Week one: check the website and save today's AI answers before you fix anything

2.1 Day one: look, don't touch, and ask the remaining nine pre-start questions

What you'll do in this section: on D1 you look but change nothing — get read-only access to the access logs, Search Console, Bing and business profiles for every domain in one go, and also request two materials that only apply to law firms (letter of engagement templates, and the written ethics guidance in the members' area of the Law Society of Singapore); then go through the nine of the "twelve pre-start questions" that chapter 0, 0.2 did not ask. When you're done, you'll have an access-request list organised by domain and the law firm's answers to the nine questions; the door, every profile and the facts page have not been touched at all.

Hard gate: until the baseline is saved, do not change the door, any profile or the facts page.

The reason to measure before you change is the same as in the General Edition: the page-type mix, the four states per URL, and the question shapes that the baseline needs to measure all have to be measured on the shelf as it stands, before the door has been fixed. Fix the door first and freeze the baseline afterwards, and what you freeze is a shelf you have already disturbed yourself — the before/after comparison in week 13 no longer holds. So on D1 you can do only three kinds of thing: request access, pull the logs (if none were kept, turn logging on that day), and record the current state. What the crawler hit table and the six-line door read look like is written in → General Edition 2.1 The five gates at a glance, and gate 0: access logs (read-only), → General Edition 2.2 Gate 1 read-only audit: what to check in nosnippet, robots and the WAF and → General Edition B.1 Door templates; not repeated here.

Law firms have one extra trap: a law firm often has more than one domain. Beyond the main site, there may also be microsites named after practice-area keywords (the "extra question" row in the nine-question table below). Do the D1 read-only audit by domain: one crawler hit table and one six-line door read for each domain. If you only check the main site, the baseline carries no record of whether the door on a microsite's side is open or closed. AI doesn't cite these microsites often, but it does cite them: in the L3 run (our 2026-09-29 measurement: 30 questions outside family law, covering 8 practice areas, plus 5 questions on choosing a lawyer and legal fees that every practice area gets asked; ChatGPT twice per question, AI Mode once), law-firm websites were cited 70 times in total across the three calls, 5 of them from microsites; in the public dataset's 22 family-law questions (2026-09-23, R1 across both engines plus R2 ChatGPT), law-firm websites were cited 37 times in total, 6 of them from microsites.

D1 access and materials to request in one go

The last two rows are materials unique to the Law Firms Edition; each gets one paragraph below.

Letter of engagement template: every fee row in the fact table must be traceable, item by item, to the same billing method, what's included and what's not, in some version of the letter of engagement template; a row that doesn't match stays off the page for now, and the version number also goes on the sign-off sheet (Appendix B.4). How to write fees is in chapter 1, 1.2; the fact table is in chapter 3, 3.1.

The written ethics guidance in the members' area: the Law Society's own material states, "From 2015 to date, the Advisory Committee has issued more than 100 written guidances to members." [Statute text; see Appendix A.5] This guidance and the two Ethics Digests are open only to members, and we did not obtain them this round [Original text not obtained; see Appendix A.6]; they are also the main source of the uncertainty left in this book's side decision, held at 75% confidence (chapter 0, 0.2). So exporting these documents is a start-of-work step, not an optional one: the firm exports them using its members' account and hands them to us to keep on file, and we read them before drafting the fees page or any page about reviews. If they set a stricter position than this book, that firm follows the stricter position; the book as a whole does not change.

The other nine questions: six are answered differently for law firms, plus one extra

Of the nine questions, 5, 6 and 10 follow the General Edition; the other six are answered differently for law firms, and there's one extra question about keyword microsites:

What the table can't show is the basis for a few of its rows.

Question 1: the name of a law corporation (LLC) must carry the suffix — the rule text reads "Every law corporation must have either the words “Law Corporation” or the acronym “LLC” as part of its name" (LPA s 154(3)) [Statute text; see Appendix A.2], and the fact table, business profiles and structured-data fields may never drop it. A Chinese firm name needs written approval — the rule text reads "The name of a law firm must be stated in the English language; but the name may, with the approval in writing of the Director of Legal Services, also be stated in any of the other official languages." (LPE Rules r 7; the rules for limited liability law partnerships (LLPs) and LLCs read the same, see r 15(1), r 29(1)) [Statute text; see Appendix A.2]; applied to Chinese pages, Chinese business profiles and Chinese directory listings, use only the approved Chinese name, and where there is none, use the English approved name — never translate it yourself [Conservative line (not statute text)]. The LSRA's field name for individual practising details was not verified this round [Original text not obtained; see Appendix A.6], so that row in the fact table reads only "LSRA register lookup link + check date", with no number written at all.

Question 9: lawyers may not pay "any commission or other form of consideration" for a referral (r 39(2)(b)), and may not share legal fees with, or pay commission to, a non-lawyer (r 19) [Statute text; see Appendix A.1]; the Law Society's Guidance Note further states that a law firm joining an Internet referral scheme that requires it to "pay a fee or share fees paid for legal services referred" breaches the Legal Profession Act (GN 6.1.1 para 35) [Statute text; see Appendix A.3]. So stop any platform that charges per enquiry, per lead, per completed engagement or by commission; for this category, not even a sign-off can release it: A written sign-off resolves only the stop-and-escalate situations; it never turns something banned as Statute text into something publishable. The kind that works as "membership fee + case assignment" — joining for a fixed fee, with the platform then referring cases for the firm — is stopped too: that a fixed entry fee also counts as "other form of consideration" is this book's reading [Conservative line (not statute text)] (80% confidence), and this cell is fixed book-wide as "do not pay" (chapter 6, 6.2). The only paid placement worth considering is a listing in the Law Society's own "Find A Featured Lawyer/Law Firm" directory (the Featured directory), which the Law Society itself sells [Conservative line (not statute text)] (70% confidence): listed by practice area, chosen by the reader, with no referral or case assignment. Its rate-card terms were not obtained this round [Original text not obtained; see Appendix A.6]; a firm already using it should get the terms first, and how to judge it is in chapter 6, 6.2.

This question also has to be asked of us: we charge law firms only a fixed service fee — never per lead, per case, per signed client, or as a percentage of legal fees. Charging per lead, per completed engagement or as a share of legal fees could amount to a criminal offence committed by us (LPA s 33(3)); on the firm's side, paying as a percentage of legal fees or a share per case breaches r 19, and paying per lead or per signed client falls under r 39(2)(b) [Statute text; see Appendix A.1, A.2]. How to write the contract is in chapter 6, 6.1.

Question 11: without Chinese-speaking clients, don't set Chinese questions and don't build Chinese pages; a Chinese page is an optional item, built only where there are Chinese-speaking clients, and accepted only against the AI Mode leg (reasons in 2.2). Indonesian questions are never set.

For the rest of the questions and how to read the answers, follow → General Edition 4.1 The full target-picking process and the twelve pre-start questions. If you can only do three things, use the General Edition's 48-hour version of shortcut 2 for the baseline instead — see → General Edition 4.8 Two shortcuts and this chapter's checklist; this 48-hour baseline must also be saved before you fix the door — everything else stays the same.

Once the baseline is saved, week one does only two things: fix the door for each domain, and build the fact table. The /facts page, each lawyer's own page, and LSRA alignment are scheduled for weeks 1–2 (chapter 3); the Law Society directory listing is done in the same period (chapter 6, 6.3); the fee guide page is written from week 2 (chapter 5, 5.1, 5.4). It's fine if the fees page doesn't fit into week one, but don't skip the baseline just to hit a deadline.

2.2 The 30 questions: 24 in English, 6 in Chinese, only for work the firm actually does

What you'll do in this section: set questions only for the one or two practice areas you are targeting this quarter, and fix the 30 questions by the firm's language ratio; fee questions, near-me questions, "is this firm good" questions and "who's best" questions must all be in the pool; questions that name a peer drop out of the pool and are used only for internal comparison. When you're done, you'll have a 30-question pool ready to sign in person; once signed, it doesn't change for the whole quarter.

How the 30 questions split by language is fixed; two people must never come up with two different denominators:

The rest of the ratios (location ≥60%, price ≥30%, bare words ≤10%) and how to do the second-person check and sign in person follow → General Edition 4.2 The question pool: where the 30 questions come from, how they are balanced, how they are signed; once signed, nothing changes for the whole quarter — changing a question is changing the ruler.

Only set questions for the one or two practice areas the firm actually handles and is targeting this quarter. The question pool is the ruler for the whole quarter: measuring work the firm doesn't do measures someone else's shelf; and once there's more than one practice area, the way AI finds sources differs sharply between them — asking ChatGPT the same way, the share of citations going to government and public-body pages is 5/11 for conveyancing and property, and 13/13 for debt recovery and small claims (L3 run 1, 2026-09-29; run 2: 5/12, 15/15). With one practice area rising and the other falling, combining them shows nothing at all, so the denominator is split by practice area (2.3).

The four question types that must be in the pool

Asking the same batch a second time lands in roughly the same place: price 6/26, "is this firm good" 2/10, "who's best" 2/14, near me 0/8, process/definition/rules/calculation 0/84, and 1 more citation landed on a checklist question (1/19, single case). Of the 22 family-law questions, ChatGPT cited a law-firm website for only 5, and all were price, near-me and "is this firm good" questions (public dataset R1, R2, 2026-09-23; entities labelled by hand, by domain). So if these four question types aren't in the pool, the ChatGPT leg barely measures the firm at all for the whole quarter. For "is this firm good", the only version that goes into the pool is "is <approved firm name> good" — questions naming a peer drop out (see the three filters below). ChatGPT rarely cites law-firm websites for "who's best": these 3 L3 questions, across both runs, drew 26 citations in total — 3 to law-firm websites, 17 to government and public-body pages (8 to the Ministry of Law, 5 to the Law Society), and 4 to ranking guides; the 1 family-law question had 4 of its 6 citations go to ranking guides. The fix for this sits off-site (chapter 6), but the sample is thin (1 family question, 3 L3 questions) — take it as a direction only.

Rules, process and definition questions can go in the pool, but keep this on the record: they feed AI Mode only. This is one place law firms differ from the General Edition — the General Edition doesn't choose educational questions, but for law firms, AI Mode still cites law firms on rules and process questions. The two legs read differently: across both ChatGPT runs, the share of citations going to government and public-body pages is 39/39 for rules, 85/92 for process, 17/18 for definition, 27/32 for checklist, and 12/14 for calculation; on AI Mode, the same 5 question types cite law-firm websites 28/76 (L3, 2026-09-29, AI Mode once per question); of the 47 law-firm-website citations across all 30 AI Mode questions, 26 are explanatory articles (heuristically labelled page type). When you read the numbers in chapter 7, only look at the AI Mode column for these questions.

The 6 Chinese questions run on both legs, but Chinese pages are accepted only against the AI Mode leg: for L3's 4 Chinese questions, ChatGPT cited a Chinese page 0/47 across both runs, AI Mode 5/22 (Chinese pages labelled heuristically, 2026-09-29); of the 28 citations on the family set's 3 Chinese questions, only 1 was a Chinese page, and AI Mode cited it (single case, 2026-09-23).

Three filters before a question enters the pool

The reason for each of the three filters:

• Drop what the firm doesn't handle: the reasoning is in the "only for work the firm actually does" paragraph above. A practice area the firm handles but isn't targeting this quarter also stays out for now.
• Drop questions that name a peer: for the two forms "which is better, Firm A or Firm B" and "is <a named peer> good", the firm has no page that can answer them without crossing the line — publicity may not make any comparison or criticism of other lawyers in relation to "the fees charged, or the quality of the services provided" (r 43(1)(c)) [Statute text; see Appendix A.1]. Answers that come back naming a peer are kept only as internal comparison material. If what the buyer is really comparing is a procedure ("what's the difference between a simplified and a contested divorce"), reword the question into a procedure comparison and then put it in the pool — a procedure-only comparison page (chapter 5, 5.7) is what answers it.
• Reword price questions: a question phrased as "how much does a lawyer cost for <matter type> in Singapore" goes straight into the pool; one that already presumes "cheapest" or "cheaper than other firms" can't be answered by the firm's own fees page — words like "cheapest" and "lowest fees" cannot be written (r 43(1)(c)) [Statute text; see Appendix A.1] — reword it to ask about billing method and amount, then put it in; one that can't be reworded is dropped and logged, for the same reason the General Edition's check 3 screens out "the cheapest X" (→ General Edition 4.5 Buyer types, annual opportunity value and the three checks).

Example (law firms) Three question phrasings used in the L3 run all pass the three filters (this only demonstrates the filter — to actually enter the pool, each one still needs evidence from source classes A–D, see General Edition 4.2): "How much does a probate lawyer cost in Singapore?" "Conveyancing lawyer near Jurong East MRT" "新加坡劳资纠纷请律师大概要多少钱？" (roughly how much does a lawyer cost for an employment dispute in Singapore?)

Two things are not decided at this pooling step: which page each question type ultimately lands on, and which page to build first (fees, near-me and "is this firm good" go first) — that's in chapter 4, 4.2; how to write Chinese pages and how to use the firm name is in chapter 5, 5.3.

2.3 Freeze today's answers, measure how much they vary, and ask AI for the first time whether it knows you

What you'll do in this section: run the web control leg once first and add in the domains unique to it, then freeze the account-level Top 20, retest 3 rounds in the same week to measure the noise band, and complete all 36 brand-six runs in the same week as the baseline. Law firms only fill in a few cells with their own values: read the law-firm-website cell as the 3 rounds combined, split the denominator by practice area, and count keyword-microsite citations back to the firm. When you're done, you'll have a frozen account-level Top 20, a noise band, and a set of raw brand-six answers you can use as the baseline.

The rulers for law firms follow the General Edition; only a few cells need the firm's own values filled in, and nothing is redefined:

Five "whys" the figure can't show.

Why the law-firm-website cell is read as the 3 rounds combined: asking the same question twice changes the list a great deal. Each of L3's 30 questions was asked of ChatGPT twice (2026-09-29): the intersection of the two runs' cited URLs over their union is 73/234 (31.2%), or 61/130 (46.9%) by registered domain, and only 2/30 questions had exactly the same URLs both times; of the 9 questions that cited a law-firm website at least once, only 5 cited one both times. The structure is stable, though: 30/30 questions cited a government or public-body page both times. So whether "the firm's own website was cited" is very unstable in any single question-and-answer, and this cell only reads two numbers: the citation count summed across the same-week 3-round retest, and the number of questions cited on at least one of the 3 rounds. Family-law questions haven't been rerun and their noise band hasn't been measured, so read them by the same rule. How the noise band itself is measured and used is in → General Edition 7.2 The noise band, the page-level signal and the ten monthly steps.

Why the denominator is split by practice area: the numbers in 2.2 show that AI's way of finding sources differs sharply between practice areas. If this quarter's pool covers two practice areas, report both numbers, seat count and on the list and cited, as one row per practice area; the totals row is for reference only, never a basis for judgement.

Why the two legs are logged separately: for the same question, the two engines cite almost none of the same pages — the shared URLs are 17/251 (6.8%, 2026-09-29) for L3 and 6/106 (5.7%, 2026-09-23) for family R1. Merge them into one score, and you can no longer tell which leg a rise or fall came from. The two legs have different denominators for share; only read "who cites what", never compare the sizes, see → General Edition 1.2 Two legs: ChatGPT looks for the source, AI Mode for second-hand summaries. The AI Mode leg's number of passes and region parameters are also fixed in the week the baseline is frozen, with the same same-week 3-round retest, and don't change for the whole quarter; the seat name list is fixed the same week too — the common forms of the approved firm name, plus every one of the firm's lawyers' names as on their practising certificate. How the two are used is in chapter 7, 7.1.

Why keyword microsites are counted back to the firm: the numbers in 2.1 show microsites do get cited. If the baseline doesn't count them in, then the day a microsite gets merged into the main site per chapter 3, 3.4, the same citations move from "someone else's" to "the firm's" — it will look like a rise, when really the accounting basis changed. So from day one of the baseline, count "the law firm, combined" — the microsite goes on the firm's domain list.

Why the web leg has to run first: a candidate pool taken only from the API leg has a systematic bias that running more rounds cannot fix. The complete method for running it has exactly one specification in the whole book, in → General Edition 4.4 The web control leg and the frozen baseline (the book's only full spec); how to do the coarse screen into piles and the full rounds is in → General Edition 4.3 Coarse screen and piles, full rounds, and the brand-six baseline. The frozen account-level Top 20 is the denominator for the entire quarter and does not change after that.

The brand six questions: ask with the approved full name

The composition of the six questions follows → General Edition 3.5 The brand six questions, and what to do when you find an error, with not one question changed. Law firms only fix the wording in three places:

• In the question text, <Legal name> is always the approved full name as it appears on the LSRA register lookup, with the LLC suffix; never a short form, never a microsite domain.
• Question 4 asks about legal fees for this quarter's main practice area ("how much are <approved full name>'s legal fees for <matter type>"); it's the most valuable of the six.
• The 2 Chinese questions follow the General Edition, run only where there's a Chinese page, and are a separate ruler (denominator 12, not folded into the 36); use the approved Chinese name where there is one, and the English approved name otherwise.

Why the brand-six run only counts these 36: the baseline's "X factual errors" figure comes only from this one run of 36. Any 2-round × 1-engine quick run (such as the two passes in shortcut 2) is only used to fix errors the same day, ahead of the queue — never as the baseline; a quick run and the 36 runs are not the same ruler, and if X came from a quick run, the before-and-after in week 13's "X → Y" would not be comparable. → General Edition 3.1 Why this comes before writing pages · The two-hour checklist also reads this same set of raw answers.

The three rulers stay unchanged, the fourth reference number never counts as a criterion, the web leg runs first, and triage must not skip layers — these four rules are the general discipline and not a word of them changes; see → General Edition 7.1 What a retest produces, and the rulers. How to read review counts is in chapter 7, 7.1.

2.4 Fix the website settings: check each domain one by one, and record the day you change them

What you'll do in this section: once the baseline is saved, fix the door through the General Edition's five gates, going through the main site and every keyword microsite one by one; record the day you fix the door as the split day; and review the words written into JSON-LD as publicity content. When you're done, you'll have one merged robots.txt per domain, one WAF allowlist rule for the layer in front of each domain, and a split day written into the work order (one per domain, where domains are fixed on different days).

Hard gate: until the baseline is saved, do not change the door, any profile or the facts page.

Exactly how to fix the five gates — the three nosnippet spots, the robots four-step merge, the three WAF allowlist steps — is all written in → General Edition 2.2 Gate 1 read-only audit: what to check in nosnippet, robots and the WAF–→ General Edition 2.6 Fixing the door: nosnippet, the four-step robots merge, the WAF allowlist; not repeated here. Law firms add just one rule: every domain is its own separate door.

Change each domain separately

A microsite might sit on a different host, a different CDN, and its template and plugins may not match the main site's either (the D1 read-only audit will tell you whether they do) — so all three things are done by domain:

• robots: each domain counts its own N: the N in the four-step merge's "Original wildcard Disallow lines = N" differs by domain; write one receipt sentence per domain, and copy each N from its own domain.
• nosnippet: grep each domain separately: the main site coming up clean doesn't mean a microsite's template and plugins are clean too.
• WAF: check each domain separately: an allowlist rule added at the main site's layer doesn't reach the layer in front of a microsite.

Whether a microsite eventually gets merged into the main site is decided in chapter 3, 3.4 — don't do it in passing on the day you fix the door: merging sites is a separate change, and the day it happens gets its own separate split day; crowd it into the same day as fixing the door, and by week 13 you won't be able to tell whether a rise or fall came from opening the door or from merging sites.

These two troubleshooting diagrams aren't only for the day you fix the door: check the door layer unconditionally once a month, and also check it when seats rise (→ General Edition 7.3 Not-moved triage and next month's three points). Law firms with microsites check every domain every month.

Door changes need no sign-off; the words in the schema are reviewed as publicity

The basis comes in two halves. The definition in PCR r 2(1) reads "“publicity” means any form of advertisement", and it states that this includes advertisements retrievable from the Internet [Statute text; see Appendix A.1]. Settings such as robots, WAF and nosnippet are access settings for crawlers, not advertising content aimed at the public; we do not treat them as publicity, and changing them does not go through a partner's sign-off [Original text not obtained; see Appendix A.6]: no prohibition found in the current rules (scope checked: PCR 2015 in full [including the r 2(1) definition of "publicity" and Part 5], the Legal Profession Act and its subsidiary rules, the Law Society Council's 12 Practice Directions and 2 Guidance Notes; check date: 2026-09-29). To the law firm we say only, "We found no rule that treats these settings as publicity."

The words written into JSON-LD are different: they are text the public can retrieve, so they are reviewed as publicity content, written per chapter 1 [Conservative line (not statute text)]. Schema only restates facts already stated in the visible HTML — not one extra word. This comes down to four rules for the fields:

1. name takes the approved full name, with the suffix for an LLC, matching /facts and the five profiles character for character (LPA s 154(3), rule text in 2.1) [Statute text; see Appendix A.2].
2. Only write expertise fields you have evidence for: in publicity, the lawyer must ensure that "any claim to expertise or specialisation can be justified" (r 43(1)(a)) [Statute text; see Appendix A.1] — how to prepare the evidence is in chapter 1, 1.6; an expertise claim you can't back with evidence goes into neither the visible HTML nor the schema.
3. Write the fees field per chapter 1, 1.2: a price in structured data is another carrier of the firm's own fees, in the same format as the fees page — only the firm's own prices, every number traceable on that same page's visible HTML to its service tier, defining conditions, billing method and what's charged separately, with the page stating "The final fee is set out in our letter of engagement."; a price not written out in the visible HTML doesn't go into the schema either [Conservative line (not statute text)].
4. Leave out aggregateRating: this is the General Edition's build rule, not a law-firm compliance judgement; the platform-rating line on your own page is written in the visible HTML, see chapter 1, 1.5.

The General Edition requires four schema items to be written correctly; law firms fill them in this way: the legal-name field takes the approved full name; the address and phone match the five profiles character for character; for the "each practitioner's registration number" item, the LSRA's field name for individual practising details was not verified this round [Original text not obtained; see Appendix A.6], so write no number — instead, write the LSRA register lookup entry point and the check date in /facts's visible HTML; identifier takes the UEN; sameAs strings together the firm's listing page on the Law Society directory (where it has one), its profile page on a ranking guide, and its business profile pages, and links the SAL Directory of Specialists only for lawyers whose accreditation is current. Where the LSRA register lookup can't give a separate URL, it doesn't go into sameAs. These fields only matter to Google's Knowledge Graph — none of AI's live fetches read JSON-LD at all — so the same words must be written into the visible HTML first, see → General Edition 2.7 Gates 3 and 4: indexing paths, and JSON-LD sealed once.

The split day goes into the work order; at the week-13 settlement, every "before fixing the door vs after fixing the door" comparison is drawn against it, see → General Edition 7.4 Holding position and the 90-day settlement. It's best to fix all domains on the same day; where that's not possible, each domain records its own split day, and at settlement each is measured against its own split day.