What is Generative Engine Optimization (GEO)?
GEO is not SEO with a new label. It optimises for a different unit — the passage — and a different outcome: being quoted inside an answer rather than listed beneath one.
Generative Engine Optimization (GEO) is the practice of structuring content so that large language models retrieve, trust and cite it when answering a user's question. Where search engine optimisation competes for a position in a ranked list of links, GEO competes to be the passage a model lifts into its answer.
That distinction sounds small. In practice it changes the unit of optimisation, the definition of success, and most of the tactics that follow from both.
The unit of optimisation changes from the page to the passage
A search engine evaluates a document. A retrieval system chops that document into chunks, embeds each chunk, and pulls the ones most relevant to the query. The model then writes an answer from whatever it pulled.
This has an unintuitive consequence: a page can be excellent and still be uncitable. If your best explanation is spread across four paragraphs, each of which depends on the one before it, no individual chunk answers the question on its own. A competitor with a worse page but a single self-contained paragraph gets quoted instead.
The practical test is simple. Take any paragraph from your page, show it to someone with no context, and ask whether it answers the question completely. If it does not, it will not be cited.
The definition of success changes from rank to citation share
There is no position three in an AI answer. There is being named or not being named, and if named, being named first or fifth, favourably or with a caveat.
The measure that replaces rank is citation share: across a defined set of prompts, run repeatedly, what proportion of answers name you? Because model outputs are non-deterministic, a single run tells you almost nothing. Ask the same question three times and you can get three different vendor lists. Sampling repeatedly and reporting the distribution is the only way to get a number that moves for real reasons.
What actually drives citations
Across the accounts we run, four properties separate cited pages from ignored ones.
1. Answer-first structure
State the answer in the first sentence under the heading, then elaborate. The inverted pyramid was already good practice for readers; for retrieval systems it is close to mandatory.
2. Factual specificity
Models prefer passages carrying concrete, checkable detail — numbers, dates, named entities, defined thresholds. "Significantly faster" is unquotable. "Reduced median response time from 840ms to 210ms" is a citation waiting to happen.
3. Entity clarity
Pronouns and implied subjects break chunking. A passage that begins "It also supports…" is orphaned the moment it is extracted. Repeat the subject more often than a style guide would prefer.
4. Third-party corroboration
Models weight sources they encounter repeatedly across the web. A claim made only on your own domain carries less weight than the same claim reflected in independent coverage, documentation, forums and reviews. This is why GEO cannot be done entirely on your own site.
What GEO shares with SEO
Roughly seventy per cent of GEO work is good technical and editorial SEO applied more rigorously than most teams bother with:
- Crawlability. If GPTBot, PerplexityBot or ClaudeBot cannot reach a page, nothing else matters. Check your robots directives — inherited blocks on documentation subdomains are the single most common finding in our audits.
- Structured data. Schema gives an unambiguous machine reading of entities, prices, ratings and relationships.
- Clean information architecture. Headings that describe content accurately, tables that are actually tables, lists that are actually lists.
- Freshness. Explicit publication and review dates help models decide whether a claim still holds.
None of this is new. What is new is that sloppiness in these areas now costs you a whole surface, not a few positions.
Where GEO genuinely diverges
Three areas have no real SEO equivalent.
Prompt research replaces keyword research. People type differently into a chat window than into a search box. "best crm for a 12 person sales team that already uses hubspot" is a normal prompt and a bizarre keyword. Your prompt set should come from sales calls and support tickets, not from a keyword tool.
Source-graph analysis replaces backlink analysis. The question is not who links to you but which domains the engines already lean on when answering questions in your category. That list is usually shorter and stranger than your backlink target list, and it frequently includes forums.
Accuracy monitoring has no SEO analogue. A search engine cannot misdescribe your pricing. A model can, confidently, and the error will persist until something in its source material corrects it.
How to start
- Build a prompt set. Fifty to two hundred questions your buyers actually ask, drawn from sales and support.
- Baseline honestly. Run every prompt across the engines that matter to you, several times each, and record what comes back verbatim.
- Diagnose each absence. Every gap traces to one of three causes: not retrievable, not citable, or not corroborated. The fix differs completely for each.
- Fix structure before you write anything new. Restructuring existing high-quality pages for citability is almost always cheaper than producing replacements.
- Measure monthly, not weekly. Model behaviour is noisy. Monthly trend lines tell the truth; weekly ones mostly tell you about sampling variance.
The honest caveat
GEO is roughly three years old as a discipline and the ground moves under it constantly. Anyone selling certainty about how a specific model weights a specific signal is guessing. What is defensible is the method: measure citation share properly, fix retrievability and citability where they are broken, earn corroboration on sources models already trust, and re-measure. That loop has held up across every model release so far, and it is the one we build on.