← Deep dives

Deep dives · 2026-11-19 · Advanced · By bombong

Generative Engine Optimization (GEO): How to Land in the 'Footnotes' of AI Answers

Generative engine optimization (GEO) is a set of techniques for editing content so that generative search services such as ChatGPT and Gemini are more likely to cite or summarize it when they write answers. Research has moved from one-size-fits-all editing rules to per-document diagnosis, structural design and intent inference. More recently it has broadened to examine the risk of manipulation and whether the effects hold up on the real web.

Contents
Key points
  • In generative search, exposure depends less on link rankings than on landing among the few citations in an answer, so the target metric should be 'citation', not 'contribution'.
  • Citations are decided through a sequence of query rewriting → search candidates → citation selection, and pages that appeared more often among search candidates were also more likely to be cited.
  • GEO techniques have evolved from applying the same rules to every document to per-document failure diagnosis (AgentGEO), structural design (GEO-SFE) and inferring intent from the document (QI-GEO).
  • The same techniques can slip defective products into recommendation lists and can also evade common detection methods, so GEO is a security and fairness issue as well.
  • On the real web, results swing widely depending on the engine, the user's location, the order of candidates and run-to-run variation, so lab numbers should not be taken at face value.

Why it was needed: from rankings to 'citations'

For a long time, search engine optimization (SEO) was a contest over 'what position you appear in on the search results list'. Generative search services such as ChatGPT, Perplexity and Gemini, however, read and synthesize multiple sources to write an answer instead of showing a list of links, and attach only a few sources below it as citations. The QI-GEO study paper sums up this shift: exposure for content creators is no longer decided by ranking alone but by whether their content is cited inside the generated answer. On top of that, generative engines are black boxes whose internals cannot be inspected. The family of techniques that emerged in this gap is generative engine optimization (Generative Engine Optimization, GEO).

There is also evidence that existing SEO instincts do not carry over directly. A large-scale study comparing AI search with Google paper ran controlled experiments across multiple industries, multiple languages and paraphrased versions of the same question. It found that AI search systematically and strongly preferred authoritative third-party sources (earned media: content written by others, such as press coverage or external reviews) over brand-owned channels or social content, while Google mixed sources more evenly. AI search services also differed substantially from one another in the diversity of domains they cited, freshness, stability when the language changed and sensitivity to how a question was phrased.

In short, in generative search the 'spot that earns the click' disappears and only 'a few footnotes in the answer' remain. GEO research rests on two pillars: what to change and how in order to win that spot, and whether those changes end up harming users.

How SEO and GEO differ

SEOCompeting for search result rankings

  • Goal is rank in a list of links
  • Users click through and read

GEOGetting cited in generative answers

  • Goal is citation or summary in the answer
  • Engine selects and synthesizes sources
  • Each engine prefers different sources

공통 Both refine content so it reaches users

The two approaches side by side: what each aims for and how exposure is decided.

Core idea: getting into a librarian's report as a footnote

Consider an analogy. A librarian receives a question, pulls dozens of books from the stacks, skims them, writes a one-page report and lists only three or four references at the end. For an author, being cited in the librarian's report now matters more than being shelved at the front of the bookcase (SEO). To get there, the author needs to organize the table of contents so the librarian can easily find and use the evidence, make the key sentences clear and fill in content the questioner would likely ask about but that is missing. This is exactly what GEO does.

The analogy brings out one distinction that matters for design. Whether my book influenced the content of the report (contribution) is different from whether my name appears in the footnotes (citation). The AgentGEO study paper points out that existing methods measured only how much a document influenced the answer, not the citations that actually send traffic back. It also flags as a problem that existing methods apply the same editing rules to every document and so cannot diagnose why each individual document is not cited. In effect, the choice of target metric sets the direction of the whole technique.

The analogy has a dark side too. If sellers could freely edit the materials the librarian reads, they could make a defective product look better supported than it really is. SafeGEO paper addresses exactly this risk for recommendation agents. GEO sits on the boundary between making content more readable and manipulating the engine.

How it works: the pipeline from question to citation

Some research has observed how citations are decided inside a generative engine. The ChatGPT citation observation study paper examined, across four configurations from ChatGPT 5.2 to 5.5 including Instant and Thinking modes, how ChatGPT rewrites the query, searches for candidate web pages and selects references from among them. The Instant configurations mostly searched only once, whereas GPT-5.4 Thinking often repeated searches several times. When the same question was run repeatedly, pages that appeared more often among search candidates were also more likely to be cited. In other words, a page must first make it into the candidates to be cited.

GEO intervenes in 'my document', the only part of this pipeline the user can change. There are three broad ways to intervene. The first is content editing, such as filling in missing information or adding statistics. The second is structural editing. GEO-SFE paper divides content structure into three layers: the organization of the whole document (macro structure), the way information is split into chunks (meso structure) and visual markers such as emphasis (micro structure). The third is per-document diagnosis. AgentGEO paper classifies citation failure types by stage of the citation pipeline and then finds where each document got stuck.

Since the engine's internals cannot be seen, optimization usually runs as a feedback loop: edit the document, query the engine or a proxy evaluator that imitates the engine to see whether it gets cited, then edit again. AgentGEO's repeated cycle of diagnosis and repair until the document is cited works the same way. Proxy evaluators, however, differ from the real engine. When the observation study above used Google Search and the OpenAI Responses API as proxy metrics, they matched better at the hostname or entity (company or brand) level than at the exact URL level. Even combined, though, they found only about one quarter of the URLs ChatGPT actually cited. A supervised learning model combining external signals reached an AUC (a 0–1 metric of how much better predictions are than random) above 0.74 on GPT-5.3 and GPT-5.5.

How generative engines decide citationsUser question → Query rewriting; Query rewriting → Web search; Web search → Query rewriting(Additional search); Web search → Candidate pages; Candidate pages → Citation selection; Citation selection → Answer with citations; GEO edits → My document; My document → Citation selection(Competes as candidate)User questionQuery rewritingWeb searchCan repeat severaltimesCandidate pagesMore appearances →higher citation odds↑Citation selectionAnswer withcitationsGEO editsContent · structure ·diagnosisMy documentAdditional searchCompetes as candidate
How generative engines decide citations The top row is the process inside the engine, and Thinking mode may repeat searches several times. The only thing GEO can change is 'my document' below, which must compete as a candidate and pass citation selection.

How the field evolved: from blanket rules to per-document diagnosis and intent inference

The starting point was measuring what had changed. Based on its measurements, the AI search comparison study paper proposed four strategic agendas: create content that machines can easily scan and use as evidence; build authority that AI recognizes by securing authoritative external sources; vary strategy by engine and language; and for smaller businesses, find ways to overcome the 'big-brand bias'. Later studies addressed 'how to edit' in increasingly refined ways.

IF-GEO paper tackled the conflicts that arise when one document has to answer many questions. The edits each question calls for clash with one another, while the amount that can be changed is limited. So it first extracts the desired editing direction separately for each representative latent query (divergence), then merges the instructions with the conflicts in mind into a single 'global revision blueprint' (convergence). It also introduced new metrics for whether performance stays stable as queries change. AgentGEO paper took a different direction: it first diagnoses 'why is this document not cited?', then selects tailored repairs from a tool library and applies them iteratively. It raised citation rates by more than 40% in relative terms while editing only 5% of the content, whereas the baseline methods edited 25%. This difference matters because the less you change, the easier it is to preserve the original meaning and quality.

GEO-SFE paper broadened attention, which had focused on semantic edits, to structure. It reported a 17.3% gain in citation rate and an 18.5% gain in subjective quality across six major generative engines, showing a path that refines only structure while leaving meaning intact. QI-GEO paper lifted yet another constraint. Most existing methods assume the user query is known, but in practice content creators do not know the queries. QI-GEO infers the range of user intents from the document itself and identifies content that is missing but needed. On GEO-Bench and its extended version, objective scores rose by up to 15.9% and subjective scores by up to 17.6%, and cases that newly gained citations outnumbered those that lost them by almost two to one.

Viewed as trade-offs, the progression looks like this. Knowing the query allows more precise edits but is less realistic (IF-GEO vs. QI-GEO). Blanket rules are cheap but require many edits and miss each document's specific circumstances (AgentGEO). Changing meaning can have a large effect but risks undermining the truthfulness of the original, while changing structure carries less of that risk (GEO-SFE).

The shadow side: ranking manipulation and risks to recommendation agents

As LLMs increasingly rank products and documents, GEO is also treated as a manipulation problem. The ranking manipulation benchmark GEO-Bench paper was the first to compare previously scattered attack methods directly on the same criteria. The comparison covered black-box methods that attack through prompts without looking inside the model (TAP, Zero-Shot), white-box methods that use the model's internal gradients (STS, RAF, StealthRank) and 10 legitimate C-SEO strategies. Rankings were produced by a fixed open-weight model (Llama-3.1-8B-Instruct), and effectiveness and stealth were measured together across five datasets.

The results were unfavorable for defenders. In adversarial attacks, effectiveness and stealth traded off against each other. Yet simply rewriting content in a black-box manner matched or beat gradient-based attacks at raising rankings, and produced more natural text. In some domains it evaded both keyword-based detection and perplexity-based detection (perplexity is a measure of how unnatural text is). Whether an attacker had access to model internals also did not predict attack strength. In other words, rewriting into plausible text alone is dangerous enough.

SafeGEO paper applied this risk to recommendation agents. In an evaluation with 22 attack variants and 600 recommendation cases, GEO attacks increased the rate at which defective products entered recommendation lists by up to 83.2 percentage points. Adding simple defenses on the agent side cut harmful recommendations by up to 39.2 percentage points but did not bring them back to the level seen without GEO. If the party that can edit content is the seller, user utility and seller exposure come into direct conflict. This is why those building engines and agents need to design defenses.

Where it is heading: validation outside the lab, on the real web

A major thread in recent research asks whether effects found in controlled experiments persist on the real web. LIVE-GEO paper published GEO-rewritten product descriptions as real web pages and applied five existing GEO strategies (CORE, C-SEO, E-GEO, AgenticGEO and adding statistics) to 80 fictional brands across 18 product categories. In most evaluations CORE had the highest rate of being picked first, at 66.3%–76.3% on real generative engines. However, the ordering of the other strategies varied by evaluator, and engines picked the same top product for the same brand only 48.8%–66.3% of the time. ChatGPT was also sensitive to the order of candidate URLs: under two opposite orderings, CORE's top-pick rate split sharply, at 73.8% and 93.8%.

Who asks and from where is also a variable. A location audit study paper posed 300 questions to ChatGPT from 19 declared locations, and to ChatGPT and Gemini from five network locations, twice each, collecting 31,200 answers. ChatGPT followed the location declared by the app down to the city level, inserting place names into its own search queries, and ignored the network location. Gemini ignored the declared location and followed the country of the network, and most of its sources changed even for questions unrelated to place. The businesses recommended in answers shifted along with the sources. Monitoring that measures from only one location shows only part of the picture.

The ChatGPT citation observation study paper showed that even within the same engine, the number of searches varies by model generation and mode. The finding that external proxy metrics could locate only about one quarter of the actually cited URLs means the tools for measuring GEO effects are themselves still crude. All three studies warn against judging GEO performance on 'one engine, one location, one run'.

Numbers from the real web

48.8~66.3%Rate at which engines picked the same top product
31,200Answers collected in the location audit
About one quarterActually cited URLs found by proxy metrics
Figures from recent studies conducted on the real web. They show how much measurements swing depending on engine, location and measurement tool.

Practical criteria and limitations

When is GEO worth the effort? It is worth investing in for domains with many informational or comparison questions, where users ask generative search and stop after reading the answer. However, given the finding that AI search strongly prefers authoritative external sources paper, polishing the wording of your own pages has limits. Building up external reviews and coverage needs to go hand in hand.

There are four common mistakes. First, measuring performance by 'did it influence the answer?'. What leads to traffic is citation paper. Second, applying the same editing rules to every page in bulk. The AgentGEO study reported that blanket optimization can actually harm long-tail content (niche content that few people search for), and that some documents cannot be fixed by optimization alone. Third, judging effects from one engine, one location and one run. Engines frequently disagree on their top pick, and results change with location and candidate order paperpaper. Fourth, treating proxy metrics as equivalent to actual citations paper.

Clear ethical and risk lines must also be drawn. Adding unsupported claims to increase exposure is of the same nature as the 'boosting defective products' that SafeGEO paper demonstrated. Being able to evade detection paper does not make it acceptable. Those building engines or recommendation agents should assume that GEO-edited sources will be mixed in and design defenses and evaluations accordingly. Finally, most of the improvement figures here come from specific benchmarks and controlled search environments, so they should not be expected to carry over unchanged to real services.

GEO checklist

Measurement

  • Do you count citations, not contribution?
  • Do you measure each engine separately?
  • Do you control for declared and network location?
  • Do you rerun the same question to check variance?
  • Do you distinguish proxy metrics from actual citations?

Editing

  • Did you diagnose the cause of citation failure for each document?
  • Did you avoid applying blanket rules to long-tail pages?
  • Did you keep the amount of editing to a minimum?

Responsibility

  • Did you avoid adding unsupported claims?
  • Are you also securing external authoritative sources?
The common mistakes and risks from the main text, turned into items to check in practice.

Further study

For an introduction, start with the AI search comparison study paper. It shows with measurements how generative search chooses different sources from traditional search and lays out a practical agenda, serving as a map of GEO as a whole. Next, reading AgentGEO paper gives you the core perspectives of 'citation, not contribution' and 'diagnosis, not blanket rules'.

To go deeper into techniques, choose by interest. For structural design, read GEO-SFE paper; for optimization when queries are unknown, QI-GEO paper; and for a single document that has to handle many questions, IF-GEO paper. On the risk side, we recommend reading GEO-Bench paper, which compares ranking manipulation attacks side by side, together with SafeGEO paper, which measures risks to recommendation agents.

Anyone designing measurement and monitoring in practice should read three recent studies: LIVE-GEO paper, which validated effects on the real web; the location audit study paper, which examines how user location changes cited sources; and the ChatGPT citation observation study paper, which covers ChatGPT's process from search to citation and the limits of proxy metrics.

Related terms

RAG Reranker Semantic search AI agent Indirect prompt injection Data poisoning Benchmark LLM-as-a-Judge

References

Foundational papers

Recent research

Written by AI from the listed references and reviewed. Please check the original sources for details.