To be recommended by an AI assistant, your company has to be easy for a model to describe: named clearly, doing something specific, for someone specific, in a place that can be checked. That is the whole job. Below is what it looks like in practice and how to test where you stand today.
The industry calls this work GEO, short for generative engine optimization. We describe it in plain words, as visibility in AI assistants, because the acronym explains nothing to the person who has to decide whether to pay for it. It is the same thing.
Your customers increasingly skip Google and ask an assistant instead. They get one answer and, with it, a short list of companies worth contacting. If your business is not on that list, you are losing clients who never learn you exist.
TL;DR
Visibility in AI assistants means being named by ChatGPT, Gemini or Perplexity when someone describes the problem you solve. In SEO you compete for a place among ten links; here you compete to be mentioned at all, because an answer usually names two or three companies. Three technical conditions raise the odds: JSON-LD structured data, answer-first writing and E-E-A-T signals. First effects show after 30-45 days, fuller visibility after 60-90, and the measurement has to be repeated, because models reread their sources on their own schedule.
How this differs from SEO
SEO is a ranking problem. Google weighs a page against hundreds of signals and returns a list of links; the user picks one. An assistant works differently: the model reads the question, synthesises what it can find and writes a single answer. Often it names specific companies.
The shift matters. In SEO you compete for a position on a list. In an assistant answer you compete to be mentioned at all, and models cite only sources they can read and have reason to trust.
- SEO: the user sees ten links and clicks one
- Assistants: the model gives one answer and names two or three companies
- SEO: you optimise for keywords and links
- Assistants: you work on whether the model understands your context at all
How models decide which companies to name
Large language models do not crawl the web afresh for every question. Part of what they know comes from training; tools such as Perplexity, or ChatGPT with browsing, also read pages as they go. So an assistant can visit your site and decide whether it is credible enough to pass on to a user.
It weighs several things at once: whether the page is technically readable, whether the text answers the question directly, and whether other trustworthy sources say the same thing about you. A model does not recommend websites. It recommends facts it can find in more than one place.
Three conditions for being cited
An analysis of AI-generated answers[1] points to three technical conditions that raise the chance of being cited:
- JSON-LD and structured data: Schema.org markup tells an assistant what your company does, where it works and what it sells. It is a business card written in the format a model reads natively.
- Answer-first writing: lead with the answer, not the preamble. Models quote fragments that address the question immediately, so each paragraph should stand on its own.
- E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness): author bios, documented experience, client references and real cases are the signals a model uses to decide whether you are worth repeating.
How to check where you stand today
The simplest test takes a minute. Open ChatGPT and ask what your customer would ask: which agency handles lead generation for small businesses, or who builds websites that AI assistants can actually read. If your company does not come up, you have a visibility problem. Repeat the same questions in Perplexity and Gemini, save the answers verbatim, and run them again a month later.
Save the wording, not your impression of it. Two assistants asked the same question give different answers, and the difference between them is the actual finding.
What to do to start appearing in answers
This is continuous work, not a one-off fix. Where to start:
- put the company facts in order first: name, base, service area, services, people
- implement structured data (Schema.org): LocalBusiness, Service, FAQPage
- rewrite the pages that matter in the answer-first format
- expand the FAQ on every service page, using questions clients actually ask
- build mentions in independent sources: trade articles, directories, reviews
- repeat the measurement on a schedule instead of checking once and assuming
Summary
- Being named in an assistant answer is a different game from ranking in Google: there is one answer, and you are in it or you are not.
- An answer usually names two or three companies, so the goal is to be mentioned at all.
- The three conditions for being cited are structured data, answer-first writing and E-E-A-T signals.
- The cheapest test is asking an assistant the question your client would ask, and saving the answer word for word.
- First effects show after 30-45 days and fuller visibility after 60-90, because models reread their sources on their own schedule.
- Nobody controls what a model answers. What can be controlled is whether the facts about you are coherent and verifiable.
Want to know whether assistants name your company today? The visibility test takes 48 hours and gives you the answers verbatim, not a score.
Sources
- Aggarwal P., Murahari V., Rajpurohit T., Kalyan A., Narasimhan K., Deshpande A., "GEO: Generative Engine Optimization", KDD 2024 (Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining), 2024. Open version: arXiv:2311.09735