You’ve published thirty articles this year. Most of them were faster to produce than anything you wrote before, because a model did the first draft and you cleaned it up. The traffic chart is flat. And when you open ChatGPT and ask it the question your best customer would ask, the question your whole business is the answer to, you are nowhere in the response.
The content isn’t bad. That’s the confusing part. You would read it and think it’s fine. The problem is that fine is the ceiling, and fine is invisible to AI. And invisible is expensive. When your buyer asks a model instead of opening a search bar, the source it names is the one that makes their shortlist, often before you knew the decision was in play.
It’s worth understanding why, because it changes what you do next. When you ask a model to write an article, it writes by averaging what already exists on the topic. It read the existing pages, most of them produced by marketers rather than the people who actually do the work, and it gave you back a clean composite of all of them. So you publish a confident summary of everything already on the internet. The model has read that summary ten thousand times. There is nothing in your article it didn’t already have, which means it has no reason to point anyone at you. You didn’t add to what it knows. You echoed it.
What a model is actually deciding when it cites you
It helps to separate two moments that get blurred together. When a model writes you a bland first draft, it is averaging what it absorbed in training, which is why that draft adds nothing it didn’t already have. When a model answers your buyer’s question in ChatGPT or Perplexity and the answer cites its sources, it is doing something else: pulling live pages at the moment of the question, reading them, and deciding which ones to quote. A citation is the output of that second moment, the model deciding that one specific source said one specific thing worth attributing. It points at the page that told it something it couldn’t get from the average. So the question stops being “how do I write a better article” and becomes “what do I know that the average doesn’t.”
Take a data migration between two business systems, the kind of project that goes sideways quietly. The average article tells you to plan carefully, audit your data, and test before go-live. True, useless, already known. The expert version says something the composite can’t: that one platform stores a customer’s full name in a single field while the other splits it into two, so unless you reconcile that before you move a single record, everything imports looking correct and every automated email goes out addressed to a blank. Nobody who hasn’t stood in that go-live morning writes that sentence. It is specific, it is earned, and it is exactly the thing a model has a reason to cite, because almost nothing else on the internet says it.
That is the whole game now. Better writing was never the lever. The lever is knowing something specific enough that a model has a reason to point at you instead of summarizing the field.
What we saw when we looked at our own data
We track this for our own category: about a hundred of the questions our buyers actually put to ChatGPT, Perplexity, and the rest, pulled every day for a quarter, with the cited sources logged. Two things stood out.
One was that the answers split hard by question type, and which side of the split you are on decides whether you have any room at all:
| The question being asked | What the models lean on | Room for your content |
|---|---|---|
| “Best agency for X,” the shortlist questions | Directories and review platforms. One directory alone showed up in over a third of the answers we tracked. Plus Reddit. | Little. The directories own this ground, and you will not out-depth them on it. |
| “How much does it cost, how long does it take, which one do I need, what goes wrong,” the questions underneath | How-to guides and articles from small and mid-size sites | This is the opening. A single strong piece can become the source the model leans on. |
The sharper finding was what separated the cited pages from the ignored ones. Take one question both answered, how long it takes to build a website, and put the two side by side. One opened with numbers a model could lift in a line: six to twelve weeks, longer for e-commerce, a revision round adds a few days, clients who come in with their content ready finish weeks sooner. The other hedged the number, narrated its own process, and padded the page with links to its other posts. The citation data tracked the difference exactly. The model pulled the vaguer page up nearly twice as often, and still cited it less than a sixth as much each time it appeared. It read both and quoted the one that gave it something to quote.
Our own results say the same thing. Our most-cited blog post is not our slickest. It answers a question plenty of competitors also answer, whether a company’s real problem is its brand or its marketing, and in that same tracking it has been cited more than a hundred times, because it does something their versions don’t. It opens by saying so: “Most articles on this topic explain what branding and marketing are. This one helps you figure out which one is actually broken.” Then it gives three named tests instead of definitions. The specificity is why it gets pulled into answers.
None of this is a law of how AI works. It is what our tracking shows right now, and it lines up with the rest: a model cites what it can attribute to you, and you earn that only by saying something specific enough to be worth attributing.
Find what the model already trusts
Start with the questions your buyers actually ask a model, not the keywords you wish you ranked for. They are different things. A keyword is “crm migration services.” A question is “what actually goes wrong when you move from one CRM to another, and how do you stop it.” Write down the real questions, in the words a real buyer would use.
Then look at which pages a model cites when it answers those questions. The tooling to see this exists now: you give it the prompt, it shows you the URLs the model pulled from. That list is the model’s current picture of your category, and you can read it the way you would have read the top of the search results fifteen years ago. Some of it is table stakes you have to cover. The more useful part is what’s thin or missing, the questions nobody answered with any real depth. Those are the ones worth your time.
Build on the knowledge that can’t be copied
The differentiator was never writing quality. It is the situational knowledge sitting in the head of whoever does the work, and the job is to get it out of their head and onto the page.
That means interviewing them, and asking the kind of question that surfaces specifics instead of summaries. Not “what’s important in a migration,” which hands you the average right back. Ask them to walk through the last one that went wrong. What did they not see coming. What changes when the client is on one platform instead of another. What’s the thing everyone forgets until it’s a problem. The stories are the gold, the specific failures and the specific fixes, because those are the parts no model already has.
You don’t need a film crew for this. You need the right questions and a way to capture honest answers. A model is genuinely good at generating the questions; it can play the sharp interviewer who keeps asking “and then what happened.” What it cannot do is invent the answer, because the answer lives in someone who has done the thing forty times.
Show up where the model is already looking
Your website is one input, not the whole game. A model assembles its answer from everywhere it has been trained to trust, and a lot of that isn’t yours: the professional networks, the video transcripts, the third-party write-ups and lists. Being a citation source means being present in those places, consistently, saying something a stranger could recognize as you.
The recognizable thing is the underlying point of view that makes your content identifiable even with the logo stripped off, not the tagline. When the same recognizable idea shows up across your own site and the places the model already reads, the model starts to associate that idea with you. That association is what gets you named.
The part that sounds like more work, because it is
This is slower than clicking a button, and that is the point. If it were fast, everyone would do it, and it would stop separating you from anyone the moment they did. The companies that win the next few years of AI visibility are the ones willing to do the unscalable thing, sit down with their expert and pull the specifics out, while their competitors keep publishing the average a little faster each quarter.
So when someone asks whether AI content is good for SEO, the honest answer is that it depends entirely on what you put into it. AI content built from AI’s own assumptions is invisible, to search and to the models. AI content built from a real expert’s knowledge, with the model doing the drafting and the human supplying the part that matters, is some of the most effective content you can make right now, precisely because so few people are bothering.
How we run it on our own articles
Concretely, on our own content the model drafts the interview questions, then whoever holds the knowledge answers them out loud, into a voice recorder, without trying to sound polished. The messy transcript is the good version. It has the asides, the “oh, and the thing nobody mentions,” the half-finished sentence that turns out to be the most useful line in the piece. Only then do we write, building from what the person actually said instead of from what the model already assumed about the topic.
The model still does plenty of the work. It just doesn’t get to make up the part that earns the citation.
None of this is new, really. The best operators always won by knowing more than the people they competed against and being willing to show it. AI didn’t change that. It made the difference harder to fake. The button gives you the average of what’s already known. Your expertise is the part the average can’t reach, and the distance between those two things is the entire opportunity.
