How to do generative engine optimization

How to do generative engine optimization: map the competitive landscape with real search and citation data, publish fact-checkable claims that build trust, and only then execute. GEO is not a campaign you buy. It is a research habit — big data sweeps first, writing second.

In What LLM SEO actually means and SEO vs GEO vs AEO, I laid out the mechanics. This is the practical half: what I actually do when I want something I build or write to show up in AI-mediated discovery — without looping it into a sales pitch for my own products.

The rule that keeps you honest

Research and data matter more than execution. Most GEO content skips straight to checklists and tools. That's backwards. If you don't know what people search, what already ranks, and what AI systems already cite, you're publishing blind. Blind work doesn't get trusted — and trust is the scarce input when answers are cheap.

Step 1 — Identify the competitive landscape with numbers

Before you write a word, pull the landscape:

• Demand — what people actually search (volume, related terms, rising queries)
• Supply — who ranks today, on what domains, with what angles
• AI surface — what shows up in AI Overviews / answer engines for those queries, and which sources get cited

Build business value around what the data confirms is valuable — not around what sounds smart in a brainstorm. Traffic and search behavior are the confirmation. If nobody searches it and nothing gets cited around it, you're decorating, not optimizing.

Step 2 — Write for fact-checkers, not for your brochure

The fastest way to lose GEO is to turn every page into "why our offering is better." Models and skeptical readers both punish that.

• Use real, checkable facts — definitions, mechanisms, measurable patterns, primary sources
• Don't invent case studies, stats, or "our edge" narratives the model can't verify
• Don't force every insight back into a product pitch — that trains readers (and systems) to treat you as marketing, not a source

Credibility compounds. Hype doesn't. If a claim can't survive a fact check, it shouldn't be on the page.

Step 3 — Sweep data at scale; don't nibble

One keyword in a spreadsheet is not research. Do a massive data sweep: clusters of queries, SERP compositions, difficulty, parent topics, overlapping pages, and where AI answers already lean. Pattern-match across the set.

You're looking for gaps: questions with demand, weak or generic top results, and thin AI citations — places where a specific, true take can still earn a seat. That's the map. Execution without the map is just content volume.

Step 4 — One person with good tools beats a rented team

You do not need a fancy ad or marketing agency to do this well right now. One person with a coding agent (Cursor, Claude Code, etc.), a few MCP integrations into real data sources, and the discipline to analyze before publishing is more efficient — and usually more cost-effective — than a team that outsources the thinking.

The bottleneck isn't headcount. It's judgment applied to data. Agencies sell process. GEO rewards someone who will sit with the landscape until it's obvious what to say.

Step 5 — Let the landscape improve the product, not just the blog

If you understand the space you're building in — really understand it from search and AI citation data — you also learn what would make the product more valuable. Gaps in what people ask, and gaps in what AI tracks well, are product signals.

GEO research isn't only for pages. It's a feedback loop into what you ship next. Expand or sharpen the offering based on confirmed demand, not on whatever the loudest competitor is blogging about.

Step 6 — Keep it practical

There's nothing fancy here. It takes dedicated time, thoughtful analysis, and proper data retrieval. That's the whole job:

1. Pull demand + SERP + AI citation landscape
2. Find the gap you can fill with something true and specific
3. Publish an extractable answer (see the Attribution Ladder) with clear entity and proof
4. Feed what you learned back into the product roadmap
5. Repeat on the next cluster — don't spray random posts

How this maps to the citation model

The four factors from the LLM SEO essay still apply — you just earn them with data, not vibes:

• Retrievability — write the answer the query actually asks, up top
• Entity clarity — one consistent author/brand the system can attach claims to
• Corroboration — facts others can verify; earn mentions the hard way
• Usefulness density — less fluff, more load-bearing claims

What is cope

• Buying a "GEO retainer" instead of learning the landscape
• Publishing before you've looked at demand and SERPs
• Invented stats and uncheckable "we're the best" claims
• Tool dashboards as a substitute for judgment
• Treating every post as a funnel for your product page

The short version

How to do GEO: sweep the data, find what's actually valuable, publish something true enough to survive a fact check, and use what you learn to make the business better — not louder. One focused operator with real data access can outwork a theater of agencies. Nothing about that is magic. It just isn't free of attention.

Discovery is only half the stack. The other half is what still holds value when answers get cheap — privacy, dollar rails, and scarce assets.