The 51% I Didn't Move

I spent a fortnight building eight landing pages for a client’s search campaigns. The night the new site went live I repointed eight ad groups at them and left the two biggest ones alone.

Those two carry 51% of the conversions in the account. They still point where they pointed before.

Every instinct in paid media says finish the job. I want to explain why not finishing it was the decision that made the work readable.

The spread that made eight of them obvious

Thirty days before cutover, this account’s conversion rate by destination looked like this:

Destination	Conversion rate
Site root (two ad groups)	11.76% and 11.06%
A legacy blog post	0.88%
A different legacy blog post	1.11%
A condition page	8.65%

A tenfold spread inside one account, and not one of those ad groups had a landing page built for it. The ones at 0.88% and 1.11% were sending paid clicks to articles published in 2010, with a navigation menu, a sidebar, an author byline, and no obvious way to book anything.

Moving those wasn’t a judgment call. There’s almost nothing below 0.88%, and the pages they were moving to were built for the search that triggered them.

Why the two at 11% are a different question

Because they were already the best thing in the account, and because the ground had just moved under them.

The root those two point at is the home page, and the home page is brand new. Their 11% was earned on the old WordPress one. The moment the site went live, their baseline was invalidated whether or not I touched them.

So the choice was between one untested change on 51% of the conversions, or two untested changes stacked on top of each other on 51% of the conversions, with no way afterwards to tell which one did what. If the number went down, I’d have no idea whether the new home page underperformed the old one or the landing page underperformed the home page. If it went up, same problem in the other direction.

There’s also a plain risk argument. A landing page that beats a 2010 blog post is a near-certainty. A landing page that beats a page already converting at 11% is a real question, and I would rather ask it with evidence than with confidence.

The eight are a free experiment

The eight ad groups I moved were the account’s worst destinations. That makes them close to costless as a test: there’s almost no downside left in a 0.88% conversion rate and a lot of headroom above it. And they’re measuring the exact thing I need to know before touching the 51%, which is whether these landing pages convert at all.

So the sequence is cheap evidence first, expensive decision second, with the cheap evidence bearing directly on the expensive decision. Moving everything at once would have spent the expensive decision to learn the cheap thing.

Write the rule before you have the number

The failure mode of holding something back is that the holding becomes permanent. Three weeks pass, the number is ambiguous, nobody wants to move 51% of an account on a feeling, and it sits there forever.

So the decision rule went into the readout script the same day, before any data existed to argue with:

• Moved-set conversion rate clearly above 11%: move the two held ad groups.
• Between 8 and 11%: run them as a 50/50 experiment instead of a switch.
• Clearly below 8%: leave them on the home page, and rethink the landing pages.

It needs roughly 300 clicks across the moved set before it’s allowed to read. That threshold is also written down, because the strongest pull in this situation is to look at the first day of data and see whatever you were hoping to see.

Write the rule while you’re still indifferent to the answer. It costs nothing on the day you make it and it’s almost impossible to do later.

What the client actually asked for

His own framing, before any of this started, was that he didn’t want to look up one day and find a sharp drop.

Holding the 51% is the honest answer to that. Half of what’s working keeps working while the other half proves itself, and when the change does come to the big ad groups it’ll come with a number attached rather than an argument.

I’ve written the rest of that engagement up in the case study, including the part where I publish no performance result at all, because there isn’t one yet.