Generative Engine Optimization

GEO treats the chunk, not the URL, as the unit of optimisation. A chunk is a 100–400 token span that says one coherent thing and survives being lifted out of the page around it. When a model composes an answer it assembles chunks from several sources at once, so the job is to make sure at least one of yours is the cleanest available statement of the fact the model needs. Schema, llms.txt and freshness signals all exist downstream of that: they get chunks found and trusted.

The term GEO is most often traded for is AEO — Answer Engine Optimization — and they are not the same thing. AEO predates the LLM era. It meant winning Google's featured snippet and the People Also Ask box, both of which are deterministic, inspectable, and attached to a ranked result you can look up. GEO targets a generated paragraph with no ranking behind it, which may blend four sources and may differ between two identical prompts. AEO's scoreboard is a SERP position. GEO's scoreboard does not exist until you build it.

The line with SEO is different again. SEO optimises a document for a system that retrieves and ranks links; GEO optimises spans of text for a system that retrieves and generates prose. They share the plumbing — crawl access, canonical tags, response times, structured data — and diverge above it. SEO rewarded comprehensive coverage of a topic. GEO rewards the density of standalone claims. GEO is additive rather than replacing: technical SEO becomes the entry price instead of the finish line.

The working metric is citation share: the percentage of tracked query-and-platform slots in which your domain is cited. The denominator has to be explicit — number of tracked queries multiplied by number of platforms, per refresh cycle. AIRank runs that grid every six hours against four engines, so a 30-query list produces 120 slots per cycle. A slot counts as a citation on a direct URL reference, a verbatim span of 15 or more tokens traceable to your content, or a paraphrase above 0.88 cosine similarity with no plausible alternative source.

Realistic ranges depend entirely on how competitive the query list is, which is why the list is the real instrument. Brand-name queries sit near total coverage and tell you nothing. A list built from buyer questions asked before anyone knows your name usually starts in the low single digits for a site that has never done GEO work. Good looks like appearing in a clear majority of the slots you actually care about, with competitors named alongside you rather than instead of you.

Take a B2B site tracking 25 buyer questions across four engines: 100 slots per refresh cycle. Suppose the baseline sweep returns your domain in 8 of them, every one on ChatGPT, and every one on a query that already contains your product category by name. That is 8% citation share and a concentration problem hiding inside it — you hold one engine's prior on the easiest third of your list and nothing anywhere else.

Split the remediation by pillar weight rather than by preference. Content extractability carries 40% of the AI Score, so the first pass rewrites the ten thinnest pages into standalone claim paragraphs with question-shaped H2s above them. Technical readiness, schema coverage and authority each carry 20%, so the second pass ships llms.txt, adds FAQPage markup to the high-intent pages, and connects the Organization entity. Neither pass touches the query list: the denominator has to stay fixed or the result is unreadable.

Two sweeps later the useful question is not whether the number went up. It is whether it went up on engines that were previously at zero. Moving from 8 slots on one engine to 20 slots spread across three is a materially better outcome than 20 slots still stacked on ChatGPT, because the first shape survives a single retrieval-layer change and the second one does not.

• The word-count reflex. A decade of SEO trained writers to pad every answer to two thousand words so it would outrank a thinner competitor. Retrieval punishes exactly that: a padded page dilutes every chunk inside it, and the model quotes whichever source states the fact in one clean paragraph. The fix is to break the padded page into the three or four questions it actually answers, and let each one be five hundred words that stand on their own.
• The brand-query mirage. Teams seed the tracked query list with their own product name, watch citation share sit near the ceiling, and conclude GEO is handled. A model asked about your brand will always find your site — that query measures the index, not your visibility. The fix is to rebuild the list out of questions a buyer asks before they have heard of you, then accept whatever ugly baseline that produces as the real starting number.
• The single-engine read. ChatGPT is the easiest engine to spot-check by hand, so it quietly becomes the only one anyone checks. The four engines resolve through different retrieval layers, and a change that reaches one index in minutes can take days to reach another. The fix is to read all four in the same view every cycle, and to treat an engine that has never cited you as a distinct problem rather than a lagging version of the one you already solved.

What the table cannot show is the direction of causation, and that matters more than the definitions. AI Score is a leading indicator; citation share is a lagging one. You move the score in an afternoon and then wait two weeks to find out whether citations followed. Citation density is not a parallel metric to the score at all — it is one of the levers inside it, sitting in the content-extractability pillar. Traditional SEO competes with none of them; it is the substrate the other three assume already works.