The first thing I check when an owner asks about AI reputation management for a service business is never the star rating. It's the gap between two timestamps: the moment the tech marked the job complete, and the moment the review request went out. At a four-truck plumbing shop I walked through last year, that gap averaged nine days. Nine days isn't a review request. It's a stranger texting a homeowner about a water heater they stopped thinking about a week ago.
That shop had 61 Google reviews and a 4.3. They were not bad plumbers. They were bad at asking.
What follows is how this gets built when it works — the trigger, the guardrails, the platform rules that will genuinely get you in trouble, and the two narrow places where AI earns its spot. It is less impressive than the vendor demo and considerably more durable.
The window is about ninety minutes, and nobody puts it on a calendar
A homeowner's memory of your tech has a half-life. At ninety minutes out, they remember the boot covers, the fact that he explained the pressure reading, and that he swept up. At nine days out, they remember that it cost money.
So the ask goes out shortly after the truck leaves. Not while the tech is standing in the kitchen — that produces a socially coerced four-star review that says "nice guy" and nothing else, and Google's relevance signals get nothing from it. Not the next morning either, when it competes with school drop-off and inbox triage. Late afternoon or early evening the same day, after the tech has left and before dinner, is where I see the response curve peak.
The harder question is what event fires it. Most owners wire the automation to "job complete." That's the wrong status. In every field service platform I've worked in — Housecall Pro, Jobber, ServiceTitan, Workiz — "complete" is the button a tech taps in the driveway before he pulls out. It fires before the invoice is settled, which means you occasionally ask for a five-star review from a woman who is at that moment on the phone with your office arguing about a $340 line item she didn't expect.
Fire on invoice paid. Payment posting means there's no open dispute. It's a slightly noisier signal — some memberships and net-30 commercial accounts break it — but for residential service it's the cleanest proxy for "this job is actually finished" that your system already tracks. If you want the mechanics of sequencing and message copy, I've covered timing a review request without sounding spammy in more depth elsewhere.
The caveat that costs people money: automation doesn't fix a broken process, it makes the broken process punctual
If your techs batch-close jobs at 9pm from the couch, your "90 minutes after completion" automation is a 10:30pm text. If your office closes out jobs on Fridays, every customer from Monday through Thursday gets asked at the same moment, which is exactly the burst pattern that trips Yelp's filter and looks strange to a human reading your profile. Fix field discipline first. The automation broadcasts whatever your process already is, on time, at scale.
The line you cannot cross: gating
Review gating means selectively asking for reviews only from customers you expect to be happy. The classic implementation is the "How did we do? 1–5" survey that routes 4s and 5s to your Google link and routes 1s, 2s, and 3s to a private feedback form. Vendors still sell this. Some of them still call it a feature.
It violates Google's prohibited and restricted content policy, which bars both gating and offering incentives in exchange for reviews. And since the FTC Rule on Consumer Reviews and Testimonials took effect on October 21, 2024, suppressing negative reviews while advertising your rating carries civil penalties that adjust annually for inflation and currently run north of $50,000 per violation. The FTC's Fashion Nova action — $4.2 million in January 2022, for blocking reviews under four stars from appearing on its own site — was the warning shot before the rule existed.
Here is where owners get confused, so let me draw the line precisely:
- You may vary sequence. You may not vary selection. Every completed job gets asked.
- If a job had a callback, a warranty issue, or an invoice adjustment, a human calls the customer first and resolves it. Then the review request still fires. You are not filtering who gets asked — you're making sure the job is genuinely finished before you ask.
- A gift card drawing for "anyone who leaves a review" is still an incentive under Google's policy, even though it's sentiment-neutral. Don't.
- Your own employees reviewing your business without disclosing the relationship is expressly covered by the FTC rule now. So is buying negative reviews of a competitor.
If you cannot survive being asked honestly, the reputation isn't the problem.
Where AI actually earns its keep: triage and reading, not volume
The volume part of this — send text, wait 48 hours, send one email nudge, stop — is plain automation. It has been possible since 2014. Nothing about it is intelligent, and any vendor selling "AI review requests" is selling you a cron job with a personality.
Three places the model is genuinely doing work:
1. Cross-platform sentiment triage. A one-star review is a live wire for roughly the first day. A classifier watching Google, Facebook, Angi, Nextdoor, and your manufacturer directories can push a 1–3 star to the owner's phone within minutes, with the matching job record attached — tech name, job type, invoice notes, call recording link. The difference between responding in two hours with the facts in hand and responding in four days with an apology template is the difference between a customer who edits their review and one who tells the story at a barbecue.
2. Reading the whole corpus at once. This is the most underused and, in my experience, the highest-value use of AI in this entire workflow. Export eighteen months of review text and ask a model to cluster it by technician name, job type, and failure mode. One roofing company I know of found "nails left in the driveway" surfacing across eleven reviews and two specific crews. That is not a reputation problem. That is a magnet-sweep SOP and a five-minute crew meeting. The reviews were already telling them; nobody had time to read four hundred of them. AI is a very fast, very patient reader — that's the actual product.
3. Drafting responses, behind an approval gate. Which brings us to the part everyone hands to AI first and should hand over last.
Responding: the failure mode is confident invention
Google's own guidance on improving your local ranking lists responding to reviews as something that signals you value customers, and prominence is one of the three factors it names alongside relevance and distance. So responding matters. The question is who writes it.
Turn an unconstrained model loose on your reviews and it will thank a customer for the great work "our technician Dave" did — on a job Dave wasn't on, in a month after Dave stopped working for you. It will apologize for a scheduling delay that never happened, in public, permanently. The model is not lying. It's pattern-matching what a review response usually contains.
The fix is boring and it works: feed the model the actual job record and constrain it hard — use no fact not present in the record; if the record doesn't contain it, omit it. Then set the gate:
- Five stars, no text: AI drafts, auto-sends. Low risk, low value, but the response rate looks good and it costs nothing.
- Five stars with text: AI drafts, someone skims for thirty seconds, sends. Mention the specific thing they mentioned.
- Anything under five stars: AI drafts, owner or GM reads every word, edits, sends. No exceptions, forever.
For the negative response itself, the shape that works in the trades is short: acknowledge the specific issue, state what you did about it, give a name and a direct line, stop. Do not argue the invoice in public. Never mention that the customer's card declined — it's a privacy problem and it reads as retaliation, and under the FTC rule, pressuring a reviewer over a review is exactly the territory you don't want to be standing in.
The platforms do not play by the same rules
This is where most automations quietly break.
- Google Business Profile: ask freely, every job, no incentives, no gating. This is the main event.
- Yelp: do not ask. Yelp's content guidelines prohibit soliciting reviews outright, and its recommendation software filters reviews that arrive in bursts — a batch of solicited reviews can earn your page a consumer alert. Affirmatively exclude Yelp from the sequence.
- Google Local Services Ads: LSA reviews are a completely separate pool from your Business Profile reviews. You can only request them from customers who came through an LSA lead, using the review link inside the LSA dashboard, and they do not cross over. If you're spending on LSAs and only automating Business Profile requests, your LSA rating flatlines while your map pack climbs — and LSA rank keys off the LSA rating.
- Manufacturer directories: GAF, Owens Corning, Trane Comfort Specialist, Carrier Factory Authorized Dealer. A homeowner comparing three roofing bids often lands there. Nearly nobody automates it. A quarterly manual push of ten happy customers to that profile is worth more than the same ten on a Google profile that already has three hundred.
- Nextdoor: hyperlocal and disproportionately powerful for landscaping, cleaning, and pest control in dense subdivisions, where the recommendation travels by neighbor rather than by search.
The carrier problem: why your review texts vanish
SMS is the channel that works for review requests and it is also the channel that dies silently. If your dashboard shows 100% "sent" and your conversion cratered anyway, you're being filtered.
Business texting in the US runs through A2P 10DLC. You register your brand and each campaign through The Campaign Registry, declare the use case, and submit sample messages plus proof of opt-in. Unregistered traffic gets blocked or throttled by the carriers, not by your vendor — which is why your vendor's dashboard cheerfully reports success. Four things that will get a compliant campaign filtered anyway:
- Public shared URL shorteners. bit.ly and its cousins are shared across thousands of senders, including bad ones. Use a dedicated branded short domain or the full link.
- No business identification in the first message. The first text in a conversation needs to say who you are.
- No opt-out language. Include it.
- STOP-only opt-out handling. FCC rules effective April 11, 2025 require honoring a revocation made by any reasonable method within 10 business days. "Quit texting me" in a reply is a revocation. If your automation only recognizes the literal word STOP, you're out of compliance.
A worked example: four trucks, ninety days
Take that plumbing shop. Four trucks, roughly six jobs a truck per day, five days a week. That's about 120 jobs a week and around 1,560 in a quarter.
Ad-hoc verbal asks — "hey, if you were happy, we'd love a review" — convert in the low single digits, because the tech remembers to say it maybe a third of the time and the customer follows through maybe a tenth of the time. A well-timed text with one nudge converts in the low double digits for residential service. Not the 40% in the demo. Call it ten percent.
Ten percent of 1,560 is roughly 156 reviews in a quarter. They had 61 total, accumulated over four years. The constraint was never demand. It was the ask.
Two things to get right on the way up. First, a burst of five-word reviews is nearly worthless — ask a question in the request ("what did the tech do for you?") so the review names the service and the neighborhood, which is what feeds Google's relevance signal. Second, don't let the star average blind you. Volume and recency relative to the other three businesses in your map pack is the number that matters, not an absolute.
What this will not fix
AI reputation management for a service business is a mirror with a megaphone. If your real service quality produces a 3.9, automation produces a louder 3.9.
And the single biggest driver of one-star reviews in the trades, in my experience, isn't botched work. It's communication — the no-show, the four-hour window that became seven, the callback nobody returned. Which means your phone system is a reputation system whether you think of it that way or not. Before you automate a single review request, look at the real math on missed calls, then at whether your front desk actually catches what comes in. Fix the field, then the phones, then automate the ask. In that order.
Done in that order, this is one of the few places where the ROI is legible enough to defend — I've laid out how I think about proving that in an honest ROI breakdown for service owners, and the broader review and reputation workflow sits alongside it. It's also the kind of build we handle as review automation at Turnkey AI, though nothing above requires us to be involved.
Questions owners actually ask
Frequently asked questions
Can I run a gift card drawing for customers who leave a review?
No. Google's contributed content policy prohibits offering any incentive in exchange for a review, and it doesn't matter that your drawing is open to happy and unhappy customers alike. The incentive itself is the violation. Ask without a carrot.
Should the review request go out by text or email?
Text, with email as a fallback about 48 hours later if the text didn't convert. Text requires A2P 10DLC brand and campaign registration in the US or the carriers will filter it, so budget a week or two for registration before you launch. Email alone will underperform badly.
A competitor is leaving fake one-star reviews. What do I do?
Report each one through Google's review reporting flow and escalate through Business Profile support, and document dates and screenshots. Respond publicly once, factually and without heat: state that you have no service record matching the review and invite them to contact you directly. Do not buy positive reviews to bury it — that is now a federal violation under the FTC's 2024 rule.
Can AI respond to my reviews automatically without me reading them?
Only for five-star reviews with no written text. Anything with text, and anything under five stars, needs a human reading every word before it posts. AI response drafts invent plausible details — technician names, timelines, apologies for things that never happened — and a review response is public and permanent.
Do my Local Services Ads reviews help my map pack ranking?
No. LSA reviews and Google Business Profile reviews are separate pools. LSA reviews can only be requested from customers who came in through an LSA lead, using the link in the LSA dashboard, and they feed LSA ranking only. If you're running LSAs you need a second request path, or that rating stops growing.
Will asking for reviews on Yelp get me in trouble?
Yes. Yelp's content guidelines prohibit businesses from soliciting reviews at all, and its recommendation software filters reviews that arrive in solicited bursts. Solicited Yelp reviews can also trigger a consumer alert on your page. Exclude Yelp from the automation entirely and let it accumulate organically.
How many reviews do I actually need?
There's no absolute number. What matters is your volume and recency compared to the other businesses showing in the map pack for your service area. Pull up the three-pack for your main service and city, look at their counts and their most recent review date, and set your target from there.