Claude Ads: Business Owners Retarget Google and Meta in One Sitting

You can build, launch, and iterate a full retargeting campaign across Google and Meta in one sitting, inside a single conversation with Claude, as long as your ad accounts are connected and your tracking is already firing. The payoff is speed and a tighter feedback loop: fewer guesses about creative, faster reads on what worked, and a next round that’s smarter than the last.

TL;DR:

• Before launch, grant AdKit campaign creation access, verify GA4 conversion events and Meta pixel activity, and test account hygiene with a $5 daily campaign.
• Use 30 to 60 day windows for engagement and 3 to 7 days for cart abandoners; exclude recent converters and customers except in upsell campaigns.
• Google search ads cap headlines at 30 characters and descriptions at 90, while Meta copy should put its hook in the first 125 characters.
• Test one variable at a time and set minimum spend or impression thresholds before naming a winner; ask Claude to flag policy risks before deployment.
• Automate low risk asset suggestions, but review budget increases, bidding changes, and broad match keywords; base each next round on one change against preset CPA thresholds.

Table of Contents

• How a conversational campaign build actually flows
• Getting your accounts and data ready before Claude touches anything
• Deciding which visitors to retarget and how far back to look
• Getting Claude to write creative that’s actually platform-ready
• Letting platform automation do the boring parts safely
• Reading results and turning them into the next round
• We started using Claude Cowork with AdKit connected for exactly the reason most owners would: the platform clicking was eating hours that should have gone to strategy. What we found confirmed something we already suspected but hadn’t had to articulate before, the clicking was never the hard part. Knowing what to launch was.
• Why the strategy gap is the real cost center
• Where Rivetline fits once the strategy gets harder than the setup
• FAQ
• Sources

How a conversational campaign build actually flows

Forget the mental model where “launching a campaign” means an afternoon of tab-switching between Ads Manager and Google Ads, copying audience IDs, and re-entering the same UTM parameters for the fifth time. When Claude is connected to AdKit, the whole thing collapses into a sequence you drive with plain language, and the platform clicking happens behind the scenes.

The flow looks like this:

• Define the goal. You tell Claude what you’re trying to do: retarget cart abandoners from the last 14 days with a $50 daily budget split 60/40 between Meta and Google.
• Connect the data. Claude pulls from your already-linked GA4 events, pixel data, and audience lists through AdKit.
• Pick the audiences. You and Claude agree on segments and exclusions in a few back-and-forth messages.
• Generate the creative. Claude writes headline, body, and CTA variants sized for each platform’s limits.
• Deploy. AdKit executes the actual account-level builds, ad group structures, and asset uploads.
• Measure. Claude reads performance from your connected dashboards.
• Iterate. The next round gets built from what the data just told you, not from a hunch.

The division of labor matters here. Claude handles language, structure, and judgment calls you’d otherwise make manually: which audience makes sense, which headline angle to test, when a metric crosses a threshold worth acting on. AdKit handles execution: the actual API calls, the campaign object creation, the asset placement. You still own the strategic calls: budget ceilings, brand voice, which risks are acceptable. Nobody is handing a credit card to an AI and walking away.

This matters because most of the “AI replaces ad management” narrative misses what was actually slow. Clicking through Google Ads’ interface to build a Performance Max campaign was never the bottleneck. Deciding what that campaign should target, and whether it should exist at all, is where hours disappeared. A conversational build doesn’t just save clicks, it forces the strategic decisions to happen up front, explicitly, instead of getting buried in interface defaults.

Getting your accounts and data ready before Claude touches anything

None of this works if your accounts are a mess, and skipping setup is the single most common way people sour on AI-assisted ad building. Do this once, properly, and every future campaign build gets faster.

1. Connect Google Ads and Meta Business Manager to AdKit with account-level access, not just read permissions. AdKit needs the ability to create campaigns, ad groups, and ads on your behalf, which means granting it standard or admin access in Meta Business Manager and the appropriate Google Ads account link.
2. Verify GA4 events are mapped to conversions before you ask Claude to build anything retargeting-related. If “add to cart” and “purchase” aren’t firing cleanly, your retargeting audience will be built on garbage.
3. Check pixel health on Meta. Open Events Manager and confirm your pixel has fired in the last 24 hours with no match quality warnings.
4. Export or confirm CRM audience lists if you’re layering first-party data on top of pixel-based retargeting, and standardize naming conventions now (campaign, audience tier, date) so Claude’s output stays organized across rounds.
5. Set up a UTM template you’ll reuse for every campaign Claude builds, so your dashboard attribution doesn’t fragment into fifty slightly different URL variants.
6. Create an asset folder structure (by product line, by funnel stage) so when Claude asks what creative exists, you’re not scrambling.

On the Google side specifically, know that the Google Ads API’s recommendation service can retrieve, apply, or dismiss optimization suggestions, and some of those require specific fields like asset group info or final URLs to even generate. If those fields are missing from your account setup, Claude’s requests through AdKit may come back empty, not because nothing’s wrong, but because the API has nothing to work with.

Pro Tip: Run your first Claude and AdKit session on a test campaign with a $5 daily budget before you trust it with real spend. You’re not testing whether it works, you’re testing whether your account hygiene is clean enough for it to work well.

Security-wise, treat AdKit’s account access the way you’d treat any third-party tool with spend authority: use a dedicated admin user rather than your personal login, and review connected app permissions quarterly.

Deciding which visitors to retarget and how far back to look

This is the part that actually separates a campaign that prints money from one that burns it, and it’s also the part most owners skip because clicking “create audience” feels like progress even when the audience is wrong.

Retargeting isn’t one audience, it’s a stack of them, each with a different intent level and a different message. When you’re briefing Claude, give it the segments by name, not by vague description:

• Engaged page viewers: people who spent real time on a product or pricing page but didn’t convert, usually the widest, cheapest audience to reach.
• Cart abandoners: added to cart, didn’t check out, the highest-intent segment outside of actual leads.
• Demo request starts: began a form, didn’t finish, which on B2B sites is often more valuable than a cold cart abandoner.
• High-intent event funnels: people who hit multiple key events (pricing page plus a feature page plus a scroll depth trigger) without converting.

Lookback windows should shrink as intent rises. A standard approach is longer windows (30 to 60 days) for upper-funnel engagement audiences and shorter windows (3 to 7 days) for cart abandoners and demo starts, where urgency matters more than reach. Pair shorter lookback windows with tighter frequency caps for bottom-funnel audiences so you’re not hammering someone who just bailed on checkout fifteen times a week.

Exclusion hygiene is where a lot of retargeting budget quietly leaks. Tell Claude explicitly to exclude:

• Anyone who converted in the last 7 to 14 days (depending on your sales cycle).
• Existing customers, unless you’re running a specific upsell or renewal campaign.
• Any suppression list required for policy compliance, especially around sensitive categories.

Once segments and exclusions are set, sequence them. A common structure is tiered: top-of-funnel engagement gets the smallest budget share and broadest creative, cart abandoners and demo starts get the largest share and the most direct, urgency-driven messaging. You can literally ask Claude to “build three retargeting tiers with a 20/30/50 budget split weighted toward cart abandoners” and have it propose the structure for your review before anything deploys through AdKit.

Getting Claude to write creative that’s actually platform-ready

Generic prompting gets generic ads. “Write me some ad copy” produces copy that sounds like every other AI-generated ad on the internet, vague benefit statements, no specificity, nothing a human would actually click. The fix is prompting Claude the way you’d brief a sharp junior copywriter, not a vending machine.

1. Specify the variant count and structure up front. Ask for 4 to 6 headline, body, and CTA combinations, not one “best” version. You want options to test, not a single guess dressed up as a final answer.
2. Give platform constraints explicitly. Google responsive search ads cap headlines at 30 characters, descriptions at 90. Meta primary text runs longer but gets truncated in feed, so ask Claude to front-load the hook in the first 125 characters.
3. Feed it landing page context, not just the product name. Claude writes sharper copy when it knows what the page actually says, what objection it handles, and what the CTA button reads.
4. Ask for image and video concepts, not just copy. Claude can describe creative direction (a before-and-after shot, a screen recording of the product, a founder-to-camera testimonial style) that you hand to a designer or pull from an existing asset library, and it can draft accessibility alt text alongside each concept.
5. Request a naming convention baked into the output (campaign-audience-variant-date) so when AdKit uploads the assets, your reporting dashboard doesn’t turn into alphabet soup three weeks from now.

For test design, treat each variant as a hypothesis, not a coin flip. If you’re testing headline angle A against angle B, keep everything else (image, CTA, landing page) constant, or you won’t know what actually moved the number. Set a minimum spend or impression threshold before calling a winner, Claude can flag when a test hasn’t reached statistical relevance yet instead of letting you declare victory on 40 clicks.

Pro Tip: Tell Claude explicitly to flag anything that might trip Google or Meta’s ad policy review, health claims, before-and-after language, superlatives without substantiation, before it goes into AdKit. Catching a policy flag in the conversation is free; catching it after a rejected ad costs you a day of review time.

If you want a reference point for how campaign architecture should actually be organized before creative gets layered on top, our Meta Ads campaign structure guide breaks down a three-campaign template that maps cleanly onto this kind of build.

Letting platform automation do the boring parts safely

Google Ads’ own recommendation engine is not a villain, but it’s also not a strategist, and treating it as either extreme wastes’ money. The RecommendationService can surface and, for supported types, programmatically apply suggestions covering budget adjustments, keyword additions, text ad changes, and bidding opt-ins like Target CPA or Maximize Conversions.

Here’s the honest breakdown of what’s safe to automate and what isn’t:

• Low-risk to auto-apply: sitelink and asset suggestions, ad strength improvements, minor keyword expansions within tight match types.
• Needs human review: budget increase recommendations (Google’s incentive and your incentive are not the same thing here), automated bidding strategy switches, and any broad match keyword suggestions.
• Worth testing in isolation first: Target CPA or Maximize Conversions opt-ins, since these shift how the algorithm spends your money and the effects aren’t always reversible without a learning period reset.

During the build itself, GenerateRecommendations can be called to request specific suggestion types, like sitelink assets or a Maximize Clicks bidding opt-in, and it returns usable recommendation objects as long as the campaign has the required fields in place: asset group info, final URL, bidding strategy type, and so on for the relevant recommendation type. Ask Claude to pull these recommendations as part of the build conversation, present them to you in plain terms, and apply only the ones you greenlight.

The practical workflow: have Claude read whatever recommendations come back, sort them into “apply now,” “queue for review,” and “dismiss,” and give you the apply or dismiss decision in the same conversation rather than making it for you. Treat every recommendation the API surfaces as a signal worth your attention, not an instruction to follow blindly, because Google’s recommendation engine optimizes for Google’s definition of a good outcome, which doesn’t always match yours.

Reading results and turning them into the next round

A campaign that launches fast and then sits unexamined for three weeks is not actually faster than the old way, it’s just as slow with extra steps. The entire value of this workflow lives in the loop: launch, measure, learn, rebuild.

What you’re watching depends on the campaign’s job. For retargeting specifically, the primary signals are:

• Conversion rate and cost per conversion, benchmarked against your cold-audience campaigns so you know retargeting is actually earning its premium.
• ROAS, where applicable, as a sanity check against the budget split you set up front.
• Audience lift, comparing the retargeted cohort’s conversion behavior against a holdout or against your site-wide baseline.

Wire GA4 and Looker Studio as the single source Claude reads from, the same dashboards you’d check manually, so the agent isn’t working from a different picture than you are. One live dashboard beats five disconnected reports every time, if only because it stops arguments about whose number is right.

One dashboard, read consistently, beats a stack of monthly PDFs that are stale before you open them.

Set thresholds before you launch, not after you see a bad number and start rationalizing it. Define what “scale” looks like (CPA below target for three consecutive days, say), what “pause” looks like (CPA above a ceiling with no recovery trend), and what “refocus” looks like (fine CPA but a sinking audience size that means it’s time to add a new segment). Hand these thresholds to Claude explicitly so it’s evaluating against your rules, not inventing its own.

The iteration instruction set that actually works is three steps: evaluate what happened against the thresholds, hypothesize why (creative fatigue, audience overlap, a landing page problem), and implement one specific change for the next round. Not five changes. One, so you know what actually moved the number.

We started using Claude Cowork with AdKit connected for exactly the reason most owners would: the platform clicking was eating hours that should have gone to strategy. What we found confirmed something we already suspected but hadn’t had to articulate before, the clicking was never the hard part. Knowing what to launch was.

Running a retargeting rebuild for a client account, we described the goal in plain English (rebuild the cart abandoner tier, tighten the lookback window, test three new headline angles against the existing control) and had a deployable campaign structure, audience logic, and creative set ready inside a single working session, where the same rebuild used to take the better part of a day split across two platforms.

The bigger shift wasn’t speed for its own sake. It was that each round got built directly from what the previous round’s dashboard showed us, instead of from an account manager’s gut feeling about what “should” work next. That loop, launch, read the result, adjust one variable, relaunch, is the actual engine behind any ad program that improves over time rather than just spinning in place.

— Chris Breikss

Why the strategy gap is the real cost center

Automation makes execution fast. It does nothing for a marketer who doesn’t know which audience to retarget first or what a winning test actually looks like, and that gap is where most ad budgets quietly bleed out.

Three mistakes show up constantly: treating every platform recommendation as gospel, testing five creative variables at once so nobody can say what worked, and never defining a threshold for pausing a losing campaign before the losing campaign defines it for you.

Conversational tools remove the busywork. They don’t hand you five years of pattern recognition for what a retargeting funnel should look like at your spend level. If that experience gap is the honest constraint, that’s a strategy problem, not a tooling one.

Where Rivetline fits once the strategy gets harder than the setup

Running Claude and AdKit yourself gets you fast, clean campaign builds. What it doesn’t hand you is the experience to know which audience tier deserves the budget first, which test actually matters, or when a “recommendation” from Google is worth taking. That’s the part we run for clients day to day.

Our Meta Ads and Google Ads, LSA & Business Profile teams build and manage exactly the kind of retargeting programs this article walks through, backed by live Looker Studio dashboards connected to GA4 instead of a PDF that’s already stale by the time you open it. The typical path starts with a short audit of your current accounts and tracking setup, moves into a pilot campaign build so you can see the loop in action, and from there into an ongoing retainer once the structure is proving itself.

If you want to try the Claude and AdKit workflow yourself first, you can use marketing apps built from your campaign data to streamline the process with Gainable. And if you’d rather have us run the whole program while you focus on the business, get in touch about a pilot.

FAQ

Can Claude integrate with Meta Ads?

Yes, when connected through a tool like AdKit, Claude can direct campaign creation, audience targeting, and creative deployment on Meta by translating plain-English instructions into the account actions AdKit executes. Claude itself doesn’t hold native Meta API access, the integration layer handles that execution.

Can Claude integrate with Google Ads?

Yes, through the same kind of connector layer, and Google’s own Ads API recommendation service supports retrieving, applying, and dismissing optimization suggestions programmatically for supported recommendation types. Claude can read those recommendations and propose which to apply, but the account-level execution runs through the connected platform.

Is Claude good for creating ads?

Claude is strong at producing multiple headline, body, and CTA variants quickly when you give it platform constraints and landing page context, which is the groundwork for a solid A/B test. It’s less useful without that input, generic prompts produce generic copy, so the quality depends heavily on the brief you give it.

How do I use Claude for ads?

Connect your ad accounts through an integration layer like AdKit, then describe your campaign goal, audience, and budget in plain language inside a Claude conversation. Claude handles the audience logic and creative generation while the connected tool executes the actual campaign build on Google or Meta.

Sources

For the API-level mechanics referenced above, Google’s own recommendation service documentation covers what can be retrieved, applied, or dismissed programmatically. If you want to see the Claude and AdKit connection firsthand, you can try AdKit here. For campaign architecture that pairs well with this workflow, our Google Ads category hub and Meta Ads delivery fix guide cover the operational side in more depth.

• Optimization score and recommendations | Google Ads API | Google for Developers

Recommended

• 5 Quick Checks Marketers Use to Pick Google vs Facebook Ads for ROI
• Google Ads and Local Services Ads Management
• Three Campaign Meta Ads Structure for Agencies: Copy Ready Templates
• Google Ads for B2B: The Playbook That Actually Works