Monthly retests: how to tell whether it is working

Read the General Edition's chapter 7 first; this chapter only covers the cells where law firms read the numbers, fall back and settle differently. This chapter's numbers use two source labels. "L3" = the 30 questions collected on 2026-09-29 (8 practice areas besides family law, plus 5 questions on choosing a lawyer and legal fees that apply across every practice area; 26 English, 4 Chinese): ChatGPT was asked each question independently twice, with run 1 in the main text and run 2 in brackets; AI Mode was asked each question once. "The family set" = the 22 family-law questions from the public dataset's R1/R2 (collected 2026-09-23, each question asked once; AI Mode was only run on R1's 12 questions). Wherever a page type is mentioned, it is labelled heuristically by URL and title, not taken apart page by page by hand.

7.1 Monthly retests: how law firms read the numbers

What you'll do in this section: monthly retests follow the General Edition's three rulers and page-level signals; law firms only change how a few cells are read: record the ChatGPT leg and the AI Mode leg separately, and for each page type read only its own leg's column; seats also count the names of the firm's own lawyers; the firm-website cell is read as the three rounds combined; reviews count only requests sent; list share is not read; the Law Society of Singapore directory listing gets a row of its own. When you're done, you'll have a monthly report reading laid out in two columns, a hit list of page-level signals, and one row of directory-listing readings.

Record the two legs separately; read each page type in its own column only

The Law Firms Edition reads two legs every month. The ChatGPT leg is the General Edition's OpenAI API leg; in reports, write it as the General Edition does, "Model-API visibility baseline · OpenAI leg". The AI Mode leg is the Google AI Mode answers and citations retrieved through a third-party SERP API, set to the Singapore region only; in reports write "AI Mode leg · third-party API sample", never "what users see on Google".

This cell differs between law firms and the General Edition. The General Edition treats the AI Mode leg only as a measured page-type test, not as a monthly ruler (→ General Edition 1.2 Two legs: ChatGPT looks for the source, AI Mode for second-hand summaries); for law firms, explainer pages and Chinese pages barely appear on the ChatGPT leg (table below) — without reading the AI Mode leg, chapter 5's explainer pages and Chinese pages would have no monthly reading at all. The General Edition's Gemini model leg and Search Console generative AI report impressions still run and still get read, but neither can substitute for this leg: the Gemini API ≠ AI Mode, and Search Console only shows this site's own impressions — no names, and no visibility into who else the answer cited.

The family set points the same way: of the 13 questions on process, requirements, definitions and rules, ChatGPT cited a law firm in 0 of them (the family set). So for explainer pages and Chinese pages the ChatGPT reading is close to 0 to begin with: that does not count as the work having no effect, and we never tell a client "Write explainer pages and ChatGPT will cite you".

The two columns have different denominators: for law-firm websites the ChatGPT column is 12/146 (run 2: 11/161) and the AI Mode column is 47/122 (L3). Read only "who cited which of the firm's pages"; never compare which column is bigger, never merge them into a single "visibility" number, and never plot them on the same bar chart. If this quarter covers two practice areas, each practice area gets its own pair of columns (2.3).

The AI Mode leg's number of passes and region parameter are fixed in the week the baseline is frozen, just as the ChatGPT leg's are, and do not change for the rest of the quarter — changing them is changing the ruler (→ General Edition 7.5 Ruler discipline: change the ruler and nothing is comparable). Its variation was not measured this round — L3 asked each question only once — so in the first quarter we measure it ourselves with a same-week retest of 3 extra runs, as the General Edition does, and for the first two months its readings are only listed side by side. Machine run time is measured in minutes (running L3's 30 questions through AI Mode once takes about 3 minutes); the extra time is a person reading the two columns, which goes into step 4 of the ten monthly steps. The AI Mode leg must save the answer text together with its citations to disk: if you only store the citation list, there is nowhere to search for page-level signals.

Seats: the firm name and lawyer names both count, one seat per firm

The definition of a seat is unchanged, word for word, from the General Edition; law firms only need to decide what counts as "your name": the common forms of the approved firm name (the full name with the LLC / LLP suffix, and the short form without it), plus the name of every one of the firm's lawyers as it appears on their practising certificate. The list is fixed in the week the baseline is frozen; a lawyer who joins mid-quarter is only added to the list next quarter — adding a name mid-quarter is changing the ruler. When a lawyer's name and the firm name appear together in the same answer, that counts as the same firm, one seat; naming only the lawyer, with no firm name, still counts as one seat. Total names in this answer follows the same rule: sort every lawyer's name to their own firm before counting. If a keyword microsite gets named, it is counted back to the firm per 2.3.

Why lawyer names count: when asked "who's best", ChatGPT often names the lawyers themselves. L3's one "who's best" question about criminal lawyers (single case) listed lawyer names one by one, following one ranking guide's tiers, both times, with the firm name trailing behind, and in a few places only the person's name was given. Counting only firm names would systematically under-count this cell.

The firm-website cell: read the three rounds together

For the 3 rounds of the monthly same-week retest, the firm-website cell reads only the total citations across the three rounds and the number of questions cited in "at least one of the three rounds"; never draw a conclusion from a single round (numbers and reasons: 2.3). The office-page row in the table above is a ready-made example: same day, same two questions, same ruler, only a few minutes apart, and ChatGPT run 1 cited law-firm websites 3/8, run 2 0/8. How to measure the noise band and judge whether a change exceeds it: see → General Edition 7.2 The noise band, the page-level signal and the ten monthly steps.

Reviews: count only requests sent

Asking clients to leave reviews [Original text not obtained]: no prohibition found in the current rules (scope checked: the current PCR in full, the revoked Publicity Rules (in full), the full category 6 list on the Law Society's ethics page; check date: 2026-09-29), 70% confidence; when asking, keep to the four conditions — offer no incentive, don't write the review for them, don't ask only satisfied clients, don't dictate the content [Conservative line (not statute text)] (1.5, 6.5). So the only task metric is requests sent: switch the metric to star ratings or review counts, and the only way left to chase the number is picking clients and offering incentives — exactly what the four conditions ban. Next to the number sent, write the number of similar cases closed in the same period, for cross-checking only: if the two numbers are far apart, check first whether clients are being cherry-picked. Star ratings, review counts, reply rate and reply time are logged per → General Edition B.4 Headers for work orders, ledgers and monthly reports, as observation only.

The reviews section of the monthly report always adds this sentence, word for word (Appendix B.3):

"The review metric counts only requests sent; star ratings and review counts are observation records, not an acceptance measure."

For family clients there are two more rules. First, the monthly report writes numbers only and never quotes review text [Conservative line (not statute text)]: if a review names a child, their school, or a detail that could piece together their identity, and we extract and forward it, we could become "the person who publishes or distributes it" — CYPA s 112(9) defines "publish" to cover even "any messaging system" [Statute text; see Appendix A.4]; the text does not go far enough to say whether a monthly report sent to the firm counts, so we never quote any review at all. Second, before every family-law review reply we draft on the firm's behalf goes for sign-off, it first goes through our de-identification check (Appendix B.4); any hit means it does not go for sign-off. The lawyer's written sign-off decides whether that reply gets published, but it cannot shield us from the publisher's own criminal liability (CYPA s 112(5)(b), FJA s 10(5)) [Statute text; see Appendix A.4] (1.7). A written sign-off resolves only the stop-and-escalate situations; it never turns something banned as Statute text into something publishable.

Law Society directory listings: a row of their own

This row logs three things every month: how many citations the firm's listing page gets on each leg, and how many questions it covers (ChatGPT combined across the three rounds); whether the 2–3 unique fact strings registered in the listing's blurb have been restated; and whether the firm name, practice areas and contact details on the listing match the firm's website.

It gets a row of its own because it is one of the few places that get into ChatGPT's answers to recommendation questions. Of the 7 recommendation questions in L3 (near me 2, is this firm good 2, who's best 3), ChatGPT cited a law firm's listing page in 3 questions (run 1: 2 questions; run 2: 3 questions), and AI Mode in 2. It is also because it is a paid placement: the directory's own page states "the listings are paid by advertising" [Statute text; see Appendix A.5]; whether every cited listing is a paid placement we judged this round only by URL and page title, without checking them one by one. It is the only paid placement worth considering for law firms [Conservative line (not statute text)] (70% confidence); we did not get the rate-card terms [Original text not obtained; see Appendix A.6] — whether to buy it, see 6.2. For firms that already bought it, whether to renew is decided by this row's quarterly numbers, not by the enquiry count.

Example (single case) For L3's "who's best" question about personal-injury lawyers, ChatGPT's run-2 answer wrote "its profile says the firm handles …", then paraphrased what one firm had written about itself in that directory listing, its practice areas and years in practice, with the citation attached to that listing page. Sentences a firm writes itself in its listing blurb do get paraphrased, so register page-level signals for it just as you would for the firm's own website.

List share is not read

There are two reasons, and each is enough on its own. First, the share is small: citations that L3 labelled heuristically as list pages are 3/146 for ChatGPT run 1 (run 2: 3/161) and 5/122 for AI Mode; of the family set's 171 citations, 2 were "best"-type articles (the family set, roughly labelled). Second, only one action is left to take: law firms do not build their own list pages [Conservative line (not statute text)] — ranking peers against each other is comparing the quality of services, which hits r 43(1)(c) [Statute text; see Appendix A.1]; for third-party editorial "best" lists, we only supply checkable facts and send correction letters, we never pay [Conservative line (not statute text)] (0.3, 6.3). Movement in this share does not correspond to any task, so this cell is simply skipped.

Ranking guides (peer-reviewed rankings such as Doyle's, Legal 500 and Chambers) do not count as list pages. They were ChatGPT's main source in two "who's best" questions (1 in the family set; 1 in L3's question about criminal lawyers; each a single case — L3's other two "who's best" questions did not cite a ranking guide), and are logged under on the list X/20 as usual; for how to take part, see 3.3.

Page-level signals: search the two columns separately

The registration rule is unchanged from the General Edition: when a page launches, register 2–3 checkable strings that only this page has, then search for them every month in the answer text already saved to disk; if one gets restated, that page was read and its content was trusted. Law firms add only one extra rule: for fee pages, office pages and directory listings, search the strings in both legs' answers; for explainer pages and Chinese pages, search only the AI Mode leg (Chinese pages register Chinese-language strings).

Example: when a single-matter fee page launches, register three strings: <the firm's service-tier name> · S$<this tier's fixed fee> · <an included item that only this page mentions>. Adjectives, or sentences that appear on other pages too (for example "The final fee is set out in our letter of engagement"), don't count — finding them tells you nothing about which page was actually read.

Page 2 of the monthly report names peers on the competitor board; it is for the firm's internal use only and never goes into any publicity [Conservative line (not statute text)]: making a point about a peer's ranking in publicity reads as the comparison of service quality that r 43(1)(c) bans [Statute text; see Appendix A.1].

How to order the ten monthly steps, how to judge the noise band and how page-level signals work: see → General Edition 7.1 What a retest produces, and the rulers, → General Edition 7.2 The noise band, the page-level signal and the ten monthly steps. The web control leg's 40 minutes are not part of the ten monthly steps; schedule them on their own day at the start of each month. Full spec: see → General Edition 4.4 The web control leg and the frozen baseline (the book's only full spec).

7.2 When next month cannot fill three tasks: thicken explainer pages first; admission and settlement

What you'll do in this section: in months when points 1 to 3 cannot fill three tasks, first assign fallback A's task of thickening explainer pages, and accept it by reading only the AI Mode column; if that still doesn't fill the quota, move to fallback C (fee-page expansion), then fallback D (a Chinese page, or a long video narrated by a lawyer); referral platforms and paid placements billed per lead are never assigned at any step. Admission depends on whether the firm is willing to publish its fee basis; at week 13, pull an honest before-and-after comparison using the split days logged under 2.4 (one per domain); renewals are quoted on a fixed service fee only.

Law firms fall under the professional content form (0.1), so when the quota can't be filled, fallback A comes first: fold the landing page for the billable intent with the lowest seat count into its intent cluster and thicken it; the task spec follows the General Edition. Splitting fallbacks by content form rests on a cross-industry query sample that does not include Singapore, at about 75% confidence.

For law firms, fallback A comes with one thing to keep in mind: in this profession, explainer pages feed only the AI Mode leg (numbers: the 7.1 table). So fallback A's tasks are accepted by reading only the AI Mode column; the ChatGPT column not moving is not a fallback-A failure. When a family law firm thickens its explainer pages, every newly written paragraph must pass the three family-law rules in 1.7: write mediation and divorce by mutual agreement as the preferred option — family lawyers have a professional duty to advise clients to consider amicable resolution (PCR r 15A(2)(b)) [Statute text; see Appendix A.1], and copy that plays up conflict, such as "fight it all the way", is not written [Conservative line (not statute text)]; to add a case study, it must first pass the de-identification check (Appendix B.4), then be signed off in writing by a partner or director; anything that could identify a child involved in the proceedings (CYPA s 112), or anything covered by a court order restricting publication (FJA s 10), is not written even with a sign-off [Statute text; see Appendix A.4].

Fallback C is fee-page expansion: add to the same table any services the firm handles that the fee guide page does not yet cover; on the same day, re-copy the court-fees column from the court's page and update the date retrieved; write every row in full per 1.2 — service tier, defining conditions, billing method, fixed fee or range, what's included and charged separately, payment stages, "The final fee is set out in our letter of engagement"; never write a comparison of fees against peers (r 43(1)(c)) [Statute text; see Appendix A.1]. Asking about price is the question type for which ChatGPT cites law-firm websites most often (L3 run 1: 5 of the 12 citations; see 4.2); of the three fallbacks, this is the only step aimed at the question type for which ChatGPT has actually cited law-firm websites.

Fallback D is a Chinese page, or pairing an existing thick page with a long video narrated by a lawyer. Chinese pages are only built when the firm has Chinese-speaking clients, and are accepted by reading only the AI Mode leg (5.3). Writing the firm's name in Chinese requires written approval (LPE Rules r 7) [Statute text; see Appendix A.2]; applied to a Chinese page, only use the approved Chinese name — where there is none, write the approved English name and never translate it yourself [Conservative line (not statute text)] (85% confidence). The video only covers procedure, how fees are made up, and what the firm does and does not handle; it never covers success rates, and clients never appear on camera (6.6).

The "never assigned" branch in the figure is fixed: for platforms that refer, assign or match cases for law firms, and placements that charge per lead, per completed engagement or by commission, the firm paying for them breaches r 39(2)(b) and r 19 [Statute text; see Appendix A.1]; the firm does not pay a referral platform a fixed entry fee either [Conservative line (not statute text)] (80% confidence). Not being able to fill three tasks is a reason to switch to a different content action, not a reason to go and buy leads.

Reddit's share is read monthly as in the General Edition; for law firms it's very low: across ChatGPT's two runs, L3 cites forums and social media a combined 1/307 (and that one isn't Reddit), AI Mode 2/122 (of which 1 is Reddit). If it ever does cross 2% and a Reddit task needs scheduling, that post is still the firm's publicity: write it per chapter 1 and put it through sign-off, and the post contains no touting (r 39(1)) [Statute text; see Appendix A.1].

Admission: will the firm publish its fee basis?

Admission for law firms asks "is the firm willing to publish its fee basis": billing method, plus a fixed fee or range, plus what's included and what isn't. This is not the General Edition's "definite, checkable price", nor the dental edition's itemised fixed price. You can write a range [Original text not obtained]: no prohibition found in the current rules (scope checked: the PCR in full [including Part 5 (r 37–49), r 17, r 18 and r 18A], all category 6 PDs and GNs on the Law Society's ethics page, the revoked Publicity Rules (in full); check date: 2026-09-29), 75% confidence; for the wording, see 1.2. Using "does it write a definite price" to judge admission would wrongly knock out a compliant firm that simply prices in ranges.

Firms that failed the second start gate at kickoff have already been downgraded (0.2): no fee pages are scheduled, the work feeds AI Mode only, and acceptance reads only the AI Mode leg. Firms still unwilling to publish their fee basis by week 13 do nothing but hold position next quarter — fee pages are one of the few cells on a firm's website that can get into a ChatGPT answer, and withholding the fee basis is giving up that column.

Settlement: split days by domain, renewals on fixed service fees only

How points 1 to 3 are picked, the five-layer triage with no skipping layers, the eight-step settlement order at week 13 (D90), and the five holding checks: see → General Edition 7.3 Not-moved triage and next month's three points, → General Edition 7.4 Holding position and the 90-day settlement; law firms don't rewrite these, they only change four things:

1. Split days are counted by domain: the main site and every keyword microsite each have their own split day, and merging a microsite into the main site gets its own extra date logged (2.4). Before-and-after comparisons are cut at each one's own split day; never cut every domain at the same date.
2. Settle the two legs separately: at the three points — the start, day 60 and day 90 — the ChatGPT column and the AI Mode column are each compared against themselves, each with its own noise band; "seats beyond the band, or cited pages +3" is also judged column by column, never adding the two columns together to make the number. Explainer pages and Chinese pages settle in the AI Mode column only.
3. Renewals are quoted on a fixed service fee only: the enquiry-source count in the settlement report (the "How did you find us?" field required on forms, calls and WhatsApp) goes into a reference column only — never into acceptance, and never into next quarter's quote. If we charge per lead, per case, per signed client or as a percentage of legal fees, we may commit the offence in LPA s 33(3); the firm paying a percentage of its legal fees, or a share per case, breaches r 19, and paying per lead or per signed client falls within r 39(2)(b) [Statute text; see Appendix A.1, A.2]. Working on a fixed service fee [Original text not obtained]: no prohibition found in the current rules (scope checked: all of LPA s 33; check date: 2026-09-29), 70% confidence.
4. In the settlement week, also recheck the reclassification triggers: whether the PCR's timeline on SSO shows a new amendment, whether the Law Society's ethics page has a new Practice Direction on publicity, and whether the written guidance the client exports from their members' account, or the Ethics Digests, contain a ruling like "fee ranges or price lists on a website are unbefitting the dignity of the profession" or "lawyers may not ask clients for reviews". If any of these hits, the whole book moves law firms to the strictly regulated side in that quarter, per 0.2; if a client firm sets out a stricter position in writing, the whole book is not changed — only that firm follows the stricter position.

To the outside world, we only state the cells that can be verified: the deliverables list, pages on the list, and factual errors (count), each with a screenshot of the original attached. Seat counts are only listed side by side; we never promise a number, and we never promise which page ChatGPT will cite.