AI Miracle didn’t start as a business. The three of us are family, my partner Ljiljana, her brother Marko and me, and we were following AI tools anyway. Our evenings kept turning into the same conversation: What did you try? What did it actually do? How far can you get on the free tier before it asks for a card? New tools were arriving faster than anyone could keep up with, and there was no shortage of "best AI tools" lists, but almost none of them told you whether the thing was still useful on day two. So, we put our notes on a site. The plan was simple enough: write what we really thought, and earn enough from it to pay for the tools we were testing. Every review we publish still carries a pros and cons section, and the cons are the ones we actually ran into, not the polite ones you add to look balanced. The earning part has been slower than we hoped. What did happen is that some vendors started giving us full testing accounts, so we can work with the paid version of a product properly instead of reading a pricing page and guessing. What makes us a bit different is that we are not covering AI from the outside. We run the whole business on these tools: research, editorial, SEO, visuals, development… So we see both sides. We know how much they can do when they work, and we also hit the hallucinations, the inconsistent output, the "time-saving" feature that costs an hour to clean up afterwards. We pay for a handful of tools and test far more than that. And when something is the best in its category and has no affiliate program at all, it still goes in, because the reader is choosing a tool, not a link.

We are a publication, so new business for us means new readers, and most of them still arrive through Google, but that’s starting to change. Some now come in because an AI assistant pointed them our way or cited one of our articles. For a site about AI, there’s something slightly surreal about being discovered by the thing you write about. The rest come from social and a few other places, so discovery is spread across more doors than it used to be.

Where we would like it to come from is direct and returning readers.

Right now, someone might search Google or ask an AI assistant which tool is best for a particular job and discover us along the way. The goal is for that person to eventually skip that step and think, "Let’s see what AI Miracle says."

That is a much harder thing to build. Search can introduce you to someone because you happened to answer the right question. Coming back is a different thing. That takes trust, and trust has to be earned by being useful more than once, by saying when a tool is not good enough, and by recommending the better option even when there is nothing in it for us.

In a way, that’s the transition we are trying to make now: from being found for an answer to being sought out for an opinion.

We have handed over a surprising amount of execution to AI. What we tried to hand over, and then took back, was judgment. Development is the clearest case. None of us trained as developers. Two of us studied a bit of PHP years ago and never turned it into anything we could actually use. Now we build and change things on the site ourselves, keep our work in sync through GitHub, and solve technical problems that used to stop us completely. We lean on it elsewhere too: research, when a claim can be traced back to a source, polishing our English since none of us are native speakers, shaping and structuring what we write, and generating a lot of our visual material. What we tried to automate much harder was publishing. The AI industry moves ridiculously fast and there is more news every day than a three-person team can realistically follow, so we tried to let AI carry the news coverage and the social posting to keep up with more of it. In theory it looked great. In practice somebody still had to read everything, check the sources, decide whether a story mattered at all, catch what the AI had misunderstood, change the angle, pick the right visual and decide whether we wanted our name attached to it. By the time you automate all of that safely, you have almost recreated an editor. The funniest version of that was a review where we needed screenshots and generated examples to show readers what the product actually does. The tool kept producing images that were considerably more revealing than anything we were willing to put on the site, and we spent more time than I want to admit getting examples that showed the product honestly and still belonged on our pages. AI could generate the images. It could not decide which images belonged on AI Miracle. That’s more or less where we’ve drawn the line. We’re comfortable handing AI research, code, language cleanup, or image generation. But the final choice, what is accurate, useful, appropriate, and worth putting our name on, has gone firmly back to the humans.

Yes, and it is one of the changes we watch most closely. We can already see traffic coming from AI assistants, ChatGPT included, so AI is not only something we write about or use internally, it is also one of the ways people find us. The route is different from search. On Google, someone types a tool name plus "review", or asks a very specific question, and lands on the article that ranks for it. With an assistant, a good part of the discovery and the comparing has already happened before they click through to us. The assistant summarises the options and then cites or recommends AI Miracle as one of the sources behind its answer. What we cannot see is that conversation. We know ChatGPT sent someone, but not what they asked, or why we were the source that got picked. That is the frustrating part, and it means we can no longer assume what a reader already knows by the time they arrive. It has changed something practical on our end too. Every review we publish now carries structured data, so an assistant reading it can lift the verdict and the answers to common questions without misquoting us. That used to be an SEO detail. Now it is closer to making yourself readable by the thing that decides whether to mention you at all. For publishers, I think that is a pretty fundamental change. We used to write for a person who would find us through a search engine. Now we are also writing for a person who may find us because an AI read us first. Which makes trust matter even more. Being clear, accurate and well sourced is no longer only for the person reading the article, it is also for the system deciding whether our work is worth citing in the first place.

I would tell them to use AI far more than they probably are, and to publish with it far more carefully than they will be tempted to. Use it behind the scenes. Let it research, organise information, analyse data, write code, clean up language, generate ideas, and take on the work your team has no time or no specialist skills for. For a small publisher that changes what is possible. But do not spend that new capacity on publishing ten times more content. We tried some of that ourselves and pulled back. The internet does not need another thousand articles saying the same thing in slightly different words. I would also start treating AI assistants as a discovery channel now, not in a year. Make your work easy to understand, easy to verify and worth citing. Do your own testing, write from first-hand experience, show your sources, and say something that is not already repeated on fifty other sites. What would I ignore? Most of the panic around every new acronym and every announcement that SEO is dead. Right now everyone is trying to reverse engineer how to rank inside ChatGPT, Gemini, AI Overviews or whatever comes next. Some of that matters, but I would not rebuild a publishing strategy around tricks for systems that may work completely differently in six months. If I had to bet on one thing surviving all of it, I would bet on being a source worth citing. AI can help anyone produce more. The next twelve months will be about proving that what you produce is worth finding.