# Chapter 8 · Using this in a new industry: four steps to work out your approach

> GEO Playbook · General Edition v1.0 · Canlah AI · CC BY 4.0 · Web page: https://canlah.ai/playbook/new-industry/
> 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.

## 8.1 Four steps for a new industry: ten questions → side check → ten axes → decision table

**What you'll do in this section**: work through the ten questions to get the raw input for the ten-axis table; then decide the side (8.2); fill in the ten-axis table (8.3–8.4); find the closest row in the decision table (8.5); and finally have this industry's ban list ready. Only when all four are in place may you open the *First work order*.

Get the starting point straight first: a one-line verdict like "this industry is hard to work" or "this industry is highly competitive" gives the people who execute nothing they can act on. What actually helps is to sort things into two groups: impressions that look like variables but change no action when they change, and real variables where a different value means rewriting tasks.

```split id=one-line-vs-playbook Figure: A one-line verdict gives you no playbook; a real playbook must give five things
Looks like a variable, but isn't || What a real playbook must give you
Industry size / search volume || Measured share of page-type mix (axis 1)
How fierce the competition is || The four states, URL by URL, in the intent-level top 10 (axis 2)
Budget size || Entity anchor layer (axis 3)
B2B or B2C || Side + switches + substitutes (axis 4 + 8.7)
Whether you have brand awareness || The row in the decision table + the first 30-line work order
```

Why each of the five items in the left column doesn't count: industry size / search volume changes no action — priorities are ranked by annual opportunity value, not search volume; how fierce the competition is is an adjective, and its actionable version already lives in axis 2 (how many reachable-enough items) and the abstain line J1; budget size changes how much you do, not how you do it; B2B or B2C is too coarse — two businesses that both wear the B2B label often differ more from each other than a B2B business differs from a B2C one, and what actually matters is axis 6 and axis 2; whether you have brand awareness doesn't change the order of actions, it only changes how high a starting point you get on the "get identified correctly" path (for the three paths, see （→ 通用版 1.3 三个关口与三条路）).

**The only test for a qualifying dimension (fixed)**: change one value, and at least 5 of the 30 lines in the *First work order* must be rewritten. If it can't force 5 lines to change, it isn't a dimension.

The order of the four steps cannot be shuffled — deciding the side comes before the ten axes: axis 4 (side and switches) and the decision table both use the side-check's conclusion, so deciding the side first avoids referencing forward to something not yet worked out.

```steps id=route4 Figure: order for a new industry; decide the side before the ten axes
① | Ten questions | Work through the ten questions, get the raw input for the ten-axis table, and judge J0/J1 along the way
② | Decide the side | 8.2's five steps: qualify → check the licence catalogue → check the regulator → check product law → reverse check
③ | Fill in the ten axes | 8.3–8.4; axis 4 directly uses step ②'s conclusion
④ | Decision table | 8.5: find the closest row, get the main battlefield and the first batch of actions
```

### The ten questions: for each, where to look, what to check and which axis it fills

Take an industry you've never seen before: ask these ten questions and you can fill in the whole ten-axis table.

| Question | Where to look, what to check | Which axis it fills |
|---|---|---|
| Q1 In this category's answers, does AI name any company? Which layer of noun does it name? | Gather 15 real questions (front-desk wording, customer-service transcripts, autocomplete suggestions from platform search boxes), run each on both engines; count whether any company is named at all, and check whether the subject of the naming sentence is an organisation, a person, a product or a course | Axis 3, axis 5; also judge J0 |
| Q2 What page type do the citations for these 15 questions' answers come from? | Open every source URL one by one and hand-label the page type, judging only by the visible page, never by the domain's brand; work out the share | Axis 1 |
| Q3 On the pages that get cited, can you get your own words onto them? | For every URL in the intent-level top 10, look for an entry point: a submission page, listing page, claim page, correction page, nomination page or seller back-office; if you can't find one, send an enquiry, and no reply within 7 days counts as not reachable; count the number of reachable-enough items | Axis 2; also judge J1 |
| Q4 Of the pages that get cited, how many are "the container is not yours but the words are"? | Ask, one by one: "Am I the supplier of this page's content?" — your product page on a platform, your vendor profile on a review platform, an app store page, an integration marketplace listing, your own entry in an official directory, your speaker profile on a platform: all of these count as "yes". This is the question people most easily miss entirely | Axis 2 |
| Q4b Does the platform itself let AI crawlers in? | For every platform you sell on, `curl robots.txt` and check the retrieval-leg crawler group; log it by date and rerun every quarter | Axis 2 |
| Q5 Who regulates this business? Is there a written advertising regulation? | See 8.2's five steps | Axis 4 |
| Q6 Which tactics get taken away? What replaces the ones that are taken away? | Go through 8.6 tactic by tactic and check each one against the rule text; for anything taken away, find a substitute in 8.7, and find an actual example of what that substitute looks like in this industry | Axis 4 |
| Q6b Can someone else publish on your behalf? Who is the "authorised person"? Is the publishing party jointly liable? | Check this industry's rule text for three things: ① how "publisher" or "cause it to be published" is defined, and whether it names the publishing party, an agency, an affiliate or an influencer among the liable parties; ② whether there's a penalty clause for joint liability when publishing on someone's behalf; ③ whether there's an authorisation document you can actually sign (its name, who signs it, which media it covers). If even one of the three can't be found, the verdict is "cannot publish on the client's behalf" — not "can" | Axis 4 sub-item (form of agency liability, see 8.6 item 7) |
| Q7 How big is one order, how many orders can you take in a month at most, and what's the bottleneck? | Get the price list and the last 12 months' sales records; being unable to answer "what's the bottleneck" = no hard cap, switch the numerator | Axis 7 |
| Q8 For the last three deals, who made the call, how many people signed off, and how many days did it take? | Get three specific deals; if cross-functional sign-off shows up even once, count it as a committee | Axis 6 |
| Q9 How do buyers find you? Can the back office see the source? | Check whether there's a source field, whether sessions coming from AI carry a referrer, and whether you can add a required self-reported-source field; where the platform side can't show a referrer, just log "can only read the brand-term curve, for reference, not attribution" | Axis 8 |
| Q10 How often does the price list change? How many buyers search in the local language? | Compare historical versions of the price list (don't take an estimate); run one pass of the local-language questions, compare the named sets and take the difference set — check whether the businesses in that difference set all have a local-language page | Axis 9, axis 10 |

🔴 **Q6b must be answered within the ten questions — don't let it surface only after work has started**: when the answer is "cannot publish on the client's behalf", the whole off-site block of work (bylined contributions, outreach letters, submissions, filling in platform back-office fields) can only be carried out, in this industry, by the licensee itself; putting someone else on it in the work order still won't get it executed.

Only when all four are in place may you open the *First work order*: one line of the *Side receipt* (8.2, written at the top of the work order), a fully filled ten-axis table (with one sentence per axis on "which point of leverage it lands on"), the matching row in the decision table, and this industry's ban list (for the strict and light sides) or platform-policy table (for the unregulated side).

## 8.2 Five steps to decide the side, and three criteria

**What you'll do in this section**: when 0.2's four questions don't give a firm answer, work through the full five steps — three questions to qualify, check the licence catalogue, check the regulator's three kinds of document, check product law, and reverse-check if you find nothing; steps 3 and 4 are each capped at 1 working day, and if either overruns or cannot be confirmed, always start work on the stricter side; then use three criteria to decide switch A (agency liability) and the cost of judging wrong; the side decision has an expiry date and is rechecked every quarter.

0.2 gives you the quick version: four questions that give you the side directly (strictly regulated / lightly regulated / unregulated) and switch E (how many tactics get taken away). For most industries that's enough. This section is the full version: five steps to gather all the documents you should read, use those documents to verify 0.2's four questions properly, and then use three criteria to work out two more things — switch A (who can publish on your behalf, and whether they're jointly liable) and the cost of judging wrong. The two sections use the same terms: the three sides plus the six switches, and no other tier gets invented. In this book, "regulated sides" means the strict side and the light side combined.

**How to split the work**: steps 1 and 2 can be finished in the same meeting (ask three questions + search the licence catalogue once); steps 3 and 4 don't belong in that meeting — reading the rule text for penalty and joint-liability clauses, and checking product law category by category for what's sold, are not things one meeting can finish; force them through in one sitting and what you get is a guess.

```steps id=side5 Figure: five steps to decide the side; steps 3 and 4 are each capped at 1 working day
① | Three questions to qualify | Does opening the business need a licence? Does what you sell need registration? Do the people doing the work need personal registration?
② | Check the licence catalogue | Search the cross-agency licence catalogue by business keyword; not found ≠ unregulated — go on to step ③
③ | Check the regulator | Advertising regulation / ethical code + FAQ + penalty clauses; capped at 1 working day
④ | Check product law | Check by product category, not by company; capped at 1 working day
⑤ | If you find nothing, reverse-check | The four base layers still apply; any one of the three reverse checks that hits moves you up one level
```

**Time limit and escalation (fixed)**: if step 3 or 4 overruns, or you checked but cannot confirm it, start work on the stricter side — never let it enter the work order marked "TBD"; leaving it as TBD means you simply can't schedule the first batch of tasks.

If all three questions in step 1 come back no, provisionally judge unregulated, but the verdict isn't complete until you've worked through step 5's three reverse checks. Not finding anything in step 2 doesn't mean unregulated: the licence catalogue is indexed by licence name, not by a business's common name, so not finding it means moving on to step 3 — never judge unregulated directly.

**One addition to step 1: check the statutory hard definition first.** Many rules mark out who's regulated with a string of conditions, and the conditions must **all hold at the same time** to count as a hit — miss one and it's not a hit.

> **Example** (education and training): one category's definition of "full-time" requires all three of "duration ≥1 month + 5 days a week + ≥3 hours a day" to hold at once before it counts as a hit; a tuition class held two hours a week doesn't hit.

### Step 3: check the regulator's three kinds of document

General-pattern sentence: **every regulated industry has three things — a regulator, a dedicated advertising regulation, and an official register lookup — fill in all three before you decide the side**; the specific agency names and rule names are an industry variable and are left to the industry edition. On the regulator's own website, check three kinds of document: the advertising regulation / ethical code (including its section on promotion), the agency's FAQ (its stated position is often stricter than the rule text itself), and the penalty and joint-liability clauses.

### Step 4: check product law to decide switch C (regulated products) and switch D (spans two sides)

In some industries the company itself needs no licence, but **what it sells** is governed by a dedicated regulation — the copy is regulated all the same, and that's switch C (regulated products). Decide the side by the product category you sell, not by the company; if any cluster of product categories or business lines under your name hits, run the whole site to the strictest cluster — that's switch D (spans two sides).

> **Example** (e-commerce): once SKUs such as supplements, medical devices or infant food hit a dedicated regulation, the whole site must follow the strictest requirement in that cluster, even if the company itself needs no industry licence at all — you can't decide the side by company; AI treats one domain as one entity, so splitting subdomains into stricter and looser sides just breaks the entity apart.

### Step 5: what to do when you find nothing

Judge it as the **unregulated side**, but the base layer isn't "no rules".

| Base layer | Hard constraint |
|---|---|
| Fair trading / consumer protection law | No misleading statements; false reviews and misleading web design have already led to enforcement cases |
| Price transparency guidelines | Covers drip pricing, strikethrough prices and discounts, "free" claims; mandatory fees must be included in the stated price |
| Personal data and the Do Not Call Registry | Sending a message needs either clear consent, or the register checked within the last 30 days; **B2B messages are exempt from this rule** — a real loosening |
| Platform terms | Often stricter and faster to enforce than the law (store suspension, review removal, ranking demotion) |

Before you judge unregulated, you must run three reverse checks — any one that hits moves the case up one level:

```mermaid id=reverse-check3 Figure: three reverse checks before you judge unregulated
flowchart LR
  q1{"Any health, efficacy or results claims?"} -->|yes| up["Move up one level"]:::warn
  q1 -->|no| q2{"Does the audience include minors?"}
  q2 -->|yes| up
  q2 -->|no| q3{"Are the buyers overseas?"}
  q3 -->|yes| cross["Move up one level and go to 8.9"]:::warn
  q3 -->|no| free["Judge unregulated side"]:::hl
```

The claim categories in the first question, in full, are six: health, efficacy, earnings, results, employment and salary; any one of these appearing in the copy counts as a hit.

### Three criteria decide switch A (agency liability) and the cost of judging wrong

Run the documents found in step 3 through three independent criteria. First fix what counts as a "hit" for each one, then look at the combination of hits — don't judge by "all hit / not all hit".

| Criterion | What counts as a hit | How to check |
|---|---|---|
| ① Written advertising rules | There's a dedicated advertising regulation, or the ethical code / industry code has a section that actually binds outward-facing promotion | Search the document for advertisement / advertising / publicity / promotion, or their Chinese equivalents; if you can't find a binding clause, it doesn't hit, however thick the document is — a document existing isn't the same as a hit |
| ② Penalties | The rule text states a fine amount or a prison term | Look for the penalties section or a "penalty" clause |
| ③ Agency liability | The rule text names the publishing party, an agency, an affiliate or the publisher among the liable parties | Look for wording like issue or publish / cause to publish / authorised person, or the Chinese equivalent of "cause to publish" |

```mermaid id=switch-a Figure: each criterion decides one thing: ① gives the four questions something to answer against, ② sets the cost of judging too loosely, ③ alone decides switch A
flowchart LR
  doc["Documents found in step 3"] --> c1{"① Is there a clause that binds promotion?"}
  c1 -->|yes| side["Decide the side using 0.2's four questions"]
  c1 -->|document exists, but no clause| n1["① does not hit"]
  doc --> c2{"② Does it state an amount or a prison term?"}
  c2 -->|yes| cost["Judging too loosely triggers the penalty"]:::warn
  doc --> c3{"③ Is the publishing party a liable party?"}
  c3 -->|yes| A["Switch A is on: agency liability"]:::hl
  c3 -->|no| noA["Switch A is not on: agency liability"]
  A -->|① and ② both miss| rare["Rare; still treat as the stricter side"]:::warn
```

Hitting only ③ (agency liability exists but there's no advertising regulation) is a rare combination, but always treat it as the stricter side: ③ decides whether this approach can be carried out at all, and the cost of judging it wrong is the most asymmetric of all the combinations. For which form of agency publishing applies once switch A (agency liability) is on, and what hard consequences each one carries, see 8.6 item 7. Check switch B (legally required fields) and switch F (referral commissions banned) off the same batch of documents while you're reading them; the conditions for hitting them are in the switch table 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）.

The cost of judging wrong is asymmetric, and that's the only reason behind the rule "when unsure, always treat it as the stricter side".

```split id=cost-asym Figure: the cost of judging wrong is asymmetric; when unsure, always treat it as the stricter side
Judged too loosely (should be strict, judged loose) || Judged too strictly (should be loose, judged strict)
May trigger penalties and joint liability || You only give up a few tactics
```

### On the unregulated side, check these three places instead

On the unregulated side there's no regulator to check, so check three other places instead; what you're guarding against also changes to three "pass-through red lines" — the only pocket of strict regulation inside an otherwise unregulated industry.

```split id=free-side-check Figure: with no regulator to check on the unregulated side, the three kinds of document become three places to check instead
On regulated sides, check these three kinds of document || On the unregulated side, check these three places instead
Advertising regulation or ethical code || Your target review platforms' review-solicitation policy
The regulator's FAQ || Your target app and integration marketplaces' listing policy
Penalty and joint-liability clauses || Your jurisdiction's consumer-protection and false-comparative-advertising clauses
```

| Pass-through red line | When it triggers | What to do |
|---|---|---|
| A certification word is a legal claim, not an adjective | The page carries a security, privacy or payment compliance-certification word | Next to every such word, spell out the form of proof (type, how long the report stays valid, who issued it, whether it can be obtained under an NDA) and the last check date; writing "compliant" when the reality is only "aligned" is actionable |
| A buyer's own regulation passes through onto your claims | The product is sold to buyers in a regulated industry or to government buyers | Any wording that touches the buyer's industry regulation must be checked by legal, and written to the strict side's standard |
| Platform terms bind harder than the law in practice | Active on review sites, app stores, integration marketplaces or package registries | Read each platform's review-solicitation policy / listing policy clause by clause before you act; delisting or profile freezes arrive faster than regulatory risk does |

### Three edge cases (fixed)

- **One company spans two sides**: treat the whole site by the strictest side; never split strict and loose by subdomain — AI treats one domain as one entity, and splitting it apart just breaks up the entity's weight.
- **A group with multiple legal entities**: decide the side by whichever entity is actually named, and treat the rest of the group's entities by the strictest one.
- **An agency or distributor publishing on the brand's behalf**: decide who's the "authorised person" first; where it's unclear, treat it as the brand's own side — unclear liability = cannot publish = this approach does not exist.

### Side decisions expire: recheck every quarter

🔴 **In an industry the regulator is actively reviewing, the side may move up within the current year.** Recheck the regulator's website for the three kinds of document every quarter — don't treat a single side decision as a permanent verdict.

**Acceptance**: the output is one line of the *Side receipt* plus a *To-verify list* (each item records three things: what's to be verified, what it's stuck on — the register page was redesigned / the rule text is only a PDF / the agency hasn't published it — and the next recheck date), plus one more check of the register screenshots; the first time you actually use a register lookup entry marked "to be verified", it must be verified on the spot and filled back in.

## 8.3 The ten axes (1): page types, containers, anchors, regulation, question shapes

**What you'll do in this section**: use seven questions to label the baseline Top 20 by page type and work out the dominant page type (axis 1); use two questions to decide the control of citation containers for each URL (axis 2); read real answers to fix the entity anchor layer (axis 3); the regulation axis uses 8.2's conclusion directly (axis 4); and pick your battle tactics by the four question shapes (axis 5).

### Axis 1 · Shelf page-type mix

Open every item in the baseline Top 20 (each intent-level top 10 counted once) and hand-label the page type — one label per item, never guess, judging only by that page's layout, never by the domain's brand. Work through the seven questions in order; stop at the first hit.

```mermaid id=page-type-7q Figure: label the page type with seven questions in order; stop at the first hit
flowchart LR
  q1{"Is it a list of multiple parallel subjects?"} -->|yes| n_lc["listicle/category"]:::hl
  q1 -->|no| q2{"Does it compare 2 to 3 subjects side by side?"}
  q2 -->|yes| n_cmp["comparison"]
  q2 -->|no| q3{"Does it cover only one subject?"}
  q3 -->|yes| n_pp["profile/pdp"]
  q3 -->|no| q4{"Is it product documentation or a spec manual?"}
  q4 -->|yes| n_docs["docs"]
  q4 -->|no| q5{"A user question plus several answers?"}
  q5 -->|yes| n_qa["qa"]
  q5 -->|no| q6{"A government statutory register?"}
  q6 -->|yes| n_reg["registry"]
  q6 -->|no| n_art["article"]
```

Three things the figure can't show: ① after hitting "listicle/category", ask one more question — "does it use ranking or better/worse wording?" — yes → listicle, no → category; ② after hitting "profile/pdp", ask "can you place an order?" — yes → pdp, otherwise → profile; ③ two people each label independently; where they disagree, don't negotiate a compromise — take whichever label comes from the lower-numbered question (the order of the seven questions is itself the priority), and anything still unresolved is logged as unclassified, listed in its own column and not spread into any label's share; if the two people's disagreement rate is >20% for that round, the whole round is void and must be relabelled — never use it directly to decide where effort goes.

```split id=page-type-fire Figure: the dominant page type decides where effort goes
Dominant page type (share threshold) || Where effort goes
article ≥40% || Push earned-media bylined contributions
listicle ≥30% || First judge whether you can touch the list (axis 4); if you can, fight for an editorial slot; if you can't and you are listed, request removal
category + profile ≥30% || Directories move from a logging action to the main effort
comparison ≥15% || Build your own comparison page / price guide page, eating the same density
registry ≥30% || The action is checking your records are correct, not submitting
qa ≥20% || Go to axis 2 and judge whether the community is platform-owned content or reachable
pdp ≥30% || Go to axis 2 and judge C or D; answer the "words" question and the "door" question separately
docs ≥20% || Build your own documentation site + integration and open-source directories
```

Write the shares as `article x% / listicle y% / comparison z% / …`; you can hit several thresholds at once, and order effort by share, highest first. comparison has the highest per-page citation density of any type in the table (1.87 citations per search), and it's counted by page type, not by who owns the domain — a comparison page you build yourself eats that same density. pdp and docs are also judged by layout: a media outlet's long buying-guide article is an article, and a brand's own product page on its own site is a pdp, regardless of whose domain it sits under.

**Page-type distribution is an industry variable, not a general rule**: you must label the page types yourself; never extrapolate from another industry.

> **Example** (dental and aesthetics): in the medical sample, article accounts for 54% of citations and listicle 0% — lists simply aren't the shelf in that industry; directories and lists are only a logging action. In a different industry these two numbers could well be reversed (for the sample, see （→ 通用版 A.1 机制、门与人这一层的证据）).

### Axis 2 · Control of citation containers

Judge who owns the container URL by URL (not domain by domain) in the intent-level top 10. Ask two questions in order: are the words on this page mine, or can I submit them (does the page have a claim, submission, correction, nomination or seller back-office entry point)? And who owns the container? The basis for the verdict must be one of: a screenshot of the entry-point URL, an email receipt, or the location of an explicit statement on the page.

```mermaid id=container-2q Figure: two questions to decide who owns the container
flowchart LR
  p1{"Are the words mine, or can I submit them?"} -->|no| D["Not reachable"]:::warn
  p1 -->|yes| p2{"Does the container belong to my own domain?"}
  p2 -->|yes| A["Own-site slot: your own site can take it"]:::hl
  p2 -->|no| p3{"Does it need someone else's approval?"}
  p3 -->|yes| B["Third-party slot: needs someone's approval"]
  p3 -->|no| C["Platform slot: fill in back office, goes live"]
```

The own-site slot has one more hard test: this page type is either already present, or can be built this quarter, on my own domain — if you can't build it, it doesn't count as an own-site slot. The line between third-party slot and platform slot is whether there's human editorial review: editorial slots, submissions, listings and nominations need someone else's approval, so they are third-party slots; if what you fill in the back office goes live straight away, it is a platform slot.

The four containers here (own-site slot / third-party slot / platform slot / not reachable) and the four states in （→ 通用版 4.6 四态、两层分母与弃权线） are two different things: the four states judge "can I get in", the four containers judge "who owns the container". How they map: own-site slot, third-party slot and platform slot all fall under the four states' "already present" or "reachable"; "not reachable" is simply the four states' "not reachable"; "to ask" isn't a fifth value — it's the state of a third-party slot where asking has not yet produced an answer. The count of "reachable enough" items is counted the way 4.6 defines it; neither not-reachable nor to-ask counts, and this is the number J1 uses.

Industry-level aggregate threshold (this is exactly what the decision table uses — never use an adjective instead): if a value appears in ≥5 items of the intent-level top 10, log it as "many / high"; ≥7 items, log "very high / saturated"; ≤2 items, log "few".

| Which cell has more | Where it points |
|---|---|
| More own-site slots | The main battlefield is your own-built pages |
| More third-party slots | Put people on bylined contributions and listings |
| More platform slots | The cell most often missed entirely: you are the owner of that page's content — changing the words costs almost nothing, goes live almost instantly, and there is no 2–6 week outreach cycle |
| More not-reachable | Count the reachable-enough items; if that hits J1, rule it out |

Three fixed rules: ① The definition of "not reachable" is "you cannot get your own words onto that page", not "this domain isn't yours" — a large share of "platform pages" ruled dead are actually C, not D; getting this wrong once means throwing away the whole main battlefield. ② The reverse also holds: it's only D when the platform itself has shut the door, and this must be judged retrieval-leg by retrieval-leg — never judge one overall state for a platform. ③ Measure platform by platform: of two equally large platforms, one may be half-blocked and the other fully open; you must `curl -L` it yourself, log it crawler-identity by crawler-identity, log it by date, and rerun it every quarter — what gets logged is "which leg is C, which leg is D", not "is this platform open"; a "to ask" that gets no reply within 7 days converts to D — never leave it hanging as a middle state.

> 🔴 **Fixed**: if the government writes about you, it's an official register — you can only check it for accuracy; if you write about yourself, and the government or the platform merely supplies the container, that's a C-type platform — judging it as D throws away the cheapest, fastest-acting cell entirely.

### Axis 3 · Entity anchor layer

Read 5 real answers and look at what noun is the subject of the naming sentence; then open three third-party pages from the baseline Top 20 and see which layer their entry titles use.

| Value | Where it points |
|---|---|
| Organisation | The anchor is the organisation facts page `/facts` (see （→ 通用版 3.2 事实表与机构事实页）) |
| Natural person | One page per person, `/team/<name>`, with every thick page, video and off-site material's byline linking back |
| Product / SKU | The anchor is the product page, not the company page; `sameAs` strings together product-level identifiers (version number, package name, GTIN, bundle id); third-party profiles are built by product, not by company |
| Parent brand | The sub-brand and the parent brand interlink and match word for word in the visible HTML |
| Course / cohort | The anchor is the course; the instructor is the second anchor |

You can have dual anchors (the parent brand feeding the answer shelf + the SKU feeding the shopping shelf), or split by sub-trade. **The real mechanism behind entity strength isn't what the schema says — it's the same set of facts appearing word for word across several independent third-party pages that can be crawled.** Pick the wrong anchor layer and you've aligned the wrong object word for word — wasted work. Fill in the word-for-word alignment table for the six public traces as usual; only the six places change to this industry's own six.

The actions in the "identity" layer are the same across all three sides, with not one step dropped: unique spelling → `/facts` page → six-trace word-for-word alignment → send a correction the same day for any mismatch → write `sameAs` while you're at it. Only one thing changes — where the strongest anchor comes from, which depends on whether there's an official register.

```split id=anchor-source Figure: whether there's an official register decides whether the strongest anchor is free for the taking, or something you have to build
Regulated sides: free for the taking || Unregulated side: you build it yourself
Official registration number + link to the official register lookup page || A general business-registration number (not an industry licence)
Write it in the first line of `/facts` || Leaving the registration-number line blank is a mistake — swap in an equivalent unique string instead
At least two of the six traces are official pages || A unique package name in a package registry or app store + your vendor-profile ID on a review platform
The anchor exists from day one || The five business profiles must match word for word; fixing the entity often comes before adding pages
```

### Axis 4 · Side and switches

This axis uses 8.2's side decision directly: the side, plus whichever switches hit — the ones that most often change the actions are switch A (form of agency liability, see 8.6 item 7), switch C (regulated products), switch D (spans two sides: run the whole site to the strictest side) and switch E (peer comparison banned). This axis produces no rankings itself; it decides which of the ranking-producing actions you're not allowed to do. The "door" layer doesn't change by a single word: the robots.txt master template, the three `nosnippet` controls, matching the four identities, the CSR check — these are identical, word for word, on the strict side and the unregulated side; for the "identity" layer, see axis 3 above. **Regulation takes tactics away; it doesn't change the building site, but it does change what goes into the "identity" cell.**

### Axis 5 · Question-shape mix

Run each of the 15 frozen questions on both engines, read the shape of each answer and count it, and write it as `recommendation x / spec y / comparison z / approach w`.

```mermaid id=intent-4type Figure: four question shapes, four battle tactics; use the tactics for the wrong shape and even a well-built page won't get in
flowchart LR
  f1{"Does the answer list a string of company names?"} -->|yes| rec["Recommendation type: seat battle"]
  f1 -->|no| f2{"Is the answer a number or a deadline?"}
  f2 -->|yes| spec["Spec type: battle to be the source of the number"]:::hl
  f2 -->|no| f3{"Does the answer explain how two differ?"}
  f3 -->|yes| cmp["Comparison type: build your own comparison page"]
  f3 -->|no| plan["Approach type: blend into the methodology corpus"]
```

The recommendation type fights over the 5–8 names in the answer; the main effort is off-site — planting your words into pages already cited, plus building your own list-style landing page (unregulated side only; the strict side does not self-build list pages, and neither does the light side — Conservative line (not statute text); on the light side the client's written reasons can relax it; see （→ 通用版 0.3 哪些页型对你有限制、改做哪一型） and 8.6). The spec type's target isn't getting listed — it's becoming the source of that number: a visible table + a stated basis and check date + ≥3 outbound links to checkable original sources (on the strict side, outbound links from your own website never go to commercial companies; linking only to government bodies and regulators is a Conservative line (not statute text), see （→ 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）). **Price and cost queries trigger AI Overviews >80% of the time, and mixed-intent queries like "roughly how much does XX cost" 97% of the time** — the richest cell, shared by every industry.

The approach type is the hardest, and the one most easily given up on: your company name only appears as an example, and the only lever you have is getting your name into the body text of methodology someone else has written (bylined contributions, being cited on documentation sites, open-source and integration directories, long community answers). The basis: among the citation sources of approach-type answers, article has the highest share, and 84% of those article citations are earned media — label it "not measured" and do it anyway. Apply recommendation-type tactics to an approach-type category, and however well the page is built, it still won't get into the candidate set.

## 8.4 The ten axes (2): decision-makers, capacity, attribution, half-life, language

**What you'll do in this section**: take the last three deals and pin down who made first contact, who made the call, how many people signed off, and how many days to payment (axis 6); pin down the monthly capacity cap and pick the right annual-opportunity-value formula (axis 7); check whether the source of a sale can even be seen (axis 8); see how many times a year the price list changes and set the half-life level (axis 9); and compare the local-language named set to decide whether to build a local-language page (axis 10).

### Axis 6 · Number of decision-makers and decision cycle

Ask about the **last three deals**: who made first contact, who made the call, how many people signed off, and how many days from first contact to payment — get three specific deals, not an average. Values: single, on the spot / single, over weeks / two people / two people + a third-party user / committee (3+ people across functions).

Single → one intent cluster, one page, no sections inside the page; **committee → the definition of a page unit must change**: a page is still one intent cluster, but is split into H2 sections by role inside the page, each section carrying its own checkable facts and sources; two people + a third-party user → a fixed two sections inside the page (a checkable-facts section for the buyer + one section for the user); high-ticket durable goods with no functional division of labour → split into H2 sections by "who suits which option". **AI picks up sections, not pages**: committee-purchase questions are already asked role by role, so multiple sections on the same page means the same URL has copyable body text against several intents at once (not measured).

### Axis 7 · Shape of the capacity cap

Ask one question: "How many orders can this line take in a month at most, and what's the bottleneck?" Being unable to answer means there's no physical hard cap — switch the numerator. Values: headcount or machine count / professional's working hours / capacity × how often a cohort runs / inventory and supply volume / no hard cap.

```mermaid id=year-value Figure: whether the question can be answered decides which annual-opportunity-value formula to use; the two formulas share the same shape
flowchart LR
  q1{"Can you answer how many orders a month at most?"} -->|yes| hard["Has a hard cap: use the capacity-cap formula"]
  q1 -->|no| soft["No hard cap: use the historical-actuals formula"]
  hard --> same["gross margin per order × orders per year"]:::hl
  soft --> same
```

Both formulas must share the same shape: the annual opportunity value of one work unit = gross margin per order × orders per year; "orders per year" must be something you can look up from the back office, a rota, a course schedule, a stock table or a sales ledger — never estimated on the spot. With a hard cap: `price per order × gross margin × monthly capacity cap × 12`, where the monthly capacity cap comes from the rota or course schedule; for goods sold, write it as `gross margin per unit × monthly supply volume × 12`, where monthly supply volume comes from purchasing or restocking records — that number is itself the physical hard cap for this cluster of product categories. With no hard cap: don't ask "how many orders a month at most" again — switch to historical actuals instead: `first-year contract value × gross margin × expected retention years × this work unit's actual number of closed deals in the last 12 months`, with no "cap" of any kind appearing in the numerator; a new work unit with no history takes the actuals of the closest matching work unit at the same business, logged in the ledger as "proxy". **Two bans**: never mix "cap" and "actuals" into the same formula; and never assign a default value just because one factor can't be found — if you can't find it, mark the whole thing "cannot be calculated" and leave it out of the ranking. **Annual opportunity value is only for ranking — not one word of it may be called a revenue forecast, and it must never be multiplied or divided by seat counts, or shown side by side with them**; changing the numerator must not change the ranking rule (see （→ 通用版 4.5 买家归堆、年机会值与三道关）).

The work unit varies by industry: one type of customer / one product or course line / one use case × one role / one category cluster. The ledger, landing pages and acceptance are all counted by number of work units, never by number of pages.

### Axis 8 · Attribution channels

Check three things in the back office: is there a source field? Do sessions coming from AI carry a referrer? Can you insert a required self-reported-source field? Values: form / phone / WhatsApp / in-store visit / free-trial sign-up / add-to-cart order / app install.

```split id=metrics-vs-channel Figure: only three rulers count towards acceptance; attribution channels are for reference only
Counts towards acceptance (the pass test) || Reference column only, never the pass test
Seat count, only counts as a rise once it clears the noise band || Attribution sources such as form / phone / WhatsApp / in-store visit
On the list and cited X/20, denominator fixed for the whole quarter || Clicks coming from AI naturally lack a referrer; any measure of this will undercount
Factual errors (count), read by hand from the brand six questions || Using it for acceptance digs your own hole
```

How to capture that reference-column number depends on the channel:

| Channel | How to capture it |
|---|---|
| Form | Add a required single-choice field, "How did you find us?" (including "AI assistant recommended it") |
| Phone | A dedicated extension number for the landing page |
| WhatsApp | A link with pre-filled text, the text carrying the page's code |
| In-store visit | A fixed front-desk question, logged in the daily report |
| Free-trial sign-up | A required self-reported source field + the distribution of sign-up email domains |
| Add-to-cart order | A discount code unique to the landing page + the brand-term curve + the product page's direct-visit curve; the platform side has no referrer, so write clearly that it's reference, not attribution |
| App install | The app store's back-office search-term report |

### Axis 9 · Content half-life

Ask: "When did this price list / feature table / course schedule last change, and how many times a year does it change?" Look at historical versions — don't take an estimate. You can set the level field by field.

```steps id=half-life-4 Figure: four half-life levels; the recheck rhythm follows the field
Very short (within 30 days) | Don't write actual figures | Write the structure, the range basis, the price variables and the last check date
Short (quarterly) | Quarterly recheck | Recheck every field when it comes due
Medium (half-yearly) | Half-yearly recheck | Or recheck earlier if an event triggers it
Long (a year or more) | Event-driven | Change it only when the regulation or the practice changes
```

The reason the very-short level doesn't write actual figures: if AI copies an out-of-date number, what it produces is one factual error, and factual errors (count) is one of your own rulers — that's marking yourself down. A decaying page both falls out of the candidate set and pours wrong words into the "get identified correctly" path; conversely, a visible time anchor gives whoever is fetching the page the confidence to copy it (78% of citations of videos with timestamps are repeat citations, and the same mechanism holds for pages).

**One rule shared by every level**: the update date must be a line of text in the visible HTML. None of the five major systems' live fetches read JSON-LD — a `dateModified` written into schema doesn't exist, as far as live fetch is concerned. For industries that need field-by-field levels, the specific field list is left to the industry edition.

### Axis 10 · Language fork

Run the same batch of intents once in the local language, compare the named sets, and check whether the businesses in the difference set all have a local-language page.

```mermaid id=lang-fork Figure: whether to build a local-language page
flowchart LR
  q1{"Is the share of local-language questions high?"} -->|low| no1["Don't build"]
  q1 -->|high| q2{"Is there a difference set in the named sets?"}
  q2 -->|yes, and those all have local-language pages| yes["Build it, alongside the English page"]:::hl
  q2 -->|no, the English page is already capturing it| no2["Don't build"]
```

**The language difference set is the cleanest causal evidence in the whole book**: the difference set is itself the evidence — no inference needed. The businesses named in local-language answers are exactly the businesses that have a local-language page. When a product is bound by territorial regulation (pay, tax, invoicing, data residency, industry licensing), a variant appears: the language doesn't fork, but a separate page is built for each jurisdiction, carrying that jurisdiction's regulatory terms in full — this is a "jurisdiction fork", another expression of the same mechanism as a language fork.

## 8.5 The decision table and rulings

**What you'll do in this section**: use the four dimensions — control of citation containers × entity anchor × side and switches × question-shape mix — to find the closest row in the decision table; use the four-level ruling to handle conflicts where "more than one row fits"; and when nothing fits any row, use the off-table fallback and add a row marked "first case, not measured".

Once all ten axes are filled in, only four dimensions actually decide the main battlefield; the other six axes govern the details: **control of citation containers (axis 2) × entity anchor (axis 3) × side and switches (axis 4) × question-shape mix (axis 5)**.

| # | Combination | Main battlefield | First actions | Notes |
|---|---|---|---|---|
| Example | Many platform slots + own-site slots high + product anchor + unregulated side + comparison type very high | Own-site pages (comparison / alternatives / category guide) + urgent pricing-page fixes + edit text in platform-slot profiles | The five door gates (robots per subdomain + docs site and trust page CSR) → product-level `/facts` + namesake-conflict check → complete the pricing page + strip `nosnippet` → mirror the trust page's visible HTML on the main domain → `X vs Y` / `X alternatives` / `best X for <scenario>` → docs site SSR → integration tutorial pages | For the values of the other industries already covered, see （→ 通用版 C.1 五个已出行业的十轴取值与第一周三件） |
| Abstain gate | J1 does not hold (fewer than 4 reachable-enough URLs in the intent-level top 10; neither D nor "to ask" counts) | **Decide whether to do it at all first** — this takes priority over every row | Count the number of reachable-enough URLs; fewer than 4 → do not do it, and you don't need to look at the methodology row or the red-ocean row | Fewer than 2, with work already under way → wrong target, switch the work unit |
| Methodology row | Approach type ≥50% (the company is only mentioned in passing) | Enter the methodology corpus | The five door gates → make the docs site and long-form articles crawlable → bylined contributions (the byline is a real person) → long community answers and open-source traces; **do not build a list-style landing page** (that is the recommendation type's tactic — it doesn't enter the candidate set here) | Not measured |
| Red-ocean row | Unregulated side + competitors' tactics all in use | Win on structure | On top of the row you've selected, add three things: fill out a full matrix of comparison pages, fill review and comparison-platform profiles to full completeness, and keep visible-update-date discipline (high-frequency updates are themselves the difference from competitors) | Not measured; this is an add-on, not a stand-alone row |

**How to use it**: for a new industry, first decide the side (8.2), fill in the ten axes, then find the closest row in the table.

```mermaid id=verdict4 Figure: The four-level ruling, working down level by level; stop once a level matches; the abstain gate is always judged first
flowchart LR
  L1{"Fewer than 4 reachable-enough URLs?"} -->|Yes| stop["Do not do it, skip the rest"]:::warn
  L1 -->|No| L2{"Two rows fit, with different side/switches?"}
  L2 -->|Yes| pick2["Go with the stricter row"]
  L2 -->|No| L3{"Two rows fit, with different battlefields?"}
  L3 -->|Yes| L3b{"Is the axis-5 share gap under 10pp?"}
  L3b -->|Yes| pick3["Go with the methodology row"]:::hl
  L3b -->|No| pick4["Take the row with the higher measured share"]
  L3 -->|No| pick1["Just use the row that matches"]
  pick1 --> L4{"Does it also hit the red-ocean row?"}
  pick2 --> L4
  pick3 --> L4
  pick4 --> L4
  L4 -->|Yes| addon["Add the three extra things"]
  L4 -->|No| final["This is the row"]
```

Two reasons the figure cannot show: at level 3, "a share gap under 10pp → go with the methodology row" is because the costs are asymmetric — using the approach-type tactic on a recommendation-type category is merely slow, while using the recommendation-type tactic on an approach-type category means the whole page never enters the candidate set; level 4 is not a stand-alone row, it is an add-on — a hit simply adds three more things on top of the row you already selected.

Not fitting into any row is normal: this table only lists out the five verticals that have been run (only one row is kept here as an example) plus three general-purpose rows — the table isn't broken.

```steps id=fallback-3 Figure: The order to fall back on when nothing fits any row
① | Use axis 2 to set the main battlefield | own-site slots high → build your own; many platform slots → edit text first; many third-party slots → contribute; many not reachable → back to the abstain gate
② | Cross out tactics | Use axis 4 / the switches; go to 8.6 and strike out, tactic by tactic, the ones you cannot use
③ | Set the first page's page type | Use the axis-5 value to decide what page type the first page is written as
Add to table | Write it down | Mark "first case, not measured", and add it to the decision table
```

A known common combination outside the table (not measured): third-party slot (B) reachable + organisation anchor + unregulated side + approach type or spec type high → the main battlefield is bylined contributions + industry directory and register listings + turning case studies into checkable-number pages, **do not build a list-style landing page**, and the entity anchor for `/facts` has to be built yourself (see axis 3 in 8.3).

> **Example** (dental and aesthetics vs K12 tutoring): on the same axis (shelf page-type mix), opposite values send the tactics in opposite directions.
> ```split id=axis-flip Figure: Same axis, opposite values, opposite tactics (example)
> Dental and aesthetics || K12 tutoring
> listicle 0% of citations || Actual search top 9: 5 editorial lists + 3 platform directories
> Best/Top pages: removal requests only || Can compete head-on for an editorial slot
> Not worth the spend, both because of the 0% and because the regulations ban self-praise and buying list spots || No professional advertising regulations, and lists are the dominant page type
> ```
> But on one point both sides reach the same conclusion: go for editorial slots only, never buy a paid slot. The reason isn't compliance — buying list spots is legal in K12 tutoring — **the reason is that paid placements and advertorials account for only 0.3% of AI citations: this is an efficiency judgement, not a compliance judgement.** "Don't buy a slot on a page that isn't already cited" governs whether you're wasting money; "if you buy it, disclose it" governs a separate line — the two criteria must not be mixed.

## 8.6 Seven tactics: how regulated and unregulated sides differ, and why each moves rankings

**What you'll do in this section**: go item by item through the seven tactics (reviews, lists, prices, testimonials and results, person entities, content boundaries, agency liability), checking whether each can be used on each of the three sides and why it moves rankings. The detail of exactly what you can write on each side is not spelled out here — see （→ 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）; the four figures below only show the strict side's "what gets banned → what fills the same slot".

### 1. Reviews

In recommendation-type questions, AI copies review text straight into the answer body (70% confidence); the rating and review count are also the sort field for the vast majority of listicle / comparison / directory pages, and listicle pages account for 19.6% of citations.

```split id=bullet-review Figure: The reviews tactic: what is banned on the strict side is replaced with checkable-fact sentences
Cannot do on the strict side || Replace with this
Actively request reviews || Passive observation only; never a task metric or an acceptance measure
Reproduce review text, screenshots, star ratings or aggregate scores || Checkable-fact sentences: licence number, product model, process duration
Confirm in a public reply that this person is a customer || Never confirm anyone's identity
```

The light side may request reviews (conditions in 5.3); the unregulated side can and must, but there the gatekeeper is the platform's terms instead.

🔴 A technical precondition that applies on all three sides and is the easiest one to miss: **AI reads visible HTML; none of the five major systems reads JSON-LD when fetching live — burying only an aggregate-rating structured-data block counts, to AI, as not having done it at all**. A review has to appear on the page as visible text; this is the mistake the unregulated side makes most often. The unregulated side isn't free to request reviews however it likes, either — review platforms have explicit incentive rules: any incentive must be token and applied equally to every invited customer; you must not invite only satisfied customers, and you must not reward only positive reviews; breaking these gets reviews purged and vendor pages actioned. The review gatekeeper has to be measured industry by industry — never inferred straight from the side.

### 2. Lists and directories

Listicle pages account for 19.6% of citations, and comparison pages get 1.87 citations per search (highest in the whole table) — the two richest shelves. Getting onto a list AI already cites and ranking first on it raises visibility and moves your position earlier in the answer; for the figures and their qualifying sentence see （→ 通用版 D.1 数字纪律：数字怎么标、什么只能自己看）; for the sample and source see （→ 通用版 A.1 机制、门与人这一层的证据）.

```split id=bullet-listicle Figure: The lists tactic: what is banned on the strict side, and what replaces it
Cannot do on the strict side || Replace with this
Pay to get on a Best/Top page || Pure directory / price-comparison pages, with no self-praise in the title or entries
Build your own list and rank yourself on it || A "how to choose" criteria-based selection guide: list the criteria, don't rank
Claim to be number one || Send a correction letter for wrong entries in pure directories
```

On the strict side, even labelling it "paid listing" does not release you from liability.

🔴 **One rule that cuts the other way (the most expensive mistake on the unregulated side)**: 84% of AI citations come from earned media; paid placements and advertorials account for only 0.3% (see （→ 通用版 1.1 买家怎么问，AI 读什么）). Buying a list spot only works if that page is already cited by AI — buying a spot on a new, uncited list drops you straight into that 0.3% bucket. Before any paid slot, use the *Citation-slot table* to confirm the page is already cited within the question pool; if it isn't cited, don't buy it. Whether a list is worth the investment is an industry variable — every vertical has to measure its own page-type distribution and must not extrapolate the previous industry's conclusion.

### 3. How prices are written

Price-type queries trigger AI Overviews >80% of the time, and 97% for mixed intent — the highest-hit-rate question type there is. **A fixed price actually comes out ahead on the "extraction" front**: a precise, adjective-free, checkable number is easier to extract into an answer sentence than a "from X" price — a "from" price carries no information, so even if AI copies it, it still can't answer the user's question. The strict side is forced to write a fixed price, so it loses nothing on extraction; what it loses is on "placement" (it can't get onto a price-comparison list).

```split id=bullet-price Figure: The price tactic: what is banned on the strict side, and what replaces it
Cannot do on the strict side || Replace with this
"From", "as low as", original price, discounts, instalments || A tiered, itemised fixed final price
Bundles, time-limited offers, "free" || Write "included in the quoted price" into the tiered price
Comparing prices with peers || See 5.2 for replacement wording
```

Prices on the unregulated side aren't a free-for-all either: mandatory fees must be included in the displayed price, a struck-through price needs a real baseline, "free" must come with no conditions attached, and fields with a very short half-life don't get written as hard numbers (8.4, axis 9). "Unregulated = nobody polices the price copy" is the most easily misread cell of all, and since the price page is the highest-hit-rate cell, misreading it costs the most. The real trap is here: drip pricing and fake struck-through prices are already enforcement targets — once a platform or regulator names you, what you lose is the credibility of the whole domain, not just one piece of copy.

### 4. Testimonials, case studies, result and efficacy claims

Short sentences with specific numbers are the material type AI extracts into answers most easily; article pages account for 23.7% of citations, and their citability rests almost entirely on the handful of checkable numbers inside them.

```split id=bullet-proof Figure: Testimonials and results: what is banned on the strict side is replaced with evidence-backed operating figures
Cannot do on the strict side || Replace with this
Customer testimonials || Process steps + duration + number of sessions
Before-and-after photos, even with a disclaimer || Brand and model of the material or product + scope of use
Success rate, efficacy or result claims || Clearly defined, adjective-free, evidence-backed operating figures
```

The strict side isn't out of cards to play — the cards just changed: the figures on the right are numbers too, just as extractable, and just as easy to back up with evidence — the regulations already require every statement to be provable, so a number you can prove is actually the safe zone. The unregulated side may make result claims, but must give the basis and sample, or the claim gets judged unverifiable and is, if anything, less likely to be copied; any "save X% / Y times faster" claim needs a clickable methodology page (sample, conditions, date) — delete any claim with no methodology page.

🔴 **A conclusion shared across every side**: a number with a denominator, a year and a stated basis is both what the strict side's rules require and the form AI is most willing to copy — so **the line "we had to weaken the content for compliance" is not allowed**; what gets cut is always the self-praising sentence that was never going to be copied anyway.

### 5. Person entities

An official registration number plus a link to the official register lookup page is the strongest, cheapest disambiguation anchor in the whole playbook — it's a single-point confirmation of you from a third-party authoritative page; the regulated sides get it for free, the unregulated side has to build one itself (see axis 3 in 8.3). This net advantage only shows the strict side's full hand of cards when read together with two others: a fixed price comes out ahead on extraction (item 3), and evidence-backed numbers are the safe zone (item 4) — grabbing only the registration-number point is playing with two fewer cards than you have.

### 6. Content form: where the limits of long educational articles lie

Comparison pages have the highest density (1.87), article pages come second (23.7%). The strict side may write long educational articles, keeping to four boundaries: ① must not lead into efficacy / results / success rate; ② must not contain anything on the ban list; ③ must be signed by the organisation or the licence holder; ④ must not recommend your own business over peers within the article.

The regulated sides are shut out of "naming and comparing competitors", and the legal way out is the "criteria-based selection guide" — list the criteria, don't rank, don't name peers. In page-type terms it still counts as listicle / article, so it still gets a share of that citation pool, without triggering either the self-praise ban or the peer-comparison ban; this is currently the only legal route for the regulated sides into that 19.6% listicle share (not measured, 65% confidence — recommend running one per vertical as a control).

> **Example** (financial advisory): if the content amounts to a product recommendation for an individual's specific situation, it may itself require a licence — adding a disclaimer does not remove the licensing requirement.

### 7. Legal liability for advertising published on someone else's behalf

This item doesn't move rankings directly, but it decides whether the tactic can be carried out at all. A tactic that can't be executed doesn't exist for this industry, so it has to stay in the playbook table — it must not be moved off into a legal appendix.

```mermaid id=distrib-4 Figure: Four forms of publishing-on-someone's-behalf liability, each with one hard consequence, deciding whether off-site work can be handed to someone else
flowchart LR
  f1["Licensee-authorised person, joint liability"] -->|publishing party = the authorised person| r1["Cannot publish without a letter of appointment"]:::warn
  f2["Principal bears it alone"] -->|treated as the principal's own publication| r2["Each outreach letter held to principal's bar"]
  f3["Platform-terms joint liability"] -->|violation by the agency running it| r3["Store closed, reviews removed, ranking demoted"]:::warn
  f4["No dedicated rule"] -->|no authorisation document required| r4["Wording-level constraints still apply"]
```

There's no downgrade path out of the "licensee-authorised person, joint liability" tier: the moment the party publishing on the client's behalf publishes, it becomes the authorised person and shares liability with the licensee, whose obligations cannot be transferred away; without a *Letter of appointment for advertising* ready to sign, nothing may be published in any capacity — no signature means a hard refusal. When the rule writes the publishing party directly into the offence as a liable party, and the business has even one such business line under its name, the whole site is run on the side of its strictest business line (switch D (spans two sides)), and the authorisation and sign-off documents come before work starts (switch A (agency liability)); if switch B (legally required fields) is on, any material published on the client's behalf that misses a legally required field is itself non-compliant material. "Principal bears it alone" also means the principal must not get a third party to publish, on its behalf, promotion that would be a violation if the principal published it itself. On the side with "no dedicated rule", the business bears primary liability, but the party publishing on its behalf may be pursued separately if it makes a misleading statement on its own initiative (inferred, 60% confidence; no direct case found). The real hard constraint comes from platform terms, which are enforced far faster than the law. If you cannot establish which form applies, judge it "cannot publish on the client's behalf" under 8.1 Q6b.

## 8.7 Cheap tactics on the unregulated side, and substitutes on the regulated sides

**What you'll do in this section**: match the eight cheap tactics on the unregulated side against their legal substitutes on the regulated sides; use the five substitutes to replace the tactics the strict side loses; and for UGC calls with a reward, once the reward is removed there are still three usable approaches. Fill in the specific form each one takes in your own industry.

```split id=cheap-tactics Figure: The eight cheap tactics on the unregulated side, and the legal substitutes in the same slot on the regulated sides
Unregulated side || Legal substitute on the regulated sides
Actively request reviews at scale || Strict side: checkable-fact sentences take the extraction slot
Buy paid Best/Top list spots || Pure directory + price-comparison pages, plus a correction letter for wrong entries in pure directories; if listed on a Best / Top page, request removal
Build your own industry list and put yourself on it || A "how to choose" criteria-based selection guide (not measured, 65%)
Comparison pages naming competitors || Compare approaches or methods only, never companies
Before-and-after / success-rate / client-results numbers || Clearly defined, adjective-free operating figures + process and duration
Discounts, original price, time-limited offers || A tiered, itemised fixed final price
Affiliate marketing / ghost-written distribution || Bylined contributions: signed, unpaid, giving checkable facts
Rewarded referrals / UGC calls || Three approaches once the reward is removed (see below)
```

Item 1 doesn't need a full substitute on the light side: you may request reviews, but without offering any benefit, without filtering, and without ghost-writing them. Item 6, on the side where switch B (legally required fields) is on, becomes a complete price list with every legally required field filled in. Affiliate marketing and ghost-written distribution move a lot of volume but at a very low citation rate (they land in that same 0.3% bucket, see （→ 通用版 1.1 买家怎么问，AI 读什么）); rate it low-efficiency and don't make it the main push (inferred, 70% confidence).

When the unregulated side uses these eight tactics itself, there are four traps: a self-built list that lists only you reads as obvious self-assessment, and both AI and editors will demote it — it only has citation value if it genuinely lists every peer; competitor information goes stale, and a stale comparison page bites back at the identity gate; results figures need a stated basis and date, or AI will keep repeating an old number as if it were the current state; and the constraints on discounts and struck-through prices are in 8.6, item 3.

### Five substitutes for the tactics the strict side loses

These five are the strict side's entire outlet for persuasive power; every industry has to fill in the specific form each one takes.

```split id=five-substitutes Figure: The tactics the strict side loses, and the five substitutes that replace them
Lost || Replaced with
Reviews + result claims || Third-party credential tier (on the strict side only technical-standard certification and registration status remain)
Self-described credentials || Official registration number + official register lookup link
Price comparison + promotions || A tiered itemised fixed-price page
Efficacy || Being listed on a manufacturer locator (they link to you — inbound)
Testimonials || A signed long-form video and bylined contributions
```

The wording and the reasons the figure cannot show:

- **Third-party credential tier**: write only the tier's name and who grants it; never "highest tier / the only one / only X on the whole island"; never "most cases". The basis: in an actual search, the one recommendation-type answer that didn't copy a review copied the manufacturer's official tier plus its official registration status instead — when a verifiable third-party tier is available to copy, AI copies the tier, not the review (sample and source in （→ 通用版 A.1 机制、门与人这一层的证据）). The strict side does not write manufacturer-granted tiers (Conservative line (not statute text); see the industry edition); "AI copies the tier" is an observation from a single actual search and has not been retested locally.
- **Official registration number**: this is also the hardest anchor on the identity path — a registration number is a unique string across the whole web, so it has the lowest disambiguation cost.
- **Fixed-price page**: captures the >80% trigger rate on price-type queries, and a self-built page gets the same density as the comparison page type.
- **Manufacturer locator**: the strict side's own page writes only the real model number and registration facts of the device, material or system, with no outbound link to the manufacturer; the payoff is on the inbound side — one application in exchange for one third-party inbound link and one brand endorsement.
- **Signed long-form video and bylined contributions**: you don't need a viral hit, just a reference that can be cited, see （→ 通用版 6.7 本人口述长视频）.

### UGC calls with a reward, once the reward is removed

What the regulated sides ban is "giving something for content", not "obtaining third-party material" — the mechanism behind this tactic (creating citable Q&A material on a third-party domain) mostly survives once the reward is removed.

```split id=ugc-no-incentive Figure: UGC calls with a reward: three approaches remain once the reward is removed, in order of effectiveness
Once the reward is removed || The specific approach
8a No-reward call for questions || Open calls via contributions, association explainers and community long-answer columns
8b Swap the reward for a byline and exposure || Speaker pages, conference Q&A transcripts, association round-table transcripts
8c Use operating figures to carry the sense of volume || Operating figures with a stated basis, denominator and year
```

8a has the licence holder sign and answer, with only checkable facts in the answer — the output is a qa page on a third-party domain, competing for the same shelf slot as UGC. 8b gives the person on camera a byline, not a benefit; the regulations govern benefits, not bylines. 8c replaces the role UGC plays in an answer — the sense that "many people have said so".

🔴 **Red line**: 8a and 8b must never involve any form of reward — including prize draws, points, discounts, priority scheduling or free gifts. The moment any reward is involved, it reverts to tactic 8 itself and the whole approach is off-limits. Platform terms often ban rewarded calls outright too, and review sites' incentive rules are almost as strict as the regulated sides' — read the target platform's policy clause by clause before starting.

## 8.8 Unregulated ≠ easy: admission criteria J0–J7

**What you'll do in this section**: on the unregulated side, competitors can use every cheap tactic too, so first use the eight testable admission criteria J0–J7 to judge whether it's worth doing at all — if J0 does not hold, don't do it; J1 must hold; at least two of J2 / J3 / J4 must hold; and if competitors have already maxed out the cheap tactics and there is no gap at either the door or the entity, don't do it.

**The intuition is "unregulated = easy to win"; the reality is the opposite**: competitors have all six tactics open too, and in a category where every tactic is fully open, recommendation-type answers default to copying whatever social proof already exists — content quality alone can rarely jump the queue, and you have to win on structure + scale + update frequency instead. Compliance substitutes can't open up a gap on this side, so differentiation comes down to just three things: a cleanly fixed door, a well-aligned entity, and thoroughly done off-site work.

The method has three layers. **L-a compliance substitutes** (getting persuasive power back under a ban) are worth the most on the regulated sides; on the unregulated side they don't drop to zero, there's just no single set of regulations you can look up — the constraints come from the four base layers in 8.2 step 5, and each layer happens to press exactly on the richest shelf slot: price transparency presses on the price-and-fees page, the ban on misleading statements presses on the results-figures page and review display, the Do Not Call Registry presses on outreach pacing, and platform terms press on review collection and filling in platform-slot (C) profiles. So on the unregulated side, L-a is "scattered everywhere — you have to collect it yourself": make a two-page sheet with links to the original text. **L-b technical door** (robots / WAF / nosnippet / CSR / multiple subdomains / platform doors) is often worth more on the unregulated side — these industries run newer tech stacks, with a higher share of SPA and headless setups. **L-c ruler and target picking** (frozen question pool, noise band, four-state list, ranking by annual opportunity value) is worth the same on both sides. **The criterion isn't "regulated or not" — it's how big the gap is in L-b and L-c.**

```mermaid id=j0-j7 Figure: J0–J7 decide whether it's worth doing; J1 is the book's only abstain line
flowchart LR
  j0{"J0: will buyers really ask AI?"} -->|Does not hold| stop0["Do not do it"]:::warn
  j0 -->|Holds| j1{"J1: ≥4 reachable-enough URLs?"}
  j1 -->|Does not hold| stop1["Do not do it"]:::warn
  j1 -->|Holds| j234{"At least 2 of J2/J3/J4 hold?"}
  j234 -->|No| stop2["Do not do it"]:::warn
  j234 -->|Yes| j7{"J7: have competitors maxed out?"}
  j7 -->|Yes, and neither J2 nor J3 holds| stop3["Do not do it"]:::warn
  j7 -->|No, or one of J2/J3 holds| go["Do it"]:::hl
```

All eight criteria are testable — never score by impression; the figure only shows the ones that decide whether to do it:

| Criterion | What it means | How to test it |
|---|---|---|
| J0 | Whether the AI entry point holds: buyers really do ask AI (holds) / AI is only used for an initial screen (half-holds) / it runs through a panel, pitch or personal network (doesn't hold) | Run 15 frozen questions through both engines; count whether any company at all gets named |
| J1 | Reachable-enough count ≥4 (fewer than 4 reachable enough in the top 10 for a given question → don't do it) | Tag the four states for every URL in the intent-level top 10 |
| J2 | The door is generally closed, and competitors haven't opened theirs either | Turn off JS and view source + the crawler hit table, sampling 5 competitors |
| J3 | Entity disambiguation is hard (namesakes, abbreviations, multiple product lines, no official registration number) | Ask each of the five major systems "what company is X", count the misidentifications |
| J4 | The key citation slots are gatekept by editors or an independent third party, and cannot be bought | Check those pages' submission-policy pages: is there a rate card, is anything marked sponsored |
| J5 | The buyer shape still fits "buys infrequently, can't judge quality" | Admission questionnaire |
| J6 | Annual opportunity value is high | Verbal + to be verified |
| J7 | Whether competitors have maxed out the cheap tactics | Sample 5 competitors, count: are they actively collecting reviews, have they bought list spots, do they have vs pages |

**Why J2 / J3 / J4 specifically**: they correspond, respectively, to the parts of the door, identity and shelf-placement gates that competitors cannot buy with money. Reviews can be bought, list spots can be bought, advertorials can be bought — but a door that won't open, an entity that can't be identified, and an editorial slot that can't be bought can only be earned by doing the work. The rankings you can actually win sit in those three cells.

> **Example** (weddings, not one of the industries covered so far): the buyer shape fits perfectly (a once-in-a-lifetime purchase, high price per order, buyers can't judge quality), and it still gets ruled dead — not because the target was wrong, but because this category has no regulatory red line, and a review-management agency can get most of the same effect for less money: that is, J7 hits while J2 and J3 do not hold.

🔴 **The most common mistake**: assuming that unregulated or lightly regulated means you can skip the door and identity layers and go straight to writing content — **regulation only takes tactics away, it does not change the building site** — skip them once, and everything after it gets multiplied by 0.

```split id=side-mirror Figure: The two sides' tactics have opposite centres of gravity, but the layer most often missed is the same one
Strict side || Unregulated side
Main effort: finding legal substitutes || Main effort: volume and structure
Technical-standard certification, registration number, fixed price, listing on manufacturer locators || Comparison-page matrix, alternatives pages, review-platform profile completeness
Main off-site push: government and association registers + bylined contributions || Main off-site push: review and comparison-platform vendor profiles + long community answers + integration marketplaces
Fastest payoff: the five business profiles + checking the official register || Fastest payoff: completing platform-slot (C) profiles + comparison pages
Most often missed: the door (all the attention goes to compliance) || Most often missed: the door ("nobody's watching anyway" so it never gets checked)
```

Government and association registers are free, lose almost nothing between effort and listing, and their share is rising; platform-slot (C) profiles and comparison pages pay off fastest because the container is already inside the cited set, so a text edit takes effect immediately. What both sides miss is the same layer — the door.

Three rules that hold across every side: ① the door and identity layers are exactly the same on all three sides — the robots.txt master template, the three `nosnippet` controls, matching all four identities, the `/facts` page, six-trace alignment word for word — not one item may be dropped; ② the compliance verdict outranks priority ranking: however rich an off-site slot is, if the compliance verdict is "correction only / request removal / do not do", you only send a correction letter or a removal request, you don't schedule it or give it priority; ③ **the side where the only constraint comes from consumer-protection law and platform policy is the easiest one to crash on**, precisely because there's no single set of regulations to check — for these categories, before starting work you must make a separate two-page sheet on "what this platform does not allow for collecting reviews / setting prices", with links to the original text.

## 8.9 Cross-border: whose law applies

**What you'll do in this section**: first answer three questions (where the service is delivered, where the offer and acceptance are made, and whether the page targets a given country); a country counts as targeted only when at least 2 of the six targeting signals are ticked. Then apply the five rules in order to take the intersection and write one version to the strictest standard — don't split domains or brands by country.

### Three questions to decide

Answer three questions before asking which law applies: ① where is the service / goods delivered (check the place of contract performance, the distribution of shipping addresses, the store or delivery location — use back-end data where you have it, don't go on what people say); ② where is the offer made and where is it accepted (check the merchant entity named on the checkout page, where the party receiving payment is registered, the jurisdiction clause in the terms); ③ does this page target a given country — this is the targeting test countries actually use for territorial jurisdiction.

For question three, tick each of six signals: ① a dedicated page in that country's language, or hreflang pointing to that country; ② prices quoted in that country's currency, or that country's local payment methods offered; ③ the page states delivery to that country / available in that country; ④ it uses that country's ccTLD or a subdirectory for that country; ⑤ it is listed or advertised on that country's platforms, app stores or ad networks; ⑥ the page names that country's local regulatory terms, local credentials or local customer-service channels.

```mermaid id=target-signals Figure: Six targeting signals; a country counts as targeted, and enters the intersection, only when ≥2 are ticked
flowchart LR
  c{"How many targeting signals are ticked?"} -->|≥2| yes["Targets this country, enters the R5 intersection"]:::hl
  c -->|≤1| no["Monitor only, does not enter the intersection"]
```

With ≥2 ticked, treat the page as "targeting that country": that country's advertising and consumer law applies. The test has to be reproducible: all six signals are facts visible on the page, so two people opening the same URL should tick the same ones; if they disagree, go with whichever side ticked more.

> ⚠️ **A common misunderstanding**: "where the ad can be seen" is not a criterion. A web page is naturally visible worldwide, so this question gives the same answer for every page — it tells you nothing — and treating it as a criterion pushes "take the intersection and write to the strictest standard" into "write to the strictest standard worldwide", which is a standard nobody can actually follow. The only thing you can judge is the six targeting signals in question three above.

### Five rules, applied in order

```steps id=cross-border-5 Figure: The five cross-border rules, applied in order
R1 | Territoriality doesn't loosen | If the licensee is local and the service is delivered locally, the full set of bans applies exactly as before
R2 | The base layers still apply | Either the supplier or the consumer is local, or the offer and acceptance are made locally
R3 | Two routes can reach you | ① platform delisting, store closure, review removal ② payment and logistics — they don't need to wait for a regulator to come knocking
R4 | Stack the stricter one | If the destination country has its own banned-word list, stack it with the local one and write to whichever is stricter
R5 | Fix the intersection | Don't write versions with different degrees of leniency for different countries; split by language using hreflang
```

R1 is inferred, 85% confidence: a page aimed at foreign buyers is still bound by the full set of bans; you may not wave through any item in the local ban list on the grounds that "this page is for foreigners". The reason: the regulations govern the licensee's advertising conduct and the licensable services delivered locally, no provision has been found that exempts by audience location, and under the regulator's own position, all of the internet (social media and websites included) counts as a regulated medium. R2 has a source (the territorial provision). R3 is inferred, 75% confidence: platforms enforce faster than the other country's regulator. R4's destination-country statute has a source; "how it actually applies to an overseas operator" is inferred, 70% confidence.

> **Example** (a Chinese-language page aimed at consumers in China): China's *Advertising Law* bans absolute terms (most / first / only / national-grade / highest-grade / best), with penalties of RMB 200,000–1,000,000; there is also the SAMR *Enforcement Guidelines on Absolute Terms in Advertising*, and, as paraphrased second-hand, not every phrase containing "最" (most) or "第一" (first) counts as a violation — wording that points at the operator itself rather than the product, that expresses an order in time or space, or that is limited to a specific scope the business can substantiate itself, may leave room for not being treated as a violation (Original text not obtained; not used as a basis for leniency). For overseas operators targeting consumers in China, enforcement in practice runs through platforms (inferred, Conservative line (not statute text)).

```split id=banlist-overlap Figure: The two banned-word lists mostly overlap; writing to the stricter one satisfies both, but not the other way round
The strict side's self-praise ban list || The destination country's absolute-terms list
Best, optimal, first, only, leading || Mostly the same set of words
Top-tier, authoritative, national-grade, highest-grade || Has exceptions (not treated as violations)
No exception clause || Its exceptions never clear a word the strict side bans
Write to this list and both sides are satisfied at once || Banned outright on the strict side does not mean penalised outright here
```

So when the two lists mostly overlap, writing the destination-country language page to the strict side's list satisfies both sides at once; the local-language page tactic (axis 10 in 8.4) can be carried straight over into both the regulated sides and cross-border cases, with no need to write a separate set of copy rules (inferred, 80% confidence).

### R5 explained: why you must not split the entity by country

Splitting domains / splitting brands shatters the evidence for the identity gate. AI confirms an entity from consistent signals across sources (same name, same address, same registration number, same account) — split into three brands and none of the three is strong enough, so none of them gets named. The correct approach: **one domain + hreflang + each language page linking to the others + identical entity information across the whole site.**

On the data side: forms, tracking scripts and remarketing pixels are configured to the strictest law among the buyers' countries — take the strictest, don't tier by country; the specific differences between each country's rules have not been measured, so this only gives the basis, not the detail.