The five gates between your page and a citation, what to do at each one, and how to tell whether any of it worked.
Getting cited by an AI assistant means clearing five gates in order: the model must fetch your page, find something quotable on it, know who you are, see you corroborated elsewhere, and you must be able to measure whether any of it changed.
Start here if you are new to this. Each section below is the short version; the deeper pages are linked at the end of each one. Work in order — the gates are sequential, and effort spent on a later gate while an earlier one is closed is wasted.
Open an assistant and ask the question your buyer would ask — the category question, not your company name. If competitors are named and you are not, you have a gap worth closing. If nobody is named, your category is not being resolved to vendors yet and this is preparation rather than urgency. Run it a few times; output varies between sessions. The full decision test →
A model has to fetch your page during the live query, using named agents that are permissioned separately from Googlebot. Two classes matter and they are not the same: retrieval agents (OAI-SearchBot, ChatGPT-User, PerplexityBot) fetch while answering; training crawlers (GPTBot, ClaudeBot, Google-Extended) affect long-term familiarity. Blocking retrieval is the expensive mistake and it is frequently done at the CDN rather than in robots.txt, which is why the file can say yes while the edge says no. The distinction in full →
Assume the model can open the page. Is there a sentence on it worth lifting? Write the heading as the literal question, answer it in the first forty words in a self-contained statement, and use tables where the content is comparative. Marketing prose built to carry a narrative gives a model nothing to quote. The extraction checks →
The model has to decide you are an identifiable thing worth naming rather than an unattributed fragment of context. That means Organization and Person schema with stable identifiers, a description that is identical everywhere it appears, and explicit disambiguation if your name collides with something else. When an assistant describes a company wrongly, it is usually repeating that company’s own structured data. A worked case →
Models lean on sources they already trust for a category — roundups, directories, trade publications. Being named in those is outreach work rather than on-page work, and it is the slowest part of this. It is also the part competitors find hardest to copy.
Freeze a prompt set of real buyer questions. Run each prompt more than once and record the spread, not the average, because assistant output is non-deterministic and a single run cannot be distinguished from noise. Do not change the prompt set when the results are inconvenient. Our full protocol, including its limits →
robots.txt for the named retrieval agents, then verify the status code each one actually receives.That is a week of work, costs nothing, and closes the two gates that fail most often.
Every gate above has a full beginner guide: getting cited by ChatGPT, AI crawlers, writing quotable content, schema, entity SEO and tracking citations. All twelve guides →
Ask an assistant your buyer's category question and see who is named. Then check crawler access for the named retrieval agents, at the CDN as well as in robots.txt. Those two steps cost nothing and identify most of the problem.
The mechanics are not. Crawler permissions, schema and writing extractable answers are all learnable in a week. The hard parts are honest measurement — frozen prompt sets and repeated runs — and corroboration, which is outreach and takes months.
In practice they describe the same work. GEO came out of academic research and emphasises content phrasing; AEO came out of industry and emphasises retrievability and entity clarity. Anyone selling one is selling the other.
Technical fixes apply immediately — unblocking a retrieval agent works the day it ships. Citation share moves over roughly eight to twelve weeks, because models have to re-encounter and re-trust the pages.
Not for the first pass. Most of the checklist is work an in-house team can do with a browser and some patience. What is hard to do by hand is per-agent status verification and repeated multi-engine measurement runs.