A practical guide to generative engine optimization
A step-by-step GEO programme: build a prompt set, measure your baseline, fix retrieval, rewrite for extraction, and track the trend.
Most generative engine optimization advice stops at principles. This guide is the operational version: what to do first, what to do next, and how to tell whether any of it worked. It assumes you already know what GEO is and want a programme you can actually run.
A GEO programme has five stages, and they need to happen in order. Skipping to the writing stage is the most common failure, because rewriting pages that engines never retrieve produces no measurable change and burns credibility internally.
Stage one: define the question set
Everything downstream depends on choosing the right prompts. You are not looking for keywords, you are looking for the questions a real buyer asks an assistant on the way to a decision.
- Start with the questions your sales team hears most often, phrased the way a person would type them into ChatGPT.
- Add category-level questions where you would expect to be one of the answers, such as requests for the best tool or provider of a given kind.
- Add comparison questions naming you and your main competitors, because these are high-intent and highly revealing.
- Add problem-first questions where the buyer has not yet named the category, since these often surface a completely different set of brands.
- Keep the set small enough to run repeatedly. Thirty to a hundred well-chosen prompts beats a thousand you measure once.
Stage two: measure the baseline
Run the full prompt set across every engine your buyers actually use, and record four things for each run: whether your brand was mentioned, roughly where in the answer it appeared, which competitors were mentioned, and which domains were cited. Repeat the run enough times to see past the variance, because the same prompt can return different answers on different days.
The baseline is worth taking seriously. Without it you cannot separate a real improvement from the natural noise of a system that rewrites its answer every time.
Stage three: fix retrieval before anything else
If a page is not retrievable it cannot be cited, and no amount of rewriting changes that. Work through the mechanical checks first:
- Confirm your key pages return their full content in the raw HTML, without requiring JavaScript to render.
- Check your robots.txt does not block the AI crawlers you want reading you, and confirm which ones you are deliberately excluding.
- Verify the pages are indexed in conventional search, since most engines retrieve through a search index.
- Remove interstitials, aggressive bot protection, and rate limiting that returns errors to non-browser user agents.
- Make sure each page has one clear subject rather than several, so an extracted passage is unambiguous.
Stage four: rewrite for extraction
Once pages can be retrieved, the question becomes whether a model can lift something useful from them.
- Put a direct, complete answer in the first two sentences under each heading, then expand. Models extract chunks, and a chunk that begins with the answer survives extraction intact.
- Use headings that state the question rather than a clever label, because heading text is a strong retrieval signal.
- Keep definitions self-contained. A sentence that depends on the previous paragraph loses its meaning when extracted alone.
- Add specifics a model can attribute to you: figures, dates, named methodologies, and first-hand observations. Generic claims get paraphrased without credit.
- Cover the adjacent questions in the same cluster, since query fan-out means the engine is retrieving for several rephrasings at once.
- Say what you are not as well as what you are. Comparison and limitation content is disproportionately cited, because models are asked to compare.
Stage five: track and iterate
Re-run the prompt set on a fixed cadence, weekly or fortnightly, and watch the trend rather than any single result. Three things are worth watching closely: your mention rate per engine, the gap between you and the competitor set, and the churn in which domains are being cited. A rising competitor in the citation list is usually the earliest signal that someone has published something better than your equivalent page.
What to expect
Movement is slower than in paid channels and faster than in classic SEO. Retrieval-driven citations can change within days of a page being recrawled, while the model's underlying disposition toward naming your brand shifts over months as training data and web mentions accumulate. Plan for both timescales, and do not judge a content change after a week.
GEO strategies that are worth the effort
If you have limited capacity, these four produce the most movement per hour spent:
- Fix the questions where a competitor is named and you are absent. Short, specific, and demonstrably winnable, because the demand is proven.
- Restructure pages that already rank but are never cited. The retrieval hurdle is cleared; only the formatting is failing.
- Publish one piece of genuinely original data. It is the only content type that cannot be sourced elsewhere, which makes attribution close to automatic.
- Correct your entries on the third-party pages that keep appearing in citations. A single fix on a widely-referenced page can change answers across several engines.
Common questions about GEO programmes
- How many prompts should I track? Thirty to a hundred is right for most B2B categories. Below thirty the week-to-week noise exceeds real movement; above a few hundred you are usually diluting with low-intent questions that inflate the score.
- How often should I re-measure? Weekly, reading the trend over at least six weeks. Daily measurement mostly produces meetings about random variation.
- Should I write new content or fix existing pages first? Fix existing pages. It is faster, cheaper, and the pages that already rank have cleared the hardest gate.
- When will I see results? Retrieval-side movement in six to twelve weeks. Changes in how models describe you from memory take two to three quarters.
Tip: keep one prompt in your set that you expect to lose. A question where a competitor is clearly the right answer gives you an honest control — when your mention rate rises on that one too, you are usually looking at engine-level variance rather than your own progress.
Still stuck? We typically reply within 1 business day.
Contact support