Have Your Agent Write Google Ads Copy

You have new Google ads written from the search terms that actually convert in your account, instead of from your gut. The agent keeps the hard character limits while doing it and puts nothing live before you approve the new ad yourself.

In the previous tutorial you cleaned up: you stopped search terms that only cost money. Now you take care of the other side: for the searches that do bring something, your ads should look good too. The problem with writing them yourself usually isn't the idea but the tightness. A Google ad (a Responsive Search Ad, RSA for short: you supply several headlines and descriptions, and Google mixes the best combination per search) gives you only 30 characters per headline and 90 per description. English is roomier than German, but a phrase like "project management software" still eats up the space on its own. By the end of this tutorial you'll have a full set of ad copy, derived from real account data, checked against the character limits, and sitting in the account as a paused ad: ready for approval, whenever you want.

What you need

• Google Ads connected, with write access (required). Reading happens for the suggestions, writing for the new ad: for that your workspace needs write access (write), not just read access (read). That's controlled by your role in the portal (app.ads.honeyfield.at). To connect the account, see Connect Marketing Accounts to Claude; for ChatGPT, Connect Marketing Accounts to ChatGPT shows the same setup.
• Conversion tracking set up (recommended). The best raw material for good ads is the terms that actually lead to inquiries for you. Without conversion tracking you only see clicks, not which of them convert. Without tracking, skip the conversion column in step 1 and derive the ad from the most-clicked terms plus your best running ad.

This builds directly on the cleanup from the last tutorial: an account without the worst budget-eaters gives cleaner signals about which terms deserve a better ad.

1. Read what works today

Before the agent writes anything, it looks at what your account already knows: which search terms convert and how your existing ad is worded. The new copy comes from this data, not from imagination.

Show me the search terms for [workspace] in the ad group [ad group]
from the last 30 days that converted. And read me the ad currently
running there, with its headlines and descriptions.

What you get back:

Converting search terms (ad group "pm-tool", 30 days)
  project management tool          48 clicks    6 conversions
  manage team tasks                30 clicks    4 conversions
  project planning software        22 clicks    3 conversions
  gantt chart tool                 12 clicks    0 conversions

Current ad (status: active, Approved)
  Headlines (15):   "Project Management Tool", "Manage Team Tasks", ...
  Descriptions (4): "Plan projects in one place ...", ...

From this the agent reads off your customers' real language: they type "project management tool" and "manage team tasks," not marketing jargon. Those become the angles of the new ad.

✓ Checkpoint: You have a handful of converting terms and see what the running ad says today.

Doesn't work? If all terms show 0 conversions even though inquiries are coming in, conversion tracking is usually missing: then use the most-clicked terms as the base and say so ("use the most-clicked terms, tracking isn't active yet").

2. Have new headlines and descriptions suggested

Now you have the new copy built. What matters is that you know what a good ad looks like: it needs many different headlines (Google recommends 15 headlines and 4 descriptions) so Google can test combinations. Explicitly demand several variants per angle.

From these terms, write me a new ad for [ad group]:
15 headlines and 4 descriptions, mixed from keyword reference,
benefit, call to action, and a clear difference from others.
Keep the Google limits (30 characters per headline, 90 per
description) and show me the character count on each line.

What you get back:

Headlines (ad group "pm-tool"):
  Project Management Tool            (23) ✓  [Keyword]
  Manage Team Tasks in One Place     (30) ✓  [Benefit]
  Up and Running in 5 Minutes        (27) ✓  [Benefit]
  Start Your Free Trial              (21) ✓  [Action]
  No Credit Card Required            (23) ✓  [Action]
  Project Management Software Trial  (33) ⚠ too long → "PM Software Trial" (17)
  The Best Tool for Teams            (23) ⚠ UWG: unproven superlative
                                          → "Built for Whole Teams" (21)
  Over 2,000 Teams Use It            (23) ⚠ only with real proof [Social Proof]

Descriptions:
  Tasks, deadlines, and your team in one place. Try it free, no credit card.  (74) ✓
  Plan projects, assign tasks, and keep every deadline in one clear view.
  Start free today.                                                           (89) ✓

The agent does two things here that easily slip when you type them yourself. First, it counts every line: "Project Management Software Trial" has 33 characters, three too many, and gets shortened to "PM Software Trial." That's exactly where ads run aground: a descriptive phrase runs a few characters over the limit. Second, it flags advertising law: "The Best Tool" is a superlative that, under German competition law (UWG), you have to be able to prove, or it doesn't belong in the ad. Without proof the agent rewords it or drops the line. Numbers like "2,000 teams" only if they're true and provable.

A quick word on advertising law (not legal advice): Superlatives ("best," "No. 1," "test winner") and specific numbers are things you have to be able to prove, otherwise they don't belong in the ad. If prices are in there, a "from" price has to be the real entry price and the total price has to include VAT. When in doubt, the agent phrases a benefit as a statement instead of a superlative.

✓ Checkpoint: You have around 15 headlines and 4 descriptions, each with a character count, no "⚠" left open (or deliberately decided that way).

Doesn't work? If the suggestions read like paraphrases of each other, demand more range: "Every headline should have a different angle, no repetitions." Redundant headlines take away Google's room to test.

3. Have it dropped into the account as a paused ad

Once the text works, you have the ad created. The agent writes nothing on its own: it shows you the finished ad as a preview first and waits for your okay.

Create this ad in the ad group [ad group]. Show me the
suggestion first, and create it paused, not live right away.

What you get back:

Suggestion: new Responsive Search Ad
  Ad group:          "pm-tool" (campaign "client-a Search")
  Headlines:         15 (all checked at ≤ 30 characters)
  Descriptions:      4 (all checked at ≤ 90 characters)
  Target URL:        your existing landing page
  Status on create:  PAUSED (won't start running automatically)

Run it? (yes / edit / no)

Only after your "yes" does the agent write the ad into the account. What matters here is your addition "paused": only then does the agent create it paused instead of serving it right away. So always say it explicitly in the prompt. A paused ad serves nothing until you set it active. Your existing ad keeps running unchanged in the meantime, so nothing surprising happens in delivery.

⚠ Don't forget "paused." New ads don't go into the account paused on their own: if the word "paused" is missing from the prompt, the ad can be served immediately. Half a sentence is enough ("create it paused"), and then you decide yourself when it goes live.

✓ Checkpoint: The agent reports "created (paused)" with an ad ID, and the old ad keeps running.

Doesn't work? If the agent reports that it isn't allowed to write, the workspace is missing write access for Google Ads (read instead of write). You change that in the portal (app.ads.honeyfield.at), not in the prompt.

4. Review and activate

The ad sits paused in the account and additionally goes through Google's own ad review. Only when both are fine do you set it active, in a deliberate, separate step.

Has the new ad been approved by Google yet? If so,
set it active, but show me what you're activating first.

What you get back:

New ad (ad group "pm-tool")
  Google review: Approved
  Status:        PAUSED

Activate? After that it runs in delivery. (yes / no)

After your "yes" the ad is set to active. The approval is deliberately a separate step: the agent never goes live on its own. If you'd rather click in the Google Ads account instead of prompting, that works just as well, the effect is the same.

Only once the new ad is approved and getting delivery do you pause the old one ("now pause the previous ad in this ad group"). That way there's never a gap with no ad. A word on directly replacing an existing ad: the new ad takes over the status of the old one. If you replace a running ad, the new one is live immediately, without the paused arc from step 3. When replacing, say that the old one should stay paused rather than removed: then the step stays reversible, nothing gets deleted.

✓ Checkpoint: The new ad is "active" and "Approved," the old one "paused."

Doesn't work? If the review still says "Under review," that's normal: Google reviews new ads, which can take from minutes to a day. Wait for approval before you pause the old one.

Variations

• "Give me five more headlines just for the price angle." (deepen one angle on purpose)
• "Build the ad for ad group [X], but in a more matter-of-fact tone, less promotional." (steer the tonality)
• "Write me the ad, but don't create anything yet, I want to read over it first." (suggestion only, no write)

Honest limit

• No click or revenue guarantee. A cleanly built ad is the better starting position, not a promise of more conversions. Whether it runs better only the real test in the account shows.
• You only see Ad Strength in the Google Ads account. The agent builds by best practice (enough headlines, variety, keyword reference), but no tool can read out the Strength value (Poor to Excellent). If someone promises you "I'll lift your Ad Strength to Excellent," that's made up.
• Google decides on approval, not your agent. Whether and how fast an ad is approved is up to Google. Until then it doesn't serve.
• The agent checks the character limits itself. The write tool doesn't enforce the 30/90: if a line is too long, it's Google that rejects it later. So watch the character count in the suggestion before you say "yes."
• The burden of proof stays with you. Superlatives and numbers only with real proof. The agent blocks unproven claims, but whether "2,000 customers" is true only you know.
• A swap resets the learning phase. If you replace an existing ad, that's a new ad to Google: its optimization starts from zero. The old one gets paused in the process (not deleted), so give the new one some time.
• New ads aren't automatically paused. Without "paused" in the prompt, a new ad can be served immediately. Don't rely on an automatic paused state, say it in the prompt.
• No automatic safety net. No server-side limit stops you. The agent showing the suggestion before writing and asking before activating is its behavior (in Claude it shows the suggestion, in ChatGPT it depends on the connector's permission setting). Whether writing is allowed at all is decided by your workspace access (read vs. write). Read the suggestion before you say "yes."

What's next

Your better ad is now running live. Next you'll want to see whether it moves anything, without checking the account every day: in the next tutorial you have the agent build a weekly report that sums up ads, costs, and conversions at a glance. → Have Your Agent Build a Weekly Marketing Report