# Chapter 4 · Which questions to go after: pick 30 buyer questions and record how AI answers them today > GEO Playbook · General Edition v1.0 · Canlah AI · CC BY 4.0 · Web page: https://canlah.ai/playbook/pick-questions/ > 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. ## 4.1 The full target-picking process and the twelve pre-start questions **What you'll do in this section**: first get the eight steps of picking targets clear: the first five come before the baseline is saved, the last three after. Then the decision-maker answers the twelve questions in person in 40 minutes; questions 2, 3 and 4 must each get a definite answer on the spot, written into the first line of the work order. This chapter is not about picking "the questions I can win", it's about picking "**buyers who pay when you win**". The door and identity gates are ×0/1 switches; only the shelf gate adds points month by month (（→ 通用版 1.3 三个关口与三条路）). This chapter only manages the targets for the shelf gate: once the door and identity are fixed, this chapter decides which few cells the effort goes into. It does not touch the door, and it does not touch identity. ```steps Figure: the eight steps of picking targets — the first five are done in D2–D6, before the door is fixed; the last three only after the baseline is saved 1 | Check the buyer type table | List price per order/margin/capacity, confirm the pool doesn't miss the most valuable type 2 | Freeze the question pool | 30 questions signed off, unchanged for the whole quarter (4.2) 3 | Coarse screen into piles | Dead pile drops out, battlefield and hold piles stay (4.3) 4 | Full rounds to measure noise | Five rounds run in full, same-week retest measures the noise band (4.3) 5 | Web leg freezes the baseline | Web control leg runs first, account-level Top 20 saved (4.4) 6 | Formal grouping and ranking | Work out annual opportunity value by buyer type (4.5) 7 | Three checks, four states | Pass the three checks; assign a state to each intent-level top 10 URL (4.5–4.6) 8 | Set this quarter's slots | Actions by state + the three combination rules (4.7) ``` Step 1 is not the formal grouping; that happens in step 6 (4.5). It is only a check before you sign the pool: copy the price per order, gross margin and monthly capacity cap from the answer to question 5 and your own books, and use them only to confirm that the buyer type with the highest annual opportunity value has a question in the pool. Skip this check and you will freeze a batch of questions, only to find afterwards that the most valuable buyer type has no question in the pool at all — and the whole quarter cannot be judged. Two premises the figure cannot show (the evidence is in （→ 通用版 1.1 买家怎么问，AI 读什么）): - The buyer's real path is: see the name → search separately → visit the website → check reviews → contact. Almost no one clicks the links in an AI summary. So only pick questions from the "who to buy from" stage; do not pick educational questions. - Most of what AI cites is pages other people write about you; paid placements and advertorials are close to zero. So what the shelf snapshot needs to record is not "am I present" but which containers currently hold the seats for this question. ### The twelve pre-start questions **Who**: the decision-maker asks in person · **How long**: 40 minutes · **Output**: a one-page answer sheet · **How to check**: questions 2, 3 and 4 each have a definite, fixed answer written down on paper. Tag every fact the other party (or you) states as "to be verified" first; the baseline (4.3–4.4) is what settles it. | # | What to ask | Why ask it | What to do once answered | |---|---|---|---| | 1 | Legal name, main market, licence/registration number | Seats and identity checks all go by the legal name | If you can't get the registration number, leave a gap in the fact table and fill it in before moving on | | 2 | Which side you're on, which switches are on | Decides which set of side rules you follow; one entity can only be on one side | Write "side + switches" down on the spot as the first line of the work order; a business with mixed activities is built site-wide to the stricter side (switch D (spans two sides)) | | 3 | How far you can go with the price (asked by side; see below the table) | The question is "can you give it", not "are you willing to publish it" | Can give it → the price page is top priority; can't → downgrade the whole job to audit + monitoring only | | 4 | Where switch A (agency liability) is on: will you sign the written publishing authorisation + the materials sign-off sheet, and who is the signatory | The publishing party shares liability with you | Won't sign → do not do it. This is the only hard refusal with no downgrade path | | 5 | How many times in a lifetime a buyer buys, and how much per order | Buy once or twice in a lifetime, high price per order, buyer can't tell good from bad — most worth doing | High frequency, low price per order → downgrade to hold or do not do | | 6 | Who can edit the website, how many days to change one page | If changing one page takes more than four weeks, the whole first month's schedule is void | More than four weeks and unwilling to change it → downgrade to hold | | 7 | Who manages the five business profiles and the two webmaster tools, and can they give read-only access | Needed for the door gate; connect the free first-party data before paying for probes | Can't give read-only access → the first two items can't be verified at acceptance; list it separately as a risk | | 8 | Is there a namesake organisation or namesake person | If AI identifies the wrong business, all the page work upstream returns zero or even less | Yes → the first batch becomes facts pages + profiles, not content pages | | 9 | Are you buying optimisation services from someone else; are you using a platform priced per enquiry, per lead or by sale commission | If switch F (referral commissions banned) is on, this is a red line | Switch F is on and in use → write it into the compliance memo; stop using it first, then start work | | 10 | Who you think your competitors are | Two reviews both found: the competitors you name yourself never show up in AI's answers | Tag all of it "to be verified"; the baseline decides | | 11 | Are there Chinese-speaking buyers, are there buyers in other minority languages | A Chinese landing page is the only lever that goes "build one page → straight into the named set" | Chinese buyers → produce Chinese questions per 4.2; minority languages, see below the table | | 12 | Whether orders mostly come in by form, chat or phone | Decides the acceptance tier (4.5) | Phone → observable tier + a mandatory front-desk field "how did you find us", never dropped from the pool for this alone | ### Three hard gates: questions 2, 3 and 4 ```mermaid Figure: each of the three hard gates puts a definite answer into the first line of the work order; if question 3 can't be met, the job is only downgraded; if question 4 is not signed, the job is not done flowchart LR q2{"Q2 Side + switches"} -->|fixed on the spot| order["First line of the work order"]:::hl q3{"Q3 Can you give a price?"} -->|can give it| order q3 -->|can't give it| down["Downgrade the whole job to audit + monitoring"]:::warn q4{"Q4 switch A, agency liability: sign or not?"} -->|sign| order q4 -->|won't sign| stop["Do not do it — no downgrade path"]:::warn ``` The full definitions of the three sides and the six switches are in （→ 通用版 0.2 先判你属于哪一侧）. Question 3 asks something different by side: on the strictly regulated side it asks "can you give a fixed final price in tiers" — this side cannot even write a range, the rule is 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 if you will only write "from", that counts as can't give it; on the lightly regulated side it asks "can you give a range + billing method"; the unregulated side has no such gate. For question 11's minority languages, run a real search first: if the phrasing points to another country's home market, that is a wrong target, not inefficiency — add Chinese questions instead of forcing that language (example in 4.2). Questions 10 and 12 are not gates: competitors are judged once the baseline is out; the order channel only decides how the acceptance tier is split, not whether a question enters the pool. --- ## 4.2 The question pool: where the 30 questions come from, how they are balanced, how they are signed **What you'll do in this section**: first spend half a day reading two sets of free first-party data and the buyer's own words from the last 15 orders, then use only the four classes of source with evidence to make up 30 questions, have someone who did not draft them count the ratios against the fixed targets, and finally have the decision-maker read every question aloud in person, make changes on the spot, and sign on the spot. Once these 30 questions are signed, they do not change for the whole quarter. The 30 questions are the ruler for the whole quarter: picking targets draws only from this pool, and the retest and the settlement use the same set of questions. **The ceiling on the question pool's quality is the ceiling on the project's results** — so source discipline comes before probing. ### Step zero: read the free data before paying for probes **Who**: the person who edits the site gets access, the person reading the data reads it · **How long**: half a day · **Output**: two screenshots + a one-page summary · **How to check**: the decision-maker has read it; only then is the probe budget released. Connecting the two back-end reports was already done in （→ 通用版 2.5 闸三只读：接上两个免费 AI 报告）; here we only cover how to read them and what to watch out for in each. ```split Figure: step zero — what you get from the two free reports and the front-desk records, and the one catch each has What you get || The one catch GSC: generative-AI impressions by page/country/date || Impressions only; already included in total impressions, don't add them again; flag blank values exported as 0 Bing: cited pages, grounding terms, citation share || Grounding terms count only as a clue, not standalone demand evidence Front desk/CRM: buyers' own words before purchase, last 15 sales || If there's no record, make it a mandatory field on the spot ``` This step does not move rankings by itself. It decides whether all the effort that follows lands on real questions, not questions you imagined. Bing's grounding terms feed straight into the question pool as a class C source below; the third row is a class A source, and the mandatory fields are "how did you find us" and "what did you most want to ask". ### Only sources with evidence enter the frozen pool **Who**: the person setting the questions drafts it · **How long**: one day · **Output**: a question-pool table, each row with a source label and the original evidence (a link or screenshot) · **How to check**: a second person checks every row, then the pool is signed in person. ```mermaid Figure: each of the four source classes needs one piece of evidence to enter the frozen pool; everything else goes only into the reserve pool flowchart LR a["A buyer's own words before purchase"] -->|screenshot/record| pool["Frozen pool"]:::hl b["B keywords in active ads"] -->|report export| pool c["C real platform queries"] -->|export/screenshot| pool d["D questions already listed on competitor pages"] -->|URL + screenshot| pool ai["AI-generated / made up"] -->|no evidence| side["Reserve pool"]:::warn ``` Class C means real GSC queries, Google autocomplete, related questions, and Bing grounding terms. Class D only takes questions that are **already listed** on a competitor's comparison page, a directory site or a list page. The four classes are mutually exclusive, each entry is tied to one checkable piece of evidence, and any row without evidence is sent back. The reserve pool can be used as a reference for topics when writing pages, but **it does not enter the frozen pool and is not a ruler**. **If you can't make up the number**: go back and find more class A words from real buyers — do not fill the gap with AI. An AI-generated sentence that gets into the frozen pool is the single most fatal, irreversible mistake of the whole quarter: it contaminates the ruler itself, not just one page's content. If you can't reach 10 class A entries, downgrade at this step to audit + monitoring: a business that can't produce even 10 real buyer questions in their own words can't support a ruler. ### The ratios are fixed **A baseline conclusion that breaks the ratios is void; redo the question pool.** ```bars Figure: the ratio lines for the 30 questions — location and price are floors, bare words are a ceiling unit: questions With a location | 18 | floor ≥60% With price or cost words | 9 | floor ≥30% Bare category words | 3 | ceiling ≤10%, for comparison only ``` Each of the three lines has one reason. The location floor: bare words can produce a fake zero — the same query with or without a location can go from "zero businesses" to "every business restored", and that "zero" in the report is one you made yourself. The price floor: price/cost-type phrasing has a high rate of triggering AI Overviews (4.7). The bare-word ceiling: a national head term with no location can't move within 90 days. Evidence in （→ 通用版 A.2 选点、写页、站外与复测的证据）. ```bars Figure: the same 30 questions split again by question shape — parallel to the ratio lines, not a separate batch unit: questions Who to recommend / which is best | 10 How much | 8 A vs B, which is better | 6 Scenario questions (including second opinion) | 4 Brand direct | 2 ``` **Where there are Chinese-speaking buyers, Chinese-language questions are the main line, not an add-on** (already decided at question 11 in 4.1). Write the count into the work order; Chinese questions always carry a location, price-type ones count into the price-word row, **and count at the same time into the location and price numerators above — they don't get their own total or a separate table**. When the ratios can't be filled, cut from the 3 bare category words first; **never cut a Chinese question**. The bar for the Chinese cell is not getting onto a list; it is whether you have a Chinese landing page (evidence in （→ 通用版 A.2 选点、写页、站外与复测的证据）). > **Example** (legal) A divorce lawyer fee question phrased in Indonesian: all 9 citation sources were Indonesian local law firms, the answer quoted a price in Indonesian rupiah, and the engine itself added a line saying the result does not concern Singapore. The buyer's intent was misaligned, and even doing the question wouldn't get you into the local named set — that's a wrong target, not inefficiency. Add Chinese questions to the pool instead. ### Second-person check and in-person sign-off **Who**: someone who did not draft it checks it + the decision-maker signs · **How long**: 30 minutes to sign · **Output**: a signed PDF of the question pool · **How to check**: there is a signature on paper, and nothing changes for the rest of the quarter. Any person who did not draft it counts through the ratios item by item and signs "ratios verified · date · name" at the end of the question pool document. Before signing the pool, check the buyer type table against 4.1 step 1 — if the buyer type with the highest annual opportunity value has no question in the pool, do not sign. The decision-maker reads every question aloud in person, makes changes on the spot, and signs on the spot; for the sign-off wording see （→ 通用版 B.2 口径句与话术）. **Changing a question is changing the ruler, and once the ruler changes, the numbers before and after can't be compared** — swapping questions partway through is the easiest way to fail the whole quarter, because afterwards no one can tell whether the number rose or the ruler changed. --- ## 4.3 Coarse screen and piles, full rounds, and the brand-six baseline **What you'll do in this section**: first run a coarse screen of 30 questions × 1 round × 2 engines = 60 runs to split the question pool into a dead pile, a battlefield pile and a hold pile; run only the battlefield and hold piles through five full rounds, then retest in the same week to measure the noise band; in the same week, also run the brand-six questions in full — 36 runs — as the baseline for factual errors. When recording answers, count "listed as a source" and "written into the answer" separately. ### Coarse screen into piles **Who**: the person running probes runs it, the person setting the questions reviews it · **How long**: half a day · **Scale**: 30 questions × 1 round × 2 engines = 60 runs · **Output**: a three-pile list. ```mermaid Figure: the coarse screen only checks whether AI named any company at all, splitting the pool into three piles — only two are worth paying to run in full flowchart LR s["30 questions × 1 round × 2 engines"] --> a{"Did it name any company?"} a -->|neither engine named any| dead["Dead pile: out, skip monitoring"]:::warn a -->|named someone, whoever it is| b{"Are you on the list?"} b -->|not on it| fight["Battlefield pile: main effort, five full rounds"]:::hl b -->|on it| guard["Hold pile: switch to hold, five full rounds"] dead -.->|up to 3 questions lifted back after a re-read| fight ``` The dead pile means neither engine names any company at all — it only gives educational content, or just tells you to consult a professional. Once out, spend no more money on it, and it does not enter the monitoring layer. The dead-pile criterion is an operational convention, not a measured finding: 70% confidence. Because a wrongful kill is irreversible, two safeguards are added: a second person must re-read the original answer text before the dead-pile list is finalised; the dead-pile list is reviewed together with the shelf snapshot, and up to 3 questions may be designated to lift back into the battlefield pile for a full round. **The approved external wording for the dead pile is fixed: n = 2, write only "no business named in the coarse screen", never "zero visibility"**. For how much the two-stage screen saves compared with running everything in full from the start, and what is lost, see （→ 通用版 A.2 选点、写页、站外与复测的证据）. ### Full rounds and the noise band **Who**: the person running probes runs it, the person reading the data compiles it · **How long**: 2–3 days · **Scale**: (battlefield pile + hold pile question count) × 5 rounds × 2 engines · **Output**: raw answers + a summary table + `noise-band.md` · **How to check**: 2 hours of deep reading by a person + checking every item against the judgement checklist. ```steps Figure: the eight steps of a full round — the baseline is all saved together in this one week; if step 1 misses the known citation, the whole round stops 1 | Validate the ruler first | Run a question you know gets cited; if the citation doesn't show up, stop 2 | Lock region and language | Don't give AI your own URL; run Chinese and minority languages separately 3 | Run the full rounds | Battlefield pile + hold pile, five rounds each × two engines 4 | Same-week retest | Same questions, same engines, same number of passes, run 3 more times in the same week 5 | Web leg + freeze | Full spec in 4.4, not repeated here 6 | Keep two tables separate | Don't merge brand questions with category questions; the brand table stands alone 7 | Sign the control agreement | Same-site pages not to touch, three competitors, the primary number to watch, when to stop 8 | Save after the deep read | 2 hours of manual deep reading can't be skipped; save it the same day, permanently ``` **When the ruler misses the known citation**: do not issue a report, do not announce any "zero". Minority languages are run separately so they don't get mixed into the main language's basis. **How to measure and use the noise band**: this month's noise band = the range (max − min) of the seat count across those 3 same-week runs, with a floor fixed at ±1 seat, written into `noise-band.md` and redone every month at the monthly retest (（→ 通用版 7.2 噪声带、页级信号与每月十步）). A monthly change in seat count may only be written as "increased" when it is larger than the noise band — report "up 1 seat" without measuring the noise band, and you are reporting noise, not results. **Once this round is done, the engines, the number of passes and the question pool are all frozen for good** — every comparison for the next 90 days is against this set of numbers. Write the two legs' official names only as given in （→ 通用版 1.2 两条腿：ChatGPT 找源头，AI Mode 找二手）; never write "real user visibility". **What to look for in the deep read**: whether a community mention is a namesake mix-up, check any negative claims once, and whether the industry wording is right. After saving, keep only the 3 most striking screenshots to make your case with; anything that names a third party is treated as internal material. Don't notarise the whole batch of screenshots. ### Two things to count separately in every record ```split Figure: one answer has two different kinds of "present" — the fix is completely different depending on which one is missing, so count them separately Listed as a source || Written into the answer The page appears in the source area below the answer || AI writes your name or number into the body text Ruler: on the list and cited X/20 || Ruler: seat count Not listed: get the page onto the list first || Listed but not written in: adding pages won't help, you need to change the sentence ``` These two numbers are not merged into one "hit" (basis in （→ 通用版 1.3 三个关口与三条路）). Also record the naming position: getting onto a list AI already cites and ranking first on it raises visibility and moves your position in the answer earlier, but this set of numbers is only used to order the work — never to promise any result. Write numbers and their qualifiers in the unified form given in （→ 通用版 D.1 数字纪律：数字怎么标、什么只能自己看）. How the four states and the intent-level denominator use this data is in 4.6. When recording each answer, also log these fields: | Field | How to record it | |---|---| | Who was named, whether you're in it, which position | Record the position | | Total names in this answer | The total count of business names named in the answer body — record the actual count, don't estimate | | What kind of site the citation comes from | own site / directory / list / review site / maps / forum / government / media; count listicles as their own category — share and sources in （→ 通用版 A.2 选点、写页、站外与复测的证据） | | Review field | Whether the answer body writes review counts or star ratings word for word; record recommendation-type and non-recommendation-type on two separate lines, never merged into one total frequency; once your own baseline is run, use your own numbers — reference numbers in （→ 通用版 A.2 选点、写页、站外与复测的证据） | | Forum source share | Only put it on the work list when Reddit's share of this month's cited URLs is >2%; at ≤2%, record only, don't invest; don't write a blanket "don't do Reddit" — reasons in （→ 通用版 A.2 选点、写页、站外与复测的证据） | | Sampling status | Keep valid/failed separate; remove failed samples from the denominator and list them separately | | Citation count | Keep URL-level and answer-level separate; during the baseline period most will be 0 or 1, URL-level detail can wait until the first retest | | Tendency | positive/neutral/negative/excluded, **a negative or excluded result jumps the queue and triggers an alert the same day** | **How to write the numbers**: write "asked K times, named X times" — never a single-question small-number form like "hit 2 out of 3". ### The brand-six baseline Run the brand-six questions in full, 36 times, in the same week: 6 questions × 3 rounds × 2 engines, defined in （→ 通用版 3.5 品牌六问与发现错误之后）. The baseline takes only this one full round, never a quick run — two runs with different composition are not the same ruler. It is the only source of the count "factual errors X": the 25-minute read in （→ 通用版 3.1 为什么排在写页之前 · 两小时清单） reads this same raw set of answers, no separate quick check. --- ## 4.4 The web control leg and the frozen baseline (the book's only full spec) **What you'll do in this section**: run the first web control leg before you freeze the account-level Top 20, and force source domains found only on the web version into the candidate pool. After freezing, ask the same fixed 5 questions by hand every month, once each, and compute only the two overlap rates, to calibrate the two API legs. Only once all three parts of the baseline are saved do you unlock （→ 通用版 2.6 改门：nosnippet、robots 四步合并、WAF 白名单）. ### Why this leg is required Both API legs measure the model API. The citation distribution on the API is not the same set of sources a person sees on the web version, and this is a systematic skew, not noise: ```bars Figure: in both comparisons, the web version is far higher than the API — the two sides cite different sets of sources unit: % # Overlap with authoritative media lists Web version | 45.5 API leg | 27.3 # Share of public-broadcast-type sources Web version | 34.6 API leg | 12.2 ``` A single domain can shift as far as position 1 on the web version and position 61 on the API. **The skew is just as large whether you run 5 rounds or 50** (sample and sources in （→ 通用版 A.2 选点、写页、站外与复测的证据）). That last sentence is the whole reason this leg exists: adding more rounds does not fix a systematic skew — only switching the path does. Skip it, and the whole quarter's off-site effort aims at the API leg's account-level Top 20 while buyers are actually seeing a different set of pages. When the effort misses, nothing warns you: the on-the-list page count still climbs, and the report still closes out clean. So two things are fixed: the ruler's official name is "**Model-API visibility baseline**", printed in the footer of every report; visibility for AI Overviews and AI Mode always uses the Search Console generative AI report's impression numbers instead (already connected in 4.2's step zero) — never pass off the Gemini model leg's numbers as a substitute. The Gemini API, AI Overviews and AI Mode are three different products. ### How to run it (the spec is fixed; change nothing) | Item | Fixed | |---|---| | Questions | Take a fixed 5 from the frozen 30: ≥2 "who's best / near me" type, ≥1 price type, the rest from the battlefield pile. Once these 5 are picked, they don't change for the whole quarter | | Path | ChatGPT web version, typed by hand in a browser. No API, no proxy interface of any kind | | Identity | Logged out, or using Temporary Chat; personalisation and memory turned off | | Region | An IP local to the target market (be there in person, or use a local exit node; e.g. an SG IP for the Singapore market) | | Passes | 1 pass per question. Don't rerun, don't pick the best answer, don't reword | | Recording | A full-screen screenshot per question (including the source area below the answer — scroll and take more than one if needed), copied by hand into `web_leg_.csv` | | Fields | question / date / businesses named in the answer (each one, in order of appearance) / source URLs listed below the answer (each one) / screenshot filename | | Who | The person running probes runs it; another person checks every row of the screenshots against the csv | | How long | 40 minutes within one day each month; not part of the monthly retest process table — schedule it as its own day at the start of the month | It produces two numbers, only these two, with no further processing: | Number | How it's calculated | |---|---| | Named-set overlap | (the set of businesses named by the web leg ∩ the API leg) ÷ their union, calculated only on these 5 questions | | Source-domain overlap | (the set of source domains listed by the web leg ∩ the API leg) ÷ their union, calculated only on these 5 questions | Neither number is mixed in with the full pool of 30 questions; the denominator is the union, not the web leg alone. They are never used as acceptance evidence: they don't go into the seat count, don't go into the on-the-list page count, are never added to any other ruler, and never produce a standalone "web visibility" score. ### The baseline run: web leg first, then freeze the account-level Top 20 The order is fixed; get it backwards and the whole quarter cannot be recovered. ```steps id=freeze Figure: the five steps before freezing — get them backwards and the whole quarter cannot be recovered; only after all five are done may you fix the door 1 | Save the API leg | The full-round baseline is done, the raw domains_cited file archived 2 | Run the web leg the same day | Fixed 5 questions × 1 pass, asked by hand on the web version 3 | Fill in the candidate pool | Domains unique to the web leg are forced in even if the API leg cited them 0 times 4 | Take the top 20 and freeze | Merge the candidate pool and take the top 20 by URL-level citation count 5 | Mark and keep permanently | web_only=Y listed separately; monthly review continues as normal ``` Use the **raw** `domains_cited` from the full-round baseline to take the top 20 — not a version with your own site and generic domains already stripped out. Once taken, a second person spot-reads 3 pages. The frozen account-level Top 20 is the denominator for "on the list X/20", and also the whole quarter's off-site target — if the target is taken only from the API leg, it is only the API's shelf, not the buyer's shelf. Freeze first and run the web leg afterwards, and the domains it turns up can no longer enter the denominator — the whole quarter's target is then wrong. ### How to read the two overlap rates ```mermaid Figure: the two overlap rates only calibrate the API leg — below 50% only changes how the report is written, never any acceptance decision flowchart LR w["Web leg, 5 questions × 1 pass"] --> n1["Named-set overlap"] w --> n2["Source-domain overlap"] api["API leg, same 5 questions"] --> n1 api --> n2 n1 --> j{"Either below 50%?"} n2 --> j j -->|yes| note["Print the non-extrapolation sentence verbatim"]:::warn j -->|two months running| stop["Not counted as not-moved; check the ruler first"]:::warn j -->|no| ok["Report using the API leg's basis as usual"]:::hl n1 -.->|forbidden| ban["Using it to calculate seats or for acceptance"]:::warn ``` The non-extrapolation sentence does not change a single word: "This month the two paths diverge (named overlap X%, source overlap Y%). The seat changes below were measured on the model API and are not extrapolated to what real users see on the web." **Never use it the other way round**: don't write "already visible" just because the web leg named you (n = 5, single pass), and don't use these 5 questions to calculate seats either. ### How to check, and common mistakes Miss any one of these and it doesn't count as done: `web_leg_.csv` has 5 rows, every row has a screenshot filename, and the screenshot opens; the screenshot shows the "logged out / Temporary Chat" interface feature and a visible date; the numerator and denominator sets for the two overlap rates are listed item by item, not just given as one percentage; in a month below 50%, the non-extrapolation sentence is present, word for word; in the baseline run, `sources-top20.csv` has a `web_only` column with at least one entry forced in — if there are none, there are only two possibilities: it wasn't run, or the two legs are completely identical (the latter must be stated explicitly, with the full domain list from both sides attached). Common mistakes: running the web version through an API or a script instead (that's still the API leg — a wasted run); running it while logged in; running one question three times and picking the best-looking answer; swapping the 5 questions every month; reporting the overlap rate externally as a results number; freezing the account-level Top 20 and only running the web leg afterwards. **The account-level Top 20, the noise band, and the 36 brand-six runs — all three saved = the baseline saved. Next, go back to （→ 通用版 2.6 改门：nosnippet、robots 四步合并、WAF 白名单）.** --- ## 4.5 Buyer types, annual opportunity value and the three checks **What you'll do in this section**: once the baseline is saved, formally group the questions by buyer type (not by question shape), rank them by annual opportunity value, then run every candidate through the three checks. When you're done, you'll have a candidate list grouped by buyer type, ranked by annual opportunity value, with every entry having passed the three checks. ### Group by buyer type **Who**: the decision-maker + the person setting the questions · **How long**: half a day · **Inputs**: the price list, the last 20 sale records, capacity · **Output**: a buyer type table · **How to check**: the price per order and capacity for each type has been confirmed in person by the decision-maker. The same question is handled differently in different industries: in AI Mode, professional services win through educational content, consumer services win through review volume (75% confidence, sample doesn't include the local market, evidence in （→ 通用版 A.2 选点、写页、站外与复测的证据）). For professional-services buyer piles, add long-form content and directory/list placements. ⚠️ But on the strict side the review-volume point must not turn into "go and ask for reviews": there, for consumer-type buyers, put the main effort into price pages, process facts and complete business profiles; reviews are only observed passively. The same playbook can't be used for every side. ### Rank by annual opportunity value ``` annual opportunity value = price per order × gross margin × monthly capacity cap × 12 ``` All four numbers are copied straight from your own price list and sale records — anyone can recompute them. This is the only ranking axis; there is no scoring. ```bars Figure: example — a clinic's own books, for ranking only, not a revenue forecast; when values are close, pick the one with less competition Buyer type 1 | 429000 Buyer type 2 | 259200 Buyer type 3 | 192000 Buyer type 4 | 158400 Buyer type 5 | 144000 ``` Four rules: fill in the numbers yourself, confirm them in person, and put them somewhere visible in the plan; **the monthly capacity cap is a hard cap** — if you can't handle that many orders in a month, don't pick a question for that many orders; annual opportunity value is for ranking only, never described externally as a revenue forecast, and it must never be multiplied, divided, or shown side by side with seat counts or enquiry counts; when values are close, pick the one with less competition. The reason for not ranking by a four-factor score: three of the factors measure the same underlying variable (which is equivalent to cubing it), and the fourth is really a proxy for search volume — it would let "ranking by search volume", the wrong method, back in through the side door, systematically pushing low-frequency, high-price work to the bottom. Dividing by cost on top of that adds only noise, because cost is roughly constant. ### The three checks ```mermaid Figure: the three checks, each with its own action when a question doesn't pass; check 2 must never be written off directly as "do not do" flowchart LR g1{"Check 1: did it name any business?"} -->|neither engine named any| out1["Drop from the pool, skip monitoring"]:::warn g1 -->|named someone, whoever it is| g2{"Check 2: ≥4 reachable-enough URLs?"} g2 -->|fewer than 4| re["Retry from another angle, never rule it dead"] re -->|still fewer| out2["Ruled do-not-do per the abstain line"]:::warn re -->|enough| g3{"Check 3: if you win it, can you deliver?"} g2 -->|enough| g3 g3 -->|won't take it or won't profit| out3["Drop it on purpose, and leave a record"]:::warn g3 -->|can deliver| go["Enters this quarter's candidates"]:::hl ``` **Check 1's definition is fixed**: naming means naming any company, not specifically you. A common misreading is "keep it only if you yourself are named at least 3 times out of 5" — read that way, every business that genuinely needs the work fails check 1, which turns away every target; and at n = 5, the difference between 3/5 and 2/5 is a coin flip. **Check 2** reads this intent's own intent-level top 10 and judges each URL's state one by one; ⚠️ it does not read the account-level Top 20. The four states and "reachable enough" are in 4.6; actions by state and "retry from another angle" are in 4.7. **Check 3** answers one question: the people this question brings in — will you take them, and will you make money? Questions like "the cheapest X" get dropped from the pool on purpose, with a record kept — write it into the plan so people can see what was turned away. Whether you can give a price was already asked once at question 3 in 4.1, so it's not asked question by question here. Whether attribution is possible is also not judged question by question in check 3: it's a site-level capability, not a property of any one question — asking it question by question always gets the same answer, and judging by channel for real would systematically cut every high-price, phone-order question. So it is handled with two acceptance tiers instead, set on the day work starts; nothing is dropped from the pool here: ```split Figure: the acceptance tier is set by the order channel; phone orders are never dropped from the pool for this alone Attributable tier || Observable tier Form/chat pre-filled, CRM tags complete || High price per order, phone orders, 3–9 month sales cycle Accepted per enquiry || Accepted by seats + position + on-the-list page count, plus manual attribution via a mandatory front-desk field ``` --- ## 4.6 Four states, two denominators and the abstain line **What you'll do in this section**: judge the state of every URL in the intent-level top 10, keep clear what the account-level Top 20 and the intent-level top 10 each govern, and count how many are "reachable enough" — fewer than 4, and this intent does not get worked on this quarter. This section is the book's only authoritative definition of the four states and the abstain line; later chapters only refer back to it and do not repeat it. ### How to assign a URL to one of the four states ```mermaid id=four-states Figure: how a URL is assigned one of the four states; "to ask" is only a transitional state and must settle into "reachable" or "not reachable" before the baseline flowchart LR q1{"Is your name on the page?"} -->|yes| a["Already present"]:::hl q1 -->|no| q2{"Is it one of the five not-reachable classes?"} q2 -->|yes| d["Not reachable"] q2 -->|no| q3{"Can you point to an entry point?"} q3 -->|yes, can point to it| b["Reachable"]:::hl q3 -->|can't tell, and it hasn't said it's closed| c["To ask"]:::warn c -->|reply to the enquiry says open| b c -->|no reply after 7 days| d ``` "Reachable" means this page accepts listings, submissions, publication, claims, nominations or corrections, free or paid; you must be able to point to the entry point: one of an entry-URL screenshot, an email receipt, or the place on the page where it says so explicitly. "To ask" is only allowed to exist at the coarse-screen stage; it must be cleared to zero by the baseline stage — send one enquiry per item. If "to ask" is still in the baseline table, that baseline isn't finished and must not be used for picking targets. The five not-reachable classes: government or statutory registers and notice pages; Wikipedia article text; pages under a competitor's own domain; platform-produced content pages (edited by the platform itself, explicitly not accepting outside submissions); closed-door ratings, or where a reply explicitly says it's not open, or 7 full days pass with no reply. Ruling Wikipedia's article text not reachable doesn't close that lead: its structured anchor, Wikidata, is reachable — whether and how to build it is in （→ 通用版 3.4 第三方资质分级、Wikidata 与五处商家档案）. Every judgement must keep four columns: state / judged on / basis / judged by. **A row with an empty basis column is always treated as "to ask"** — you can't tell whether it was "judged too generously at the time" or "that page really did redesign and close submissions", and the next action for these two is exactly opposite. **Judge by URL, not by domain**: under the same domain, some pages are open and some are not. How the four states map to the glossary is in （→ 通用版 D.2 术语表）; how they map to the judgement axes is in （→ 通用版 8.3 十个轴（一）：页型、容器、锚点、监管、问法） — "who controls the container" there judges a different thing; don't mix the two up. ### Two denominators, each with its own job ```split Figure: the two denominators each have their own job — never mix them Account-level Top 20 || Intent-level top 10 All intents merged, taken as the top 20 by URL-level citation count || Each intent takes its own top 10, from its own answers Frozen, unchanged for the whole quarter; it judges no single intent || Grouped by intent_id; only exists after the baseline Used only for the one number "on the list X/20" || Read for the abstain line, check 2, and the reachability side of not-moved triage ``` Why the two layers must be split: one shared account-level list always gives the same count, so using it to judge "do or don't do this intent" gives only two outcomes — every intent judged dead together, or none judged dead at all; the safety valve is effectively not installed. The data already exists; all that changes is how the summary script groups it — the raw citation records already carry "which question's answer this came from", it was just being merged away before. Intent-level takes the top 10, not the top 20, because a single intent's source pool is naturally shallow — URLs past rank 10 are mostly cited only once, and putting them in the denominator would only dilute the judgement. ### The abstain line The criterion counts the number "reachable enough": a URL only counts as reachable enough if it meets all four conditions **at the same time**. ```mermaid Figure: the abstain line — all four conditions must hold at once to count as reachable enough; 4 entries must also span two classes flowchart LR u["A URL in the intent's top 10"] --> c1{"State is reachable or already present?"} c1 -->|no| no["Doesn't count as reachable enough"] c1 -->|yes| c2{"Allowed on this side?"} c2 -->|no| no c2 -->|yes| c3{"Has an email or a public entry point?"} c3 -->|no| no c3 -->|yes| c4{"Free, or the rate card is already public?"} c4 -->|no, needs asking first| no c4 -->|yes| yes["Counts as one reachable-enough URL"]:::hl yes --> cnt{"4 reachable-enough URLs reached?"} cnt -->|fewer, or all in the same class| stop["Ruled do-not-do for this intent"]:::warn cnt -->|4 or more, spanning two classes| go["This intent continues"]:::hl ``` Whether it's "allowed on this side" is checked by side: on the strictly regulated side, list pages whose title is itself laudatory (the Best / Top N kind) don't count as reachable enough; the verdict is "request removal": if you are already listed, send a removal request, and if you aren't, leave the page alone; the full rule is 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）. If all 4 fall into the same action category, count them as 3. ```steps Figure: how "4" was derived — anyone can recompute it; don't substitute a different number 1 | Denominator | Intent-level top 10 2 | Split into four classes | already present / reachable · lists and directories / reachable · maps and reviews / reachable · own site 3 | At least two classes | Bet on only one class, and if that class won't move, the whole quarter has no fallback 4 | At least two per class | Only 1 = betting on a single page — one redesign and it drops to zero 5 | Multiply | 2 classes × 2 entries = 4 ``` ### Which list to read at which stage Which of the three to read at each stage is fixed; never mix them up (the ban on writing a bare "Top 20" or "top 10" is in （→ 通用版 D.2 术语表）): | Stage | Which list to read | How "to ask" is counted | Can it be reported externally | |---|---|---|---| | Deciding whether it's worth starting (4.8, shortcut 1) | A separate coarse-screen top 10 for each of the three questions, not merged | Counted as reachable (this half-day exists to talk you out of bad projects; better to re-judge at the baseline than to rule a URL out here) | No — n = 1, single engine, only serves the single decision "do it or not" | | Picking targets and off-site | The baseline's intent-level top 10 | Must already be cleared to zero by now | Yes | | "On the list X/20" | Only reads the account-level Top 20, unchanged for the whole quarter | — | Yes | --- ## 4.7 Actions by state, the ten question types and this quarter's slots **What you'll do in this section**: assign an action to every URL by state — send an update letter first for anything already present; decide, for each of the ten question types, whether it means writing a page, going off-site, or dropping from the pool; break each priority question down into the things AI will check, and match each one to a page or an off-site entry; then use the three combination rules to lay out this quarter's slots. When you're done, you'll have a list of actions by state and a finalised table of this quarter's buyer-type slots. ### What action each state gets ```mermaid Figure: the action for each of the four states — already-present goes first; the own site is not the default main battlefield flowchart LR al["Already present"] -->|do this first| fn["Send an update letter"]:::hl jr["Reachable · lists and directories"] --> pitch["Rate card, submit, watch nomination windows"] jm["Reachable · maps and reviews"] --> profile["Fill out the business profile's fact fields"] jo["Reachable · own site"] --> page["Write a page / expand an old page"] no1["Not reachable · government or big platforms"] --> retry["Policy question → execution question"] retry -->|still dominated| deny["Counted against the abstain line"]:::warn ``` For "already present", send the editor an update letter: new price (written per side), new credentials, new photos, an 80-word description. Adding content into a container that's already cited shows results in days; building a new page yourself takes 2–8 weeks (inferred, 65% confidence). For "reachable · lists and directories", ask about the rate card, submit, apply for a listing, and watch for the entry window on nomination-type lists — schedule it separately; on the strict side, first apply the separate rules for payment and for wording, see （→ 通用版 6.2 发信闸门、付费收录与夺取表）. For "reachable · maps and reviews", only fill in fact fields, and also claim Apple Business, Foursquare and Yelp: all free, about 2 hours in total. ChatGPT's local answers don't read Google Business Profile directly, and retrieval doesn't go through Bing alone either (70% confidence, evidence in （→ 通用版 A.1 机制、门与人这一层的证据）). Only for "reachable · own site" do you write a page or expand an old one: most citations are on third parties, so the own site is not the default main battlefield — it becomes one only when this cell says so. "Not reachable · dominated by government or big platforms" must not be ruled dead outright — doing so would wrongly kill a batch of regulated industries. Retry from another angle once first, turning the policy question into an execution question; only if it's still dominated after retrying does it count against the abstain line. > **Example** (local services) "Can this subsidy be used for this service" is a policy question; turned into an execution question it becomes: "If you do it with us, how much do you pay after the subsidy is deducted, and do we help you apply?" ### Where each of the ten question types goes ```split Figure: where each of the ten question types goes — shelf questions stay, educational questions drop from the pool Question type || Where it goes Second opinion / remedy || Priority; scheduled into the first batch (wording below) How much / cost breakdown || Priority; write price by side Symptom + option + price, three-part || Priority; this is a shelf question A vs B, which is better || Priority; lands on a qualifying comparison page Who to recommend / which is best || Priority; off-site gets you on the shortlist; your own pages get you through the check Which one nearby / near me || Priority; lands on business-profile fact fields Brand question / correction || A category of its own, takes slot 1 this quarter Scenario question (does this suit my case) || Do it once it passes the three checks Pure symptom type / what is it, should I see someone || Drop from the pool; this is an educational question, not the "who to buy from" stage ``` Price-type questions have a high rate of triggering AI Overviews, and mixed-intent questions an even higher one — numbers and sources in （→ 通用版 A.2 选点、写页、站外与复测的证据）. For "A vs B, which is better", the main action is "make a qualifying comparison page appear under this question": comparison pages' citation density is counted by page type, not by domain ownership — whose domain the page sits on is a secondary question; compare options, not businesses. Full rule 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）. "Who to recommend" works in two steps: AI first searches up a shortlist, then checks the businesses one by one (see "Question breakdown" below). The shortlist comes mainly from off-site — best-of blog posts already take up a large share of ChatGPT's citations (A.2), and directories and business profiles come in at this step too; for the check, AI searches with the named business's name or `site:` its website, and what it uses there are that business's own practitioner pages, service pages and price pages. So this cell needs work at both ends: off-site gets you on the shortlist; your own pages get you through the check. On the strict side, never seek a listing on Best/Top N pages or pay for one; if you are already listed, ask the publisher to remove you, see the same section 5.3. Brand question / correction is the only one where "stopping means falling back", and also the cell you can see with your own eyes and reproduce yourself. **The wording for second opinion / remedy questions is fixed** (applies together to the H1, H2, body, title and off-site material): the page's subject must be the buyer's situation, not another business's conduct. Write only the checks and how long they take, the available options and the conditions each one suits, itemised fixed prices, and what to bring. Do not write success rates, do not narrate what happened, do not show before-and-after comparisons; do not point at or disparage other businesses — named or unnamed is equally banned. In external documents, call this group only "second-opinion clients", never a label like "failed-case clients". > **Example** (dental and aesthetics) Second opinions such as orthodontic relapse or a loose implant have a price per order of S$8k–18k, are the least price-sensitive, and the buyer is bound to search on their own — so schedule them into the first batch. **How to describe the current state of second opinion / remedy questions**: in industries already measured, this type of phrasing shows no business named at answer level, and its citation slots are already occupied (internal ranking basis: region not locked, single engine, one-off, 85% confidence, evidence in （→ 通用版 A.2 选点、写页、站外与复测的证据）). This is "no one has been named in the answer yet", not "no one is doing this" — it is the question with the best chance of an exclusive seat within 90 days, but the work must count those citation slots into the citation-slot table too, not just write an own-site page. For your own industry, go by what your own baseline shows. **Say only this one line externally**: "Who AI names for these questions today, and whose pages it cites as sources, we will count on the spot in the baseline." Never say "no one's doing this" or "competition is near zero", and never give a number of businesses — the moment the other side searches for themselves next week and sees several businesses, you lose their trust. For symptom-type questions, keep only one distinction: a symptom question that includes an option and a price is a shelf question; a pure symptom question that only describes the symptom gets cited from authoritative educational sources by AI and almost never names a specific business — drop it from the pool directly, don't sit it in the monitoring layer first. > **Example** (dental and aesthetics) "My gums are swollen: do I need the tooth pulled, and how much does an extraction cost" is a shelf question — priority; "why are my gums swollen" is a pure symptom question — drop it from the pool directly. ### Question breakdown: what AI checks for each priority question Before it answers, ChatGPT searches the web, and it doesn't just search the buyer's own words once. From 225 stored ChatGPT answers we pulled out the searches it actually ran (101 multi-industry, 60 law firm, 64 dental and aesthetics): for "who's best" questions, the median answer ran 11–16 searches (dental 11, multi-industry 16); for information questions, 6–8.5 (dental 6, multi-industry 8.5). Of the 743 URLs cited in answers to "who's best" questions, only 49 (7%) ranked in Google's top 10 for the buyer's own words, while 363 (49%) ranked in the top 10 for one of AI's own searches (Google is only a proxy; for how this is counted, see （→ 通用版 A.2 选点、写页、站外与复测的证据）). The words AI adds are its own checklist, not the buyer's words. Of the 995 multi-industry searches for "who's best" questions (one search can fall into several categories), 373 were about credentials and official sources, 238 about price, 208 added a year, 105 were about reputation, and 247 used `site:` to limit the search to one website (about 150 of them to one business's own website); only 79 still carried best or top. The process roughly has two stages: first search up a shortlist, then check the businesses one by one. The named business's own pages are used mainly in the check stage. So once each priority question on this quarter's list is settled, break it down once more: ```split Figure: every category AI checks needs a page or an off-site entry to answer it What AI checks || Which page or off-site entry answers it Credentials and official registers || Registration fields on the practitioner page + the official register lookup Price || Price guide page, single-service price page Trade terms (the buyer's words swapped for professional ones) || Service pages and definition pages written in trade terms Reputation || Business profiles; handle reviews by your side Checking each named business (by name or site: website) || The named business's /facts, practitioner page, price page Finding candidates || Off-site: directories, profiles and third-party pages Year || The page states its update date ``` How to break it down: run the question through the API at least 3 times (the ChatGPT web app doesn't show everything it searched), save the raw web-search queries from each answer, and sort each one into a category in the left column above; then check whether each category has a matching page or off-site entry in the right column, and fill whichever cell is missing. The baseline round has to be run in full anyway (see （→ 通用版 4.3 粗筛分堆、满轮与品牌六问基线）); save the searches together with the answers, and no extra probe budget is needed. For which pages each industry should have ready, see chapter 4 of your industry edition. This is correlation, not causation: when the named business also has its own pages cited, it may be that AI first picked it on credentials and name recognition, then went to its website for material. Whether adding pages gets you onto the list has not been shown by any intervention experiment. So the question-breakdown table covers "when AI comes to check, do you have something to show it" — it doesn't promise a place on the list. ### This quarter's combination rules and slot allocation **Who**: the person setting the questions produces the list, the decision-maker signs · **Output**: this quarter's list (intent / buyer type / annual opportunity value / four-state action / acceptance tier) · **How to check**: the three rules are checked off one by one. Once the list takes shape, check the three hard rules off one by one; it counts as final only if all three pass: | Rule | Why | |---|---| | Must have ≥1 hold-pile question | The hold pile's baseline isn't 0, so a change in position can be measured within 90 days | | Must have ≥1 brand / correction question | Shows results fastest (what AI says often changes first once the fix is made, see 3.1), and you can reproduce it yourself | | National head term with no location ≤1 | Can't reach a determinable threshold within 90 days; each extra one you pick is one more entry guaranteed to end up written as "not demonstrated" | **These three combination rules are the minimum for the quarter to reach any verdict, and they come before the annual-opportunity-value ranking** — a quarter in which not one question can reach a verdict is an empty ledger, however valuable its ranking looks. But once the three are satisfied, every remaining slot follows annual opportunity value in descending order, and the order may not be adjusted for any reason. ```steps Figure: this quarter's slots are ordered this way — the combination rules come before annual opportunity value 1 | Slot 1 | Brand / correction question, filled as N/A · hold item 2 | Slot 2 | The hold-pile question with the highest annual opportunity value 3 | From slot 3 on | Taken in descending order of annual opportunity value 4 | Last slot | A national head term with no location, at most 1 and only in the last slot ``` Slot 1 is filled as N/A because a brand / correction question is not a buyer type, so it simply has no annual opportunity value to fill in. A high-value question squeezed out by the combination rules is logged into *Next quarter's first picks* by its annual opportunity value and taken up first next quarter — you don't go back and swap questions for it. Only when the three-pile list genuinely can't produce a hold-pile question or a brand/correction question do you go back to the three-pile list and swap in a question — that is not starting over. Every row in the list must state which buyer type it belongs to. --- ## 4.8 Two shortcuts and this chapter's checklist **What you'll do in this section**: with zero probe budget, use shortcut 1 (1 hour 40 minutes) to first decide whether it's worth starting; when pressed for time or able to do only three things, use shortcut 2 (48 hours) to produce a baseline and a first work order. The output of both shortcuts never enters the contract or the acceptance record. Only once every item on the checklist at the end of the chapter is ticked off do you move into writing pages. ### Shortcut 1: half a day, no probe budget, decide whether it is worth starting Use this when you haven't yet decided whether to put this quarter's effort in at all. The criterion in one line: at the coarse-screen stage, take one top-10 source-URL list per question (three lists, not merged), and count the "reachable enough" entries per 4.6: fewer than 4 rules that question out, 4 or more always continues; at this stage, "to ask" is counted as reachable. ```steps Figure: shortcut 1 — four steps, 1 hour 40 minutes in total, to decide whether it's worth starting 1 | 30 minutes | One cost question, one recommendation question, one scenario question; ask by hand, logged out, no probe tool 2 | 20 minutes | Copy a top-10 source-URL list for each question, three lists, not merged 3 | 40 minutes | Check each URL for a name and an entry point; don't send letters, don't ask about price 4 | 10 minutes | Count "reachable enough" and "to ask" for each list, write them onto the conclusion page one list at a time ``` The three lists in step 2 are called coarse-screen top 10, not the account-level Top 20 used as the acceptance denominator; merge them into one and step 4's three separate counts simply can't be done. ```mermaid Figure: shortcut 1's conclusion — how the three questions are judged together flowchart LR n{"Questions with ≥4 reachable-enough URLs?"} -->|≥2| go["Do it, following the full process"]:::hl n -->|exactly 1| one["Do it, starting from just 1 buyer type"] n -->|0| retry["Retry from another angle once"] retry -->|still 0| stop["Don't do it, or downgrade to audit + monitoring"]:::warn ``` If you start from just 1 buyer type, add more types once the baseline has run. Retrying from another angle follows 4.7's "not reachable · dominated by government or big platforms" cell: turn the policy question into an execution question. None of the numbers this step produces ever enter the contract or the acceptance record: the coarse screen is n = 1, single engine, by hand; a one-off, single-engine observation with n ≤ 3 must not be used as a result or as acceptance evidence. Any conclusion-page reference to this step's numbers must carry, in the same sentence, "coarse screen, n = 1, single engine, , not for external use". ### Shortcut 2: a 48-hour quick start for picking targets Take this route when things are urgent, or when you're starting with a small pilot first. **The output can only be used to pick targets — it can never be the acceptance baseline.** ```steps Figure: shortcut 2 — 48 hours to a baseline and a first work order Day 1 morning | Write 12 questions | 3 of each of the four shapes, classes A/B/C only, ≥7 with a location Day 1 afternoon | Three things in parallel | 12 questions × 3 rounds × 1 engine, a deep competitor review, brand-six × 2 passes Day 2 morning | Identity check | If identity or facts are wrong, fix the facts page and profiles first Day 2 morning | Door, half-day version | Read-only: the 12 retrieval UAs, the three nosnippet spots, the logs Day 2 morning | Shelf | Out of the 12 questions × 3 rounds, keep any where a peer business was named in ≥2 of 3 rounds Day 2 morning | Entry ticket | One top-10 list per question, judge the four states, count reachable-enough, decide where the effort goes ``` The four shapes are how much, who's best, A vs B, and which is nearby — never write an educational question. The half-day door version is read-only; don't touch anything. The deep competitor review takes the raw `domains_cited`; a person reads the brand-six answers for 15 minutes, used only to jump the queue and fix errors the same day — it is not a baseline, and it is not used to top up the full set either; the real baseline is run in full — 36 times — separately, in the freeze week (6 questions × 3 rounds × 2 engines, see 4.3). Day 2 morning's four signals take 2 hours in total; at the entry-ticket step, own site >40% of sources → work on your own site; lists and directories make up most of it → move the effort off-site. The output is a one-page *Target selection and first work order*, with three coarse-screen top-10 tables and three "reachable enough" counts attached. Shortcut 2 is the baseline for the "only three things" path (（→ 通用版 0.1 全书一句话与 90 天翻书顺序）), and it gives up a few things: n = 3, so a single question's conclusion is a direction, not a number; the dead pile can only say "no business named in the coarse screen"; there's no Chinese or minority-language coverage; there's no full baseline and no noise band — the full baseline and the noise band are filled in, in parallel, during the first week of work. **Shortcut 2's baseline must also be saved before the door is fixed** — this discipline doesn't relax just because you took the shortcut. ### Checklist for this chapter - [ ] **Mechanical gate: if an intent has no "before" sample, its page may not launch** - [ ] The ruler is validated: a question known to get cited was run first; if the citation didn't show up, the run stopped - [ ] The question pool is signed: all four source labels present, the 30-question ratios met, checked by a second person, signed in person by the decision-maker - [ ] The dead pile was re-read: a second person re-read the original answer text before it was finalised, up to 3 questions lifted back into the battlefield pile - [ ] Full rounds, all five run: battlefield pile + hold pile, five rounds each × two engines; a single screenshot on a single day does not count as a baseline - [ ] The noise band was measured: same-week retest, the value written into `noise-band.md` - [ ] The web leg ran first: run before the account-level Top 20 was frozen, unique domains forced in, `sources-top20.csv` has a `web_only` column - [ ] Counts kept separate: listed as a source and written into the answer recorded separately, URL-level and answer-level separate, failed samples listed separately, no single-question small numbers - [ ] The control agreement is signed: same-site pages not to touch, three competitors, the primary number to watch, and when to stop are all written down - [ ] The shelf snapshot was reviewed page by page: anything naming a third party is marked internal, not put up for display - [ ] The buyer type table's price per order, gross margin and monthly capacity cap were confirmed in person by the decision-maker - [ ] The three-pile list has been reviewed, the three checks completed, and every entry in the four states has a judged-on date / basis / judged-by name - [ ] This quarter's list passed all three combination rules, every row states its buyer type and carries its annual opportunity value and check-2 action, anything dropped from the pool has a record, and the decision-maker has signed; if it falls short, go back to the three-pile list and swap in a question — that is not starting over Coarse-screen numbers never enter the contract or the acceptance record; any reference to them must carry, in the same sentence, "coarse screen, n = 1, single engine, ". Only once every item on this checklist is ticked off does writing pages unlock — the first page starts with the price guide page, see （→ 通用版 5.4 三十天写页顺序与三道机械闸）.