Answer Engine Optimization (AEO) Strategy: The Earned vs Owned Decision Tree

Answer engine optimisation (AEO) strategy is a way of balancing two kinds of signal across generative discovery platforms such as Google AI Overviews, Perplexity, ChatGPT Search, Claude and Gemini. Earned signals are third-party authority: independent benchmarks, peer evaluations and industry datasets that shape which documents an engine retrieves. Owned signals are structured, machine-readable content on your own site that makes the facts an engine extracts precise. This article sets out a decision tree for choosing between them, based on where a brand stands in generative answers today.

Earned Authority and Owned Precision in Generative Retrieval

Visibility in generative discovery platforms comes from a multi-stage neural pipeline. Candidate retrieval begins with dense vector representations produced by dual-encoder transformer models. Platforms index very large corpora with vector quantisation, and the retrieval paradigms involved include neural generative retrieval, dense passage retrieval, Hierarchical Navigable Small World (HNSW) indexing, reciprocal rank fusion (RRF), click-feedback reranking, entity salience and authority priors, and ALCE citation grounding.

Third-party earned authority behaves like a prior during the initial retrieval. Generative engines sample candidate documents on the basis of cross-corpus consensus, so citations across independent technical benchmarks, peer-reviewed evaluations and industry datasets shift the distribution from which candidates are drawn.

Owned structured content does a different job. When an engine reads the top candidate documents, unstructured web text adds semantic noise and makes entity resolution ambiguous. Owned assets counter that noise with explicit, machine-readable models: Schema.org JSON-LD 1.1 graphs (see the W3C Recommendation under References) and concise, self-contained answer capsules give the engine predicate-and-object facts to ground a response in, which reduces the chance of a fabricated attribute.

The two channels depend on each other. Owned structured content deployed alone does little when engines have not yet established any trust for the brand in the retrieval space. Earned digital PR without owned semantic anchoring tends to produce plenty of mentions that carry wrong attributes or inaccurate technical specifications. With a finite budget, the decision is therefore one of sequencing: which investment to make first, and when to switch.

One simple way to express the combination is as a probability that a document is cited for a query, built from three terms: the similarity between the query and document embeddings, the document's earned authority, and the strength of its owned structured grounding.

$P(\text{citation} \mid q, d) = \sigma \left( \alpha \cdot \langle \mathbf{e}_q, \mathbf{e}_d \rangle + \beta \cdot \log \pi_{\text{earned}}(d) + \gamma \cdot \Phi_{\text{owned}}(d) \right)$

Here $\sigma$ is the logistic function, $\mathbf{e}_q$ and $\mathbf{e}_d$ are the query and document embeddings, $\pi_{\text{earned}}(d)$ is the document's earned authority, $\Phi_{\text{owned}}(d)$ is a score for its owned structured grounding, and $\alpha$, $\beta$ and $\gamma$ are weights. This is a conceptual model for reasoning about the trade-off, not a published formula used by any engine.

A Four-Phase Diagnostic and Allocation Protocol

1. Phase 1: Generative retrieval diagnostic. Run automated probes across several answer engines to estimate how often the entity appears, $P(\text{mention})$, and how often it is recommended when it does appear, $R_{\text{rec}}$.
2. Phase 2: Gap triage. Classify the brand into one of the diagnostic states in the next section, using the mention rate, the recommendation preference ratio and the rate of factual errors in what engines say about the brand.
3. Phase 3: Dual-channel allocation. On the earned side, seed authoritative third-party technical benchmarks, peer evaluations and industry datasets. On the owned side, publish JSON-LD entity graphs and short answer capsules, and keep server response times healthy so pages remain easy to crawl.
4. Phase 4: Aggregation and continuous verification. Combine candidates from both channels with reciprocal rank fusion, and keep monitoring for semantic drift in how engines describe the brand.

When to Invest in Earned Proof or Owned Content

The decision tree has five starting states. Each one points to a first move and explains why other moves would be wasted at that stage. The state boundaries are thresholds that each team sets for itself from its own probing data. This article does not claim universal values for them.

When a brand is rarely or never mentioned, it is in an awareness and trust gap. The entity is largely absent from what the engines retrieve, so on-page content or schema optimisation produces little return: there is no retrieval foothold for the owned content to refine. The right first move is earned proof, meaning third-party evidence that puts the entity into the corpora engines draw on.

When a brand is mentioned often but seldom preferred, the gap is one of proof and preference. Engines know the brand but do not choose it. Third-party comparative evaluations and technical benchmarks are the move, since they supply the comparison material that recommendation depends on.

When engines recommend the brand but state incorrect facts about it, such as wrong prices, architecture details or feature capabilities, earned coverage is already sufficient to secure retrieval. The problem is semantic noise in ungrounded web text, and owned schema injection and atomic answer capsules directly restore factual precision. This is the answer to the question of when to prioritise owned assets over earned media: when attribute distortion, rather than absence, is what holds back conversion.

When a competitor dominates comparative answers, the gap is a comparative proof deficit, and the move is a hybrid: earned head-to-head evidence together with owned feature matrices that machines can parse.

When the brand is already recommended accurately, the work shifts to compounding and defence: owned conversion pages, continuous monitoring of how engines describe the brand, and regular schema updates.

Written as a rule, with thresholds $\tau_{m}$, $\tau_{r}$ and $\tau_{e}$ chosen by the team for the mention rate, the preference ratio and the factual error rate:

$\Omega_{\text{alloc}}(S) = \begin{cases} \text{Earned Proof Priority} & \text{if } P(\text{mention}) < \tau_{m} \\ \text{Earned Trust Priority} & \text{if } P(\text{mention}) \ge \tau_{m} \land R_{\text{rec}} < \tau_{r} \\ \text{Owned Precision Priority} & \text{if } \text{Error}_{\text{param}} > \tau_{e} \\ \text{Hybrid RRF Expansion} & \text{otherwise} \end{cases}$

Quantifying Information Gain and Fusing Channels

Generative answer engines evaluate a document's usefulness by how much it reduces uncertainty about the answer. Relative information gain is the reduction in conditional generative entropy achieved by conditioning the response on a verified document. Let $Y$ denote the generated response and $d$ a candidate document. The expected information gain is the reduction in Shannon entropy:

$I(Y; d) = H(Y) - H(Y \mid d)$

where $H(Y)$ is the entropy of the answer distribution before conditioning on $d$. A retrieval pipeline that selects context documents to maximise this quantity favours documents that make the answer more determinate, and it penalises redundant or ungrounded passages.

To merge candidates from different channels, pipelines commonly use reciprocal rank fusion. Let $\mathcal{M}$ be the set of retrieval channels (dense embeddings, lexical BM25 postings and knowledge graph entity lookups), $r_m(d)$ the rank of document $d$ in channel $m$, and $w_m$ the weight of that channel. A fusion score modulated by semantic certainty can be written as:

$\text{RRF}(d) = \sum_{m \in \mathcal{M}} \frac{w_m}{k + r_m(d)} \cdot \left[1 - \frac{H(Y \mid d)}{H_{\max}}\right]$

The first factor is standard reciprocal rank fusion, in which $k$ is a smoothing constant that is commonly set to 60. The second factor is an illustrative veto: when the semantic entropy of a response conditioned on the document approaches a critical threshold $H_{\max}$, the factor attenuates the document's score. This stops high-variance passages from contaminating context synthesis. The weights $w_{\text{earned}}$ and $w_{\text{owned}}$ express how much trust the system places in each channel and should be calibrated against a team's own query workload.

A Four-Phase Information Gain Protocol

1. Prior entropy sampling. Sample several stochastic generations from a language model to estimate the baseline answer entropy $H(Y)$.
2. Conditional entropy quantification. Condition generation on candidate document $d$ and measure the residual semantic entropy $H(Y \mid d)$.
3. Reciprocal rank weighting. Compute the composite fusion score using calibrated channel weights and the smoothing constant $k$.
4. Coverage verification. Check that the document's vector projection covers the region of semantic space occupied by competing sources.

Diagnosing Omission, Cannibalisation and Distortion

Diagnosing weak performance in answer engines means moving beyond keyword rank checks to multi-stage fault classification. Underperformance falls into four failure modes: dense retrieval omission, attribution cannibalisation, entity ambiguity and generative hallucination. They must be separated, because applying an owned on-page fix to an earned retrieval failure produces no lift.

Dense retrieval omission occurs when a brand's core offerings do not appear in the top candidate set of the engine's index. Automated probing across engines shows low citation recall. The usual root cause is sparse third-party co-occurrence of the entity with its topic in training and retrieval corpora. The remedy is earned proof: independent technical evaluations, open-source repository activity and peer-reviewed or benchmark datasets.

Attribution cannibalisation occurs when engines credit a competitor for features the target brand pioneered. The signature is frequent appearance together with a low preference ratio, and the root cause is a competitor's dominance of third-party comparative review content. The remedy is joint earned syndication together with head-to-head technical evaluations.

Factual parameter distortion is the case where unstructured prose induces wrong attribute values in generated answers. The remedy is owned: a JSON-LD entity graph and atomic answer capsules that state the facts unambiguously.

Semantic entropy instability appears when conflicting assertions about the brand exist across public sources. The remedy is hybrid: canonical schema claims on owned pages and an audit of third-party citations that contradict them.

A Four-Phase Remediation Protocol

1. Telemetry probing. Query target answer engines with a standardised set of technical prompt variants and capture citation recall, the preference ratio and factual precision.
2. Failure isolation. Classify the failure: low citation recall indicates omission, a low preference ratio indicates cannibalisation, and a high error rate indicates entity ambiguity.
3. Targeted channel execution. Dispatch earned syndication for retrieval deficits and JSON-LD graph injection for entity inaccuracies.
4. Entailment verification. Run natural-language-inference entailment audits to confirm that engine statements about the brand are supported by the owned sources.

Machine-Readable Standards for the Owned Channel

Informal layouts and unstructured prose force an answer engine to rely on heuristic parsing, which degrades factual precision. Establishing machine-readable authority means following open standards. The foundation is the W3C JSON-LD 1.1 Recommendation. Schema should not be deployed as isolated fragments. A connected @graph is the better model, in which every entity node (for example Organization, SoftwareApplication, TechArticle and WebPage) declares an unambiguous @id, typed relationships, and sameAs links that connect owned assets to authoritative third-party knowledge repositories.

The following Schema.org configuration is an illustrative payload for the dual-channel model. Replace the host with your own domain.

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "TechArticle",
      "@id": "https://example.com/aeo-strategy#article",
      "headline": "Answer Engine Optimization (AEO) Strategy: The Earned vs Owned Decision Tree",
      "dependencies": "RFC 9110; W3C JSON-LD 1.1",
      "proficiencyLevel": "Expert",
      "about": {
        "@type": "Thing",
        "name": "Answer Engine Optimization",
        "sameAs": "https://example.com/glossary/answer-engine-optimization"
      }
    }
  ]
}

At the transport layer, HTTP Semantics (RFC 9110) governs content negotiation, caching and conditional requests. Origin servers help crawlers by answering quickly, sending clean ETag headers and declaring canonical URLs. Combining deterministic transport behaviour with JSON-LD entity graphs and third-party citation anchors is what links the owned and earned channels into one strategy.

References

• JSON-LD 1.1 (W3C Recommendation) — W3C, 2020. https://www.w3.org/TR/json-ld11/
• HTTP Semantics (RFC 9110) — IETF, 2022. https://www.rfc-editor.org/rfc/rfc9110
• Retrieval-Augmented Generation for Large Language Models: A Survey — Gao et al., 2024. arXiv:2312.10997
• FActScore: Fine-grained Atomic Evaluation of Factual Precision in LLMs — Min et al., EMNLP 2023.