Install a Review Ask + Human-Sent Reply (Never Fake a Star)

You're the owner. The shop across the street has 400 Google reviews. You have 40. Your work is not four times worse. Nobody asks at the right time, and the replies you do post sound like a hotel chain. Reviews are how a stranger picks a dentist, an HVAC company, a store, or — carefully — a law firm. That is Effectiveness. The person who just paid you should not get a ghost. That is Quality.

This playbook installs two loops: an honest ask after a real job, and a human-sent reply to whatever lands. We never write fake Google reviews. We never pay for them. We never put words in a customer's mouth and hit submit from the shop. If a vendor demo shows you a pile of auto-generated five-stars, walk out.

AI's job is timing, drafting, and classifying. A human still sends. Always.

Who this is for

Dental practices, HVAC / home services, law firms, and multi-location retail. Agencies can steal the reply queue for their own GBP. You have a Google Business Profile, a way to text, and one person who can own a queue for 15 minutes a day.

Law is narrower. Some matters should never be on Google. A model does not get to decide that.

What done looks like

• A one-page quality rules file the helper reads every time.
• An ask that fires only on tagged "smooth" jobs/visits, with the correct Maps link, sent by a human.
• Every new review classified: praise / complaint / legal-adjacent.
• Draft replies in your voice. Human hits post. Owner or office manager owns anything under four stars or anything legal-adjacent.
• A 48-hour SLA on replies you can see without a new dashboard.

Time and tools

Afternoon one: write the rules, the never-ask list, the three-class labels, and the prompts. Week one: run the ask on real completed work and the reply queue on whatever is already sitting on GBP. You do not need a new review platform to start. Google, a texting tool, Claude, a sheet.

Prerequisites

• Login to Google Business Profile for the correct location(s). If location B's link points at location A, you are donating stars to the wrong pin. Fix that first.
• A tag the desk or tech already understands: ask / don't ask / wait.
• One named owner of the queue. "We'll all reply" means nobody replies on Sunday.

Step 1 — Write the quality rules before you touch a prompt

Why. If the rules live in someone's head, the model will invent a coupon, a diagnosis, or a "we'll have a tech there in 20 minutes." Write them. Treat the file like it could leak, because it will get pasted.

Clicks.

1. Open a Google Doc titled review-rules.md (plain text in a doc is fine).
2. Fill the skeleton. Owner writes the never-list. Office manager adds when we ask vs wait.
3. Pin it. Every prompt in this playbook starts with "read review-rules.md first."

SHOP: __________
GBP NAME / MAPS LINK: __________
VOICE: short, named, specific. Never "Dear Valued Customer." Never "We strive for excellence."

WHEN WE ASK
- Job/visit completed, outcome clear, tagged SMOOTH
- Same day or next morning
- Direct Maps link for THIS location
- Human sends the text

WHEN WE NEVER ASK
- They just complained
- Callback likely / parts on order / treatment conversation unfinished
- Insurance fight, billing dispute, clinical complication
- Law: any matter we would not want described on the internet
- The visit ran long and they left angry, even if we think we were right

NEVER
- Write, buy, or "seed" Google reviews
- Auto-post a reply
- Invent a discount, diagnosis, legal position, or arrival time
- Argue medicine or the law in public
- Copy location A's five-star reply onto location B

WHO SENDS
- Praise (4–5 star, no sting): [front desk / office manager]
- Complaint (1–3 star, or 4–5 with a sting): [owner or office manager]
- Legal-adjacent (injury, discrimination, HIPAA-ish, threat, named staff in a fight, anything a lawyer would want): owner only, and maybe don't reply until counsel glances

SLA: every review touched in 48 hours, even if the touch is "parked for owner."

Definition of done. The file exists. The never-ask and never-fake lines are in writing. One person is named for each class.

Step 2 — Classify praise vs complaint vs legal-adjacent

Why. A five-star that says "great" and a five-star that says "great, except the hygienist was rough" are not the same job. A one-star about wait time is a complaint you can own. A one-star that describes a procedure, an accident, a threat, or a named employee in a way that smells like a claim is legal-adjacent. The model classifies. A human can override. Classification happens before anyone drafts a public sentence.

Clicks.

1. Add three columns to a sheet: date | location | stars | raw text | class | draft | sent? | who sent.
2. For the last 15 reviews, run the classifier prompt. You are training your eye, not building a product.
3. Put the class labels on the wall: PRAISE / COMPLAINT / LEGAL-ADJACENT.

Classifier prompt.

Read review-rules.md (pasted below).

Classify this Google review into exactly one:
- PRAISE — positive, no operational sting
- COMPLAINT — wait time, price feel, no-show on our side, rude, mess left behind, "they never called back"
- LEGAL-ADJACENT — injury, clinical outcome, alleged discrimination, HIPAA/privacy, threat of a lawyer, review that names a matter or a minor, accusation of fraud, anything I should not answer in public without a human brain

Review:
Location: [A / B / only one]
Stars: [n]
Text: """
[paste]
"""

Return:
CLASS:
WHY (one sentence):
PUBLIC REPLY ALLOWED? (yes / draft-only after owner / do not reply yet)
ONE RISK if we answer badly:

HVAC example of legal-adjacent: "they cracked my slab and won't admit it." Dental: anything about a nerve, an extraction gone wrong, a child's visit. Law: any description of a case. Retail: slip-and-fall, food illness, accusation a cashier stole. When in doubt, LEGAL-ADJACENT. You can always downgrade. You cannot unsay a public paragraph.

Definition of done. Last 15 reviews are classified. You and the office manager agree on at least 13 of them. The three misses become examples in the rules file.

Step 3 — Install the ask (timing is the whole trick)

Why. People leave reviews when the work is still warm and the outcome is clear. Not during the mess. Not six weeks later. The failure in most shops is never sending the link.

Clicks.

1. In your job/visit system, add or use a tag: SMOOTH / WAIT / NO ASK. HVAC tech leaves it on the job. Dental front desk leaves it on the visit. Retail manager leaves it on a resolved special order. Law: default is NO ASK unless the client independently offers, and even then a lawyer decides.
2. Trigger: job complete + SMOOTH → a draft text with the right Maps link sits for a human.
3. Human sends. If you cannot get a true tag yet, the office manager reviews yesterday's completed jobs each morning and sends the ask to the SMOOTH ones only. Ugly, works.

Ask-draft prompt (human still sends).

Read review-rules.md.

Write a SMS (under 320 characters) asking for a Google review.
Shop: [name]
Voice: [one line — "direct, like a neighbor" / "warm, not cutesy" / "plain, not TV-lawyer"]
Job type: [tune-up / crown seat / consult we CAN mention / special order pickup]
Maps link: [url]
Rules:
- Do not offer a discount or a gift for the review
- Do not say "leave us 5 stars"
- Do not mention clinical or legal details
- One ask, one link, thank them for the actual job
- If WAIT or NO ASK is in the notes, output ASK: NO and why

Notes from the job: """
[smooth / delayed / callback likely / they complained about X]
"""

Output:
ASK: YES or NO
TEXT:

Vertical timing (write this into the rules file).

• HVAC. After the house is back together. Same day or next morning. Only SMOOTH. Callback likely = human follow-up first, no ask.
• Dental. After they have walked out. Not from the chair while they're numb. Not during a statement fight. Hard treatment conversation = WAIT.
• Retail. After a real moment — special order arrived, warranty swap that didn't make them wait — not after a gallon of milk. Correct pin per store.
• Law. Usually no ask from the firm. If you have a policy, a lawyer owns it. The model does not.

Never auto-send the ask to someone who just complained. Tag them out. That's a rule, not a hope.

Definition of done. Yesterday's SMOOTH jobs got an ask this morning. You can count the number of asks that actually left the building. You cannot control how many reviews arrive. Stop pretending you can.

Step 4 — Draft replies that sound like the shop (not like every shop)

Why. The model will cheerfully write: "We're so sorry for your experience! We strive for excellence and would love to make this right." Every zip code has that paragraph. Specific is human. Named is human. "Sarah at the front desk will call you this afternoon about the crown that didn't seat" is Quality. "Our team values your feedback" is noise.

Clicks.

1. Paste review-rules.md into Claude.
2. Paste three replies you were proud of and three you killed. Label them GOOD and BAD.
3. For each new review, run the matching vertical prompt. Edit. Human posts.

HVAC reply prompt.

Read review-rules.md and these GOOD/BAD examples:
[paste]

Class: [PRAISE / COMPLAINT / LEGAL-ADJACENT]
Stars: [n]
Location: [name]
Review: """[paste]"""
Job notes if we have them (do not invent): """[paste or NONE]"""

Write a Google reply in our voice.
- Under 80 words
- Name the tech or the desk if we know them and it's praise
- If COMPLAINT: acknowledge the specific thing, take it offline, own what is ours. Do not blame the homeowner in public. Do not promise a 20-minute arrival.
- If LEGAL-ADJACENT: do not admit fault, do not diagnose the house, do not argue. Output a two-line "please call the office, asking for [owner]" AND a private note for the owner. Flag DO NOT POST until owner says so.
- Never invent a coupon or a free revisit

Dental reply prompt.

Read review-rules.md and GOOD/BAD examples:
[paste]

Class: [PRAISE / COMPLAINT / LEGAL-ADJACENT]
Stars: [n]
Review: """[paste]"""

Write a Google reply.
- Under 80 words. Warm, not cutesy. No spa language.
- No clinical claims. No "your x-ray looked fine." No treatment advice.
- If they mention wait time, we can own the wait. We cannot diagnose in the reply.
- Complaints: apologize for the part that's ours, offer a named person and a phone, take clinical detail offline.
- LEGAL-ADJACENT (nerve, extraction, a child, alleged injury): DO NOT POST. Two-line holding language plus a private owner note only.
- Never invent a discount on treatment

Law reply prompt.

Read review-rules.md.

Class: [PRAISE / COMPLAINT / LEGAL-ADJACENT]
Review: """[paste]"""

Default: LEGAL-ADJACENT unless this is generic praise with no matter facts.
Rules:
- Never discuss a case, a result, or another party
- Never say we will win, got a result, or "we believe you"
- If the review describes a matter: DO NOT POST a substantive reply. Output "thank you" only if counsel would allow a thank-you, else output SILENCE + private note
- If generic praise: short thank-you, no client facts, no outcome claims
- Complaints about callbacks or fees: acknowledge process, take it to a named person, no legal analysis

Retail reply prompt.

Read review-rules.md.

Location: [store]
Class: [PRAISE / COMPLAINT / LEGAL-ADJACENT]
Review: """[paste]"""

- Mention the store by the name customers use ("Oak Street") if it helps
- Specific > generic. If they named the red couch or the cashier, you may name the thing, not a different personality per store
- Complaints: own the wait / the stockout / the return if it's ours; offer a manager name; no "Dear Valued Customer"
- LEGAL-ADJACENT (slip, illness, theft accusation): DO NOT POST. Owner + maybe counsel.
- Do not copy a reply from another store

Definition of done. You have four prompts in a doc. You ran them on last week's reviews. The office manager had to rewrite fewer than half. If she rewrote most of them, the voice examples are wrong, not the staff.

Step 5 — Human-send queue, 48-hour SLA

Why. Auto-post is how a bad public sentence becomes expensive. Drafts are cheap. Send is a person.

Clicks.

1. Route new GBP reviews into the sheet (email forwarding from GBP, a Zapier "new review → row," or the office manager pastes them at 9 a.m. — pasting is allowed).
2. Run classifier, then the matching reply prompt.
3. Front desk posts PRAISE the same day.
4. Owner/OM posts COMPLAINT after a two-minute edit. Same day if you can, 48 hours max.
5. LEGAL-ADJACENT sits in a labeled tab until the owner says post, hold, or silence.
6. Save the edit: "We cut 'delighted.' We named the tech. We offered a call, not a coupon." That lessons line goes back into the rules file on Monday.

You do not need Zapier. A morning paste plus Claude is a real queue. If you already have Birdeye or Podium, use it as the inbox. Do not turn on auto-reply.

Definition of done. Nothing on GBP older than 48 hours without a touch. Legal-adjacent items have an owner note, even if the public reply is silence.

Step 6 — Count a month, not a campaign week

Why. Campaigns die on Friday. A month tells you if asks are actually leaving and if drafts are getting easier.

Clicks. At the end of four weeks, fill this. No vendor dashboard required.

Asks tagged SMOOTH: ___
Asks that actually sent: ___
Reviews that arrived: ___ (you do not control this number)
Replies posted inside 48 hours: ___ / ___
Drafts rewritten from scratch: ___  (if most, the voice file is wrong)
Legal-adjacent parked for owner: ___
Fake-review temptation we killed: ___ (write the vendor name so you remember)

Effectiveness: more real asks after real jobs. Efficiency: the office manager stops composing from a blank box. Quality: replies sound like the shop, and the angry ones got a human.

Definition of done. The scoreboard exists. You know whether the leak was "we never ask" or "we ask and sound fake." Those are different next months.

Hypothetical worked example

Hypothetical. Not a client. Napkin numbers.

Two-location dental. 22 reviews on the main pin, 6 on the second. Hygienist is "supposed to mention reviews." Most days the chair runs late. Owner writes review-rules.md on a Tuesday. Classifier on the last 15: 11 praise, 3 wait-time complaints, 1 legal-adjacent ("they ruined my nerve"). That last one stays silent in public; owner calls.

Ask: front desk tags SMOOTH on hygiene that ran on time and on crown seats that didn't bounce. Same-day text after they walk out, Maps link per location. NO ASK on the nerve conversation, on anyone who argued a statement, on the parent who waited 40 minutes.

Replies: office manager posts praise before lunch using the dental prompt. Owner takes the three wait-time complaints — owns the wait, names the desk, no clinical claims. Month one hypothetical: 40 asks sent, some reviews arrive, 100% of new reviews touched in 48 hours, two drafts rewritten because the model said "delighted." They add "never say delighted" to the rules file. They do not buy a fake-review app. Location two stops donating clicks to location one's pin.

Failure modes

You buy a tool whose pitch is five-stars by Friday. That's fabrication with extra steps. Kill it.

You auto-post. One wrong clinical or legal sentence will cost more than a year of office-manager minutes.

You ask during the mess. HVAC with the system in pieces. Dental while they're numb. That's how you farm one-stars.

You use one Maps link for three locations. You are dressing the wrong pin.

The model copies the same paragraph 80 times. Google and humans can smell it. Force a specific detail or don't post yet.

Law firm runs the HVAC prompt. Matter facts in public. Don't.

Nobody owns Sunday. Reviews sit. SLA was a wish.

FAQ

Can we offer a gift card for a review? Don't. It trains fake-feeling proof and can violate Google's rules. Ask after real work. That's enough.

What if we know the one-star is wrong? You may own your part and take it offline. You may not argue the patient's mouth or the homeowner's house in public. If it's legal-adjacent, silence plus a private process is often the adult move.

Do we need Birdeye / Podium / a review platform? No to start. GBP + texts + Claude + a sheet. If you already pay for a platform, use it as the queue. Turn auto-reply off.

A vendor said nobody will know the reviews are generated. Can we? No. I will know, the customer who never wrote it will know if they see it, and you will know when you get caught. We don't do it.

Screen-recording outline (7–9 minutes)

1. 0:00–0:35 — "Never fake a star. Ask after real work. Human sends the reply."
2. 0:35–2:00 — Fill review-rules.md live. Highlight NEVER ASK and NEVER FAKE.
3. 2:00–3:30 — Classifier on three real-looking reviews: praise, wait-time complaint, legal-adjacent. Park the third.
4. 3:30–5:00 — SMOOTH tag on a fake Jobber/Open Dental job. Run the ask prompt. Show the text. Do not auto-send.
5. 5:00–7:00 — Dental or HVAC reply prompt on the complaint. Edit out "delighted." Post as a human (or show the GBP reply box and stop).
6. 7:00–8:00 — Month scoreboard. Close: "free Mini Margin Check. Bring a week of real reviews. No fake-review machine. No free listing audit."