What is generative engine optimization and how do I get started? An agency primer

Generative engine optimization, or GEO, is the practice of making sure AI systems mention your brand, cite your pages and describe you correctly when someone asks a question in your market. The systems in question include ChatGPT, Perplexity, Gemini, Claude, Copilot and Google's AI Overviews and AI Mode. People also call it AEO, AI SEO or LLM optimization. We do not fuss over the name.

We get asked how to start almost weekly. This is the version we give prospects, and it is the same one we follow on client programs, which we run on Promptwatch.

The definition we use with clients

We define GEO by outcomes, in four parts:

Visibility. The brand appears in answers to the questions that matter.
Citation. The pages cited as sources include ours, not only competitors' and third parties'.
Access. AI crawlers can reach and read the pages.
Accuracy. The answer describes the product correctly.

A fifth part is the outcome: visits, leads, sales. But that depends on the other four, and it lags behind them.

How it differs from the SEO work you already know

Most SEO fundamentals still apply: crawlable pages, useful content, credible sources. Google's Search Central documentation on AI features is the official reference for how its AI experiences treat publishers, and we send clients to it directly.

What changes is the unit of work. In SEO we think in keywords and positions. In GEO we think in prompts and share across many runs. A buyer asks a long question, the engine splits it into several searches behind the scenes (query fan-outs), and the answer pulls from sources that are often not the brand's own site. Forums, review sites, news and video carry weight. And the same prompt can return a different answer tomorrow, so a single check means little.

What we do in the first month

Week one: build the prompt panel

We gather 30 to 50 prompts from the client's own sources: sales calls, support tickets, Search Console queries and People Also Ask. We group them by topic and by persona and add competitor comparison prompts. Promptwatch's Prompt Explorer helps by suggesting prompts from Search Console, ranking keywords or any page, and Prompt Volumes gives search volumes and difficulty. We edit them by hand afterwards.

Week two: baseline

The panel runs daily per model, country and language. We record Visibility Score, Share of Voice, Position in Answer and Sentiment, and then open the citation reports to see which pages and domains the engines use. We write the Brand Book and aliases at this stage so mentions are counted properly.

Week three: access

We check robots.txt for rules that block GPTBot, OAI-SearchBot, ClaudeBot or PerplexityBot, and we run the free Agentic Readiness Check. When the client can connect a log source, Agent Analytics gives IP-verified crawler logs, with errors and failure explanations. Crawler logs start on the Professional plan, and we scope the log connection with the client's developer.

Week four: one fix

We choose one gap and fix it. It might be a topic with no real answer on the site, a page the bots never fetch, or a third-party source that never mentions the client. Content gap analysis and the Action Board help us choose, since the Action Board gives a prioritized list with evidence attached. We write down the date, and we do not change anything else on that topic for a few weeks.

What tends to help

Answering the question in the prompt, plainly, near the top of the page.
Facts, data and quotes that only the client has.
Clean structure, and structured data that matches the visible text.
Presence on the third-party sources the engines already cite.
Fast, error-free access for the AI crawlers.

None of this is secret. It is ordinary good publishing for a new kind of reader.

What we tell clients it cannot do

It cannot guarantee a mention or a recommendation. The engines sample, and they change. What we can offer is a baseline, a trend, a list of sources and a prioritized plan. We say this in the first call because clients who expect a guarantee are disappointed by everyone.

We also warn them about fixed formulas. If someone claims a single trick that works on every engine, ask for evidence on your own prompts.

Where to begin on your own

You can start without paying. Promptwatch has a free Explore plan with 10 prompts on ChatGPT, plus free tools such as the AI Visibility Report, the Agentic Readiness Check, robots.txt and llms.txt generators, and a glossary of GEO terms. Write your prompts first, record a baseline, check access, and make one change. When you want citations, crawler logs and content in one place, a paid plan starts at Essential, $95 a month.