AI Wrote Our SEO for 4 Months. It Failed.

Introduction

Lambert finished the new website and the design system, then moved onto the problem that decides whether any of it matters: acquisition. He's spent 4 months on SEO. This is the honest accounting of what that bought us.

Every number below comes from Google Search Console, pulled while writing this.

Prefer to watch or listen? Here's the full episode

The 4 months that didn't work

The pipeline we built

We started SEO in the first 3 weeks of building Kai. Right instinct, wrong plan. It was as AI-driven as we could make it: agents researched the competitor tools, agents drafted the pages, agents generated the screenshots. A human barely touched any of it.

The targets were competitor comparison and pricing pages. "Notion pricing", "Otter.ai review", "Akiflow pricing". Low keyword difficulty, supposedly the easy way onto page one for a domain with no authority.

We shipped 42 of them.

What it actually returned

42 articles, 3 months, 18 clicks.

Search Console only has data on us from 13 April, so 3 measurable months rather than 4. In that window: 26,014 impressions, 18 clicks, a click-through rate of 0.07%.

None of that was news to us. We'd known for months that the blog wasn't working. What changed a couple of weeks ago is that Lambert stopped living with it and started taking the pipeline apart.

The assumptions that turned out wrong

Every one of them, starting with the load-bearing one: that we could lean this hard on AI-generated content and still rank.

1. We underestimated domain authority. We thought a new domain would rank in 3 or 4 weeks. Turns out it takes months of consistent activity before Google ranks you at all. We ran a free-tools play alongside it, a tactic that's brought us real acquisition before, and on a domain this young that barely registered either.
2. We thought volume was the lever. 10 articles a week, on the theory that more shots means more hits. Turns out 10 forgettable pages are worth less than 3 good ones. The new pipeline spends the same hours on 3.
3. We picked the play everyone else picked. Comparison pages have low keyword difficulty: X vs Y, X alternatives, X review, X pricing. Turns out an easy keyword is easy for everyone, so we shipped what every other growth team shipped, in the same quarter, and Google had a flood of it to sort through. Reading a March article back now, it was beautiful AI slop.
4. We thought good research was enough. The agents weren't lazy. They ran real deep-research passes, pulled from Reddit and review sites, and came back with good material. Turns out all of it already existed on the web. We were reorganising what anyone could already find, so nothing in a Kai article could only have come from Kai. Google wants evidence you know the thing you're writing about, and a tidy summary of other people's experience isn't that. Lambert's read is that this is the one that cost us most.

The pipeline we run now

Three weeks ago Lambert threw it out and rebuilt it. The headline change is 3 articles a week instead of 10, all written in one focused day rather than spread across the week.

The change that matters more is what AI is pointed at. It no longer writes in Lambert's place. It researches, and then it interviews him.

1. Build the roadmap first. Claude pulls Ahrefs, finds topics with a real opportunity, and maps each one back to something Kai actually does. The output is months of slots, so no writing day starts on a blank page.
2. Deep research. 10 to 20 minutes per article, cross-referenced against our own knowledge base and product docs.
3. Claude interviews Lambert. The move worth stealing. It asks the questions: have you used this tool, who around you works this way, what happened when you tried it. He answers out loud for 20 to 30 minutes. When he doesn't know, he goes and asks a teammate who does.
4. Then it drafts, and tells him which screenshots it needs. He captures them in CleanShot, from the real product.
5. Human review, about an hour. He challenges the claims, cuts what's generic, and sends the ones he's unsure about to me so we can argue about the angle.

A day for 3 is far more human time than the old pipeline spent on 10. That's the point. Lambert onboarded onto Fyxer and Superhuman himself, took his own screenshots, and asked the team how the tools actually felt. None of that can be regenerated.

Two of those gates are worth stealing. The keyword pull can veto a piece before it's written, so a topic that can't win never gets a writing day. And nothing ships without a first-hand asset, which is the rule that would have blocked all 42 of the old pages.

We also finally have a backlink process. For 4 months there wasn't one. Not a bad one. None: no task, no owner, no step anywhere. The first exchange started this month. It's the one lever we have on domain authority, and we ignored it for the whole first run.

The same trick works outside SEO. I run my LinkedIn posts this way: dump raw material at Claude, get interviewed for 15 minutes until my actual opinion is on the table, then write. Same shape as this podcast. The conversation is the raw material, and the article is downstream of it.

Why we're investing more in SEO, not less

The obvious objection is that we're rebuilding a channel that's dying. Everyone searches with AI now. Google puts an AI Overview above everything, then Reddit, then YouTube, then ads, and your blue link is 2 scrolls down.

Our take: AI search is the biggest opportunity we have. Two reasons.

Deep beats broad. The keyword worth having is the narrow one that describes an actual workflow, not the high-volume head term. Someone who asks an assistant a very specific question about how they work, and gets Kai back, arrives already qualified. So the pages Lambert is writing this week target keywords with almost no search volume at all.

Your blog is no longer the only surface. LLMs pull from LinkedIn, Instagram, Reddit and YouTube. When I ask an AI what it knows about Kai, some of what comes back is sourced from David's LinkedIn posts. The job stops being "rank a blog post" and becomes "be present everywhere someone might look, saying the same thing." Otherwise none of it corroborates the rest.

Which is the honest answer to why this series exists. Google wants proof you know what you're talking about, and a weekly show where we walk through our own numbers is that proof. Lambert's framing, which I liked: people call it GEO or AIO, but it's still SEO. Only the practice changed.

The honest caveats

• We don't know if the new pipeline works. This is brand new, and we're still testing. What we can judge is output quality, our proxy for future ranking, and it feels better.
• The recent numbers look good and we can't take credit for them. Impressions on the old pages have more than tripled in 3 weeks, and average position went from 36 to 21. But that's the old pages, the ones we called slop, finally maturing. It says SEO takes months. It says nothing about the new approach.

What's next

Lambert is writing use-case pages during this hackathon, the backlink exchange is running for the first time, and we'll report back on whether any of it moves. If it doesn't, we'll say so here.

On the product: Kai now wears the design system we built for the website, we're in open beta for another month and a half, and 200 to 300 people are actively using it. The team is deep in the memory system, which is what makes Kai feel personal rather than clever.

The number that keeps the rest in proportion: the site went from 0 to about 2,000 visitors a month across every channel, and we're not halfway through July. In the trenches you never feel that. It only shows up when you zoom out.

Next episode, Lambert interviews me about the hackathon.

If you want to follow along, the whole series lives at How We Grow, and the changelog is the receipts of what we ship week to week. Sign up and each issue lands in your inbox, along with early access to what we're building.