbers in the conclusion block match the table, number for number; every fact has a source and a check date, and anything out of date is marked historical or deleted; Ctrl+F in the page source finds your own prices, results and specs; the fact-check file is complete, and prices, registration numbers, rule text and statutory deadlines have been read back word for word by a non-author; the organisation's operating figures have a ledger export and a written sign-off | Wording problems: fix on the spot. Any item missing from the fact-check file: does not go live | | Compliance | First decide the side, then scan item by item down your column in （→ 通用版 5.2 三侧速查（一）：先判身份；价格、促销与赠送、结果数字） and （→ 通用版 5.3 三侧速查（二）：证言与评价、比较、榜单、头衔、外链、FAQ 与图注）, including a word-by-word scan against the Chinese–English banned-word list; scan the FAQ and the end-of-page Q&A as well, because they are not an exempt zone. Also scan for the general red lines: traffic multiples, ROI, conversion lift, guaranteed first place, exclusive, "independent review", and method leaks such as small numbers from a single question, how the probes are run, or the question list. The page passes professional review (pages that touch personal safety carry a reviewer's signature; remedy pages follow the order in （→ 通用版 5.19 补救与二次评估页（pt08））; statutory deadlines cite the original text). The compliance-reviewed draft carries the licence holder's written sign-off, and the banned-word scan shows zero hits; before launch, grep the whole site once against the banned-word list | Delete every hit and fill the same slot with the replacement wording from 5.2–5.3; the delivery note states "deleted X, replaced with Y"; after deleting, go back to the positive gate and recount. No written sign-off: does not go live | | Launch | A non-cached load shows the new H1; view source shows the body text (no reliance on JS); fetches as the four identities match; IndexNow has been pinged and the receipt kept; the visible "Updated" line matches `dateModified`; the page register has logged the intent cluster, buyer type, URL, split day and version; the off-site target is logged; a test enquiry has been recorded | Roll back | | Chinese pages: extra checks | The three Chinese-page hard gates (see 5.9); whether each H2 is in the Chinese buyer's own words or a literal translation of the English heading; whether the answer sentences were written independently or translated | Any hard gate missing: send it straight back, no fixing on the spot. An H2 or answer sentence that was translated: send it back to step 2 to be rewritten | > **Example** (education and training): grep the whole site for the six "approval"-type words "approved / accredited / endorsed / supported / validated / recognised", and replace each instance with a sourced factual sentence or with "registered". ### The red line on lying ```mermaid Figure: The red line on lying: better to leave a fact out than keep one that hasn't been checked flowchart LR f{"Price, reg. no., rule or deadline?"} -->|yes| a["Non-author reads 100% back word for word"] f -->|no| b["Automated full scan + 30% sample read-back"] b -->|sample finds a false item| c["Recheck the whole page in full"]:::warn b -->|no false item found| g["Go to compliance review"] a --> d{"Does it check out?"} c --> d d -->|no| e["Delete it, don't keep it"]:::hl d -->|yes| g g --> h["Licence holder signs, no exceptions"] ``` Neither the non-author fact check nor the compliance review may be skipped under any scheduling pressure, and not a minute of either may be cut. The sign-off has limits too: **A written sign-off resolves only the three stop-and-escalate situations; it never turns something banned as Statute text into something publishable** (see （→ 通用版 0.4 合规句的三档标签与停笔规则）). ## 5.32 Master table of common mistakes **What you'll do in this section**: go through this master table before handing over. Fix, one by one, the mistakes in the figure below that aren't written up elsewhere; for mistakes already covered in the page-type cards and neighbouring sections, this section keeps only one line and a link and does not repeat them. ```split Figure: The mistakes that recur most often when writing pages, and how to fix them Wrong way || Fix The conclusion block says "this article will introduce", and the first screen is all lead-in || Give the own sentence directly under the H1, and a number in the first sentence of the first H2 The first-screen summary and the table numbers don't match || Take the conclusion-block numbers from the table, and check them number by number before handing over Prices and results appear only in images, JS or a booking widget, or switch by IP || Put all of it in the first-screen HTML The product detail page body gives no actual price || Show the real price on the first screen, matching the feed, structured data and checkout Writing the page without logging an off-site source target || Every page states the third-party page that should cite it, who handles the correction or submission, and the compliance approval status A separate page for every SKU and every phrasing variant (not measured) || One intent, one URL; variants go into H2s; one question, one page only for genuinely independent questions Changing only the update date, not the content (rule of thumb) || Keep the price check date, regulation check date and date of substantive change separate; only a substantive change moves `dateModified` Machine-translating the Chinese version directly || Validate Chinese-language demand first; rewrite the question phrasing and answer sentences, and put it through compliance separately Calling it done because ChatGPT, asked once on day 7, can quote the page back (rule of thumb) || Freeze the questions, run baseline and retest over several rounds, and log times named, position and cited URLs separately ``` The two rows marked rule of thumb are not measured, the one row marked "not measured" is inferred, and the other six rows have measured evidence behind them. Evidence: see （→ 通用版 A.2 选点、写页、站外与复测的证据）. ### Mistakes covered elsewhere: one line and a link | Mistake | Fix | See | |---|---|---| | Padding the word count to reach a minimum | Set the length by page type; stop when the page is complete | （→ 通用版 5.7 所有页型都成立的七条） | | Writing the H1 as a question, or every H2 as a question | The H1 is a statement; only the first H2 is a question | （→ 通用版 5.7 所有页型都成立的七条） | | A comparison page that compares only your own models, or puts the verdict mid-page | Competitors must be included; the verdict goes in the first sentence of the first H2 | （→ 通用版 5.15 对比页（pt04）） | | Turning one question, one page into one page with ten questions | One question, one URL | （→ 通用版 5.17 一问一页（pt06）） | | Several H1s on one page | One H1 per page | （→ 通用版 5.19 补救与二次评估页（pt08）） | | A remedy page that starts with the cause, is written as a self-test checklist, merges the two skeletons, or has an H1 blaming "another business did it badly" | Personal-safety type: order it "what to do now → red flags → what not to judge yourself → examinations → pathway → cost"; split the personal-safety type and the rights-deadline type into two pages; the subject is the buyer's situation | （→ 通用版 5.19 补救与二次评估页（pt08）） | | 301-redirecting a byline after the person leaves | Keep the person profile page and mark it "formerly", or 301 the whole page to the directory; on business pages, only change the name to the person currently responsible | （→ 通用版 3.3 个人实体页与六处痕迹对齐） | | Writing your own "best" list to answer "who's best" | Neither leg cites a self-built list, and on the regulated sides it also breaks the rules; use business profiles, third-party rankings, the product feed and the organisation facts page instead | （→ 通用版 5.6 两条腿，两种句子） | | Putting a market-range sentence into the feed or the Offer schema (not measured) | The feed and schema carry only your own real price | （→ 通用版 5.2 三侧速查（一）：先判身份；价格、促销与赠送、结果数字） |