I Outsourced our Digital Marketing to AI. Here's What Happened

Ahrefs| 00:12:09|Mar 25, 2026
Chapters13
The team aims to boost efficiency by using AI to handle marketing tasks while acknowledging the CMO's distrust of AI content.

Ahrefs tests an AI-powered marketing assistant to generate blogs, videos, and newsletters, but AI falls short on real product knowledge and authentic voice despite huge time-savings.

Summary

Ahrefs’ experiment, led by the team’s CMO and documented by the creator, pushes an AI bot to replace core product marketing tasks: blog posts, monthly video scripts, and a mass newsletter. The goal is clear: 80% publish-ready output with minimal human touch, preserving time savings without sacrificing quality. Tim, Andre, and Rebecca shape the process by gathering a year of blog posts, video scripts, and past newsletters to train and calibrate ChatGPT’s instructions. Early drafts show promise: HTML-to-Google-Docs conversion, iterative feedback from Andre to tighten tone and add concrete examples, and a template-driven approach for the newsletter. When it comes to the video scripts, the AI-generated version initially fools some viewers, but comparing two videos reveals the human-authored version communicates product context more effectively. A critical turning point comes when the creator adapts by re-uploading documents to coax better outputs, yet the AI still lacks core product knowledge and the brand voice. Tim and Andre conclude that while AI can massively cut time (3 minutes versus about a week for all three tasks), a 60-80% threshold isn’t enough for a reliable ship—the business needs human nuance to truly convey the product. In the end, the experiment emphasizes a practical boundary: AI can accelerate work, but it can’t fully replace seasoned product marketers who understand the product and audience.

Key Takeaways

  • Training data matters: 12 months of blog posts and past scripts were used to bootstrap the AI's output.
  • Initial drafts improved after explicit style guidance and concrete examples from Andre, showing the importance of human-in-the-loop feedback.
  • AI can produce publish-ready content quickly, but risk-free deployment required iterative human review and edits.
  • Video scripts exposed a gap: AI lacked real-world product context, making the human-authored version clearer and more trustworthy.
  • A practical benchmark emerged: automating three tasks saved roughly 3 minutes vs. a week of manual work, but quality/control limited deployment.
  • A workflow hack helped recover from broken instructions: re-uploading the source document allowed the AI to regenerate the correct script.
  • Human review remains essential for branding and audience-appropriate tone, especially for complex features or technical content.

Who Is This For?

This is essential viewing for product marketers, content teams, and AI practitioners at B2B SaaS companies who want to know where AI can cut time without sacrificing brand voice or product knowledge.

Notable Quotes

""The goal isn't to replace anyone. It's to prove AI can save our team a massive amount of time without significantly lowering the bar.""
Statement of the mission and success criteria for the AI experiment.
""AI got us close, but it couldn't communicate product knowledge in our voice.""
Tim’s assessment of AI’s limitations in capturing domain-specific language.
""If it's not 80% publish ready with minimal human intervention, it's a failure.""
The success threshold for the AI-driven outputs.
""I downloaded the document, re-uploaded it, and asked it to generate the video script. And finally, we had success.""
A workaround that fixed AI output for video scripts.
""AI can save so much time, but the quality isn’t reliable enough yet for release without human touch.""
Overall takeaway on when AI is helpful vs. when humans are needed.

Questions This Video Answers

  • How close can AI come to writing B2B marketing content without a human editor?
  • What are the concrete steps to train ChatGPT for a product marketing workflow?
  • Can AI-generated newsletters and videos ever match human-crafted product storytelling?
  • What are the biggest pitfalls of using AI for technical product content?
  • How do you measure publish-ready quality for AI-generated marketing materials?
AhrefsAI in marketingChatGPTProduct marketing automationBlog post generationVideo script generationEmail newslettersBrand voiceHuman-in-the-loopContent quality control
Full Transcript
Our CMO challenged me to build an AI bot that does our marketing team's job. What we're looking for is efficiency and time savings. Our team not having to do it manually. But there's one big problem. He hates AI content. AI content is not original in any way. Skeptic. He gave me three real tasks to test it. Generate a product updates blog post, a YouTube script for a monthly product updates, and a newsletter going out to tens of thousands of people. If I could get it to 80% publish ready with minimal human intervention, it's a success. If not, I failed. The goal isn't to replace anyone. It's to prove AI can save our team a massive amount of time without significantly lowering the bar. Right away, it feels AI and AI share a voice in one view. Okay, now I'm not sure. This is how I used AI to try and replace the work of some of the world's best B2B [music] product marketers. The first thing that we're going to do is we're going to gather the different blog posts for the new [music] feature update. That's going to be the lowest hanging fruit. So, I downloaded 12 months of our product marketing team's blog posts and created a chat GBT project and let's give it some instructions. What better way to come up with instructions than to have ChatGBT generate it for you. And after a few minutes of giving chat GBT context about the project, I was almost ready to create our [music] first AI generated blog post. But I wanted to give chatbt one more big challenge. As a bonus goal, I want Tim to say that the AI article script or newsletter is better than a humans. And with that, our GPT was ready to be put to the test. I copied all the product updates from Slack, pasted them into our custom project, and sent the message not knowing how things would turn out. Come on. Okay, this has to work. Our first draft had a couple of simple errors, but after [music] updating the instructions with more details on our writing style, we were getting very close. Yeah, that's pretty good. All right, I converted the AI generated blog post from HTML to a Google doc. This is freaking awesome. I think this [music] is actually close. Okay, so I'm done the first iteration of the blog post. So, I'm gonna send these over to Andre. He's [music] in the product marketing side of things with Rebecca. I'm going to ask him to look at it and give me his [music] honest feedback. And if there's one thing you need to know about Andre, he will always tell you the truth. Is there anything in here that you feel like we could do better? Uh, it's a bit technical. Uh, [music] probably tone down the language. Explain me like I'm five. The benefits could be more specific. [music] Use examples for features that are hard to explain or hard to digest. [music] from text only. And yeah, that's the only comment I have. Andre's instructions were clear. Be more specific, and add more examples. We had a path forward. I updated the instructions using Andre's advice, spent an hour implementing it, had ChatGpt generate the HTML, converted it to a Google doc again, and just like that, we had our new draft. Oh, that's beautiful. I think we're pretty good with this one. I then wrote up some quick instructions on how to generate our post. And bing, bang, boom, we were done. At this point, I was confident Tim would see the value in how AI would save our team tons of time without sacrificing quality. Sweet. Now, I'm going to go and do the video scripts. I reached out to Rebecca and asked her to send me a year's worth of video scripts from our product updates. Then, all I had to do was [music] rinse and repeat. I downloaded them, added them to our project, and updated Chat GPT's instructions so it would know to repurpose the video script from the blog post. It's generating again. All right. So, this is good. And page of monthly product updates. We have quite a few updates to introduce this month. Let's start with site explorer. This sounds exactly the same. [music] Yeah. You know what? This is it. Let's stop this here and let's say final video script. But then the worst thing possible happened. I went to generate a new blog post and [music] everything felt broken. This is confused now. No, this sucks. The new instructions I added messed it all up. I was back at square one. Yeah, look how many things went missing. Why? I tried to rewrite the instructions to make things work properly. Over. Well, that's kind of annoying. They don't put it in a code box anymore and over. What? They don't have placeholders now for images. No. And over again. Please. It's still generating. Do I look tired? I am. Right as I was about to give up, I had a breakthrough. Instead of trusting Chat GBT to remember what was written in my blog post, I downloaded the document, re-uploaded it, and asked it to generate the video script. And finally, we had success. Yes, it looks exactly as I wanted. I think we're done. I'm so done. At this point, all that was left was the newsletter. [music] Okay, so I just finished doing the newsletter and it looks pretty good, but I did it a little bit differently this time. [music] So, because the newsletter follows a very specific format, I created a template this time and then using the content from our blog, I was able to [music] generate it quite easily. And so, yeah, I think it looks pretty good. I'm ready to show it to Tim, see which one he likes better or to see if we're close. The work was done. It was time to meet Tim where I find [music] out if the mission was successful or a complete failure. Tim, what's up? How's it going? All right. So, about a month ago, you and I talked about using AI to automate a lot of the product marketing stuffs, the blog post, video scripts, and a newsletter. And so, I'm done. And I think you're actually going to be pretty impressed. So, what I'm going to do is I'm going to show these to you. And you told me that if it's going to save the product marketing team a ton of time, then it would be considered a success as long as people aren't able to identify that the quality has gone down. Yeah. So, let's look at the blog post, read through it, and when you feel like you know whether it's human or AI, just let us know. I don't know. Right away, it feels AI. He looks unimpressed. So formatting, I don't like the formatting. I don't Yeah. Okay. The AI citations widget in overview now maps cleanly to brand radar. If a human person wrote this, we're going to need to talk with that human. What's wrong with that statement? I don't know. It just feels weird. Maybe maybe because I'm not a native speaker, but maps cleanly. Yeah. Uh, I'm not a fan. Not a fan. Not a fan. Okay. Let's look at the second blog post. Immediately for me, it is better. This feels like our regular update. It delivers the point across more clearly without using any extra words. Halfway through reading the paragraph, I already see what the person is trying to tell me. While in the previous example, It was convoluted. This second one uh choose better. I would be surprised if the first bot was a person and the second one was AI. You're right. I'm right. Yes. Tim clearly knew which one was AI and he didn't [music] like it. But would he feel the same way about the video script? What happened next surprised both of us. So I'm going to show you these two videos. If you can just Oh yeah, just hit play whenever you're ready. We have 20 updates to introduce this month. Let's start with site explorer. First up, the backlinks report now has one-click presets. These make it a lot faster to clean up link. This is us. I don't need to watch further. Show me the second one. Hey, and welcome to HS monthly product updates. We have 17 updates to introduce this month. Let's start with Brand Radar. Brand Radar now tracks YouTube. There's a new video visibility section and a full YouTube index so you can see exactly where your brand shows up across [music] high traffic. Play the first one again. Hey, and welcome to Inress monthly product updates. Page counts in overview. Now open the cited pages report and brand radar. It's a small update, but it connects organic. Okay. Okay. The the first one is AI. The second one is us. Again, same same thing like it's kind of more convoluted. It's harder to follow. It doesn't feel like the person knows the product. Well, what Tim was saying made a lot of sense. Our AI bot didn't actually know the product. It only knew what I told it. It didn't have the realworld context of how Hrefs is actually [music] used. And with a tool this powerful, that context matters. Without it, the AI struggled to write the way a marketer at Hrefs would. In the beginning, I liked the phrase one click presets and I thought that this was our copyrightiting. But then as I started looking this second video, it also was giving the information in a very efficient way. First one is AI, second one is us. Yeah, right. Damn. But but in the very first 5 seconds, you're like this is us already. Yeah. And so this is us. Yeah, I don't need to watch further. So, it seems like it's coming down to an understanding of the product. Do you think that someone reading the blog posts or watching the videos will be able to tell that anything has changed? So, it's not about them being able to tell, it's about us being able to communicate what we want to communicate. And I don't feel that AI has a goal to communicate something. Sometimes it would use the language that is overly complex and like very professional. Other times it would use the language that we don't use in our kind of professional environment. But there is no middle ground where you know you're talking to a person who understands the industry and understands what kind of language we use but also doesn't want to make it overly complicated just in case some newbies are watching. Okay, fair enough. The last thing that I want to show you is the newsletter. We have not [music] sent out this newsletter yet. So, I want you to take a look through it. Uh, okay. Right away, it says, "In this quick 5 minute video, we walk you through blah blah blah. I don't expect our newsletter to pitch the video. I expect our newsletter to pitch the product." We're also sitting here analyzing exactly what's wrong, whereas someone else who's actually looking at it would just kind of be skimming through it, right? This is literally my job to make sure that our marketing like has a certain standard like if we drop this standard that then why am I having my job? It would be great to also go through the process and see like where exactly and we can use some human touch to push it to kind of 90% good. Right now I feel it's like 60% good. I wouldn't even give it 70. So then would you consider this to have been a success? No, but it's saving so much time. Yeah, according to Andre, roughly to do all three tasks would be about a week's worth of work and this would take 3 minutes. Okay. Overall, it's not bad. I think I would agree that we would send something like that. If we can push it to like uh 8090, then I would be comfortable shipping it. AI got us close, but it couldn't communicate product knowledge in our voice. Not for something as powerful and versatile as Adri. That gap, that's literally our job.

Get daily recaps from
Ahrefs

AI-powered summaries delivered to your inbox. Save hours every week while staying fully informed.