How Does Generative Engine Optimization Work? The Pipeline, Stage by Stage
How does generative engine optimization work? It runs through five stages: query decomposition, retrieval, reranking, synthesis, and attribution. Citevora models each stage as a separate failure point, because the edits that help a page survive retrieval do nothing at reranking, and the edits that win attribution are different again. What follows is the mechanical version, with an evidence label attached to every claim.
Key takeaways
Five stages, not one. Query decomposition, retrieval, reranking, synthesis and attribution each remove different pages for different reasons. A tactic aimed at the wrong stage is wasted work, however well executed.
Retrieval is index-specific. Googlebot's index feeds AI Overviews, OpenAI documents OAI-SearchBot for ChatGPT search, and Perplexity documents PerplexityBot against its own index. Visible in one engine tells you almost nothing reliable about another.
Reranking is the stage most advice skips. A cross-encoder or LLM judge scores individual passages against generated sub-questions. Passages that only make sense inside their surrounding section lose here, and no amount of keyword coverage recovers them.
Synthesis punishes undated, unsourced specifics. Generated answers flatten several sources into one response. Claims that need context, a date, or attribution to stay true get compressed or dropped.
Attribution is not the same as inclusion. A source can be retrieved, reranked, used, and still never named. Cited, used, and cited are three different states, and the third one is the only one that compounds.
Only the synthesis stage has causal evidence. The single controlled experiment on this pipeline - Aggarwal et al., KDD 2024 - supplied retrieved results to the engine as context. It measures how content edits change visibility inside synthesis. It does not measure retrieval or reranking.
The SERP problem: nine explainers, no mechanism
Search for how does generative engine optimization work and you get nine organic results, every one of them an agency or vendor explainer. Walker Sands holds rank one, Conductor's academy is second, Power Digital's service page is third, followed by Optimizely, Go Fish Digital, Zeta Global, Kleecks, CrafterCMS and Periscopify. There is no vendor documentation in the set and no engineering writeup. Nine pages, one genre.
The content underneath the rankings is thinner than the positions suggest. Walker Sands carries an H2 titled "How do Generative Engines Work?" that covers LLM training data and periodic refresh, then moves on - no retrieval, no reranking, no attribution. Conductor's page has sections on what GEO is, why it matters, examples of strategies, and GEO versus SEO, with no pipeline section at all. Power Digital mentions retrieval and synthesis, but only inside FAQ prose, never as a sequence of steps a reader could audit or a contractor could implement.
That is the gap this page exists to fill. Answering what is generative engine optimization is a definition task, and the top three do it competently. Explaining how geo works is a systems task, and none of them attempt it. Demand is modest: 140 monthly US searches for the head query, 70 for how does geo work, and 10 each for three variants, at low-to-medium difficulty. A mechanical explainer does not need volume to be the best answer available. It needs to be the only one actually describing the machine.
The pipeline at a glance
|
Stage |
Mechanism |
What it drops |
The edit that changes the outcome |
Evidence status |
|---|---|---|---|---|
|
1. Query decomposition |
The prompt is split into subtopics, each issued as its own search |
Pages that answer the headline question but none of the generated sub-questions |
Add one section per sub-question, titled the way a buyer would type it |
Documented by Google for AI Mode; may apply to AI Overviews |
|
2. Retrieval |
Hybrid search - lexical BM25-style matching plus dense embeddings - against one engine's index |
Blocked, unrendered, unindexed, or duplicate pages |
Allow the AI crawlers, server-render the claims, canonicalize duplicates |
Crawler identity documented; ranking inputs not |
|
3. Reranking |
A cross-encoder or LLM judge re-scores candidate passages against each sub-question |
Passages that are not self-contained claims |
Rewrite every section opener as a complete, standalone assertion |
Inferred - no vendor documents the reranker |
|
4. Synthesis |
An answer is generated from retrieved context spanning several sources |
Claims needing surrounding context, undated assertions, unsourced specifics |
Date every assertion; attach source and unit to every number |
Measured by the KDD 2024 experiment |
|
5. Attribution |
A subset of the sources actually used is cited in the output |
Sources that were read but not credited |
Make naming cheap: unique phrasing, explicit source lines, clean markup |
Observed; selection logic undocumented |
The stages are sequential in the sense that a page must clear one to reach the next. They are not sequential in cost. Most teams spend their entire budget on stage-2 fixes - crawl access, rendering, technical hygiene - which is necessary and almost never sufficient. When someone asks how does geo work and gets an answer about schema and site speed, that is a stage-2 answer to a five-stage question.
Stage 1: Query decomposition (fan-out)
What happens mechanically. A single prompt rarely runs as a single search. Google documents fan-out in AI Mode: the question is broken into subtopics, multiple related searches run across those subtopics and data sources, and the results are combined into one response. The documentation also states that fan-out may be used in AI Overviews. It does not disclose how many sub-queries are generated, how they are phrased, or how they are ordered. Everything past that is inference.
What it drops. Pages that answer the headline question and nothing adjacent. A page titled "What is GEO?" that defines the term in 300 words can satisfy the literal prompt and still be excluded, because the generated sub-questions include things like how it differs from SEO, what it costs, whether it works, who offers it, and how to measure it. The page was never competing on the prompt. It was competing on the fan-out.
The edit. Build one section per sub-question, with the H2 or H3 phrased the way a buyer would type it. For a B2B SaaS page, that means the sub-questions the category actually generates: seat-based versus usage pricing, migration risk, security review, time to value. The page does not need to be long. It needs to be complete against the fan-out pattern rather than complete against a keyword list.
Evidence status. Documented at the level of existence, inferred at the level of detail. Google confirms fan-out happens and declines to specify the sub-queries, which means anyone selling you a fan-out simulator is selling a model of a process they cannot observe. Useful model, still a model.
Stage 2: Retrieval
What happens mechanically. Retrieval is hybrid. A lexical pass does BM25-style term matching; a dense pass does embedding similarity; the two are merged and truncated to a candidate set. The algorithm matters less than the index being queried. Googlebot's index feeds Google Search and, by extension, AI Overviews. OpenAI documents OAI-SearchBot as the crawler behind ChatGPT search results, and separately documents GPTBot for training and ChatGPT-User for user-initiated fetches - three agents, three jobs, and access to one says nothing about access to another. Perplexity documents PerplexityBot and runs its own index.
That distinction is where most cross-engine assumptions fail. A page that is well indexed in Google, with strong traditional rankings, may never enter ChatGPT's candidate set, because OAI-SearchBot was disallowed in 2023 when the robots.txt file was written to keep training crawlers out. The page did not decline. It was never a candidate.
What it drops. Anything the crawler cannot reach, cannot render, or has chosen not to retain. That includes pages behind a robots rule written for a different bot, claims injected client-side that a fetch never executes, parameterized duplicates competing with each other for the same claim, and content that lives behind an app shell. Retrieval also drops pages whose passages are too thin to compete in the dense pass - a 200-word page against a 3,000-word page covering the same sub-question rarely survives the merge.
The edit. Audit per engine rather than in aggregate. Confirm the AI user agents are allowed on the specific paths that carry your claims, not just on the homepage. Confirm the claims are present in server-rendered HTML. Confirm a canonical resolves to one URL per claim set, because two URLs splitting the same claim weaken both. This is unglamorous work, and it is why AI search visibility services engagements usually open with an access audit rather than a content brief.
Evidence status. Crawler identity is documented by each vendor, including which agent handles search versus training versus user-initiated fetches. Ranking inputs inside retrieval are not documented anywhere. Nobody outside Google, OpenAI or Perplexity knows the weighting between the lexical and dense passes. Any agency presenting that weighting as fact is guessing at a number.
Stage 3: Reranking
What happens mechanically. The candidate set is small in relative terms and still far too large to place in a prompt. A reranker - typically a cross-encoder that scores query-passage pairs directly, sometimes an LLM acting as a judge - re-scores individual passages against each sub-question and keeps the top handful. Note the unit of evaluation: passages, not pages. A page is not judged as a page. Sections of a page are judged, and the losing sections do not drag the winning ones down.
This is also where the fan-out and the reranker interact. Each generated sub-question gets its own scoring pass. A page with one excellent section can be retrieved for that sub-question and absent for the other four, which is why partial coverage produces partial visibility that looks like randomness in a dashboard.
What it drops. Passages that are not self-contained claims. A section that opens "This is why it matters" and resolves the antecedent two paragraphs later scores badly against a sub-question the reranker cannot resolve context for. So does a table with no caption, a bullet list of features with no stated consequence, or a paragraph that only parses after the one above it. Dense prose loses to shaped prose at this stage, not because the model prefers simple writing, but because extractability is the scoring criterion.
The edit. Rewrite each section so its first one or two sentences state a complete claim with the entity, the attribute, and the value all present. "Reranking re-scores passages individually against each sub-question, which is why section openers have to stand alone" is extractable. "This is the part most people miss" is not. Apply the same discipline to headings: a heading that names the claim outperforms a clever heading that only parses in sequence. A section answering a pricing question should open with the price, the unit, and the comparison, then explain.
Evidence status. Inferred. No major engine documents its reranker, its scoring inputs, or where it draws passage boundaries. This is the stage where the industry's generative engine optimization best practices are least grounded and most consequential. Nearly every practitioner recommending short paragraphs and question-shaped headings is reasoning backward from output patterns rather than forward from a published spec. Treat those recommendations as strong hypotheses with a plausible mechanism, not as rules.
Stage 4: Synthesis
What happens mechanically. The model generates an answer from the retrieved context, which spans several sources at once. It compresses, reconciles conflicts between sources, and fills gaps where the context is silent. Crucially, it does not preserve your framing. A carefully qualified claim becomes a one-clause summary, and if the qualification does not survive compression, the summary can be wrong in a way you have no channel to correct.
What it drops. Three categories, consistently. Claims that depend on surrounding context to remain true. Undated assertions, because "recent studies show" is unusable when the model has no anchor for when. Unsourced specifics, because a number with no attribution is the first thing a model hedges, rounds, or omits. Each of the three has the same root cause: the claim was only true inside your document, and synthesis does not carry your document.
The edit. Date assertions in the sentence that carries them, not in a footer or a "last updated" badge. Attach a source and a unit to every number the page depends on. Restate the qualification inside the claim rather than in the sentence after it. If a statistic drives the argument, put the sample size and the method in the same paragraph, because the model will summarize the paragraph, not the page. The KDD benchmark below explains why these three edits specifically, and not topic coverage, are the validated ones.
Evidence status. Measured. This is the only stage with a controlled experiment behind it, and the section below covers precisely what that experiment did and did not test.
Stage 5: Attribution
What happens mechanically. The response is assembled from several sources and a subset of those sources is cited. The three states are genuinely distinct. A source can be retrieved without being used, used without being cited, and cited without being the primary basis for the answer. Citation selection does not mirror retrieval rank, which is why rank-one pages sometimes disappear from an answer while rank-five pages are named.
What it drops. Sources that informed the answer but were not credited. Two causes recur. Aggregation: your data point appears inside a competitor's roundup, and the model cites the roundup because it is the cleaner object. Paraphrase loss: your distinctive phrasing is smoothed away during synthesis until nothing in the answer identifies the origin, and the model credits nobody or credits the most quotable neighbour.
The edit. Make attribution cheap. State the finding in a form that survives paraphrase, with the entity named in the sentence rather than implied by the section heading. Include an explicit source line where you are citing someone else, because it gives the model a clean object to carry forward. Keep structured data accurate, since it is the least ambiguous machine-readable statement of what the page is about and who published it. Then measure attribution separately from inclusion, which is what AI citation analysis exists to do: tracking which domains are named for which prompts, and where you are present in the answer but missing from the source list.
Evidence status. Observed. Vendor documentation confirms citations exist and describes how they are rendered - inline links, source panels, footnote clusters. None of it describes selection logic. Studies of answer composition consistently find heavy overlap and heavy skew toward a small set of domains, but overlap is not causation, and a domain being cited often does not tell you which edit caused it.
What the only controlled experiment actually measured
There is one benchmark worth citing here, and it has a boundary that most summaries of it ignore.
Aggarwal et al., "GEO: Generative Engine Optimization," arXiv 2311.09735, published at KDD 2024, built GEO-bench: 10,000 queries drawn from nine datasets and evaluated across generative engines. The authors applied content transformations to source documents and measured the change in visibility - the share of the generated answer's words and citations attributable to a given source.
The headline result is a visibility lift of up to 40%. The tactic-level numbers are more actionable. Adding statistics lifted visibility by 21.0%. Adding quotations lifted it by 28.9%. Adding primary-source citations lifted it by 22.5%. Keyword stuffing performed no better than the unmodified baseline.
The rank interaction is the finding that should reframe how you budget. Under the citation-adding transformation, a source ranked fifth in the underlying search results gained 115.1% visibility. A rank-one source lost 30.3%. Low-ranked sources had the most to gain because the transformation gave the engine something extractable that the ranking had not already rewarded, while top-ranked sources were already being used and had their existing advantage diluted by the new competition.
Now the limitation, stated plainly. The experiment supplied retrieved results to the generative engine as context. That isolates the synthesis stage. It measures what happens to visibility when content changes and the candidate set is held constant. It does not test whether your edits change what gets retrieved, and it does not test reranking at all. Every causal claim about retrieval-stage optimization that cites this paper is overreaching. The paper validates content-level edits inside synthesis - statistics, quotations, citations - and is silent on the other four stages.
That is why a rigorous generative engine optimization process has to be labeled by evidence class. Synthesis tactics have a causal experiment. Retrieval tactics have vendor documentation of crawlers and nothing public about ranking. Reranking tactics have output patterns plus a plausible mechanism. Attribution tactics have observation. Building a program without those labels produces confident recommendations with wildly uneven backing.
Where the pipeline breaks in practice
Four failure modes account for most of the invisible pages we audit. Each maps to a stage, and each needs a different fix.
The fan-out gap (stage 1). The page answers the question in its title and nothing else. It earns traditional rankings on exact-match intent and never surfaces in generated answers, because the sub-questions were never addressed anywhere on the domain. The tell is a prompt set where you appear for the head term and for nothing adjacent - strong on "what is GEO," absent on "GEO cost," "GEO versus AEO," and "does GEO work."
The index exclusion (stage 2). The page is fine and the crawler cannot see it. A robots rule blocks OAI-SearchBot while allowing Googlebot. A pricing table renders only after a click. Two URLs serve the same claim set and split whatever authority either had. This failure is invisible in a rank tracker and obvious in a log-file review or a fetch test, which is exactly why it persists for quarters.
The reranker loss (stage 3). The page is indexed, retrieved, and still loses, because its sections are narrative rather than claim-shaped. Teams usually respond by adding words. Adding words to a passage that is not a self-contained claim makes it longer and no more extractable. This is the most expensive misdiagnosis in the discipline, because it consumes content budget while the actual constraint stays untouched.
The attribution gap (stage 5). You are in the answer and not in the citation list. The content was useful enough to synthesize and not distinctive enough to credit. The fix is partly rewriting - entity named, claim stated flatly, source line attached - and partly measurement, since most dashboards conflate "used" with "cited" and report a single presence number that hides the difference.
How to diagnose which stage you are failing
Diagnose in pipeline order and stop at the first failure. Optimizing a later stage while an earlier one is broken produces clean data about nothing.
Step one: define the prompt set. Twenty to forty prompts a buyer in your category would actually type, split between head questions and sub-questions. Not keywords. Full sentences, including the comparison, cost and proof questions that buyers ask out loud.
Step two: test reachability per engine. Fetch your key URLs the way each crawler would. Check robots rules, rendering, canonicalization and indexation status for Google, ChatGPT search and Perplexity separately, because they are separate systems with separate agents. If a stage-2 failure is present, stop and fix it. Nothing downstream can be diagnosed through a closed door.
Step three: check fan-out coverage. For each prompt, list the sub-questions a competent answer would have to address, then map them against your pages. Gaps here are stage-1 failures and they are usually structural: no page exists for the sub-topic, rather than a page that covers it badly. Structural gaps cannot be fixed by editing existing pages.
Step four: score passage extractability. Take the top five pages for a priority prompt and read only the first sentence of each section. If those sentences do not each state a complete claim, stage 3 is your constraint. Rewrite before you expand, because a broader page of unextractable passages loses to a narrower page of clean ones.
Step five: separate used from cited. For each prompt where your content appears in the answer, check whether the domain is named. Where it is not, treat it as an attribution failure and rewrite for distinctiveness. This is where how to measure AI visibility becomes practical: prompt-level presence tracking, citation share by domain, and the delta between the two reported as separate metrics.
Step six: instrument before and after. Record the prompt set, the engine, the date and the outcome before any edit. The best generative engine optimization tools are the ones that hold that baseline steady while you change one stage at a time. Most cannot, which is why serious teams run their own prompt panels alongside a vendor dashboard rather than trusting one blended score.
The sequence takes about a week for a focused B2B set and produces a diagnosis you can act on, with a ranked list of edits. It also tells you quickly whether you need AI search optimization services at all, or whether the problem is a ChatGPT SEO services issue confined to one engine's crawl access - a two-day fix, not a program.
Frequently asked questions
How does generative engine optimization work in one paragraph?
It works by surviving five stages in order: a prompt is decomposed into sub-questions, candidate passages are retrieved from one engine's index, a reranker re-scores those passages against each sub-question, an answer is synthesized from the survivors, and a subset of sources is cited. Each stage drops different pages for different reasons. Underperformance is usually one stage failing, not a general content weakness.
How does GEO work differently from SEO?
SEO optimizes for a ranked list of URLs. GEO optimizes for inclusion in a generated answer, which is a different target with different units. Traditional rankings evaluate pages; the pipeline evaluates passages. A rank-one page can lose visibility when it is rewritten and a rank-five page can gain it, which is what the KDD experiment's rank interaction demonstrated.
How generative engine optimization works when a page is indexed but never cited?
That pattern points to stage 5, not stage 2. The page was retrieved and used, and the citation went elsewhere - usually to a roundup that aggregated your data, or to nobody, because the claim was too generic to attribute. The fix is distinctive, entity-named phrasing plus an explicit source line, not additional coverage.
Is there a documented generative engine optimization process I can follow end to end?
No. Vendors document fragments. Google documents fan-out and AI Overviews eligibility. OpenAI documents its crawlers. Perplexity documents its crawler and its index. None documents retrieval weighting, reranking inputs, or citation selection. Any end-to-end process you buy is a model built from documented fragments plus inference, and it should be labeled that way on the page and in the deck.
What is answer engine optimization (AEO), and how does it relate to GEO?
What is answer engine optimization (AEO) is a narrower question than it looks. AEO targets answer surfaces - featured answers, People Also Ask, voice responses, AI Overviews - and historically emphasized structured, question-shaped content. GEO targets generative synthesis across multiple sources. In practice AEO tactics help at stages 1 and 3, since question-shaped headings map onto sub-questions and clarify passage boundaries. They do not address attribution, which is the stage that produces compounding visibility.
Which generative engine optimization best practices survive reranking?
Three survive reliably: one claim per section with the claim in the opening sentence, headings that name the claim rather than gesture at it, and numbers with units and sources in the same paragraph. Practices that do not survive: keyword density targets, page-length minimums, and intro-paragraph keyword placement. The KDD benchmark found keyword stuffing performed no better than an unmodified baseline, which is as close to a controlled refutation as this field has.
Do the best generative engine optimization tools measure all five stages?
Almost none do. Most tools measure outputs - mentions, citations, sentiment - at stages 4 and 5, then infer earlier stages from those outputs. That is a reasonable proxy and a poor diagnostic, because the same output can come from a retrieval failure or a reranking failure. If you need stage-level attribution of cause, you need LLM search optimization instrumentation: prompt panels, crawl access testing, and passage-level scoring, assembled rather than bought.
How do AI search engine optimization strategies differ for B2B SaaS and professional services?
The pipeline is identical and the sub-questions differ. B2B SaaS fan-out skews toward integration, security review, pricing model and migration risk. Professional services fan-out skews toward engagement scope, credentials, jurisdiction and conflict checks. Both need passage-level claims; the SaaS set usually needs more comparison content, because the generated sub-questions are comparative. The full set of AI search engine optimization strategies differs by industry for that reason and not because the mechanism changes.
How long before a pipeline fix shows up in AI answers?
Stage-2 fixes can appear within a recrawl cycle for one engine - days to weeks - and stage-1 additions take as long as indexation takes. Attribution changes are the slowest, because citation selection appears to favor established domain patterns and cited-versus-used ratios move gradually. Measure per engine and per prompt. A blended score across engines will hide the change you just made.
Where does Citevora start with a new client?
With the diagnostic sequence above, run on your own prompt set before any content is written. The output is a stage-level diagnosis and a prioritized edit list, which is a more useful artifact than a visibility score. You can see a client account on our testimonial page, and the same method applies across industries with different sub-question sets.
What this means for your next quarter
The useful conclusion is not that GEO is complicated. It is that GEO is specific. Five stages, five failure modes, and only one stage - synthesis - has causal evidence behind its tactics. Teams that treat the discipline as a single lever spend a quarter on technical hygiene and then wonder why the citation list did not move. Teams that diagnose by stage spend the same quarter on fewer edits with a measurable effect.
Citevora runs the diagnosis first, per engine and per stage, and writes second. That ordering is the difference between a content program and a visibility program. If you want the stage-level version - which sub-questions you are missing, which passages fail reranking, which prompts use you without citing you - start with a pipeline audit. Book one here and bring your prompt set.