There is a moment, and it arrives without warning, when a competent executive at a competent company types the name of the company they run into ChatGPT and receives, in return, a polite paragraph that is confidently, breezily, mostly wrong. The founding date is off by two years. The product is described in the vocabulary of a competitor. A funding round that closed last April is either missing or attributed to a different firm entirely. The tone of the paragraph is the tone of a friendly waiter describing a dish he has heard about but never tasted.
This is not a rare event. It is, at the moment, the default event. It happens because the model is not being lazy and it is not being adversarial. It is doing exactly what a language model does, which is compressing whatever public text has been written about your company into a plausible summary. If the last thing anyone wrote about you was a funding note from a year ago and a warm profile in a regional newspaper from before the pandemic, that is the raw material the model has, and the summary you get back is going to look like it.
If the last thing written about you is a funding note from a year ago - that's your whole story.
The Cadence Problem, Priced In
The unpleasant thing about how large language models decide who to cite is that it is not, mostly, a quality question. It is a cadence question. A model reading the internet in 2026 encounters Salesforce roughly every day, because Salesforce publishes roughly every day. It encounters Microsoft several times per day, because Microsoft is essentially a wire service that occasionally makes software. The company that publishes once per quarter shows up once per quarter, which, in a corpus measured in petabytes, rounds to zero.
This has been a known dynamic in search for two decades. The novelty is that the machines have stopped forwarding the reader to your website and started answering on your behalf. The old bargain, in which you optimized a page and Google sent a curious human to read it, has been quietly rewritten. The new bargain is that the model reads the page for you, forms a view, and delivers that view to the human directly. If you are not in the training corpus and not in the retrieval index, you are not in the conversation. You are, functionally, a rumor.
A Short Taxonomy of the Invisible
Companies vanish from AI answers in a few characteristic ways, and it helps to name them, because the fixes are different. There is the freshly-launched company, which is invisible because it is genuinely new; the corpus simply does not know it yet. There is the quiet middle-of-the-market company, which has been operating for eight years, is profitable, and has decided that press releases are undignified. There is the rebrand-in-progress company, which has changed names and now competes, mostly against itself, in the models' memory of who it used to be. And there is the heavily-covered-but-inconsistently-described company, whose problem is not silence but a kind of narrative static: everyone talks about it, no two people agree on what it does, and the model averages the noise into a description its own founder would not recognize.
Each of these has a recognizable smell in the ChatGPT output. If your paragraph is short and hedged, you are quiet. If your paragraph confidently describes a business you exited two years ago, you are stale. If your paragraph is fluent nonsense, you are being averaged.
The Four-Move Loop
The mechanics of getting cited are less mysterious than the acronyms suggest. GEO, AEO, LLMO - the industry has, in the manner of industries, minted three letters for what turns out to be one behavior: publishing often, being cited by other people, and structuring the resulting pages so the machines can retrieve them cleanly. The loop, stripped of jargon, has four moves.
- Publish. Weekly, ideally. Long enough to be substantive, short enough to be finished.
- Get cited. Journalists, analysts, adjacent operators. Citations behave, for the models, the way backlinks behaved for Google in 2007.
- Get retrieved. The models prefer deep pages, fresh timestamps, clear headings, and prose that answers the question a human actually typed.
- Repeat. Compounding is not a metaphor here. It is the mechanism.
The trick, and it is a trick that catches out serious operators, is that none of these steps produces a satisfying result on its own. Publishing once is not publishing. Getting cited once is a nice email. Getting retrieved once is a coincidence. The loop only works when it closes, and it only closes when it runs long enough that the corpus starts to notice the shape of a voice.
A note on the honest version
There is a dishonest version of this loop, and it involves generating a great volume of low-density text and hoping the models cannot tell. They can tell, or if they cannot tell today they will be able to tell by the next model release, and the reputational cost of being caught contributing to the sludge is higher than the cost of being quiet. The honest version of the loop involves a real person, at your company, saying real things, on a schedule, in public. There is no shortcut, and the ones being sold are, in the technical sense, scams.
Every company is becoming a media company - whether it wants to or not.
Why This Feels New When It Isn't
The uncomfortable joke about generative search is that it rewards the same behavior a regional newspaper rewarded in 1955: showing up, on time, with something to say, every single week. The stack has changed. The economics have changed. The distribution has changed. The muscle is identical. What is new is that the penalty for skipping a week is now enforced by a system that has no sentiment about you, will not answer a call from your PR firm, and cannot be persuaded over coffee. It just quietly downgrades your position in the answer.
This is the part that tends to alarm executives when it lands, which is the correct response, because it is alarming. The old media landscape had gatekeepers, and gatekeepers could be worked. The new landscape has a retrieval index, and a retrieval index can only be fed. There is no lunch to buy, no relationship to warm up, no favor to trade. There is only the question of whether the last time you published was recent enough and interesting enough that the machine would rather quote you than guess.
What to Do on Monday
The intervention, mercifully, is boring. Pick a cadence you can hold - weekly is the honest floor, monthly is the honest ceiling. Assign one person to own it, because content owned by a committee is content owned by no one. Publish under a real byline, ideally the operator's, because the models are getting better at distinguishing house voices from anonymous marketing prose. Structure each piece around a question a human might actually ask an AI. Link out generously; the corpus reads outbound citations as a signal that you have read the neighborhood. Do this for a year. Then look at your own name in ChatGPT again.
The paragraph will be different. The founding date will be right. The product description will use your vocabulary. The funding round will land in the correct April. And, in a small way that matters more than it should, when a buyer types your category into a language model at 11pm on a Tuesday, the answer will contain you.
Fifteen Ways to Say the Same Thing
Because this argument is unusually resistant to being made once, an incomplete list of ways it has been rephrased, largely by people who have watched a competitor's paragraph get longer while their own got shorter:
The Small, Quiet Point
The small quiet point underneath all of this is that being a great company that nobody has heard of used to be a viable business model. It was even, for a while, a kind of aesthetic. The machines have made it obsolete. Not because the machines are cruel, but because the machines are, in the end, an index of what has been said. Say nothing, get indexed as nothing. Say something, at a rhythm the corpus can hear, and you become a company that exists to a language model at 11pm on a Tuesday. Which, increasingly, is a decent proxy for whether you exist at all.