Before the argument, one thing about where I’m standing.
I’ve spent most of my working life on the wrong side of a machine’s opinion. It was Google suggesting to me what to do- and I almost never listened. My niches varied from SaaS to crypto, from casino to payday loans. Categories where an algorithm decided whether I ate that month. When these systems change how they choose sources, I notice it sooner than is good for my blood pressure. Something has shifted. And the strange part isn’t that machines are answering questions about your company. It’s where they go to find out.
– Ayush
Go and type your own brand name into an AI assistant. Ask it what your company is.
Now read the answer. Then look at what it cited.
For most companies, that citation list is a set of places nobody at the company has ever logged into. A review site. A Reddit thread. A comparison page a competitor wrote to steal your traffic. Somewhere down there, if you’re lucky, your own website.
That’s an odd result when you sit with it. You wrote thousands of words describing exactly what your company does. On a domain you control. Structured properly. Updated last quarter.
The machine read it. Then it mostly went somewhere else to find out about you.
So why does it ignore the page you wrote about yourself?
First, the numbers, and who paid for them.
Almost all citation research is published by companies selling AI-visibility services. The study we’re about to use is no exception.
It comes from Omniscient Digital, an organic growth agency whose business is helping B2B software brands earn citations. Their finding is that most citations come from off your site. Which is also an argument for hiring them. Dental advice from a toothpaste company, essentially.
So I read it with one eyebrow up. I’m using it because the direction shows up in other datasets, not because a vendor said so.
Omniscient ran 240 branded prompts through ChatGPT, Perplexity, Gemini, AI Mode and AI Overviews. They collected 23,387 unique cited sources.
When somebody names your brand in a query, earned media supplies 48% of citations. Commercial content, meaning competitor pages and review aggregators, supplies 30%. Your own website supplies 23%.
More than three-quarters of what a machine says about you comes from outside your company.
That’s the puzzle. There are two obvious explanations for it, and both are wrong.
It isn’t that your site is badly built.
The first instinct in any marketing team is technical. Bad schema. Thin content. Crawl issues. Something an agency can bill for by Thursday.
That explanation runs into a problem. Your site does get cited. It just gets cited for a particular kind of question.
Omniscient’s breakdown by content type puts product pages and commercial-intent content at 12% of citations. Small. But their read of the data is that those owned pages act as the source of truth. Technical specifications, use cases, the things other sources go on to echo.
So when the question is what the product does, your page is in play. When the question is whether the product is any good, it mostly isn’t.
A crawl problem doesn’t know the difference between those two questions. Whatever is doing the sorting here does.
It isn’t that you just need more authority.
The second instinct is that the old game still works and we simply need more of it. Build authority. Earn links. Get big enough. The citations follow.
Now, this one is contested, and I’d rather show you the fight than pick a side. I don’t think either camp has it settled.
On one side, analysis of B2B citation behaviour finds that topical relevance beats raw domain authority. A niche DR30 blog covering one category properly can out-cite Forbes when it fits the query better.
On the other, a separate 2026 analysis finds most ChatGPT citations still come from domains above DR60. Sites with 32,000 or more referring domains are 3.5 times likelier to be cited than sites under 200. The reason given is mundane. High-authority sites rank better in the search engines these models retrieve from.
Both are probably true, and here’s how I’d reconcile them. Authority still gets you into the room. Relevance decides who gets quoted once you’re in it.
But notice that neither one explains the split we started with. If relevance decided everything, your own site would win everything. No page on the internet is more topically relevant to your product. You wrote it about your product.
It doesn’t win everything. It wins one kind of question and loses the other. That pattern is the thing I think we have to explain.
Your website is the reference document, not the verdict.
A model building an answer is doing two jobs. It uses two different kinds of source to do them.
For anything checkable, it needs a reference document. What the product does. What it costs. What it integrates with. Which plan includes the thing.
Your site is the best source in the world for that. You know your spec sheet best. The information comes only from you.
For anything evaluative, it needs a verdict. Is this good? Who is it for? Is it better than the alternative? Would somebody like me regret buying it?
Your site is close to the worst available source for that. Not because it’s inaccurate, I want to be clear. Because it’s yours.
Every page on it exists to make you look good. The model has read enough marketing copy to know that. A company saying it’s high quality is like a job applicant claiming to be hardworking.
So it goes and asks somebody else. Reviews. Forums. Comparison pages. The trade press. Four thousand people arguing about you in a thread you cannot edit.
And that’s the answer to the puzzle. The machine isn’t ignoring the page you wrote about yourself. It read the page. It took the facts. Then it went to find out what those facts were worth, from somebody who wasn’t being paid to like you.
You write the reference document. Everybody else writes the verdict.
What follows from that is uncomfortable.
Remember who writes the verdict. Reviews, forums, support threads.
Which means the people writing your AI visibility are product, support and sales.
Product decides what the reviews say. Support decides what the angry Reddit thread says. Sales decides whether the G2 review mentions the thing that annoyed them at renewal. Add a few thousand customers describing you in ways no one approved. That makes up most of the citation pile.
Marketing writes the reference document. That’s the one source with the most obvious commercial motive attached. Which is exactly why a model discounts it on any question of worth.
For most of my career, marketing could paper over the rest of the business. I’d say that was half the job. A weak product with strong positioning could outrun a strong product with weak positioning for years. There was a gap between what a company was and what it looked like. A lot of us made a living in that gap.
That gap is closing now. The system writing your reputation reads the parts of the internet marketing doesn’t control.
An industry is already appearing to sell you a fix for this. So let’s draw the line carefully rather than sweepingly.
We can influence what gets extracted from our own material. That part is real and documented. Structure a claim so it can be lifted. Source it. Put a real number in it. It gets pulled more often.
What we cannot purchase is the verdict. Those sources carry weight precisely because we didn’t write them. The moment a thread becomes purchasable it stops being worth citing.
And that defence is adversarial rather than structural. It holds because humans are watching those threads and reacting badly to astroturfing. Not because anything in the architecture stops somebody from trying.
Which leaves the least satisfying version of the advice. It’s the only one I can defend.
Go and find out what the machines say about your category. Which sources they pulled it from. Where somebody else’s version of you is winning.
Then go and fix the thing itself. Usually that’s the product, or the support, or the fact that nobody has any particular reason to mention you.
It’s no rocket science. Go and try it, and see what comes back.
For anything else, you know how to reach me. Though I’d like to believe it’s biological neurons reading this and not the silicon ones.
Until next time.
