Google Ads has a terrible interface, so I stopped using it

For weeks, Google thought every sale I made was worth $1. I only found out because I'd stopped using the Google Ads dashboard.

If you've ever tried to answer a simple question in Google Ads, like which searches actually produced a customer last week?, you know why. Six menus, three date pickers, a report that doesn't match the other report, and a banner urging you to raise your budget.

So in July, running my first campaigns for The Modern Shrine, I connected Google Ads to Claude and made it my ads manager. Instead of clicking, I typed what I wanted:

• "Duplicate this campaign for the Southeast Asia market."
• "Test this offer against that one."
• "Put more budget behind the keywords that convert."
• "Add negative keywords for the searches that waste money." (Negative keywords tell Google which searches not to show your ad for.)

Claude did the work through Google's official connection for software (its API), showed me what it found and what it changed, and I made the decisions. Typing a sentence and getting a result beat an afternoon of clicking.

What the first campaign showed

The first campaign ran for 11 days on a budget of SGD 20 a day. Final numbers, pulled once late conversions had settled:

Spend	SGD 443
Clicks	199 (7.9% click-through rate)
Free-report sign-ups	23
Cost per sign-up	SGD 19.28

The interesting part was which searches converted. I'd spent weeks debating whether to market the product in plain "career test" language or lean into its traditional Chinese framework. The data settled it in days. About 60% of conversions came from generic career searches and 40% from searches for Bazi (Chinese astrology) and astrology, and the cost per conversion for the two themes was almost identical: SGD 18.15 against SGD 18.39. The best single keyword, "bazi chart analysis", converted at SGD 12.29.

So I kept both. The market answered a question I couldn't answer by thinking about it.

Why the dashboard was the problem

None of those answers were hidden, exactly. They were just expensive to reach. Each one meant building a custom report, exporting it, and cross-checking it against my own app's analytics, because what Google calls a "conversion" and what my business calls a customer are different things.

With the API connected, the same work became a conversation:

• "Pull search terms for the last 14 days with cost, clicks and conversions. Which terms spent money and never converted?"
• "Compare these two campaigns. Their settings are identical, so why is one converting at three times the cost of the other?"
• "List every setting that tells Google a sale happened, how each is set up, and whether it has ever actually counted one."

The last question found the most expensive problem.

What the interface was hiding

Google Ads relies on a few settings that tell it when a sale happens. Two of mine were broken. In the dashboard, the only sign was a small grey "Misconfigured" label that was easy to miss. Asking Claude to list them made it obvious: one had never counted a single sale, and another was set to count every plan sale as a flat $1, whatever the customer actually paid.

Google's automated bidding optimises towards whatever you tell it a conversion is worth. For weeks I had been telling it my sales were worth a dollar.

We replaced the broken actions rather than trying to repair them, created proper sign-up and purchase tracking for all four products, and shipped the matching code the same day.

Nine campaigns, and what they taught me

With tracking fixed, I expanded to nine campaigns: four products (career, wealth, relationships and parenting), each in two markets. A few findings I would never have found by clicking through menus:

• Identical settings, three times the cost. Two parenting campaigns had the same bids, budgets and keywords. One converted at SGD 8.41 and the other at SGD 26.26. The search-terms report showed why: in one market the astrology keywords were attracting people with no interest in parenting at all. That cluster took 63% of the clicks and spend and produced one conversion.
• Copied settings can switch a campaign off. A new campaign inherited a day-of-week schedule from an older one and delivered nothing for six days, because the days it was switched on weren't in its allowed window.
• Missing tags cost a week of data. The new campaigns were built without the tracking tags the first one had, so 166 assessments from their first week can never be attributed to a keyword.
• Relationships were the surprise winner, at SGD 4.16 to 5.64 per conversion in their first week, and they produced most of the week's checkout clicks.

I also kept a log that ties every batch of results to the exact version of the website that was live at the time. That way a result can't be misread as "the new keywords worked" when the real change was that the checkout page had been updated underneath the campaign.

The bigger shift: from clicking through software to talking to it

Google Ads isn't unusual. Your accounting software, CRM, booking system, email platform and ad accounts are mostly the same thing: a useful database behind an interface designed for the average user, which is nobody.

More and more of these tools now let an AI assistant plug straight in, so you can skip the interface entirely. Instead of learning where a vendor hid the report, you say what you want in a sentence. The assistant does the clicking and fetching, and you stay on the part that matters: reasoning about the results and deciding what happens next, while the work is being done.

That's the shift under way right now, from software you operate to software you talk to. For a small business it changes three things:

1. Questions get cheaper. When a question takes 30 seconds instead of 30 minutes, you ask it weekly instead of never.
2. Cross-checking becomes normal. The biggest problems live between systems, such as ad conversions against real payments, and a dashboard can only show its own side.
3. Configuration errors surface. Reading settings as plain text exposes problems that a small grey label in a busy interface will always hide.

The guardrails I'd insist on

Giving an AI access to a live ad account is giving it access to your money. What I'd recommend to any business owner:

• Start read-only. Let it pull and explain data for a few weeks before it's allowed to change anything.
• Approve every change. Budgets, bids and pauses should be proposed, shown to you, and applied only after you say yes.
• Keep a written log of what changed, when, and why, so you can connect results to decisions later.
• Check the platform's numbers against your own. Google is not a neutral scorekeeper of Google's performance.

These are the lower rungs of what I call the AI Trust Ladder: read, then suggest, then draft for approval, and only then act within limits.