You cannot rank in ChatGPT, because there is no ranking. There is only presence: when a buyer asks a question in your category, the assistant names a handful of companies, and you are either among them or you are invisible. Presence can be earned, and the way to earn it is more knowable than most of what gets written about it. This page is the honest version.
How assistants choose the brands they name
Two mechanisms, and the difference matters for everything you do about it.
From memory. The model learned about brands during training. If your company is written about consistently across the public web, the model may know you the way it knows any fact. Training happens in cycles, so this moves slowly, and nothing you publish today changes it this quarter.
From retrieval. For questions that sound like research, assistants search the live web, read a few pages, and compose an answer from them, usually with citations. This is where presence is winnable on a timescale of weeks, because the assistant names whoever the pages it retrieved happen to support. Roughly a third to two-fifths of cited sources change from day to day. Unstable means open.
The six things that move presence
- Be retrievable. If Google has not indexed a page, no assistant can cite it. Boring, foundational, checkable in Search Console today.
- Answer one question per page, canonically. Assistants lift passages, not sites. A page that opens with a clean, self-contained answer to one specific buyer question is liftable; a clever page about three things is not.
- Say what you are, plainly, where machines look. Your homepage, your about page, and consistent naming everywhere. When an assistant is asked about you by name, its answer comes from what it can retrieve about you. Empty or contradictory pages produce empty or wrong answers, and those lose deals you had already earned.
- Exist where assistants already read. Answers to "best X for Y" questions lean heavily on comparison pages, directories, and industry lists. Being accurately listed in the places that already get cited is faster than convincing the assistant to cite you directly.
- Publish facts nobody else has. Original numbers, real methodology, dated statements. Generated answers reach for concrete, citable claims the way journalists do. Ten adjectives lose to one measured figure.
- Write the comparison yourself. Buyers ask assistants "A vs B" constantly. The vendors named in those answers almost never wrote the comparison; a third page did. Be the third page, honestly, with a table.
What does not work
Keyword stuffing for machines, invisible text, "AI optimization" tricks, and publishing thirty thin pages in a week. Retrieval favors the same things careful readers favor. There is no technical shortcut that outruns being the best specific answer, and vendors selling one are selling weather. What the honest end of that market sells, and what it actually charges, is documented in answer engine optimization services, priced and decoded.
How you know it worked
One conversation with ChatGPT proves nothing: answers vary run to run, day to day. Measurement means asking the same buyer questions repeatedly across a window, counting mentions, and comparing windows, with the sample size stated. That discipline is what a GEO audit is: the baseline of where you stand before you spend a quarter trying to move it, and the ruler afterward.
The short version: get indexed, answer real questions one page at a time, make your own identity legible, put facts where machines read, and measure like you mean it. That is how you "rank" in a system that has no rankings.