I gave three AI engines the same Google Ads brief. Which did the best job?

Why this is a whole new world

You have been able to ask an AI to "write me a search campaign" for two years now. What comes back is an average of everything written online about search campaigns: generic keywords, generic ads, no connection to your business.

What changed is access to data. When the engine reads a year of real search terms from the account, sees what you actually paid per lead, and which ad assets worked and which didn't, it stops guessing. It reasons from evidence. The difference between a generic "write me a campaign" and a campaign built on what people actually clicked and converted on in this account is the difference between a template and real work.

Back to contents

What each engine gave

Reading the data, and the core insight

This is where the gap was widest. Fable didn't start from keywords, it started from what the data shows: the "obvious" head term everyone assumes people search for has almost no real volume. The actual demand was hiding in adjacent phrasings, and in variations the account had barely covered. From there came the conclusion: build the campaign like a funnel, not like a list of words. The exact term at the center, and around it everyone searching for the same intent in a different phrasing.

Opus caught a different angle, no less sharp: the location setting. The difference between advertising to people physically in the target area (Presence) and people merely "interested" in it (Presence or interest). The second setting, which is the default, pulls in irrelevant traffic before you have added a single negative keyword. It's an easy mistake to miss, and an expensive one.

Grok kept a lean line, but got the principle right: separate different search intents and don't mix them.

What to take from this: don't assume you know what people search for. Open the search terms report and look, and check the location setting before you touch keywords.

Campaign structure

All three understood you need one campaign with ad groups split by intent, not a separate campaign per topic. They differed on how many: Fable built seven groups, Opus six, Grok five.

The reason you don't split one campaign into several small ones is worth understanding: Smart Bidding learns from conversions. On a reasonable daily budget, splitting into two campaigns leaves each algorithm with too few conversions a month to learn anything. One campaign, one model, all the conversions flowing into it. You split later, once you have enough data.

Keywords and negatives

On one point all three converged, and it's the approach I use in practice: start with Phrase and Exact, and expand to Broad only after enough conversions have built up, roughly 30 to 50 a month, with a tight negative list.

And here is where you have to be precise, because it's easy to get confused. This is not Google's official recommendation. Google itself pushes Broad with Smart Bidding from the start, because the algorithm weighs intent and the account's history. My logic is the opposite: as long as there aren't enough conversions for the algorithm to learn from, uncontrolled Broad burns budget on queries you never wanted. So you start narrow, prove what converts, and only then open up.

And it's worth looking ahead now. Google is gradually moving parts of Search into AI Max, and DSA campaigns, for example, are set to auto-migrate to it in February 2027. AI Max leans on exactly that, Broad plus Smart Bidding plus your data, so "accumulate clean conversions, then expand" is also the right preparation for what's coming.

The nicest move showed up with Fable: it recommended adding negative keywords to the other campaigns already running in the account too, so they wouldn't compete with the new one for the same queries. Two campaigns running on the same word drive up each other's price, and a seasoned manager thinks about the whole account, not just the campaign in front of them.

Responsive search ads (RSA)

In an RSA, Google assembles the headlines itself. You can "pin" a headline to a fixed position, but do it sparingly, only when a message truly has to appear in a specific spot, like the brand name in the first headline. And if you do pin, it's better to pin two or three headlines to the same position rather than one, otherwise the ad turns static and loses its testing advantage. Read Ad Strength as a compass, not a sacred score: the quality of the message matters more.

And one more constraint I set in the prompt: the ad may not promise anything that isn't on the landing page. Google penalizes a gap between ad and page, and the visitor simply bounces. Message match between the ad and the page isn't a nuance, it's what holds up ad quality and conversion rate.

Assets and conversion improvers, and the mistake

Here two opposite things, and both came from the same engine, Opus.

The smart move: Opus noticed the account was running conversion actions unrelated to this campaign, and that Smart Bidding might learn from that noise. Its recommendation was to narrow the campaign's conversion goals to the relevant actions only. Correct thinking. The algorithm learns from whatever you feed it, and if you feed it garbage, it learns garbage.

The mistake: that same Opus recommended adding a sitelink pointing to WhatsApp. That recommendation violates Google's policy: sitelinks must lead to the advertiser's own site, or to a narrow list of approved third-party destinations that doesn't include it, otherwise they get disapproved. And beyond policy, it's an operational mistake too: a redirect like that pulls the visitor out of the funnel into a chat you don't measure or control. That's exactly the example behind the second iron rule: a recommendation that sounds reasonable, phrased with confidence, and is simply wrong. Ship it AS IS and you find out when Google disapproves the asset.

Honesty about what can't be known

One sign that Fable's output was serious: it stated outright what it couldn't know from the account alone. The real volume of keywords needs a check in Keyword Planner. Seasonality needs a wider view than the campaign's own numbers. Instead of sounding certain about everything, it said "this far the data goes, from here you check." An engine, like a manager, that draws a line around what it knows is worth more than one that always sounds sure.

Back to contents

The ranking: who won, and why

First place, Fable 5.1. The deepest and most grounded. It opened from the data, built a funnel instead of a list, thought about the whole account, and honestly flagged what it didn't know. If I had to hand one of these outputs to a client as a starting draft, this is the one.

Second place, Grok. The leanest and fastest to run. It got the principles right, with fewer moving parts. The downside: it threw out a performance figure without showing where it came from. It's easy to take a number like that as fact, so always ask to see the source.

Third place, Opus 4.8 Max. It actually had the sharpest strategic ideas, the location angle and cleaning up the conversions, but that same engine also returned the longest output. The most interesting thinking, and the one that needed the most correction.

Back to contents