Content optimisation cannot guarantee that a page will appear in, influence or be cited by an AI-generated answer.
That is an important place to start.
Generative search involves several stages between publishing a page and receiving a citation. A source may need to be discovered, indexed, retrieved, selected into context, used during generation and then attributed by the system. Even after that, a citation may contribute little to the substance of the answer and may produce no website visit at all.
However, this does not mean content optimisation is irrelevant.
Existing academic research and search-engine documentation indicate that factors including relevance, evidence, information structure, originality and technical accessibility can influence different parts of this process.
The practical implication is less dramatic than many claims made about generative engine optimisation, or GEO, but potentially more useful:
A modern content audit can assess whether a page is well prepared to be discovered, understood, selected and reused by generative search systems.
I call this AI Citation Readiness.
It is not a ranking score. It is not a probability of being cited. It is an editorial framework for identifying weaknesses that may make useful information harder for retrieval and generative systems to use.
What is AI citation readiness?
AI citation readiness is the extent to which a piece of content is technically accessible, semantically relevant, well organised, evidence-supported and sufficiently useful to function as source material within a generative search system.
The distinction between readiness and probability matters.
A page could satisfy every item in an AI citation audit and still never receive a citation. Another page could be poorly structured yet be selected because it contains a unique fact required for a particular query.
Citation depends on more than the content itself.
The query matters. Competing sources matter. The retrieval system matters. The generative model matters. The composition of the available context matters. Different platforms may behave differently, and the same platform may produce different outputs across repeated runs.
AI citation readiness therefore asks a narrower question:
Have we removed avoidable obstacles that could prevent useful information on this page from being discovered, understood, supported and reused?
That is a question a content audit can reasonably attempt to answer.
Where content optimisation fits into the generative search process
One reason discussions about GEO become confusing is that several different outcomes are treated as though they were the same thing.
They are not.
A simplified process looks something like this:
Discovery/indexation → retrieval → context selection → generation → citation → citation absorption → downstream visit/conversion
Each stage represents a different problem.
Discovery and indexation
The system first needs some mechanism through which the source becomes available.
For Google’s generative Search features, Google states that pages must be indexed and eligible to appear in Search before they can appear as supporting links.
That establishes a basic principle: content quality cannot compensate for technical ineligibility.
Retrieval
A system then has to decide that a source is relevant enough to a query, or to one of the related searches it conducts, to retrieve it.
Google’s documentation describes generative Search features as being rooted in its existing Search systems and discusses retrieval-augmented generation, or RAG, in which relevant information is retrieved and supplied to a generative system.
Google also documents query fan-out, where multiple related searches may be issued to develop an answer.
This means the retrieval opportunity may be broader than the exact words entered by the user.
Context selection
Being available to retrieval does not necessarily mean a document becomes part of the context used to produce an answer.
A system may have more candidate sources than it can meaningfully use.
Generation
The model generates its response from the information available to it, combined with its underlying capabilities and whatever system-specific processes are involved.
Citation selection
A source may be used, selected or presented as a citation.
This is a distinct outcome from merely being retrieved.
Citation absorption
A citation can also vary in importance.
A source might receive a link while contributing only a minor supporting fact. Another source could substantially shape the generated explanation.
Researchers Zhang Kai, He Xinyue and Yao Jingang describe this distinction as citation selection versus citation absorption.
Visit and conversion
Finally, even a prominent citation does not guarantee a click, lead or sale.
Visibility within an AI answer, citation by the system, influence on the response and commercial performance should therefore be measured separately.
Performance at one stage does not guarantee performance at the next.
That distinction should also shape how we interpret GEO research.
What existing research tells us about content and AI citations
Research into generative search optimisation is still developing.
Some studies provide useful evidence that content characteristics can affect generated answers, but the strength of that evidence varies considerably depending on what is being tested.
GEO: Generative Engine Optimization
One of the foundational studies is GEO: Generative Engine Optimization, published at KDD 2024 by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande.
The researchers developed GEO-bench, involving 10,000 queries across multiple domains, and tested different modifications to source content.
Among the reported results:
| Modification | Position-adjusted visibility | Relative to baseline |
|---|---|---|
| No optimisation | 19.3 | — |
| Keyword stuffing | 17.7 | -8.3% |
| Fluency optimisation | 24.7 | +28.0% |
| Cite sources | 24.6 | +27.5% |
| Add statistics | 25.2 | +30.6% |
| Add quotations | 27.2 | +40.9% |
Adding quotations produced the strongest result for this particular metric, while statistics and source citations also produced substantial improvements.
Keyword stuffing did not improve the principal visibility metric.
These findings are often simplified into claims such as “adding quotations increases AI citations by 40%”.
That interpretation goes too far.
The experiment retrieved the top five Google results and supplied those sources to the generative system. It then tested what happened when content within this existing source set was modified.
The defensible conclusion is:
The study found that modifying already-retrieved sources could materially change their visibility within generated answers.
It does not demonstrate that adding statistics or quotations to an arbitrary webpage will increase its probability of being organically retrieved from the wider web by 30 or 40 per cent.
That difference is fundamental.
The study nevertheless provides useful evidence that the characteristics of content can matter after retrieval has already occurred.
Structural optimisation and citation behaviour
A 2026 preprint by Junwei Yu, Mufeng Yang, Yepeng Ding and Hiroyuki Sato examines the relationship between content structure and citation behaviour.
The researchers divide structural optimisation into three levels:
- Macro-structure: the overall architecture of a document.
- Meso-structure: how information is grouped and organised within sections.
- Micro-structure: the way individual information units are presented and emphasised.
Their structural feature engineering approach was tested across six generative engines and reportedly produced a 17.3% improvement in citation rate and an 18.5% improvement in subjective quality within their experimental methodology.
The study is a preprint, so those findings should not be treated as settled evidence of a universal ranking mechanism.
They do, however, provide support for a practical hypothesis: how information is organised may influence how easily generative systems can identify and use it.
This does not mean every paragraph should be rewritten into an artificial “AI chunk”.
In fact, Google’s own guidance specifically rejects the idea that publishers need to split content into tiny pieces purely for generative Search.
Good structure and artificial fragmentation are not the same thing.
Citation selection versus citation absorption
The 2026 preprint From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms provides another useful distinction.
The study covers:
- 602 controlled prompts;
- ChatGPT;
- Google AI Overview/Gemini;
- Perplexity;
- 21,143 valid search-layer citations;
- 18,151 successfully fetched pages; and
- 72 extracted features.
Rather than treating every citation as equivalent, the researchers distinguish between whether a source is selected and how substantially that source contributes to the generated response.
Their analysis found an association between high-influence citations and pages that were more semantically aligned, more structured and richer in extractable evidence.
Examples included:
- definitions;
- numerical facts;
- comparisons; and
- procedural steps.
This does not establish that inserting a definition or table into a page causes an AI system to cite it.
The result is observational.
It does, however, provide a reasonable basis for including extractable information as one dimension of a content audit.
Limits of the current GEO evidence
The limitations become especially important when individual experimental findings are converted into broad SEO recommendations.
A 2026 critical survey by Olivier Martinez reviews 45 GEO studies published between 2023 and 2026.
The survey argues for separating generative visibility into distinct stages including:
retrieval → reranking/context → citation → prominence → absorption → user behaviour
This framework helps explain why a technique can perform well in one experiment without establishing a reliable organic visibility strategy.
For example, the original GEO study showed that modifying content already supplied within a fixed retrieval environment could alter its representation in generated responses.
That is evidence about visibility within that environment.
It is not equivalent to proving that the same modification will make an arbitrary page more discoverable across ChatGPT, Google, Gemini, Perplexity and future AI systems.
Martinez identifies topical relevance and context position among the more reproducible effects in the literature, while warning that generic optimisation heuristics frequently transfer poorly between systems.
The survey’s broader conclusion is particularly important:
No technique reviewed demonstrated a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behaviour.
The survey itself is also an arXiv preprint, but its caution is consistent with the methodological limits visible across this emerging research area.
There is evidence worth acting on.
There is not yet a universal GEO formula.
What Google says about optimising content for generative Search
Google’s July 2026 guidance, Optimizing your website for generative AI features on Google Search, provides a useful first-party reference because it explains how Google wants publishers to think about optimisation for its own generative Search experiences.
The guidance makes several points that cut through some common GEO claims.
Traditional Search foundations still matter
Google says its generative Search features are rooted in its core ranking and quality systems.
Pages still need to be accessible, indexable and eligible for Search.
AI visibility is therefore not presented as a replacement for normal technical SEO.
Retrieval remains important
Google describes the use of retrieval-augmented generation to find relevant material from its Search index.
It also describes query fan-out, where a user’s information need can result in several related searches.
This reinforces the distinction between producing a good passage and being retrieved in the first place.
Useful, original material matters
Google recommends creating unique, non-commodity content and incorporating original viewpoints and experience where appropriate.
This gives content audits a useful question beyond “Is the keyword present?”
A stronger question is:
Does this page contribute something useful enough to justify retrieving it instead of another page saying essentially the same thing?
Organisation should primarily help readers
Google recommends clear paragraphs, sections and headings.
That supports sensible information architecture.
It does not support the more prescriptive claim that publishers must convert all writing into tiny LLM-friendly passages.
Google explicitly says there is no requirement to break pages into artificial “AI chunks”.
There is no special AI-search markup requirement
Google also states that publishers do not need special schema markup specifically for its generative Search features.
Likewise, its documentation says llms.txt does not provide a visibility advantage in Google Search.
Structured data can remain useful when it accurately represents content and supports existing Search features, but it should not be presented as a hidden GEO switch.
The same caution applies to Google’s guidance more generally.
It tells us how Google describes its systems.
It should not automatically be generalised to ChatGPT, Perplexity or every other generative search platform.
The AI Citation Readiness Content Audit
The following framework translates the research and documentation above into six practical audit dimensions.
It is an editorial framework, not a validated ranking model.
1. Discovery and eligibility
What it means
Before content can influence retrieval, the underlying page needs to be technically available to the relevant system.
For Google, that means satisfying the normal requirements for crawling, indexation and Search eligibility.
Why it might matter
Improving the quality of information on a page does little for Google’s generative Search features if the page cannot enter the Search index or is otherwise ineligible.
What to audit
Check whether:
- the page is crawlable;
- the intended URL can be indexed;
- canonicalisation reflects the intended version;
- the main content is accessible;
- robots controls are intentional;
- useful content is not dependent on problematic rendering;
- the page meets normal technical SEO requirements.
This is not specifically “AI SEO”.
That is the point.
Generative retrieval still depends, at least in Google’s documented implementation, on foundational Search eligibility.
2. Query and semantic alignment
What it means
The page should clearly address the information need for which you expect it to be retrieved.
Why it might matter
The research reviewed by Martinez identifies topical relevance as one of the more reproducible factors associated with generative visibility.
The Zhang et al. citation-absorption study also found that highly influential cited pages tended to be more semantically aligned.
What to audit
Identify:
- the dominant entity or subject;
- the primary user intent;
- secondary questions the page answers;
- terminology needed to interpret the subject;
- ambiguous pronouns or references;
- sections that drift materially away from the page’s purpose.
This is not a keyword-density exercise.
A page can repeat a phrase frequently while remaining conceptually vague.
The objective is to make the relationship between the page, its entities and the information need unambiguous.
3. Information structure
What it means
Information should be organised so that the relationship between a question and its answer is apparent.
Why it might matter
The structural GEO preprint reports changes in citation behaviour following macro-, meso- and micro-structural optimisation.
Google also recommends clear organisation through paragraphs, sections and headings.
What to audit
Look for:
- descriptive headings;
- logical section order;
- definitions positioned near the concepts they define;
- explicit comparisons;
- sequential procedures;
- tables where tabular presentation genuinely improves understanding.
Good structure is not synonymous with short content.
Nor does it mean every idea needs its own 40-word block.
A 1,500-word technical explanation can be highly structured. A page containing 50 disconnected snippets can be poorly structured.
4. Evidence density
What it means
Important factual claims should be supported by evidence appropriate to the claim.
Why it might matter
In the Aggarwal et al. experiment, adding citations, statistics and quotations improved visibility metrics among sources already supplied to the generative system.
This does not prove that inserting references automatically improves organic retrieval, but it suggests that evidence-rich content can affect how retrieved material is represented.
What to audit
Identify claims that could be improved through:
- primary sources;
- academic research;
- official statistics;
- attributable expert quotations;
- documented methodology;
- external references.
Evidence should make a claim stronger for the reader, not merely add decorative links.
Ten weak citations do not necessarily make a page more authoritative than one strong primary source.
5. Extractable information
What it means
Some useful information can be expressed in forms that remain meaningful when encountered within a broader generated answer.
Why it might matter
The Zhang et al. preprint found that high-influence cited pages were associated with information including definitions, numerical facts, comparisons and procedural instructions.
What to audit
Look for genuinely useful information units such as:
- concise definitions;
- explicit factual claims;
- numerical evidence;
- comparison tables;
- criteria;
- ordered procedures;
- clearly named entities;
- conclusions that retain sufficient context.
The objective is not to produce robotic fragments.
It is to avoid forcing readers, search engines or generative systems to infer a conclusion that could have been stated clearly.
For example:
Weak:
“These options differ in a number of important ways.”
More explicit:
“Option A has a lower upfront cost, while Option B provides the longer warranty.”
The second statement carries its meaning with less inference.
6. Information gain
What it means
The page should contribute something beyond a generic restatement of material already available elsewhere.
Why it might matter
Google’s current generative Search guidance emphasises unique, non-commodity information, including original viewpoints and experience.
What to audit
Potential information gain might include:
- practitioner interpretation;
- a useful original framework;
- synthesis of research scattered across multiple sources;
- calculated comparisons;
- expert judgement;
- analysis of primary sources;
- templates or tools;
- a new taxonomy;
- first-hand examples.
Originality does not require proprietary survey data.
This article is itself an example.
The underlying studies are public. The AI Citation Readiness Content Audit is the additional contribution: a framework that translates those findings into an editorial process.
AI Citation Readiness scoring framework
Each of the six dimensions can be assigned a simple score:
| Score | Meaning |
|---|---|
| 0 | Absent or serious weakness |
| 1 | Weak |
| 2 | Adequate |
| 3 | Strong |
Maximum score: 18
A simple interpretation is:
| Total | Interpretation |
|---|---|
| 0–6 | Low citation readiness |
| 7–12 | Moderate citation readiness |
| 13–18 | Strong citation readiness |
Important: The AI Citation Readiness score is an editorial content-audit framework, not a search-engine ranking score or statistical prediction of citation probability. It synthesises factors identified in existing research and platform documentation into a practical auditing methodology.
A page scoring 18 is not “three times more likely” to receive a citation than a page scoring six.
No such probability is implied.
The score exists to prioritise improvement.
For example, an otherwise excellent page receiving zero for discovery and eligibility probably has a more immediate problem than a page receiving two instead of three for evidence density.
AI Citation Readiness checklist
A more granular audit can use the following questions:
| Audit question | Yes | Partial | No |
|---|---|---|---|
| Can the page be crawled and indexed normally? | □ | □ | □ |
| Is the page clearly aligned with a defined user information need? | □ | □ | □ |
| Are major entities explicitly identified? | □ | □ | □ |
| Do headings accurately describe the information beneath them? | □ | □ | □ |
| Are important definitions understandable out of context? | □ | □ | □ |
| Are important claims supported by credible sources? | □ | □ | □ |
| Does the page contain useful numerical or factual evidence where appropriate? | □ | □ | □ |
| Are comparisons presented explicitly rather than implied? | □ | □ | □ |
| Are procedures presented as clear sequences where appropriate? | □ | □ | □ |
| Does the content add information beyond generic web consensus? | □ | □ | □ |
| Is authorship or expertise clear where relevant? | □ | □ | □ |
| Has the page been tested across a repeatable set of AI/search queries? | □ | □ | □ |
The checklist is deliberately broader than a single AI platform.
Testing should also be treated as measurement rather than proof of causation. If a page begins appearing after an optimisation, that observation may justify further investigation, but it does not by itself prove which change caused the result.
Examples of content optimisation
The following examples are illustrative only. They are not experimental findings and should not be interpreted as tested citation interventions.
Example 1: vague definition
Before
Citation readiness is an important part of modern SEO and can help websites perform better in new search experiences.
The sentence says very little about what the term actually means.
After
AI citation readiness is the extent to which content is technically accessible, relevant, well organised, evidence-supported and usable as source material within generative search systems.
The revised version supplies an explicit definition.
Example 2: unsupported claim
Before
Structured content receives more citations from AI systems.
That statement implies a broad causal relationship not supported by the available evidence.
After
A 2026 preprint testing structural feature engineering across six generative engines reported a 17.3% improvement in citation rate within its experimental methodology.
The claim is narrower, attributable and appropriately qualified.
Example 3: unclear comparison
Before
The approaches produced different visibility results, with some techniques working better than others.
After
| Aggarwal et al. treatment | Position-adjusted visibility |
|---|---|
| No optimisation | 19.3 |
| Cite sources | 24.6 |
| Add statistics | 25.2 |
| Add quotations | 27.2 |
The table makes the comparison explicit while retaining the need to explain the experimental limitations in surrounding text.
Example 4: poorly organised procedure
Before
Audit the content, make sure the technical setup is right, look at evidence and headings, decide what questions the page answers and then test it.
After
- Confirm discovery and indexation eligibility.
- Define the intended information need.
- Review semantic alignment.
- Assess information structure.
- Identify unsupported claims.
- Review extractable facts, definitions, comparisons and procedures.
- Assess information gain.
- Test the page against a repeatable query set.
The underlying advice is similar, but the procedure is easier to follow and evaluate.
What content optimisation cannot guarantee
The framework becomes less useful if it is presented as more certain than the evidence allows.
Several limits need to remain explicit.
Retrieval dependence
A well-structured passage cannot influence an answer if the source never enters the system’s relevant retrieval or context process.
This is one of the key limitations when extrapolating from experiments in which sources have already been retrieved.
Run-to-run variability
Generative systems are not necessarily deterministic.
Repeated tests may return different sources, wording or citations.
A single observation should therefore be treated cautiously.
Platform differences
Google, ChatGPT, Perplexity and other systems do not necessarily retrieve, select or cite sources in the same way.
Evidence about one platform should not automatically be turned into a universal rule.
Query wording
Small changes to an information request can alter what information is needed and which sources are useful.
Optimising a page for a broad topic therefore does not guarantee that it will be the best source for every related prompt.
Source competition
Citation is comparative.
A page is not evaluated in isolation from the other information available to the system.
A strong page may still lose selection to a source that is more relevant, more specific or uniquely informative for that particular query.
Citation and information use are different
The citation-absorption framework highlights another complication: being cited does not tell us exactly how important the source was to the generated answer.
Likewise, information may potentially influence an answer differently from how citations are displayed.
Citation count alone is therefore an incomplete measurement of source influence.
Traffic and conversion are separate outcomes
Even successful citation does not necessarily create commercial value.
Users may obtain the information they need from the generated response without visiting the source.
For businesses, a complete measurement framework may therefore need to distinguish:
- retrieval or visibility;
- citation frequency;
- citation prominence;
- citation absorption;
- referral traffic;
- leads;
- conversions.
An optimisation associated with improvement in one metric should not automatically be credited with improvement in all of them.
Conclusion
The strongest case for AI-aware content optimisation is also the least sensational one.
Current research does not establish a universal formula that makes pages appear in ChatGPT, Google AI experiences, Perplexity or other generative systems.
There is no validated “GEO score” that predicts citation probability.
What the evidence does suggest is that content characteristics matter at different stages of the generative search process.
Technical accessibility affects whether content can become eligible for retrieval. Relevance affects whether it fits an information need. Structure can make relationships between pieces of information clearer. Evidence can strengthen factual claims. Definitions, numerical facts, comparisons and procedures can make useful information explicit. Original analysis can give a system a reason to retrieve one source rather than another generic summary.
Those are legitimate content-audit concerns regardless of whether an AI citation follows.
A modern content audit therefore should not attempt to write for an LLM.
It should identify and remove obstacles preventing useful information from being discovered, understood, supported and reused.
That is a more modest goal than promising AI citations.
It is also one that the current evidence can support.
References
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 5–16. DOI: 10.1145/3637528.3671900.
https://arxiv.org/html/2311.09735
Yu, J., Yang, M., Ding, Y. & Sato, H. (2026). Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior. arXiv:2603.29979. Preprint.
https://arxiv.org/abs/2603.29979
Kai, Z., Xinyue, H. & Jingang, Y. (2026). From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms. arXiv:2604.25707. Preprint.
https://arxiv.org/abs/2604.25707
Martinez, O. (2026). Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026). arXiv:2607.14035. Preprint.
https://arxiv.org/abs/2607.14035
Google Search Central (2026). Optimizing your website for generative AI features on Google Search. Updated 10 July 2026.
https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
