The argument in one line.
A business function becomes a loop the moment you give an AI agent a repeatable task, an objective metric to check against, and a stop condition, letting it build, verify, and learn on a schedule without you in the room.
Read if. Skip if.
READ IF YOU ARE…
- You run a business with an existing website, ad account, or product and want to automate the repetitive optimization work an agency or freelancer would normally do.
- You're comfortable directing Claude Code, Codex, or a similar coding agent and want a concrete pattern for turning one-off AI tasks into recurring, self-improving ones.
- You want to cut agency or freelancer spend on SEO, paid ads, or product analytics by handing the grunt work to an agent with real data access.
- You're deciding whether to trust an agent with an ongoing, unsupervised task and want to see the cost, cadence, and guardrails first.
SKIP IF…
- You have no existing website traffic, ad spend, or user base yet -- a loop needs real data (rankings, clicks, feedback) to optimize against, so there's nothing to feed it.
- You're looking for a fully autonomous AI business builder that needs zero oversight -- even the enthusiasts here still check in monthly and approve changes.
- You want deep technical implementation detail (API auth, code, prompt engineering specifics) rather than the conceptual walkthrough and live demo shown here.
The full version, fast.
Loop engineering means giving an AI agent a build step, a verify step tied to one objective metric, and a stop condition, then letting it run on a schedule -- the same build-measure-learn loop from the Lean Startup and Toyota manufacturing, just automated. Elie Steinbock demonstrates it live on Draft Fantasy's Google Search Console: the agent reads rankings and click data, edits the site, and checks back in a month, for under five dollars in tokens per run. The same pattern extends to Facebook ad copy testing (objective metric: revenue) and a product feedback loop that reads analytics, Sentry logs, and user feedback to prioritize fixes (objective metric: DAU or revenue). The advice for starting: pick one channel, tie it to a small verifiable metric like impressions or ten likes, and let it compound rather than aiming a loop at something unmeasurable like 100,000 followers.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Who's talking.
00:00hostGreg Isenberg
02:54guestElie Steinbock
Where the time goes.
00:00 – 02:54
01 · Intro and episode promise
Cold open with the episode's promise, then Greg asks Elie to commit to explaining loops clearly enough that listeners can copy the workflows themselves.
02:54 – 06:51
02 · What is Loop Engineering
Elie traces the term to Boris (Claude Code) and Peter Steinberger, plus a joking Dmitro Krasun tweet about software that builds itself, then grounds the idea in the Lean Startup's build-measure-learn loop and Toyota's manufacturing process.
06:51 – 11:17
03 · Loops with AI agents: build and verify
Elie maps build-measure-learn onto an agent: a build step, a verify step, and a stop condition, using Inbox Zero's evals (categorization accuracy above 90%) as the working example.
11:17 – 15:29
04 · Example of Loop: SEO as an objective-metric loop
SEO becomes the flagship example: Google ranking is a clean, objective metric, and a loop can run once a month for years, learning from a markdown memory file each time.
15:29 – 25:27
05 · Setting up the SEO loop and tools
Elie shows the live Draft Fantasy Google Search Console, explains giving the agent access to Search Console and DataForSEO, runs the loop in a terminal, and points to the Atom Eve prompt as a copyable template.
25:27 – 29:05
06 · Cost and token economics
Greg raises a friend's skepticism that loops just burn tokens; Elie argues a monthly SEO loop run costs under five dollars, cheaper on a max-tier plan, with GLM 5.2 as a budget option.
29:05 – 33:10
07 · The Paid ads loop
The pattern extends to Facebook ads: the agent tests copy and creative variants against a revenue metric, working best as a mix of human-shot raw content and AI-driven editing and budget reallocation.
33:10 – 36:25
08 · The product feedback loop
Elie describes the 'ultimate loop' -- an agent reading customer feedback, PostHog analytics, and Sentry logs to prioritize and ship fixes or features, verified against DAU or revenue, with bugs split into their own uptime-based loop.
36:25 – 39:21
09 · A minimal viable loop for every channel
Elie argues nearly every business function -- social, cold outreach, support -- could run as a loop, but recommends starting with a minimal viable loop tied to a small verifiable metric like ten likes rather than an unmeasurable goal like 100,000 followers.
39:21 – 39:43
10 · Closing Thoughts
Greg thanks Elie for sharing the examples and points listeners to his social links.
Lines worth screenshotting.
- A loop is just three parts: a build step, a verify step tied to one objective metric, and a stop condition -- without all three it's just a task, not a loop.
- The lean startup's build-measure-learn cycle and Toyota's plan-do-check-act manufacturing loop are the same pattern AI agents now run automatically.
- An SEO loop connected to Google Search Console and a rank-tracking API can run once a month for years, using a markdown memory file to remember what it already tried.
- One SEO loop run costs under five dollars in tokens on a $100-200/month max-tier AI subscription -- far cheaper than hiring an SEO agency.
- A site ranked 4th for a search term with 120,000 clicks could see roughly half a million clicks by reaching position 2 or 3, which is the kind of gap a loop targets.
- Inbox Zero's product loop pushes an AI model or prompt to keep adjusting itself until it clears a 90% accuracy threshold on email-categorization evals.
- A Facebook ads loop works best as a mix: humans generate the raw hooks and footage, AI edits variants and reallocates budget toward whichever one wins.
- The most defensible loop separates bugs from features: a bug loop should verify against uptime, while a product-feedback loop should verify against DAU, retention, or revenue.
- The safest way to start a loop is with a small, easily verifiable metric -- ten likes or weekly impressions -- not an unmeasurable goal like 100,000 followers.
- None of these loop experiments are irreversible: if a change tanks a metric, the loop (or the operator) can simply revert it, so the downside of trying is low.
- Tight-budget builders should reach for cheaper open models like GLM 5.2 for loop runs rather than frontier models, since the task itself doesn't require peak reasoning.
A loop is a build step, a verify metric, and a stop condition
WHAT TO LEARN
Automating a business function with AI isn't about one clever prompt -- it's about wiring a build step to a single objective metric and a clear stop condition so the cycle can run unattended on a schedule.
02What is Loop Engineering
- The Lean Startup's build-measure-learn cycle and Toyota's plan-do-check-act loop are the direct ancestors of today's AI agent loops -- the pattern isn't new, only the automation is.
- A joking tweet about software building itself and reaching product-market fit on its own is what crystallized 'loop engineering' as a term, even though the underlying idea predates it by over a decade.
03Loops with AI agents: build and verify
- An agent loop needs three parts to actually be a loop: a build step, a verify step tied to one objective metric, and a stop condition -- skip any one and it's just a task.
- Product evals work the same way engineering tests do: define a pass threshold (like 90% categorization accuracy) and let the agent keep adjusting the prompt or model until it clears it.
04Example of Loop: SEO as an objective-metric loop
- Google ranking is a clean example of an objective metric because it's unambiguous -- a loop can read Search Console and a rank-tracking API monthly and know exactly whether it helped.
- SEO loops don't need to run continuously -- a monthly cadence is enough, since ranking changes take weeks to show up anyway.
05Setting up the SEO loop and tools
- A markdown memory file that records what the loop already tried is what lets it improve across separate runs instead of repeating the same experiments.
- Giving an agent access to real business data (Search Console, a rank-tracking API) is often the single highest-leverage setup step, independent of whether you build a full loop.
- A monthly or biweekly automation trigger (a scheduled routine) is what turns a one-off agent run into an actual loop that picks up where it left off.
06Cost and token economics
- Cost is not the blocker it sounds like: a monthly SEO loop run can cost under five dollars in tokens, and heavier subscription tiers have enough headroom to not think about it.
- Budget-constrained builders should reach for cheaper open models like GLM 5.2 for loop runs rather than assuming a frontier model is required.
07The Paid ads loop
- AI-generated ad variants are weaker than human-shot ones on their own, so the strongest ads loop mixes human-created raw hooks with AI-driven editing, testing, and budget reallocation.
- Paid ads is fundamentally a volume game of testing angles and hooks against an audience -- which is exactly the kind of repetitive work a loop can absorb.
08The product feedback loop
- Splitting a bug loop (verified against uptime) from a feature/product loop (verified against DAU, retention, or revenue) keeps each loop's objective metric unambiguous.
- Reading customer feedback, analytics, and error logs together is what lets an agent prioritize the highest-impact fix or feature rather than guessing.
09A minimal viable loop for every channel
- None of these changes are permanent -- if a loop's edit hurts the metric, it (or the operator) can simply revert it, which is what makes starting one low-risk.
- Start small: tie a new loop to an easily verifiable metric like weekly impressions or ten likes rather than an unmeasurable target like 100,000 followers, then let it compound.
Terms worth knowing.
- Loop engineering
- Giving an AI agent a repeatable task, an objective metric to measure progress against, and a stop condition, so it runs build-verify-learn cycles on a schedule instead of a single one-off prompt.
- Objective metric
- A single measurable number a loop is judged against, such as Google ranking, ad revenue, or daily active users -- it has to be unambiguous so the agent knows whether a change helped.
- Stop condition
- The defined result that ends a loop's iteration on a given run, such as 'evals score above 90%' or 'the feature works in the browser' -- without one, an agent could loop indefinitely.
- Evals
- Automated test cases that score how well an AI model or prompt performs a task, used as the verify step in an engineering-focused loop (e.g., email categorization accuracy).
- Google Search Console
- Google's free tool showing a website's search rankings, impressions, and clicks per query -- the data source an SEO loop reads to know if it's improving.
- DataForSEO
- A paid SEO data API, similar to Ahrefs or SEMrush, that shows how a site ranks against competitors for a given search term.
- Minimal viable loop (MVL)
- The smallest version of a loop worth starting: one channel, tied to a modest and clearly verifiable metric, that can be launched immediately and expanded later.
Things they pointed at.
15:29toolGoogle Search Console
19:04toolDataForSEO
20:44toolAtom Eve (atomeve.dev)
27:30toolGLM 5.2
33:10toolPostHog
33:10toolSentry
06:51productInbox Zero
00:45bookThe Lean Startup (Eric Ries)
Lines you could clip.
04:37
“In 2026, you don't prompt anymore. Your software should be able to build itself and achieve product-market fit on its own. Your only job should be to find money to pay for tokens and take care of yourself.”
19:04
“This is 4.4 right now. If I can get this up to three or two, imagine -- this wouldn't be a 120,000 clicks, this might be a half a million clicks.”
26:40
“I wouldn't be shocked if this like cost you less than $5 in tokens to basically go and run this one time.”
30:00
“The humans are becoming the API layer... let AI go into that folder and edit it from there versus going and creating a fully AI ad, less context, less human layer.”
33:30
“I would actually do a bug loop separate from a feature loop. The bug loop would be around uptime... the product feedback loop might be around core metrics like DAU over MAU or retention or virality.”
38:00
“You are that agent starting your loop. You're thinking today, how can I improve my business? It's the same for the AI.”
Where the conversation goes.
02:54 – 06:51steadyOrigins of loop engineering hype
06:51 – 11:17denseBuild/verify/stop-condition mechanics
11:17 – 25:27denseSEO loop concept + live demo
25:27 – 29:05steadyCost and token economics
29:05 – 33:10steadyFacebook ads loop
33:10 – 36:25denseProduct feedback loop
36:25 – 39:43steadyMinimal viable loop / closing advice
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00You might have heard of engineering loops. They've been going viral on Twitter and everything like that. And I think they're really interesting, but they're way more interesting to use to actually run your business.
00:11There's a way to use loops to actually get customers, get SEO, be seen by LLMs, and actually improve your product twenty four seven.
00:20Now, I haven't seen anyone cover how to actually implement these loops, so I created a tutorial how loops work, how you can use it to run your business, and how you can use Claude code or Codecs to actually implement it. In this episode, I share everything with my friend, Ellie, and you'll see and understand completely how to do it yourself so that you can get traffic, get customers, build a startup today.
00:46My favorite loop is actually the last loop that we share. Enjoy the episode, and I'll see you at the end.
01:01Ellie, welcome to the show. By the end of this episode, what are people gonna learn?
01:06Yeah. So you're gonna learn how to use loops to better automate your business. Loops have been really popular over the last few weeks.
01:13People are using loop engineering to better develop products, but it can go a lot further than that. You can use it for SEO, for Facebook ads, really to automate almost every part of your business.
01:24So that's what we're gonna talk about today. Okay. Cool.
01:27So, yeah, people are using Loops to basically build products,
01:30but you're basically saying there's a way to use Loops that could you know, you can run your business on it basically. And that's gonna help you get customers. That's gonna help you, uh, build a more efficient business.
01:41And what I'm asking for you, Ellie, is if you can clearly explain how to actually do this thing, and then show some examples so that people can actually just copy some of these workflows. And then by the end of the episode, they're gonna understand loops for for, you know, how to how to use rookes to run run your business, but also how they can get started today.
02:04Can you make that commitment, Eli?
02:06Yeah. Yeah. Sure.
02:09I'm gonna show you how to use loops to run your business. We're gonna talk about it at a high level, like sort of what the concept is, where you can potentially use it but then I'm also gonna show you how it actually runs in practice.
02:22So it's not just theoretical. I'll show you how you can actually improve your SEO massively using loops.
02:28It's the sort of thing a lot of you might be doing today if you have anything running on a schedule that's a form of loop. But we're gonna sort of really push it far and I think the state of AI today you can really do quite a lot with a loop over a long period of time. Most of the time we're talking about loops.
02:45Maybe, you know, they run-in half an hour. Now we're here. We're talking about loops that might last for months or even years.
02:51Let's do it. Alright. Let's get into it.
02:54Cool. So, yeah, around a month ago, loop engineering got really popular. Boris from open well, Boris from Claude Code started tweeting about it.
03:05Also, Peter Steinberger from OpenClaw started tweeting about loop engineering and everyone sort of was like, wow, what is this loop engineering thing? Overall, it's quite a simple concept but it's really blown up.
03:17And I guess it's nice that sort of it's got a term now, loop engineering before someone could have described this concept and it didn't have sort of a one word explanation. Now it does. Shortly after this whole hype cycle started, a friend of mine, Dimitro, he went and tweeted this.
03:33In 2026, you don't prompt anymore. Your software should be able to build itself and achieve product market fit on its own.
03:41Your only job should be to find money to pay for tokens and take care of yourself. So he was definitely joking when he wrote this. I think he was making fun of this whole idea of loop engineering, how we had like, you know prompt engineering, context engineering, harness engineering.
03:57Every month we've got another hype cycle. But if you read the tweet, I I found it funny. I thought it was a great tweet.
04:03But then the question is wait, could you actually do this? What would it look like to actually have everything running on a loop in your business? So that's what we're gonna speak about today.
04:14And the idea of loops honestly, it's not that new. Maybe even like ten, fifteen years ago, the lean startup book was pretty popular and a big part of that book was this loop where you build something, you'd measure how it does, then you learn from it, then you build a bit more.
04:31But basically if you break down a business or you think about Dmitry's question, how could an entire business run as a loop? It's basically let's go build something, let's get feedback, let's improve it, and just keep that cycle going of like build and learn, build and learn, and I guess measure as part of that as well. And you can do the exact same thing with AI and it's not just sort of high level for like the the business to to build and get feedback that might be on the product but you can do this for so many parts of your business and it's actually what you do already.
05:03If you're improving your SEO you're seeing okay where do I rank today? Where do I want to rank? What are the things I can do to improve it?
05:10Who is ranking above me? You do all these experiments and then you try and rank higher.
05:15And you you have accurate measurements from Google coming back to you and so that would be an example of a great loop that you can run. It's a loop that I'm running today in production. If you're like sort of familiar with like lean manufacturing where or to the sort of the Toyota story where a lot of this stuff became popular as well, That's also a loop where basically you're just constantly iterating and trying to make things better.
05:41And so these aren't new concepts. I think we're all familiar with them. When it got paired with loop loop and loop engineering, it was like wow what is this?
05:48But I think it's something we all understand quite well, and it's just how do we take these ideas and get our agent to do the same thing.
05:56Right. Yeah. And the Toyota example, I think that was the basis of the Lean Startup book.
06:01Right? I think Yeah. Eric looked at the Toyota example and basically said, like, hey, there's this Japanese company and the way they manufacture.
06:09How are they able to create such reliable, consistent cars? And it was through the the the loop mechanism that they had this assembly line that was just highly efficient, and because of that, they were able to just create incredible products at at a good price.
06:26What Eric looked at, he said, okay. Well, you can actually build a startup in that same way. You know?
06:31Before that, people weren't building startups in that loop way. They would sort of it was more artistic. They would kind of like just put out a product and change it, you know, as they go.
06:41But I think the, you know, Toyota slash lean manufacturing process, applying that to startups was one of the reasons why we had such successful startups post, you know, 2005. So what what are we talking about when we're talking about loops with AI agents?
06:58Yeah. So I think maybe the best way to explain it is to jump into that first example I spoke about. Well, I guess let's talk about a sort of a loop engineering first and then jump into specific examples.
07:10But I would say if the loop for the lean startup is build, measure, learn, we have very similar steps here with an agent. We have this build step which is like telling an AI, hey go build me at my new SaaS for example.
07:23Then we have this verify step where if you're building product, the verify step might be that all tests work or that the agent has used a browser to make sure it can click through everything or there's some other agent that's running if you're doing this in Claude code and you use slash goal.
07:41So that's basically running this loop and it's got this other agent checking has it actually finished or not, is it working. And if it's not working it's just gonna tell the sort of the main builder agent to just keep looping and looping and looping till it's fully working. Some other examples of this within engineering.
08:00So you always have this stop condition. You don't want the AI just to loop infinitely. There needs to be some sort of result that you converge on.
08:09So at some stop condition examples, one is like the feature works in the browser. I want sign up like the goal is to have sign up working. You know, you slash goal would make sign up work and then once it's working in the browser it's what you know, it's passed and that's sort of the end of the loop.
08:25Another one for anyone building AI products, I run a product called inbox zero. So this they're like it manages it's an AI that manages your inbox and this one is super important for me. Basically, have evals.
08:37The evals are sort of tests for AI, like how well does it do. In the case of inbox zero, it would be how well does this model categorize emails. So for example, I just got a newsletter email that came in.
08:49Does it get categorized as newsletter? So evals are basically like the test the evaluations to to check how well the AI is doing.
08:58Different models will perform better or worse. Depending on what prompt you have it will perform better or worse. So your goal is to sort of get your evals really high.
09:08It might be choosing the right model or adjusting your prompt and so you can run this as a loop as well. And what that would look like is tell your agent hey I want our tests or evals to pass like get a score of 90% and above and so it can keep running the prompt over and over and it can keep adjusting it and if it sees oh I'm only passing 88% of the time it can try again and each time it will try and do a bit better till eventually it gets past 90% accuracy.
09:40And so if we take this that's sort of on the engineering side how you're always sort of building and verifying. But it's verify step. It doesn't just have to be related to product.
09:51It could really be anything. And really what you want is some sort of input back into the system, some sort of objective metric. And so in the case of SEO, which is the first example I mentioned, the objective metric is where do you rank in Google search.
10:08So right now if you search for the term inbox zero on Google, inbox zero ranks first. But some other term, AI email assistant, we really wanna rank high for that as our business. But we're ranked, I don't know, position 30 or so.
10:22So what we can do is run a loop that runs every month for example and tries to push us further and further up until basically we're on that first page.
10:32And honestly this is a loop that never really has to end. Maybe this loop ends when we're in position one. But this isn't a loop that's running let's say for half an hour straight.
10:41It's running once. It's taking its step and then a month later it will continue its process and try and push us further up.
10:49And so we can go into detail like what this actually looks like and what is needed to make a loop like this work because this example I think like it's a good example because it applies to a lot of other things in the business. Facebook ads for example, you know, you you're spending a $100 a month on the ads or a $100 a day.
11:07You know, you want it to get to profitability. So it's it's all the same ideas. So how can you get an AI agent to get there is basically the ideas the idea here that we're trying to understand.
11:17Yeah. And with with SEO, I think, you know, the way you would typically do this is you would hire an agency or you'd hire a freelancer to essentially do this loop.
11:27Right? So what you're suggesting is you kinda don't need to hire that person. At least to start, you can, know, hire slash build a loop that has a KPI, in this case, Google ranking, which is, you know, isn't gray.
11:45It's black or white. Either you moved up this month or you moved down or you stayed the same. Right?
11:50And you're able to basically say, okay. Am I doing a good job? Am I not doing a good job?
11:55And then based on that, actually, you know, perform actions. So the only question mark is can agents at the time of recording, can agents are they smart enough to actually work as good, if not better, than hiring an agency or a freelancer?
12:14Because, you know, ultimately as a business owner, you care about the, you know, moving up in the rankings. Right?
12:20So you you don't wanna like have a loop just for the loop's sake.
12:25Yeah. Exactly. And I would say also even if you do try this experiment and it doesn't work, you haven't necessarily lost anything.
12:33A lot of us aren't necessarily going to be hiring the SEO expert anyway. So it's just like you know you could run this experiment and worst case scenario you see, oh, it's actually had negative impact. What a loop like this would do would be like let's say we move from position 20 to position 30 in like sort of Google rankings.
12:51You could just undo the change basically. So none of this is really set in stone and it's sort of just experiments that we're running and hopefully like sort of long term will push us up. If anything goes wrong, we can always revert.
13:03Not nothing is set.
13:05Right. I mean so I guess, do you think that agents are good enough today such that they can actually impact Google ranking and and get you more traffic?
13:18So having run it myself, it's definitely having positive results. It's going to take a few months to sort of really have the impact that I want. But yeah, for sure, like I I before that we did this recording I took a look at the numbers and I can see a whole bunch of numbers are going in the right direction.
13:36Some of them you know I might be moving from page three on Google to page two for a certain term. So it's I guess it's valuable. It's getting there but obviously the ultimate goal is to get to first page ranking.
13:48Um, I do think it depends on lots of different factors like you know inbox zero, the main rating might be like 63 or so. Last I checked 64 something like that for a new business that sort of has a super low domain rating maybe you know it would work out differently. But yeah to me I'd happily run this whether it's like an established business or you know, a a new business that you're starting to to set up basically.
14:15Well, that's the thing with SEO is don't expect to do SEO and it works in twenty four hours, you know? Yep.
14:22SEO is something that takes months, not days. That's just in general. So this is the type of loop that you kind of want to have working in the background while you're doing other things too.
14:32Right? Yeah. Exactly.
14:34So that way, you know, you might wake up on month on month four, like nothing's really happening. Then month four, all of a sudden, bam, bam, bam, you know, you're on page one.
14:44And that's happened to me in the past where it's just like SEO wasn't really working for, you know, some amount of time, But, you know, you're still you're doing the things necessary to rank well. And then all of a sudden, that compounds and it starts to really, you know, bear fruit. So, yeah, let's let's go deeper into this and see some examples.
15:05Yeah. Exactly. And into it, it's definitely something that compounds over time.
15:09And I think everything you do marketing wise is all gonna have an impact. Also, you know, we've been speaking about SEO here, but all of these things obviously benefit your your LEO or GEO for ranking in search and in LLMs as well.
15:25So it's still super valuable even if you're not someone using Google search that much anymore. Like going sort of deeper into this like how would you actually set this up?
15:35So the example you brought of having an agency that's sort of running your SEO, I'm not an SEO expert but what they would likely do is run certain experience.
15:48They do an audit of your website. This is something I think you should get AI to do for you regardless. Just get an audit done.
15:55It will say like, oh, we should improve these meta tags or you know you've got these JSON LDs which could like give you a small boost. So go and do all of that. Maybe you don't have a sitemap.
16:04There are a lot of things AI can just get fixed immediately and quick win for most websites I would say. But after that what sort of that agency might do is start to experiment with certain terms.
16:17It's seeing okay you're ranking quite well for AI email assistant but you're not on the first page yet. What is you know what what is happening there? What what can we fix?
16:26And so this whole thought process you don't even need to worry about it too much. The AI will go into it and be like okay you might be cannibalizing your own links because you're sharing you know the the link power between two different links on your website. But like whatever the AI comes out with, um, that or the SEO agency, what they're gonna do is make those improvements and they're not gonna see results immediately.
16:46They're gonna come back a month later and see, okay, we have moved up, we have moved down. And so the exact same thing that the SEO agency is doing, that's what we want our own agent to do for us.
16:58And so the first thing we need to do is give it access to all the tools it needs. The main ones I would say one is Google Search Console that that way you can basically see all your data and Google Search Console has an API so it will show you exactly where you're ranking for Google rankings.
17:18I'm gonna go to that.
17:21So here we're looking at my Google search console for draftfantasy.com. This is a business I started around twelve years ago.
17:29It still runs today. It's not my main focus but because of the World Cup it's had quite a lot of activity recently.
17:37And here you can sort of see how it's ranking. It's had 10,000,000 impressions over the last three months on Google search.
17:44Down here we can see sort of some of the queries that it's ranking for. For the thirty eight zero search term right now it's had a 120,000 clicks which is actually quite insane.
17:56You can see it's got a million impressions. This is actually not a business that I've been running this sort of this loop agent on.
18:03I didn't wanna go into the numbers behind inbox zero. But I'm happy to sort of share what's happening with drawfantasy.com right now.
18:10And you can see it's ranking well for a bunch of terms. But like what I did around two days ago is basically tell my, you know, my Claude code, go and do the same loop engineering thing we're doing for SEO for inbox zero. Let's just have it run for draft fantasy as well because why not?
18:26It will run-in the background. I don't really need to think about it. It will make, you know, good updates over time and it will remember what it's done and then go and make more improvements.
18:36And so here you can sort of see like lots of data around like you know where your search terms are ranking.
18:44Where is it? Let's say average position this is like a big one. So for example over here you can see I'm ranked fourth for the term thirty eight zero but let's say I wanna push that up to one like the AI can basically look at all of this data that I have here on screen.
18:59It can connect via the Google API and all this data will come into it and it can make a really smart decision. Honestly, a lot better than me even and decide, okay, these are the terms I'd like bringing a ton of traffic right now. How can we change things so we can rank even higher.
19:15So this is 4.4 right now. If I can get this up to three or two imagine this wouldn't be a 120,000 clicks.
19:21This is might be a half a million clicks. So it can drive just a ton of value and there might be some really low hanging fruit that it can go and sort of fix up and make it work.
19:32So yeah, the first thing to do and then just across your business whether you're doing loops or not, I think one of the easiest tricks is just connect your AI to your different tools, your real data. And the tools here would be Google Search Console. Another one would be data for SEO.
19:46It's like an SEO API similar to Ahrefs and SEMrush I believe. And it will show you how you're ranking against competitors. So Google search console will just show you okay you're ranking fifth over here but like what are the four articles that are ranking higher than yours for this term that you're really after.
20:04And so the more information you can give to your AI obviously the better it can do. And so what this loop actually then looks like is it makes improvements. It can check an objective metric which is your Google ranking where you're ranked.
20:17It can learn from that which you know you can do immediately and it it can continuously iterate.
20:24And the idea is every month or maybe every two weeks, it looks back at what it's done. It's noted everything down. This is another important part of it.
20:31Like have it remember, have let's say a markdown file with everything that's happened than the last time it made improvements and then it can basically check its experiment.
20:40Did it do well or not? I decided to change the description of the page. Did it you know, did that description change?
20:47Did it rank our article higher or lower? And so it can look back at what was tried, what wasn't tried, and it can iterate on that the same way as an SEO agency would do for you.
20:58So if someone wants to actually create this SEO SEO loop today, is the easiest way to do it basically screenshot this, paste it into your, you know, Claude code or codex, and be like, I want to create an SEO loop.
21:19I want to give you access to my Google Search Console slash data for SEO, and I want you to be judged on the objective metric of a of the Google ranking.
21:32So, like, check check the metrics.
21:34Is that what people should be doing? Or or how would you optimize that? Yeah.
21:39I think if you did that, honestly, you could go quite far with it. I can show you an example. If you go to atomeve.dev, this is another website I put out not so long ago.
21:49But here, there's actually a real example of this SEO improver or just improve this people don't need to use this as if and if people are familiar with Eve, which is a Vercel project that just came out or Flue framework by the ASTRO team.
22:05So this is sort of like, you you don't need to use these to build agents, but this is one way of building agents. But either way even if you don't use this you could honestly copy and paste this URL into your Claude code or codex and just say hey I want you to go and sort of copy the ideas here.
22:21But here you'll see basically a prompt that does the same thing which is like, you know, this is my Google search console. This is, you know, the the data for you to get into the API key for you to get into data for SEO.
22:35And then here's sort of a third prompt that you can go and copy if you want.
22:39Oh, wow. This is great. This is awesome.
22:42So this is yeah. This is basically a more this is a expanded upon version of basically what I just said.
22:48So this is basically like you're the SEO improver. You're, you know, you're gonna be judged upon these three metrics. And and it's yeah.
22:55And instructions on MD file for the specific
22:59job. Right? Yeah.
23:01Exactly. So you you can see for example, it's saying when you apply changes, select the subset of this week's recommendations and clean meter files and the blob repo. You can read through it if you want but the basic idea is exactly what we said and you know if you're if you wanna play with the CLI you can even run this command or even copy this prompt honestly into Claude code.
23:21This is a prompt and it will set that up for you. You don't need to use this. It it might actually complicate things for some people like using Eve or Flow.
23:29If like I might just show you this in my own Claude code quickly. Yeah.
23:34Cool. So here's my own codecs just running it in a terminal on my machine. I've actually gone and like taken the idea we had here and just taken a screenshot and of the chart we had before and that's the loop we basically want to have running.
23:49But yeah, if I say hey I I want to set this up for myself. I want to create an SEO loop.
24:00Basically, honestly, even with that, we should be able to get quite far. Maybe, like, if you're doing this yourself, speak to the AI a little bit more about it.
24:10In terms of what what actually needs to happen, maybe you can use plan mode. But like it literally is as easy as that. It will guide you through like how you have to connect Google Search Console.
24:22If you're running it on your own computer, that's the easiest. There's a CLI you need to install or use the Google API. So there's like a few steps you need to go through like to to give access to your data.
24:33But once you've done that honestly it should be quite easy and you know say something like we want to improve our SEO. That would sort of be the main thing.
24:42Maybe even do it on your own repo. It depends where your blog is, how this is done exactly. If you have a WordPress blog, maybe you wanna give access to WordPress.
24:50If it's you know on some other system, if it's GitHub then you could do that differently. But you give AI access to your blog everything you're doing, all your data, and then honestly from there it should be able to run on its own. The one step afterwards what you really want is to have some sort of automation set up.
25:06So like if you're you if you're a Claude user they have I think are they called routines on Claude right now? Yes. And Cursor has automation so and I think Codex is also called automations.
25:16So you can run one of those and the idea is just every few weeks it should pick up where it left off basically. And, yeah, if you want a much deeper example, then use what we showed for Atom Eve basically.
25:27Cool. So I had my friend Ross Mike on the pod recently, and we talked a lot about loops.
25:36And his his perception about loops is you know, he's an engineer.
25:41He's a front end engineer. So he's looking at it from an engineering perspective. He basically was like, I don't really believe the hype around loops.
25:47I think the people that are gonna get rich from loops are the token providers because people are just gonna be burning tokens. Now we didn't talk about any business use cases.
25:57We were talking specifically around engineering. If I were to implement an SEO loop, would it be smart to basically say, a click to me is worth 3¢ or a customer to me is worth a $100?
26:13You know, stop. Like, you know, stop stop basically, stop the loop if these things happen.
26:22Right? Because you basically you what's gonna happen with these loops is that they're it's gonna cost money.
26:27And it might be $50 a month, $100 a month, $200 a month depending on what you're actually doing. And you might just decide, like, it's not worth it. So I'm curious how you think about cost benefit analysis for loops.
26:41Yeah. So I I watched Mike's video, your guys' video together, and it was great. And I, like, I definitely agree with a lot of what he's saying.
26:49That's like, you know, the unnecessary hype around these terms. Also in terms of cost for sure, like he mentioned that Peter Steinberger works for OpenAI now spending $1,300,000 a month on AI credits.
27:03You know, it might even be more at this point. So I fully agree with that. For this loop, would actually say it's quite cheap.
27:10So you should very much do it. You really shouldn't worry about cost especially if you compare it to what this would cost if you hired an SEO agency. The reason I say it's so cheap is like it's not each run-in this loop it's happening once a month for example.
27:25I wouldn't be shocked if this like cost you less than $5 in tokens to basically go and run this one time right now like what I I just ran it in the background. So each of these runs they're not that deep.
27:38It's not that it's in like sort of an AI guessing itself into an infinite loop. It sort of is but like it's infinite over time meaning it will run once every month for the next two years or five years. Honestly for me, I'd be happy for it to just keep going, do that once a month thing.
27:55You might even want to sort of you you might want to have the AI update you in between. This is something else I do myself.
28:05Every time one of these runs, I need to know it's running. So I'll get it to ping me on Slack basically whenever it's done a run and then I can look over things and I can sort of give a quick approval if I like it or don't like it.
28:17And so I'm very happy to get these like once a month updates for things we can improve in our SEO and yeah the overall cost is gonna be small. The other thing I mentioned that Mike didn't is that if you're on a max plan you are getting tens of thousands of dollars per month in your like you know a 100 or $200 per month subscription.
28:38If you know you're really tight on budget and on a $20 plan then yeah you like you need to be much more wary of tokens and I think about like using open source models that are cheaper for this sort of thing like GLM 5.2 type thing. But if you are lucky enough to be on sort of a 100 or $200 per month max fan, you really you've got thousands and thousands of tokens there.
28:57Um, and so I I wouldn't be worrying about cost for something like this. It should be fairly cheap honestly.
29:04Cool. Alright. So we looked at SEO loops.
29:07What are other loops that people could be thinking about?
29:11Yeah. So another really good one would be a Facebook ad loop. So you're running you're running ads on Facebook, maybe even the AI is generating its its own ads, and it's just it's looking at the data.
29:23It's put out like an experiment with three different variants. It sees, you know, variant a is doing super well, and so it pushes more in that direction. And so this is exactly what you'd be doing if you're hiring an ads agency as well.
29:36They're gonna be experimenting with lots of different copy, lots of different, you know, graphics and, you know, images or videos and so on. And so you could run the same thing basically with an AI. Where this might get a little bit challenging is that the content that gets created by the AI is not always going to be amazing.
29:58Um, if you're doing video content generation with AI or graphics being generated with AI, it won't necessarily, you know, be as good as what a human can put together. Um, I'm sure there are some very good AI generated ads running by right now but I'm I I if I had to guess the human generated ads are running better but things like changing a line of copy for example that is very easy for an AI to go and change and then, yeah, see how it's performed and then improve on it.
30:26Or if we're talking about Google Ads where you know you don't have images necessarily, you're just trying to rank on Google search ads, the AI can very easily change the copy basically.
30:36Yeah. We're you know, it's funny because, like, in the sense of, like, create a folder and every day have, like create a new ad where you're yapping for thirty seconds and then let AI kinda edit it.
30:51And let AI go into that folder and and and edit it from there versus going and creating a fully AI ad, less context, less human layer.
31:03You know, I think my belief is the best ads are actually I mean, if you have millions of dollars to spend, yes, the the best ads are hiring the best humans on the planet to go and do that.
31:16But not everyone has millions of dollars to spend or hundreds of thousands of dollars to spend on the best agent you know, ad agencies on the planet. We're not making Super Bowl ads here. So the way to do it is a mix of humans plus AI to get you to a really, really quality level.
31:32And and and I just think that, yeah, if you just integrate this into your ads loop, you kind of get the best of both worlds. You're getting the human feeling of an ad, but you're getting the AI optimization around it.
31:46And the game around Facebook ads in general is a game of volume. You know, people forget this, but, you know, it's it's really this this game around a bunch of different narratives and hooks and seeing which one works.
32:05So it's basically taking your one product, but trying different angles and hooks and different types of, you know, people, a female, a male, an older person, a younger person, and then seeing how the algorithm reacts to it, and then cutting the losers, doubling down on the winners.
32:22And I could see how this loop could optimize this.
32:26Yeah. Exactly. And frankly, frankly like we are we are doing this loop regardless whether you're doing it yourself or the AI is doing it.
32:32Like you might even have like a thing in Todoist. Like I often put schedules in Todoist like every three days remind me to look at this thing. You're basically doing that exact same thing with AI.
32:42Like go look at Facebook ads in three days from now. You don't need to be on top of it every hour of the day. You know every day or two you do need to look back at what just happened and try different angles.
32:52And so you know if you wanna try a thousand angles as a human that's difficult. As an AI, it's pretty easy to do to just, you know, try as many variants as possible. Obviously, budget plays an impact plays a part of it as well.
33:04You need to give enough budget to each variant to sort of make a a decision as to whether it worked or not. Ellie, do you have time to show one more loop? Yeah.
33:15Like, the ultimate loop which I sort of interesting, like, product feedback loop, like like, if you actually wanted to have, like, your entire,
33:22like, business run on AI, like, just like, you know, an AI that builds itself and also gets feedback from users and then builds itself. That will be something like that. Yeah.
33:31May maybe That's really cool. Okay. So what you're saying here, I'm just looking at this.
33:35So this is really cool. This is basically you have an AI agent that's reading customer feedback, that's looking at your analytics, like your post hoc, looking at your logs, your Sentry. And based on that, it's prioritizing, it's finding out the biggest pain points, it's learning, and it's prototyping features, fixing bugs, and then it looks at the actual you know, I don't know if it's DAU or revenue.
34:02Like, I guess I mean, you you could decide. Yeah. You can decide.
34:07You know, sometimes it's NPS. Sometimes it's, you know, retention.
34:14So sometimes it's virality. So you can decide, or you can even let the agent decide. Basically say, like, for each feature, pick the best possible KPI, and maybe you have to approve it.
34:29But, you know, I think that could also make sense because there's certain features actually, the way I would think about this, Ellie, and and and tell me if I'm wrong here, I would actually do a bug loop separate from a feature loop.
34:44So the bug loop would be around
34:47Yeah.
34:49Like, uptime, you know. Like, the the the objective metric would be more around uptime and and things like that.
34:56But the product feedback loop might be around core metrics like DAU over MAU or retention or virality, stuff like that.
35:05Yeah. For sure. Yeah.
35:06I think that would be a great way to look at things. Yeah. There this this loop is sort of I it's almost like sort of the ultimate loop.
35:14It's the Yeah. Yeah. The loop may be the the lean startup loop.
35:17But everything you would do to run a business is like how can we give as much information back to the AI to sort of build itself. I think this would be like sort of a true Pulsia like a true company builder where it's like the idea and everything is like on the agent itself.
35:32I think this would be risky to do on a real business but I'm sure we're gonna see a lot of companies come out which try and do something along these lines. You just throw in a line like, hey, go build me a business and you know, that helps real estate agents. It starts building something and you know, if if it had access to enough tools to market itself to get feedback from users, you know, and that feedback might just be in the analytics or in the database or, you know, whatever it has access to, that that would sort of be be the ultimate loop.
35:59And I'm sure we'll start to see some really good businesses built like this in the next year. I've even seen like early experiments of it happening right now. I assume none are doing incredibly well but yeah like this does feel like the future.
36:11Like anything that can be done at a computer and AI can do so why can't it like even why can't it decide on its own features and, you know, experiment
36:21and, yeah, adjust its product over time the same way humans do. Okay. So we've done, you know, product feedback loop, the holy grail loop.
36:29We've done the ads loop. We've done the SEO loop. You know, Just take us home, Ellie.
36:36You know? What are other types of loops that we can we can use this for? Is the sky the limit?
36:44Yeah. I I think so. I mean, there are limitations to AI, but it does feel like every part of your business you could potentially set on a loop.
36:52You as sort of the founder of your business, you wake up every day, you know, you've got your schedule, the alarm clock goes off. . You're thinking today, how can I improve my business?
37:03It's the same for the AI. How can we get it to sort of be in that same mode? Um, you might be doing social media, video content, cold outreach, whatever it's support.
37:12All of these things that you're doing and you know checking every few hours or improving it and looking at some objective metric for example on social media. How many likes did I get? How many impressions did I get?
37:23You know, how many conversions did I get? You that all of that could theoretically be fed into the AI to help it improve and iterate on itself, learn from it, and do better next time.
37:32You know, there are definitely things here which it won't do incredibly well. I'd be skeptical that you could get an AI to get to like a 100,000 Twitter followers. But, you know, there are a lot of parts of the business where I'm certain it can have massive impact and, you know, you don't really lose anything for trying.
37:49Well, yeah. I think to me, like, you know, I wouldn't give it a loop around go find a 100,000 x followers.
38:00You know? You kinda wanna start with a smaller loop. Right?
38:03Like, the the minimal viable loop, the MVL, in the sense of first start by just creating incredible posts.
38:11Yeah. And
38:14and just optimize around the posts. And and maybe the verifiable outcome isn't a 100,000 followers, but it's 10 likes.
38:29Yeah. No. I agree a 100%.
38:31The outcome should not be a 100,000 followers. I think even for me, it would be I mean, impressions you're guessing on a post, for example, would be what what, like likes, impressions, something like that.
38:44Like it yeah. Every piece of content you put out, how well is it performing? Obviously, the number of followers should go up over time.
38:52It's difficult to go backwards. But like how many views are we getting on average per week that's sort of the metric I'd be trying to push up. And it's the same thing I do for myself you know I put out 10 tweets this week nine of them didn't do very well one did do well Why did that one do well?
39:08How can I do it better next time? And you're you're obviously great at this. You have a much larger social following.
39:14And you know, must be doing the exact same thing. So it's like, could could we get an AI to sort of run that same process itself?
39:21Ellie, thank you for coming on, for explaining Loops, for for opening our eyes, for sharing examples.
39:29I'll include links for where to follow Ellie on social media in the description, in the show notes. And Ellie, thanks again for coming on, being generous with their sauce, and I'll see you see you next time.
39:41Yeah. It's been great speaking. Thank you.
The Hook
The bait, then the rug-pull.
Loop engineering has been trending on Twitter as a buzzword for self-building software, but Greg Isenberg and Elie Steinbock use this episode to make it concrete: an AI agent given a task, an objective metric, and a stop condition can run SEO, ad testing, and product feedback as an ongoing loop instead of a one-time prompt.
How they asked for the click.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
