Generative engine optimization (GEO) is the engineering discipline of making a brand's pages retrievable, quotable and attributable inside AI-generated answers. Classic SEO optimizes for a ranked list of ten links. GEO optimizes for a single synthesized answer that may cite three to eight sources and may or may not say your brand name out loud.
This guide treats generative engine optimization as a pipeline you can instrument, not a bag of tricks.
What generative engine optimization actually optimizes
Every answer engine — ChatGPT Search, Perplexity, Google AI Overviews, Google AI Mode, Gemini and Claude — runs some version of the same loop:
- Query understanding and fan-out. The user prompt is rewritten into several sub-queries ("best commercial dough mixer", "dough mixer 60 quart warranty", "Hobart vs used mixer").
- Retrieval. Each sub-query pulls candidate documents from a search index or a live fetch.
- Grounding. Passages are extracted from those documents and ranked for relevance and trust.
- Synthesis. The model writes an answer conditioned on the extracted passages.
- Attribution. Some passages become visible citations; some brands are named in the prose.
A brand can win or lose at each step independently. That gives you three outcomes worth tracking for every commercial prompt:
| Outcome | What happened | What it is worth |
|---|---|---|
| Retrieved | Your URL was fetched or indexed for the sub-query | Necessary, invisible to buyers |
| Cited | Your URL appears as a source link | Referral traffic, weak brand recall |
| Named | Your brand appears in the answer text | Recall, shortlisting, assisted revenue |
The gap between cited and named is large enough that we treat it as its own failure mode, Cited-Not-Named.
The GEO pipeline: retrieval, grounding, attribution
Retrieval: be fetchable by the right crawler
AI products use different user agents for different jobs. Search-index crawlers (such as OAI-SearchBot, Claude-SearchBot and PerplexityBot) decide whether you can appear in answers at all. Training crawlers (GPTBot, ClaudeBot, the Google-Extended robots token) affect future model knowledge. User-fetch agents (ChatGPT-User, Claude-User, Perplexity-User) retrieve a page live when a person asks about it.
Blocking a training crawler is a legitimate business choice. Blocking a search crawler by accident — usually through a WAF rule or a blanket Disallow — silently removes you from AI answers. Audit both robots.txt and the live HTTP response each bot receives; a 200 in your browser says nothing about a 403 served to a data-centre IP.
Grounding: be extractable
Models quote passages, not pages. A passage is extractable when it is self-contained, factual and close to the question. Practical rules:
- Put a two-to-three sentence direct answer immediately under each question-shaped heading.
- State specifications as plain text, not only inside images, tabs or client-rendered widgets.
- Keep one claim per sentence, with the number and unit together ("60-quart bowl, 2 HP motor").
Attribution: be nameable
The model names a brand when the grounded passage itself names it. If your product page says "our mixer" and the brand name only appears in the logo, the passage carries no entity. Write the brand into the answer sentence: "The Northwind M60 is a 60-quart spiral mixer built by Northwind Supply."
Measuring GEO: metrics that survive scrutiny
Track a fixed prompt set per market and engine, polled on a schedule, and compute rates rather than anecdotes:
citation_rate = answers citing your domain / answers polled
named_rate = answers naming your brand / answers polled
cnn_rate = answers citing but not naming / answers citing
share_of_voice = your brand mentions / all tracked brand mentions
Two rules keep these numbers honest. First, a failed poll (timeout, empty answer) is excluded from the denominator rather than counted as a loss. Second, every metric is segmented by engine, because Perplexity and Google AI Overviews cite very differently.
Engineering the page: answer-first structure
An answer-first commercial page looks like this:
<article>
<h1>Commercial Spiral Mixers for Bakeries</h1>
<p>Northwind Supply sells rebuilt 20–120 quart spiral mixers with a
12-month parts warranty and nationwide freight.</p>
<h2>Which spiral mixer size does a bakery need?</h2>
<p>Most retail bakeries producing under 200 kg of dough per day
choose a 40–60 quart Northwind mixer.</p>
<!-- specification table, FAQ, Product JSON-LD -->
</article>
The first paragraph names the entity, the category and the differentiator. Each H2 is a question a buyer actually asks, answered in the first sentence beneath it.
Structured data that grounds claims
Structured data does not make a model cite you, but it removes ambiguity about who you are and what you sell. Ship an Organization node with sameAs links to registries and profiles, and a Product node whose offers match the visible price exactly:
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Northwind Supply",
"alternateName": ["Northwind", "Northwind Supply Co."],
"url": "https://northwind-supply.example",
"sameAs": ["https://www.linkedin.com/company/northwind-supply-example"]
}
Never mark up a price, rating or stock level that the page does not show. Mismatched markup is the fastest way to lose trust with both search engines and the answer engines that reuse their indexes.
A 30-day generative engine optimization rollout
- Week 1 — Baseline. Pick 50–150 commercial prompts, poll every engine, record citation, named and Cited-Not-Named rates.
- Week 2 — Access. Fix crawler blocks, publish
/llms.txt, confirm search crawlers receive200responses. - Week 3 — Extractability. Rewrite the ten highest-value pages answer-first and add Product and Organization JSON-LD.
- Week 4 — Gaps. Publish evidence-backed buying guides for prompts where competitors are named and you are absent, then re-poll.
Common mistakes
- Optimizing one engine and assuming the others follow.
- Counting citations without checking whether the brand was named.
- Publishing AI-written pages with claims nobody verified; one wrong spec costs more trust than ten new pages earn.