The most useful thing I built this year is also the most boring, and I mean that as a compliment. It doesn't photograph well, and there's no demo where a robot does something clever and everyone nods along approvingly and clapping. It just sits off to the side and stops money falling down a hole you never knew was there—which I honestly figured out the hard way, which is also why I say everything I do is "battle-tested", because it literally is.
What happened was this: I put the strivefortone search campaign on broad match to see what it would do, and it bit me. Burned through budget pulling in AI SaaS product names, ecommerce gurus, competitor agencies, and the general churn of people who were curious but were never going to buy a thing. Useful lesson. Annoying receipt. And the trouble with a search-terms report is that by the time you're reading it the money's already spent, so the report is really just the itemised proof of a decision nobody got to make.
So I built the thing that should have been watching while it happened. A negative keyword guardrail, which is a deeply unglamorous bit of plumbing that watches the search terms and asks one fairly simple question on your behalf—should this query ever be allowed to cost money again?—and then puts the call in front of a person before anything changes in the account.
The problem isn't that the platforms are stupid
It's that they're hungry. If you run Google Ads, especially on broad match, you already know the feeling: you sell one thing, Google discovers a wildly generous interpretation of that thing, and before long the budget's going on adjacent nonsense, competitor names, research queries, software tools, people hunting for jobs, and the occasional search term so far off the mark you sit there staring at it like it has personally betrayed you. The platform will find traffic, and it's good at finding traffic, and whether any of it is useful to your business is the bit it would rather get to later, in a quieter room, once the spend's already gone through.
A guardrail's whole job is to catch that earlier—not in the monthly report, not after the fact, but early enough that you can look at a cluster of waste with the evidence attached and say, plainly, no, not that. It doesn't need to be dramatic. It needs to check the account, group the rubbish, and put the call in front of someone who can actually make it.
What it looks for
- product categories you don't sell
- people after free tools, templates or tutorials
- competitor terms you've no interest in chasing
- locations well outside your service area
- research intent when what you need is buying intent
- brand names that look relevant to a machine and are obviously wrong to a human
The real trick isn't spotting the individual bad term, it's seeing the pattern, because one junk query is a flyspeck and fifty related junk queries are the platform quietly telling you it's found a brand new way to spend your money. That's the bit worth knowing before it compounds, and it's exactly what bit me on strivefortone—not one weird search, but a whole adjacent category the machine had decided I was in.
Why this is a good fit for AI
A spreadsheet can show you the search terms and a human can read them, sure, but the work is repetitive and judgement-heavy and exactly the kind of thing that's easy to keep putting off until next week, which is how the leaks become permanent. AI is good at the first pass here because it can read the language, infer the intent and group the related terms far faster than a tired operator does on a Friday afternoon—and then the human still makes the actual call, which matters, because negatives applied carelessly will block traffic you wanted. It recommends; it doesn't go and mutate the account on its own. A good guardrail doesn't make the decision for you. It just gets the expensive nonsense into your line of sight before it hardens into a habit.
If this is your problem
This kind of thing is usually a bespoke guardrail build, scoped to your account and your data, from $5,000 — the shape depends on how many accounts you're running and which failure modes are actually costing you. If you'd rather start smaller, a $2,500 roadmap sprint is the low-commitment way to map where the leaks are and what's worth building first. And because the real value of a guardrail comes from it watching continuously rather than once, it tends to live inside the Intelligence Retainer at $3,000/month, feeding what it finds into the 06:00 briefing so the waste turns up next to everything else worth your attention. Either way, the first version recommends changes for approval — it doesn't quietly mutate the account. Full pricing's on the AI implementation page.
That's the kind of AI I keep coming back to. It isn't trying to replace the strategist, it's just handing them a better morning. None of it is spectacular. It just keeps the money where it's meant to be, which on most weeks is the more useful thing to have—and I'd know, because I'm the one who watched a week of it disappear before I went and built the fix.