# Chapter 7 · Monthly retest: how to tell whether it's working

> GEO Playbook · Dental & Aesthetics v1.0 · Canlah AI · CC BY 4.0 · Web page: https://canlah.ai/playbook/dental/measure/
> 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.

## 7.1 Monthly retest: how dental reads the numbers

**What you'll do in this section**: At each monthly retest, read the numbers against the reference table below: use the three rulers and the page-level signal exactly as in the General Edition; log review counts and star ratings but never judge them; don't read listicle share; and add one fixed sentence to the review section of the monthly report. When you're done, you'll have monthly-report numbers read on one consistent basis, and a list of this month's page-level-signal hits.

```split Figure: The three rulers and the page-level signal are used exactly as in the General Edition; dental changes only the review and listicle-share cells
Reading || How dental reads it
Seat count; only counts as up once past the noise band || Use as is; noise band uses 3 rounds
On the list and cited X/20 || Use as is
Factual errors (count) || Use as is; baseline is the 36 runs from 2.3
Review counts and star ratings || Passive observation only; never a metric, never used for acceptance
Page-level signal || Log 2–3 strings unique to each page
Listicle share || Not read: dental can only request removal from lists
```

Why reviews are logged but never judged: in this industry's data, review volume really does matter, but actively asking patients to write reviews, giving a benefit in exchange for a review, and reposting platform reviews onto the clinic's own website are all explicitly banned (**Statute text**; see Appendix A). All you can do is stay completely out of it and record the numbers that arise naturally, so this cell never goes into a task metric or the acceptance measure. If a client wants to ask for reviews, or wants to make review count a KPI to chase, a written sign-off from the compliance officer cannot unlock it: a written sign-off resolves only the three stop-and-escalate situations; it never turns something banned as Statute text into something publishable (see 1.1).

Add this exact sentence to the review section of every monthly report, word for word (template in Appendix B):

> "This clinic does not request reviews. The figures above are a passive record of reviews that arose naturally and are not used as a work target or acceptance measure."

Why listicle share isn't read: by citations, lists make up 0% of AI citations in healthcare (from deltaV's page-type breakdown by industry); but by who gets named, the picture is different — of the clinics named under a generic "best", 36/62 of the names overlap with third-party best-of lists (rough control: of the clinics in Google's top 10 for the same procedure's price and detail questions, 27/83 are also on these lists, and only 9/40 once the clinics AI named are removed; evidence in （→ 通用版 A.2 选点、写页、站外与复测的证据）). So only one reason for not reading it is left, and it is enough on its own: in dental, Best / Top lists get no submissions and no payment, and if the clinic is listed you must also ask the publisher to remove it (6.2); you never build a list page of your own either (see 0.3); so this share cannot be chased as a target. Skip this cell entirely; don't try to fill it.


The factual-error baseline accepts only the full 36 runs from the week the baseline was frozen (2.3). The quick run of 2 rounds × 1 engine under the General Edition's shortcut 2 is only for fixing an error the same day, ahead of the queue, and never counts as the baseline: the two runs are made up differently, so a before-and-after comparison between them gives a difference that means nothing. Run the same make-up every month (6 questions × 3 rounds × 2 engines), unchanged for the whole quarter.

> **Example**: When a page goes live, log checkable facts unique to that page, such as `crowding 3–6mm` · `S$6,800` · `5 follow-up visits + 1 set of retainers` · `brand XX model YY scanner`, and every month search the saved raw answer text for them. If one is repeated back, that page was read and trusted.

For how to schedule the ten monthly steps, how to judge the noise band and how the page-level signal works, follow （→ 通用版 7.1 复测的产出与量具） and （→ 通用版 7.2 噪声带、页级信号与每月十步）. The web control leg's 40 minutes are not part of the ten monthly steps: set aside a separate day at the start of each month to get it done. The specification is in （→ 通用版 4.4 网页对照腿与冻结基线（全书唯一完整规格））.

## 7.2 When next month's three points fall short: fallback B; admission and settlement

**What you'll do in this section**: In a month when points 1 to 3 can't fill three tasks, assign profile and fact tasks under fallback B, and never a review task. For a client who will never publish itemised fixed prices on its own website, only hold position. At week 13, produce an honest before-and-after comparison, split at the split day logged in 2.4.

```mermaid Figure: When dental can't fill all three points, it takes fallback B only; the review task is never assigned, at any step
flowchart LR
  t["Triage → point 1 to point 3"] --> q{"All three points filled?"}
  q -->|Yes| done["Assign tasks from three points"]
  q -->|No| b["Fallback B: profile and fact tasks"]:::hl
  b --> c["Fallback C: expand the price page"]
  c --> d["Fallback D: Chinese page or long-form video"]
  b -->|Never assigned| r["Review task"]:::warn
```

By content form, dental counts as consumer, so when the three points can't be filled it takes fallback B (the evidence for splitting the fallback by content form doesn't include Singapore; about 75% confidence; see 0.1). Do all five items of fallback B in full:

- Fill in every fact field on all five business profiles (how: see 3.4)
- Expand the `/facts` page
- Complete the itemised prices on the price page (how to write them: see 1.2)
- Complete the regulatory outbound links: never link to manufacturers or any other commercial company [Statute text; see Appendix A.2]; link only to government and regulator pages [Conservative line (not statute text)] (see 1.5)
- Manufacturer locator listings: get manufacturer locators to list you; they link to you, and your own pages don't link back (how: see 3.3)


Review tasks are never assigned because the strictly regulated side may not actively request reviews. In dental, this kind of task does not exist at any point in the quarter; it is not a case of "not assigned only because this month came up short".

Admission criterion: whether the client is willing to publish a fixed, checkable price on its own website. A client who is never willing to, or whose website can never be changed, gets hold position only. The price page is one of the few slots on an owned site that can genuinely get into an answer; not giving a price is the same as giving up that slot (evidence: （→ 通用版 A.2 选点、写页、站外与复测的证据）). For dental, admission is judged on **itemised fixed prices**, not on "whether they're willing to publish a range": on this side, writing a range is itself non-compliant, so treating "won't give a range" as "won't give a price" would wrongly rule out a client who is in fact compliant and simply can't write prices to the general template.

For how to pick points 1 to 3, the rule that the five-layer triage never skips a layer, the ten-step settlement at week 13 and the five holding checks, follow （→ 通用版 7.3 没动分诊与下月三个点） and （→ 通用版 7.4 守位与九十天结账）; the dental edition doesn't rewrite them. At the week-13 settlement, the before-and-after comparison splits at the split day logged in 2.4 (the day you fixed the door): before that day counts as before the door was fixed, and after it counts as after.