How Our AI Content Pipeline Writes Expert Articles Without Losing the Expert

Is outbound really dead, or just badly targeted?

More and more often I hear how marketing is the new B2B sales. In other words, the new deals will be generated by prospects researching the market and contacting sellers rather than the other way around. Exactly like B2C model functions.

The underlying hypothesis is simple - “outbound is dead”. I don’t know about you, but I get cold calls weekly, with aggressive salesmen trying to sell me life insurance or telecom services. With cold emails it is even worse - I get those daily. And they all sound the same. Sometimes even the competition offers us the very same agentic AI services we provide.

I sometimes used to joke with them answering: “Please provide the contact details of decision makers in your largest client companies".

But nowadays the process is so machine driven that instead of human getting the joke I get ChatGPT generated answer:

“We cannot provide the contact details due to bla bla, but we’d be happy to share some case studies”

Here I would like to note that I personally DISAGREE that outbound is dead. In my example, the spammers made a couple of mistakes. Poor targeting for their message. Or we can put it other way around - poor messaging for their targeting. If you message the competition, you should offer them some sort of partnership. Rather than trying to implement their very own services under their roof.

The reasons why people do it are beyond the scope of this article. However, targeting, messaging and timing CAN all be drastically improved. Maybe we’ll share a secret or two in one of our next articles, but for now, let’s get back to the inbound channel we discuss here - marketing!

How do buyers find you when they research your services?

As you can conclude from my introduction, we want to get contacted by the people researching products and services we provide. In order to do that, they must first find us. There are several ways we maximize chances of that happening in the digital World. We are going to discuss two: SEO and GEO.

SEO stands for Search Engine Optimization and does exactly what it says. When people research our services, we want to be on the first page of search engine results. If possible, in the very first place.

My simplified recipe for how to do that:

What does a working SEO checklist for a B2B site contain?

1. For starters, you should have the website in the first place. It should be professionally designed and working properly. Otherwise we send very wrong messages to our visitors. It should also emit authority and target multiple keywords. In other words, it should be rich in content. We will cover this part shortly.
2. You should technically optimize the site, so the crawlers can read it properly. This is where typically IT companies will obsess, optimizing the page to perfection on tests like PageSpeed insights. Sadly, this is where many stop as well.
3. You should select the keywords for which you compete. Keywords are search phrases users enter in search engines. The keywords you choose should have decent traffic but moderate competition. You don’t want to compete with blue chip vendors, whose sites carry enormous authority.
4. Next, you should create high quality content which includes the selected keywords in a natural way. Before, people used to repeat their keyword endlessly in order to cheat Google. Nowadays, they replace it with hyperproduction of AI generated low quality content. Not so surprisingly, the guys in Google are smart and work hard on penalizing such deceitful tactics. So the content should still be original. And yes, did I mention quality?
5. Last but not least, you should get high quality backlinks. Backlinks are links from 3rd party websites pointing to your site / content. Again, people try to cheat the search engines by buying a lot of links for little money. This will NOT improve your site rankings. Links from sites of blue chip companies, universities and reputable conferences will. Think of it in this way: backlinks from high authority sites improve your own authority. Links from spammy sites - I guess you can finish the sentence yourself.

Is SEO still the foundation for Generative Engine Optimization?

When it comes to GEO (Generative Engine Optimization), it is a super hot topic right now. Basically, it is optimizing for being mentioned by LLMs when people are researching services or products you provide. Much like outreach, many people claim SEO is about to die. Nobody will Google anymore - you will ask ChatGPT instead. At least so they say.

Trends are here - people are using AI for research more and more. But still, SEO is the foundation for GEO. Let me give you an example of our own.

What happened when I asked ChatGPT who builds ETL migration products in Croatia

We have recently completely re-engineered our website. New technology, new design, new concepts. Google understood this is a new site and needed to reindex our content.

I asked ChatGPT who in Croatia has ETL platform migration product, knowing our EPIC Ingenioso is the only of its kind. It said there are no such products in Croatia, but there are companies providing platform migration services. Interesting for our topic, ChatGPT mentioned 5 companies. I heard about only one of them, despite being deep in the Croatian IT scene. If this doesn’t convince you how important SEO / GEO optimization is, nothing will.

But let us get back to why ChatGPT didn’t mention us. I asked ChatGPT why - and the reason is simple. We were only on the third page of results for the search term it used. So AI is as superficial as we humans are - nobody searches beyond the first page, unless being forced to.

As soon as the service page got indexed, ChatGPT mentioned us and only us in my next inquiry. No more never-heard-of companies. Conclusion is simple - if you want people researching your services to find you, you have to do SEO and GEO.

Why is expert content so hard to publish regularly?

How did it work for us before agentic AI? We wanted to have high quality, original, expert content. So the experts wrote it, not marketing. I myself probably took the brunt of it.

The main problem? Time. The content was short and often published in bursts rather than regularly. Keyword research and back linking was done by our marketing, but it was also time consuming. Other, more urgent matters often took precedence.

When it comes to video blogging, it was even worse. Filming me narrating took a lot of time and repetitions - well, I am not an actor. :-) Editing the materials took even more time for our marketing colleagues. So video publishing was even less regular than blog posts.

Let us now examine how our AI-powered content generation pipeline addressed all these problems.

How does an AI interviewer get an article out of a busy expert?

The first agent in our AI content creation pipeline is an interviewer. I give it whatever I have – sometimes a few Croatian bullet points, sometimes half an article like this one. It reads what I already wrote, works out what is missing, and asks me about it. In English, of course.

An interview takes up to an hour. Writing a 2,500-word article by hand would take me significantly longer. But honestly, time was only half of the problem. The other half was inspiration. Staring at an empty page after a long working day is not how good content gets written. Answering questions about something you built is a different story.

What surprised me was how tough the questions were. The interviewer quite often asked the kind of question I would rather not hear. It turned out to be great preparation – a real prospect asks exactly the same questions when I pitch the solution, and now I have already answered them once. It also gave the articles a more critical, more realistic view. In my opinion, that beats just showing off every time.

The interviewer’s most valuable output is not the draft, though. It is the fact ledger – every number and claim I stated, together with where I stated it. Every agent after it is checked against that list. If a number is not in the ledger, it does not get published.

Can two AI models fact-check each other?

Next comes the technical editor. Its job is to check the facts and to remove sentences that sound good but say nothing. Marketing fluff, basically.

Here we use a trick. Two models from two different vendors, one from Anthropic and one from OpenAI, review the draft independently. Only the corrections both of them agree on are applied automatically. Everything else lands on my desk as a report.

Why bother? Because one model alone is confidently wrong too often to be trusted unsupervised. And if one of the two is unavailable, the editor applies nothing at all. A single model’s opinion is a suggestion, never an edit.

How does an AI agent pick keywords people actually search for?

Then comes the part our marketing used to spend so much time on – keyword research and SEO / GEO optimization.

Remember my recipe from above? The agent follows it. It pulls real search volumes and competition from Google Ads Keyword Planner, checks who is on the first page of results, and skips keywords where blue chip vendors dominate. The goal is a keyword we can realistically bring to the first page. Ideally, to the first place.

Real data is not a detail here. Before we got access to Keyword Planner, the model picked keywords using its own judgement. They all sounded perfectly reasonable. When we priced them later, six of seven primary keywords had zero search volume. Zero. “Travel expense report automation”, for example – a phrase that describes our article perfectly, and that nobody types into Google. LLMs invent search volumes with total confidence. That is exactly why our agent does not guess. It asks Google.

How do you structure an article so a generative engine will cite it?

Keywords are only the beginning. The agent plans the headings so that every section answers exactly one question a buyer might ask, and still makes sense when it is cut out of the article. That is exactly how generative engines quote content. It adds an FAQ with real long-tail questions, picks the quotable numbers and chooses which of our service pages to link.

Then it does something I really like. It asks both models, independently, whether they would cite this article when answering each of those buyer questions. Where both say no, we know we still have work to do. Finally, it lists backlink opportunities, weighted towards directories and industry roundups – because those are what AI models cite.

How do you keep your own voice when AI helps you write?

After SEO come styling, branding and verification.

The style agent rewrites the text in my voice, learned from the articles I have written before. It splits the sentences that run too long and quietly fixes the typical mistakes of a non-native speaker. For articles like this one, where I wrote most of the text myself, we skip that step altogether.

The brand agent does something marketing people will appreciate – it keeps brand mentions in check. Two to four per article, and each one has to prove a point. Not too much, not too little. A few links to relevant services, one call to action at the end, and that is it. We deliberately separated it from styling. When one agent did both, it kept inserting marketing phrases, which the technical editor then kept removing.

A small example of AI going a bit crazy. At one point my articles started describing me in the third person. On my own blog. Instead of “I”, the text talked about some mysterious third person who had apparently written my articles. The reason was banal: in the interview transcripts, my answers were labeled “Author”, and the model simply echoed the label it was given. Typical of what you run into when developing agentic workflows – the agent does what you told it, not what you meant. Once we found the cause, the fix was simple.

What stops an AI pipeline from publishing something wrong?

Verification is the last automated step, and here no AI decides anything. Eight checks, all plain code. Is every fact from the ledger still there? Does every number trace back to something I actually said? Are the keywords present, but not stuffed? Is the heading hierarchy correct? Did anything from our confidential internal documents leak into the text? Is the company positioned correctly? Are internal figures kept out of the title and the opening? Is the article ready to publish? The same article always gets the same verdict.

Why is human review still the last step?

Then it is my turn. The review takes about an hour.

What do I fix? Mostly small inaccuracies. For example, our coding agent article talked about the “first line ticket resolution rate”. But we were measuring the resolution rate of all tickets, not only the first line. So “first line” was not only slightly misleading, but also not understandable for most readers. So I took it out.

My corrections are stored as a list of find-and-replace pairs and applied at the very end, when the article is exported. If a correction no longer matches anything in the text, the export stops instead of quietly skipping it.

Could an agent take over that hour as well? Perhaps one day. But I am the one signing the article, so I am the one who approves it.

Can an AI content pipeline publish straight to your CMS?

Our website runs on Sanity. The last step produces an HTML version of the article, which I open in a browser, copy and paste into Sanity’s editor. Sanity turns it into proper headings, lists, links and tables. Pasted markdown would just show up as a bunch of hash signs. Next to it, I get the field values – SEO title, meta description, related services – ready to paste.

Automatic publishing to Sanity is possible, and it is the next logical step. We have not implemented it yet. Since human review is needed anyway, there was no rush.

How much does an AI content creation pipeline cost to run?

Here comes my favorite part. In model calls, one article costs us between about 1.35 and 6.60 US dollars. Not per hour – per article. Where an article lands in that range depends mostly on how often we re-ran a stage while tuning it.

Compare that to the expert and marketing time an article like this used to take. That is what I mean when I say AI agents can be dramatically cheaper than human labor. If they are developed properly, that is. Developing them properly is exactly what we do at VIS Solutions – for our clients and, as this pipeline shows, for ourselves.

And the development itself? Building the pipeline, including fine tuning it on the first five or six articles, took two weeks during summer vacation. Not full time – rather part time work during hot summer afternoons.

By the way, this very article went through the same pipeline – as did every showcase article on the VIS blog this summer. The interview was a short one this time, since I had already written most of it myself.

Writing expert articles by hand versus running them through the pipeline

What comes next after written content?

If you like what we did with written content generation, subscribe to the newsletter to see what we did with video. We will reveal a secret or two, so stay tuned.