How I think about paid ads

A campaign has to make enough money to fund the next round. That’s the whole game. Ecommerce works on paid because purchase margins recycle immediately: what you sell today buys tomorrow’s ads. Subscriptions work when the first month covers acquisition and the rest is profit. If your monetization is weak, you’re throwing money at empty leads; paid ads as a one-time push rarely works.

There is one deliberate exception: venture-funded, winner-take-all markets. Uber could overpay for early riders because the point was to kick-start a flywheel, not to break even on a cohort. If that isn’t your situation, your loop has to close on its own.

So before you write a single ad, do the math on whether the loop can close: break-even ROAS is the loop stated as a number. Business model comes before ads. If the loop can’t close, the fix isn’t better creative; it’s the business.

That’s also the honest answer to “should I run ads at all.” If you can’t say what a customer is worth and how the money comes back, paid traffic doesn’t fix a weak model. It just makes the weakness show up faster and at higher cost.

The offer is what you sell. A paid campaign is not an experimentation ground for it. Angles, the ways your offer meets a real pain, are where all the work is.

I learned the difference as the client, expensively. At a homebuying platform I led growth for, we hired an agency that spent six figures of our budget on experiments. The experiments optimized for CTR, not leads, not qualified buyers. A copy variant would “win” and tell us nothing about which angle actually sold homes. That budget bought us zero learning, because the tests were never designed against the goal. An experiment that answers “which ad gets clicked” instead of “which angle produces buyers” is money spent on vanity: a good CTR is not learning, and if you’re hiring an agency, this failure mode is the first thing to screen for.

Here is what an angle change looks like with the offer held fixed. We offered homebuyers a $10,000 rebate. It underperformed, because the rebate wasn’t the job. When we reframed the same money as help with your down payment (the thing the rebate actually accomplishes for a buyer), it worked. We also dropped “AI-powered home buying” feature-speak for outcome language: find them a better home. Same offer, same dollars. Different angle, different result.

The lesson from both isn’t “never hire help.” It’s that experiments must be designed against the end goal you are optimizing for. If the goal is qualified buyers, a test that can’t tell you which angle produces buyers is not a test, whatever its dashboard says.

Take a watermelon juice company. Sold as a fun drink for kids, the buyer is unaware: no parent is searching for it. They have to discover it while scrolling. That’s Meta and TikTok. Sold as hydration for fitness geeks, the buyer is problem-aware and actively searching. That’s Google.

Search costs more and converts better; discovery is cheaper and colder. Neither is “the good channel”: an expensive click from someone already looking and a cheap impression on someone who didn’t know they had the problem are different purchases, and which one scales is an economics question. The answer depends on the angle you chose.

So the order matters: pick the angle, the angle tells you how aware your buyer is, and the awareness stage picks the channel. Not the other way around.

Marketing used to be four rights: the right message, to the right people, at the right time, in the right channel. The platforms ate the middle two. Meta’s models now decide who sees your ad and when. You don’t.

That wasn’t an accident. Platforms took real heat for offering minute-level targeting, so they hid it. The models are still powerful; the controls are gone. Interest stacking and manual audience building, the old craft, are the part of the job that no longer exists. Let Facebook do its thing.

Your leverage moved to the two rights you still own: the message and the channel.

No amount of variations will save a bad angle. If people aren’t interested in drinking watermelon juice for hydration, showing that same juice fifty different ways won’t work. The media buyer’s real job today is angle discovery: finding the framing of your offer that a specific buyer actually cares about.

To be precise about the word: an angle is not a headline. It’s the pairing of a real pain with the outcome your offer delivers. Kids’ fun drink and fitness hydration are two angles on one juice. A $10,000 rebate and help with your down payment are two angles on the same money, and only one of them worked.

Test angles first; variations come after an angle proves out. Run it backwards (ten variations of one unexamined angle) and you’ll conclude “Facebook doesn’t work for us” without ever having tested an idea.

You don’t have to guess angles from zero. Meta’s Ad Library shows every ad an advertiser is running, and it leaks two honest signals: ads with many variations (proof the advertiser found an angle that converts and is scaling it) and the longest-running ads, because nobody keeps paying for losers.

Read the signal, not the creative. Copying an ad gets you a worse version of someone else’s brand; reading which angles a market keeps funding tells you which pains buyers have already voted for with money. I wrote a full walkthrough of the Ad Library method that’s free and needs no account.

Once the angle is chosen, the ad has one shape: name the pain, position the product against it, use your angle to state the outcome, say how the product achieves it, and give one simple action.

Then run it through the stranger test. An ad is a promise. A known brand is a trusted friend: loose promises land. A new brand is a stranger on the street: buyers need specifics. Exactly how will you deliver? What if you mess up? What guarantee? Write the ad that answers the skeptic.

Specificity does double duty, because creative is the targeting now:

The test is one second long: if someone in your target market glances at the ad and it speaks to them, you did your job. If it’s not clear it’s for them, you didn’t.

An angle test needs enough budget to mean anything. If your target is $100 per lead, spend $300-400 on an angle before judging it. Results are lumpy: $100 can buy zero leads and the next $100 buys three. Leave room for the average to show up.

Judging at $150 does the worst of both: it kills angles that would have worked and keeps you cycling through creative for reasons that were just variance. This is also why Meta’s learning phase punishes impatient edits. The platform needs volume to find signal for the same reason you do.

Charlie Munger said to take a simple idea and take it very seriously. That’s the entire system: you don’t need complicated campaign trees with piles of ad sets for paid ads to work. You need a proper way to run experiments: learn first, then scale on signal.

Variations are for scaling, not discovering. Only after an angle proves out do they matter; then scale the winners with variations that carry the message better or hook faster. Complexity feels like diligence. Mostly it’s a place for losing ads to hide.

The whole account, drawn on a napkin: a handful of angle tests, enough budget on each to read the result, winners scaled with better-carrying variations. If your structure doesn’t fit on the napkin, it’s not a strategy. It’s a place losers go to survive.

Everything above is the doctrine Kelpi runs on. The free audit checks exactly what this essay argues matters: are you running real angles, and is the account set up simple enough to learn from? Complexity hides losers; simplicity exposes them.

The channel thinking comes from a lived funnel. At the homebuying platform, Meta wasn’t our demand source: it was our demand warmer. Cold prospecting ran on Google search, tightly targeted at niche home hubs; Facebook earned its keep on retargeting and secondary lead magnets, like a homebuying guide that gauged interest and turned cold traffic warm before we converted anyone. Give each channel the job it’s actually good at for your buyer’s awareness stage, and judge it on that job, not on someone else’s.

And the product thesis follows from the loop. Creative tools are designed to make one-off ads. But growth work has to compound: the more context you gather about a business (its positioning, its angles, what already worked), the better every next campaign gets. Those platforms aren’t learning your business as you go. Kelpi is built to be an AI growth employee: it understands the business, knows what to research, walks the same learning path a real hire would, and grows with you after onboarding.

If you want this doctrine applied to your own account, start with the free audit. If you want to see angle-first thinking on your own brand, the ads generator drafts five distinct concepts: five angles, not five copy variations. And if you run ads yourself with an AI assistant, the same doctrine is packaged as copy-paste skills.