Start by dropping the word "rank"
A ChatGPT answer is generated text, not an ordered list. For any question a brand is either named, cited as a source, recommended for that question, or absent — and the order in which brands appear changes between two identical runs. Tools that report a "ChatGPT rank" are reporting that order from one sample, which is not stable enough to trend. The useful question is not "how do I rank higher" but "what has to be true for ChatGPT to name and recommend us" — and that has four parts, in a strict order. Working on part three while part one is broken is the most common way to waste a quarter.
1. Let the right OpenAI agent in
OpenAI documents three agents with different jobs. OAI-SearchBot builds the index that surfaces websites in ChatGPT's search features — this is the one that decides whether you can appear at all, and it is the agent sites most often block by accident while meaning to opt out of training. GPTBot crawls content that may be used to train models; blocking it is a defensible position and does not remove you from ChatGPT search results. ChatGPT-User fetches a page because a person asked, and may not follow robots.txt. Name each agent explicitly in robots.txt instead of one wildcard aimed at OpenAI, then verify by fetching your own pages as each agent: a CDN or WAF rule can return 403 to a named bot while the file says Allow. Agent names and IP ranges: OpenAI bots documentation. See also OAI-SearchBot.
2. Make the answer liftable
ChatGPT assembles a reply from passages, not pages. The test for any page is: if a hundred words were cut out of it and pasted into a reply, would they still be true and still make sense? Pages fail that for structural reasons — the sentence that answers the question depends on the heading above it, arrives after three paragraphs of preamble, or lives behind an accordion that only renders after JavaScript runs. The rules are short: name the subject in the sentence carrying the fact, answer in the first sentence under each heading, put facts in tables with units and dates, and keep the substance in server-rendered HTML. The full house style is in writing so an answer can be lifted out.
3. Make the brand an unambiguous entity
To name you, the model needs a stable idea of who you are. One canonical name used identically everywhere, one canonical one-sentence description, one canonical URL form, and an Organization block in structured data whose sameAs links point at your profiles elsewhere. A brand spelled three ways across its own site, or colliding with an unrelated company, has nothing stable to be named by.
4. Get the claim corroborated somewhere you do not own
A recommendation is a claim the model has to stand behind, and a claim sourced only from the vendor's own marketing page is the weakest kind. What moves recommendation — as opposed to mention — is the same claim appearing somewhere the model treats as independent: a directory listing with every field completed, a comparison page you did not write, a community thread, documentation written by a customer. This is the slow layer and the one that decides outcomes; see third-party corroboration.
How to know whether any of it worked
Fix a set of twenty to forty buyer questions with no brand names in them, run each more than once, and record mentioned, cited and recommended as three separate rates. Keep the raw answers. Two consecutive runs moving in the same direction is a finding; one run is noise. The method is spelled out in brand share in AI answers, and why a ChatGPT rank tracker measures something that is not there explains what to ignore. Nothing here promises a result — these are the conditions under which a result becomes possible, and the order in which to establish them.