We paid Meta to acquire customers we'd already acquired. Then we ran a Haus holdout to see if it was worth it. 1.93x incremental ROAS against a 1.5x target. It passed. Comfortably. Full breakdown in the article.

This is exactly why I almost never suggest working with generalist agencies. You’re paying $15k+/month to have some junior $2k/month strategist throw a bunch of shit at a wall with your own ad dollars. If an agency underpays on talent, they will under service your account.

I've been so AI pilled, that I've done a complete 180. AI is a money incinerator. It's a collective hallucination. Everyone sees everyone else shoveling money into the incinerator and starts doing the same. Everyone just keeps shoveling money hoping that one day the engine starts running. And maybe it will. But it isn’t. AI is expensive. Especially when you have to actually pay for tokens. And the real systemic work, the stuff that replaces people because it is autonomous, runs on tokens. The question isn’t whether your AI systems are cheaper than Americans. It’s whether they’re cheaper than people in Latam, Southeast Asia, or Africa. They aren’t. AI also suffers from broken clock syndrome. When we task AI with doing a specific task, it can often give a quality or correct output. But the problems arise when we ask an AI to produce that type of output with different inputs. If you don't know the process it used to give you the initial correct output, there's no way to know if it can take your input type and generate a quality output repeatedly or if it is just a broken clock that happened to be right. So in order to build durable results with AI, you need to build the process the AI is going to use to generate the output. You need to build in oversight. And the more varied the inputs, the more complex the output, the more granularly you need to control the functions within the black box. Otherwise, your fail rate is going to be high. And this failure rate increases whenever there is nested processing and thinking. Then there is the cost of building these systems in the first place. If you actually want something that works, you need to invest considerable resources into the build. Senior talent to develop it, token costs during calibration, ongoing maintenance when it breaks, etc. The common argument against this is to “let AI do the building on its own.” The problem with this is AI doesn’t think. It doesn’t reason. It guesses the next most likely thing in a series. So it will build something for you that will look like a house, but you go in the basement and the foundation is upside down. It’s got wires running through water pipes. The hot and cold water taps are reversed. It looks right from a distance, but any deeper level of scrutiny the illusion falls apart. Nobody is making money from the “work” that AI is doing. If they were, we would have such a rapid consolidation of wealth nobody would know what is happening. Every industry would be consolidated overnight into unassailable monopolies. I hear from lots of people who say they are making money from using AI. And my question to them is how much did it cost to build and what’s the payback period. Large companies are pulling back on AI investment. These are the companies that should be most able to realize a positive ROI on AI. So the question is whether they are uniquely incompetent at development and deployment or are smaller businesses bad at calculating cost and ROI? This results in our current environment: The only people making money in AI are the people who are selling it. It remains to be seen when using AI is profitable and how much work it can profitably replace.

just thought you'd like to know that I've been plugging your social content on AI into AI (and all of the comments and replies of people arguing with you) to see what it thinks (it names you the "winner" with the caveat that you go too far 🏆 😁). Anyway, I learned stuff...

A Trybe creator messaged us last week: “That's a damn fine brief. It’s always a relief to get a brief like that.” A relief. That’s the word that stuck with me. Most briefs are a product page pasted into a Google Doc (or no brief at all!). The creator guesses, the footage comes back unusable, and the brand decides Trybe doesn’t work 😡 The brief is where Trybe is won. Here’s what goes in ours. 1. How it works, on one page The monthly loop: get the deck, film, upload, we edit and run it as ads. Creators shouldn’t have to work out the process. If they have to ask, the brief failed. 2. What’s selling right now, with numbers We pull it from the account data every month. “Opening on the problem stops 2.5x more scrollers.” “Our top video includes XYZ and sells at 2x the average.” Creators film better when they know what works, and they want to know. Their commission depends on it. 3. The product, limited to what matters Six features, one line each. Then a rule: only mention a feature you’ve used yourself. 4. The offer, said exactly “$149 for the pair, regularly $249.” Plus a do and don’t list: no invented bundles, no countdowns. If the price in the video doesn’t match the landing page, the ad can’t run. 5. Non-negotiables The name spelled right. The product name as it appears on their box. The compliance wording, with the exact phrase they can use and the ones they can’t. 7. Two briefs, pick one A gift to yourself, or a gift for someone. Each has an angle, who’s on camera, the setting and what to record. One brief per submission. 8. Visual hooks, as short clips Six 3-to-5-second shots: product on a fingertip, opening a tiny gift box. Silent is fine. The first 2 seconds of an ad are a picture, not a sentence, so we ask for the pictures separately and pair them with the spoken hooks in the edit. 9. The spoken hooks, written out Four per brief, each tied to a visual hook. Creators change the details to fit their life, but they start from a line that’s already been tested. 10. The script, in beats Problem, putting it off, the gift, the quality, the price, the close. Each beat is one or two sentences. 40 to 45 seconds total. We cut it down. 11. The shot list and the b-roll What to film talking to camera, what to film of the product, and 8 everyday clips of 5 to 10 seconds. Phone at eye level, window light, no logos, no copyrighted music. 12. How to label the upload Brief, version, hooks filmed, model. “Brief 1 · Hooks 1, 2, 4 · V1 to V6 · Script · 3.0.” The editor opens the folder and knows exactly what they’ve got. 13. A final checklist Ten boxes. Name spelled right, price said exactly, no cure claims, uploaded by required date. That’s 20-odd pages. It takes us a day to produce each month, per brand. It gets us creators who film what we need first time and want to work with us again - at scale, ads that pass compliance, and an edit team that isn’t guessing. The alternative is 40 videos a month you can’t use. If you’re running Trybe and the content coming back is hit and miss, the brief is almost always the reason.

notice.. > Delta is blocking AI agents from booking flights > JPMorgan Chase and Charles Schwab block agents from accessing accounts > Yelp requires paid data licensing for agents > Amazon has blocked Grok bot and other shopping agents. the internet is starting to block AI agents that were supposed to do everything for us

If you only do four of these before November, do these. The credit line one is an agencies worst nightmare... client: "yo who turned our ads off?" 😡

You are underestimating how much people out there hate AI I was at a wedding this weekend and the consensus take among people I met was that using AI for anything is morally wrong and generally inexcusable Guess I’m living in a bubble bc I was shocked…?

Agency vs in-house is back on the timeline 🤗 Good ol' Twitter has a way of recycling topics, like clock-work. Every. Single. Year. This time the argument goes: fire the agency, run the ads yourself with Claude, it just takes an hour a week, and your business grows. Agencies bad, me and Claude good. An hour a week... A founder who gives an hour a week to anything and then just walks away is a cog in that thing. The founders I know and get the pleasure to work with... eat, sleep, and breathe whatever compounds the business. They don't check on it on Sundays and then go about their week. Especially if paid media is an engine. You're in it every day or you've handed it to someone who is. An hour a week with Claude is not a business. At that point, you have a hobby with a budget. And AI is making this lovely little debate even louder. I read @Superhuman 's AI Maturity Report, where they surveyed professionals and... you guessed it... marketers came back as the function getting the most from AI. WHILE also absorbing the most friction. Thought it was the perfect little resource for this post. So, here's some fun facts to tack onto this wall of text. 73% said AI creates extra work to review or fix, 13 points above the broader workforce. In high-friction environments that number hits 95%. 65% say new tools show up faster than they can learn them and 69% feel pressure to use AI even when they're not sure it helps... I strongly believe that this is the pressure engine behind "me and Claude can just handle this, an hour a week." Okay, cool Jake. Nice stats but why is this a problem? Within marketing specifically, AI amplifies the environment it enters. Hand it to a strong operator and it compounds. Hand it to a weak one and it manufactures slop at a speed no human could match. AI is still adding work before it removes any. That's the AI problem on both sides of this debate. The worst agencies are slinging AI slop at clients. The worst in-house teams are slinging it right back. For me, I get the added benefit of using our pipeline at the agency as a proxy. Example: Our pipeline is full of accounts that were run on AI slop, and they arrive in the exact same way as a brand leaving a bad agency did. You know, the ones that just rip branded search and call it a day. That said, zoom out. Way out... You quickly realize that tools are making both sides faster at building the wrong things. And, when I zoom out, I use that a reminder to keep it simple. Always. In this case, what predicts the outcome is the operator. Killer operators are killer operators, and they show up on both sides. I've watched agencies compound, almost infinitely. I've also watched in-house owners run playbooks better than most shops I've competed against. Let me tell you... Those owners weren't giving it an hour a week and the AI in their stack is clean. It stays grounded in the brand, carries context forward, and takes the repetitive work off the operator so the operator can spend the day on judgment. At this point in my ramble, let's highlight a recent change of mine when it comes to AI slinging. It might help you... If I write something to a brand or if I get something from a brand, I now expect both parties to be able to answer: "What do you think we should do with this, and why?" If you OR the sender can't answer that line, the message or whatever you're sending isn't ready. It will be AI Slop. Yep... Congratulations, you just tossed a Slop Grenade and became a stat. I've been doing this for over a decade. I'm entering like my 12th or 180th BFCM (if you know you know), so I'll be honest with you about the other side of it. There will always be agencies, and in-house people, who do not operate this way. You have to go find the ones who do, and you might eat 1 or 2 absolutely shit experiences on the way. But we're out here. Taikun keeps 95% of its partners year over year. Our founders who've been with us for years could go the MCP route tomorrow but they put that time back into the business instead, because they've done the math on what an hour of their attention is worth. They also know that we wake up absolutely obsessed with making them profit. Obsessed with growth and saving every penny we can so the founder has more to reinvest, with no price gouging. And even then, we still hand as many resources as we possibly can to the founders who want to run it themselves. Because the best operators on either side want the brand to win more than they want the credit or an invoice. So... Agency vs in-house? Both have terrible versions. Both have operators who'd run circles around the other side. Find the one that compounds with you and your business. The one that tests before it scales, shows you the number it's embarrassed by instead of hiding it in AI slop. The one that fights for the win instead of phoning it in, and the one that wakes up every day trying to leave your account (and your business) better than they found it. Agency or not, that's the hire.

I've been thinking a lot about client comms, and I love that @tobi coined slop grenade, because it represents something we're seeing more and more in our client engagements. Situations where clients have automated systems that they use to judge or monitor our work and then paste the output from the AI directly for us to address. Most of the time I'm skeptical they even read it. We then need to spend human hours responding to questions the client doesn't actually have and are generally missing context, incorrectly analyzed, or straight up irrelevant. I totally understand the desire to be on top of things. To solicit a second opinion on something you may not be fully informed on. Curiosity is good! We love questions from clients, and we love sharing our knowledge. But your AI monitoring and asking questions isn't curiosity, it's surveillance. Time spent on comms is up 40% over the last 6 months almost entirely due to my team having to take time to respond to slop grenades. The easy answer is "use AI to answer their AI" but at that point what are we even doing? What's the point? I'm strongly considering writing into our contracts that we will not respond or engage with raw AI outputs. If there are questions it prompts you to have and you share them, we're happy to address. But we're not going to go through and respond point by point to an 8 page report your AI generated that you didn't even read.