It's the 5th of the month. Google just charged the card. The HVAC owner is looking at last month's search terms on a laptop in the office that still smells like solder. The ad says "Trusted local HVAC." The landing page is a handshake photo and "Contact us for a quote." Click-through is fine. Booked diagnostics are not. Someone on a forum told him to "use AI for ads." He is one prompt away from forty headlines that still don't name the $89 diagnostic.
AI will not rescue a weak offer, a landing page that overpromises, or a service nobody asked for. It will help you test faster — if you point it at the parts that actually print money. That's Effectiveness. Ads exist to make money. Not to look modern. If a change doesn't raise booked jobs, new-patient calls, or in-store visits, it's a hobby.
I use AI on ads the same way I'd use a junior media buyer who types fast and has no taste yet: variants, negative keywords, and offer clarity. Then a human still ships the customer-facing version. You. Your office manager. Your agency, if you have one. The model drafts. It does not hit publish on a promise you can't keep.
The offer is the ad
Before you generate forty headlines, write the offer in one sentence a tired person can repeat. HVAC: "Same-week AC diagnostic, $89, waived if we do the repair." That's an offer. "We care about comfort" is a vibe. Google does not pay you for vibes. Dental: "New-patient exam and x-rays for $99, evening hours on Tuesdays." That's an offer. "A practice that feels like family" is a poster in the lobby. Fine for the wall. Useless in a 90-character headline. Multi-location retail: "This weekend only: buy two bags of mulch, third is free, in stock at all four stores." That's an offer. "Your neighborhood hardware destination" is what the landlord put on the lease.
If you can't say the offer without a semicolon, the model will invent one. It will invent discounts you don't run, warranties you don't write, and "same-day" you can't staff. Faster creative on a fuzzy offer is how you buy more clicks into a lie.
We write four lines on a pad. What the customer actually buys (the job, the visit, the SKU). What they pay, and what they don't. What happens in the next 48 hours after they click or call. What you will never claim. That last line is the one shops skip. HVAC: we do not promise "new system by tonight" in July. Dental: we do not promise "insurance covers everything." Retail: we do not promise "in stock" if the back room is a coin flip. Put the refusals in writing. Then you can let AI draft variants without it wandering into a one-star review.
Three jobs AI is actually good for
Not twenty. Three.
First: variants of a true offer. You already know the offer. You need a dozen ways to say it for Google Responsive Search Ads, Meta primary text, and the first line of a landing page. Feed the model the four lines plus three ads you were proud of last year. Ask for tight variants, not "creative concepts." Kill anything that adds a claim.
Hypothetical (not a client, not a promise): you run three HVAC trucks and spend $4,000 a month on Google. Your current ad says "Trusted local HVAC" and the landing page is a handshake photo. I'd rather have eight headlines that all name the $89 diagnostic and the waiver than eighty headlines that name your city and "family-owned." The model can spit the eight in five minutes. You still pick three. You still make sure the landing page matches. If booked diagnostics don't move, the offer is the problem, not the adjectives. I will not invent a ROAS from a blog.
Second: negative keywords and junk queries. This is where HVAC and dental money quietly dies. HVAC: you bid on "AC repair." You also pay for "AC repair DIY," "AC repair training," "AC repair jobs," and "how to recharge home AC." Dump last month's search terms into a sheet, feed the junk to a model, get a first-pass negative list in twenty minutes. You still scan it. You still add "DIY," "salary," "course," "parts only" if that's not your customer. The model is a highlighter, not the media buyer of record. Dental: "emergency dentist" is a money query. "Dental school," "dentist salary," and "free dental clinic" are not. Law firms have the same leak: "personal injury attorney" vs "personal injury attorney salary" vs "how to file without a lawyer." Agencies buying ads for clients forget this because the dashboard looks busy. Busy is not booked. I would not let a model auto-add negatives in the Google account. That's a Quality call. One bad negative ("near me," "open Saturday") can mute the people who pay you. Human reviews the list, then pastes it in.
Third: offer clarity on the landing page. This is where most shops waste the creative they just generated. The ad finally names the $89 diagnostic. The page still says "Contact us for a quote" and a form with eleven fields. The visitor was promised a next step. You handed them homework. Use the model to rewrite the above-the-fold block so it repeats the ad. Headline, price, what happens in 48 hours, one button. Then write the "what this is not" paragraph so you don't attract the wrong click. HVAC: this is not a $89 full system. Dental: this is not a full treatment plan. Retail: this is not a raincheck on a SKU you ran out of at noon. If the page still lies, stop generating ads. Fix the page. Screenshot the ad next to the page. If a stranger can't match them in five seconds, you are buying clicks into a story you don't tell.
Faster creative is a trap if the loop is broken
Agencies love this trap. The pitch is "we'll produce more assets." More assets on a dead offer is a more expensive way to learn nothing.
Run this check before you ask a model for images. Does the landing page say the same number as the ad? Can the front desk actually book what you advertised this week, or are you sold out and still spending? Do you have a human who answers in a timeframe a paid click deserves? If you spend $40 to get the call and nobody picks up until voicemail, AI copy will not save you. That's a routing problem. Ads just make the leak louder. Are you measuring booked jobs, consults, or register rings — not likes, not CTR as a vanity score?
HVAC in July: if you cannot staff the diagnostic this week, pause the "same-week" claim. Don't generate twenty new images of sweaty attics. Dental: if hygiene is booked out six weeks, stop advertising a $99 exam you cannot seat. Retail: if the SKU is gone at two stores, geo-split the ads or kill them. Spend only where you can take the money.
Quality sits right next to that. A dental ad that implies sedation is included when it isn't will get you a consult and a bad Google review. A law-firm ad that sounds like a guarantee will get you a bar problem. An HVAC ad that says "we beat any quote" when you don't will train the neighborhood to use you as a free bid. The model will happily write all three. That's why a human still sends the customer-facing message. Always.
How I'd actually run a week of this
Not a strategy offsite. A week. Monday: owner writes the four lines. Offer, price, next 48 hours, never-claim. Office manager adds the three phrases customers actually say on the phone. HVAC: "it's 82 in the house and the unit is blowing warm." Dental: "I haven't been in three years and I'm nervous." Retail: "I don't want to drive across town if you don't have it." Tuesday: dump last 30 days of search terms and last 10 ads into a folder. Ask the model for headline variants that stay inside the four lines, a negative-keyword first pass, and a rewrite of the landing-page hero that matches the winning ad. Do not ask it to "rebrand" you. Wednesday: you and one other human kill anything that added a promise. You paste three headlines. You add ten negatives by hand. You change five sentences on the page. You do not rebuild the website. Thursday through Sunday: watch booked calls and form-to-book, not just CTR. If the page matches and the offer is real, you'll see whether the new copy is doing a job. If nothing moves, the offer is the problem, not the adjectives.
I would not train a custom model on your ads. I would not buy an "AI ads platform" until this loop is boring. A $20 chat tool, a true offer, and a human with login access beats a new vendor dashboard. Agencies running HVAC, dental, or multi-location retail accounts: same week, their four lines. Stop them from asking for "more creative" when the offer is mush.
I will not auto-publish. I will not let it invent star counts. I will not let it clone competitor claims you can't honor. I will not connect it to the ad-account card. Law firms: same rules, tighter. No outcome promises. Human review on every character that hits Google. If your bar or E&O carrier would wince, it doesn't go live.
Paste this with your four lines and last month's search terms:
Stay inside these four lines. Do not add claims.
1. What they buy:
2. What they pay / don't pay:
3. What happens in 48 hours:
4. Never claim:
Shop: [HVAC / dental / retail / law]
Last 10 ads:
Last 30 days of search terms:
Return:
A) 8 headlines and 4 descriptions that stay inside the four lines
B) A first-pass negative keyword list (DIY, jobs, salary, school, parts-only, etc.)
C) A landing-page hero rewrite that repeats the ad (headline, price, 48 hours, one button, "what this is not")
Flag anything that invented a discount, warranty, wait time, or star count. Do not publish.
Filter it one more time. Effectiveness: more jobs, patients, or tickets? Efficiency: true variants and a cleaner negative list without a two-week creative round? Quality: does the customer get what the ad described, in a voice you'd say at the front desk? If a toy fails all three, don't buy it.
FAQ
How should a local business use AI for advertising? Write a true offer in four lines, then use the model for variants, a negative-keyword first pass, and a landing-page hero that matches the ad. A human still publishes.
Why won't more creative fix my ads? Faster creative on a fuzzy offer is more clicks into a lie. If the page doesn't repeat the number in the ad, stop generating headlines. Fix the page.
Can the model add negatives in Google for me? No. It highlights junk queries. You scan the list. One bad negative can mute the people who pay you.
If you want this on your shop, bring one live campaign, last month's search terms, and the landing page that eats the clicks. That's the free 15-minute Mini Margin Check. We write the four lines, kill the lying sentences, and leave with variants you can paste. Paid seat. No free audit of your account.