When we build negative keyword lists, it tends to go something like this. You scan the search terms report and ask whether each term belongs. Is it a competitor? A service you don’t offer? Someone looking for a job rather than a supplier? If it doesn’t fit, out it goes.
That’s an in-principle decision, and a perfectly reasonable one. What often gets bypassed is whether each term actually does more harm than good in practice.
Principle and practice don’t always agree. And when they don’t, it’s the data that deserves the final word.
Reading the term, not the results
LLMs make this tendency very visible. Ask one to suggest negatives from a search terms report and it tends to start exactly where we usually do, with relevance.
In one gas safety certificate account, I gave it the search terms data, conversions included, and asked for a list with a deliberately high bar against excluding genuine buyers. It came back with seven negatives. Among them were several British Gas terms, which had converted at an aggregated CPA around 40% lower than the campaign average. It suggested these were probably people looking for British Gas who had landed on the page and happened to convert anyway.
The conversion pattern was real. Rather than questioning the principle, it explained that away.
Of the seven negatives, only one held up once performance was properly checked (a ‘commercial’ term which had a better home elsewhere in the account).
A good proxy
Principle is the natural starting point, and for good reason. It’s intuitive, and it rests on an assumption that’s close to foundational in PPC: that we can read intent adequately from the search term itself.
Most of the time, we can. A judgement made on principle is a proxy for a judgement made on performance, and the correlation between the two is high. Most terms that look wrong are wrong.
But the correlation isn’t perfect. Looking straight at the data closes that gap, provided there’s enough of it to produce reliable patterns.
There’s another in-principle approach to negatives, which pulls in the opposite direction. It holds that negatives now do more harm than good, because machine learning works best with the freedom to find patterns across a wide field of play.
There’s always been some truth in the idea that Google’s algorithms perform better given more room to work with. That doesn’t make the extra room good for the advertiser…
In the same account, search terms containing ‘register’ took around 5% of spend over a year, at well under half the average conversion rate and twice the average CPA. In the latest 30 days, their share had risen to 9%. This is a mature campaign that has had years to learn, and its spend on the term was still going up.
In cases like this, the human still needs to step in, and the negative is the intervention.
Earning a place
‘Performance as well as principle’ is a useful habit well beyond negatives.
Relevance isn’t sufficient. A keyword or search term can sit exactly on target and still fall short on a performance basis.
In a recent post, I looked at a single query reaching one account through thirteen different keywords. The most expensive route by some distance was the exact match keyword added specifically to catch it.
On principle, that keyword belonged. In practice, it was the worst place for the query to land.
So principle is a fine place to start. It’s intuitive, it’s quick, and it’s right most of the time.
But before a negative goes in, or a term gets the benefit of the doubt, look at what it’s actually doing. Where there’s enough data to say, it’s the data that gets the final word.
