What generative engine optimization actually involves
A practical walkthrough of GEO: what changes when a model composes the answer instead of ranking a list and the specific work that follows from that difference.
Generative engine optimization is the work of becoming one of the sources an AI system draws on when it writes an answer. This lesson covers what that means concretely and what you would actually change on a site.
Prerequisite: the glossary if the acronyms are new.
The one difference that drives everything else
A ranked result and a generated answer fail in different ways.
In ranked search, being mediocre makes you lower. A motivated searcher scrolls. In a generated answer, being mediocre makes you absent. There is no scrolling past the summary to find source number seven.
Everything below follows from that.
Step 1: Confirm the systems can reach you
Before anything else, check that AI crawlers are not blocked. They are controlled
separately from Googlebot:
User-agent: GPTBot
Allow: /
User-agent: OAI-SearchBot
Allow: /
User-agent: ClaudeBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Google-Extended
Allow: /Blocking these can be a deliberate publisher decision. Blocking them by accident, through an inherited robots.txt, a security plugin default or a staging
config promoted to production, is common and entirely self-inflicted.
Also verify your content renders without JavaScript. Many AI crawlers do not execute it. If your service descriptions only exist after a client-side fetch, they may as well not exist.
Step 2: Make your claims extractable
A model writing an answer needs a specific, attributable statement. Compare:
We’re passionate about delivering world-class results for our clients.
Nothing in that sentence can be quoted in an answer to any real question. Now:
We work with businesses across the United States. Engagements run month to month with 30 days’ notice and the audit is delivered within ten business days.
Every clause is a fact a model can lift and attribute.
The practical test: read a sentence and ask whether it could appear, verbatim, as the answer to a question someone would actually ask. If not, it is brand atmosphere, fine in moderation, useless for retrieval.
Step 3: Structure for extraction
Models parse structure. Help them:
- Answer first. Put the direct answer in the first paragraph under the heading, then expand. Do not build to a conclusion.
- Headings that match real questions. “How much does an AI visibility audit cost?” beats “Investment”.
- Tables for comparisons. A model extracting “X vs Y” will take a table over six paragraphs of prose every time.
- Numbered steps for processes. Ordered lists survive summarisation.
- One idea per paragraph. Long paragraphs mixing three claims are harder to quote cleanly.
Step 4: Fix your entity data
The model needs to know what kind of thing you are before deciding you are relevant.
- Describe your company identically across your site, your profiles and third-party listings, same name, same category, same service area.
- Implement
Organizationschema with consistentname,url,logo,description,areaServedandsameAspointing at profiles you genuinely control. - Keep structured data consistent with visible content. Marking up claims absent from the page violates search guidelines.
- Where an author is real and identifiable, attribute content to them. Unattributed content is weaker evidence.
Step 5: Give models a reason to prefer you
This is where most of the durable advantage is and it is the slowest part.
- Publish original material. Research, benchmarks, documented methods, things that exist nowhere else and therefore have to be cited to be used.
- Write the comparisons. Balanced, factual comparison content is disproportionately retrieved, because comparison is exactly what buyers ask assistants about.
- Define things properly. Clear definitions of category terms get pulled into explanatory answers constantly.
- Earn independent mentions. A synthesised answer weighs what multiple sources say. You cannot self-declare your way into being recommended.
Step 6: Measure, then iterate
Build a prompt set, run it on a schedule, record who gets named. Without a baseline you cannot tell whether any of the above worked and in a category this noisy, that is the difference between a programme and a hobby.
The method is covered in detail in How to measure AI visibility.
What not to do
- Do not keyword-stuff for models. They are not matching strings.
- Do not publish unreviewed AI-generated volume. It produces exactly the unverifiable content these systems are being tuned to discount.
- Do not fabricate claims to be quotable. A model repeating your false statistic is a liability, not a win.
- Do not expect
llms.txtto do anything. It is an unadopted proposal. Publish it if you like; do not build a strategy on it.
Realistic expectations
Technical fixes show up in crawling within weeks. Movement in generated answers is slower and noisier, read it over two to three months. Nobody can guarantee citation and anyone who does is describing something that does not exist.