Why “rank” is the wrong verb, mechanically
A classic search engine computes an ordered list for a query and shows everyone roughly that list; rank is a real property of that system. ChatGPT composes an answer. When it answers from model weights alone, your brand’s presence depends on how it appeared in training data nobody can inspect or petition. When it decides the question needs fresh information, it runs a search, fetches a handful of sources, and writes a synthesis, sometimes with citations. Two runs of the same prompt can search differently, select differently, and phrase differently.
That variability is not a measurement problem to engineer away; it is the system’s actual behavior, and any vendor or consultant promising a “number one spot in ChatGPT” is selling a property the system does not have. The honest reframe: stop asking “where do we rank” and start asking “in a fixed, recorded sample of buyer questions, how often do we appear, as what, and sourced from where.” That question has an answer you can defend, and the difference between the two questions is the difference between GEO and SEO as disciplines.
The four layers you can actually work on
The honest version of how to rank in ChatGPT maps onto four layers, and the order matters because each depends on the one before it. This is the same structure as the four-layer model of AI search visibility; here it is applied to ChatGPT specifically.
| Layer | The question | The work |
|---|---|---|
| 1. Crawl access | Can OpenAI’s crawlers fetch your pages at all? | robots.txt review, rendering check, firewall and bot-management rules |
| 2. Source eligibility | When ChatGPT searches, is your content usable as a source? | Indexable, specific, attributable pages that answer buyer questions |
| 3. Representation | When you appear, are the facts right? | Consistent entity facts everywhere the systems read |
| 4. Measurement | Can you show any of this changed? | A frozen prompt panel, recorded runs, honest deltas |
Layer 1 is where cheap, real wins live. OpenAI operates distinct crawlers with distinct jobs, and sites regularly block the wrong one for the wrong reason. OAI-SearchBot fetches pages for search-backed answers with links; GPTBot collects content for model training; ChatGPT-User acts on live requests a user triggers. Each is documented, with IP ranges, in OpenAI’s bot documentation. Blocking GPTBot is a policy decision about training; blocking OAI-SearchBot is a commercial decision about being findable, and inheriting one from the other is the single most common self-inflicted wound in this space. The full separation, and how to audit it, is in OAI-SearchBot vs GPTBot.
Layer 2 is mostly work you already understand. When ChatGPT searches, it needs what any search-backed system needs: pages that load without executing a fragile JavaScript path, say one specific thing checkably, and carry the facts near the claim. OpenAI’s own publishers and developers FAQ frames appearing in search-backed answers in terms of being crawlable and useful as a source, not in terms of any special markup. There is no secret ChatGPT meta tag, and anyone selling one is charging for Layer 2 SEO hygiene with a new label.
Layer 3 is the layer buyers feel. A brand can appear in answers and be described wrongly: an old price, a retired product, a competitor’s feature attributed to you. Assistants assemble descriptions from whatever sources they trust, so contradictions between your site, your directories, and third-party profiles become contradictions in answers about you. The fix is unglamorous: one set of entity facts, kept identical everywhere the systems demonstrably read.
Layer 4 is what makes the other three billable. Because answers vary, single observations prove nothing in either direction. A defensible measurement is a frozen panel of buyer-decision prompts, run on schedule, with answers recorded before anyone changes anything; the method is specified in how to build an AI visibility prompt set, and the reporting discipline in how to measure AI search visibility.
If it cannot be scored, it is a sloganBefore spending anything on ChatGPT visibility, score the protocol you would actually run against 47 published checks covering scope, access, prompts, evidence, and handoff. Two minutes, no account, and it grades the method rather than the brand.
Score your protocol freeThe myth table
| The claim you will hear | What is actually true |
|---|---|
| “We got you to #1 in ChatGPT” | There is no persistent position to hold; an answer observed once is one observation, not a ranking |
| “Add this schema and ChatGPT will cite you” | No markup is documented to control assistant citations; structured data helps machines parse facts, and cannot compel selection |
| “Block GPTBot, it’s stealing your traffic” | GPTBot affects training, not search answers; blocking OAI-SearchBot is what removes you from search-backed answers, and the two are separate decisions |
| “ChatGPT hallucinated your brand, nothing you can do” | Representation often traces to real, fixable source contradictions on pages the assistant can fetch |
| “Our dashboard shows your AI visibility score” | A single blended score hides which layer moved; ask what was observed, on which prompts, from which accounts, how often |
What an honest ChatGPT engagement promises
Written scope for this work, whether internal or sold, should promise exactly four deliverables: a crawl-access audit with the specific rules found and fixed, a source-eligibility review of the pages that answer buyer questions, a representation baseline recording how the brand is currently described against approved facts, and a measurement panel with its freeze rules and schedule. It should decline, in writing, to promise rankings, citations, traffic, or revenue, because no one controls the model’s selection.
That refusal is not modesty; it is the commercial position that survives. The buyer who wanted a ranking guarantee will eventually get one from someone, watch it fail to materialize in any verifiable way, and remember who declined to sell it. Meanwhile the four deliverables above produce observable before-and-after evidence: rules unblocked on a date, contradictions reconciled on a date, appearance rates in a frozen panel moving between observation windows.
This page cannot guarantee that any of it makes ChatGPT mention you, and neither can anyone else; the model’s selection is the model’s. The work above changes what the system can fetch, verify, and quote about you, and it produces the records that show whether the needle moved. In a market full of ranking promises, being the one with records is the durable advantage.