GEO / AEO Specialist

Dan Petrovic Portfolio

Managing Director of Australian AI SEO agency DEJAN and a two-time Google award winner for discovering exposed internal Google systems, who has coined "Selection Rate Optimization" (SRO) as his own framework for AI-answer visibility.

Answer Engine OptimizationSelection Rate OptimizationAI Search ResearchBrand Visibility Tracking

Location

Not listed — probably too busy shipping to update the bio.

Skills

Answer Engine OptimizationSelection Rate OptimizationAI Search ResearchBrand Visibility Tracking

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

A self-coined, named framework distinct from the industry's shared terms

Rather than adopting "GEO" or "AEO" wholesale, naming his own concept (Selection Rate Optimization) gives a specific, ownable point of view that's easy to differentiate from other practitioners using the same two acronyms.

2

Independent, verifiable technical credentials predating the AI-search wave

Two Google awards for discovering exposed internal systems are checkable achievements from before generative search existed, lending weight to newer AI-search claims.

3

Specific measurement methodology named on the page

Describing the actual technique — bidirectional prompts, frequency and rank analysis, and knowledge graphs — for tracking brand visibility in LLMs is concrete enough to be evaluated, unlike a vague "AI monitoring" claim.

What GEO / AEO Specialists can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • If you can name your own distinct concept within a crowded emerging field, do it — it's more memorable than repeating the same buzzwords everyone else uses.

  • Surface older, unrelated technical credentials if they're genuinely impressive and predate your current specialty.

  • Describe your actual measurement method in specific technical terms rather than a generic "we track AI visibility" line.

Case studies

Dan Petrovic works on AI visibility by building instruments: he probes the models, logs what they do, and trains small classifiers to replicate the decisions. The talk, the grounding model and the SRO method all come out of that loop.

Case StudyadvancedPresented at SEO Week 2025, New York

Beyond Rank Tracking: Analyzing Brand Perceptions Through Language Model Association Networks

Bidirectional promptingStructured JSON outputKnowledge graphsPageRank / centralityCross-model aggregation

A framework for measuring how LLMs perceive a brand. It prompts both ways, brand to associated entities and entity to associated brands, forces JSON output, and normalizes entity variants. Rankings are aggregated weekly across models and scored on frequency plus average position. The collected associations become a graph, where PageRank surfaces the brands most central to a topic.

It turns "AI visibility" into a repeatable measurement: defined prompts, normalization rules, a weighted score and a graph metric. Anyone can rerun it, including on competitors.

  • Brand-to-entity and entity-to-brand prompts, run in a loop and aggregated weekly

  • Volatility read as a signal: low oscillation in mentions means stronger brand authority

  • PageRank over the collected association graph to rank brand influence

Case StudyadvancedOngoing

Selection Rate Optimization (SRO)

Grounding snippet reconstructionCross-encoder re-rankingHypothesis testingGoogle AI Mode / GeminiChatGPTClaude

DEJAN's named method for AI search. It optimizes Selection Rate, the probability that a model picks your snippet once your page is in the retrieved candidate set, rather than rank position. The pipeline reconstructs the grounding snippets AI Mode and Gemini would extract, scores them with a cross-encoder trained on Bing search data, then tests edits, keeping those that lift selection and discarding those that don't.

It splits AI visibility into two contests, getting retrieved and getting chosen, and puts a metric and a test loop on the second one, which most GEO advice leaves vague.

  • Selection Rate defined as the AI analogue of CTR

  • Grounding-snippet engine tuned to Google AI Mode and Gemini extraction

  • Edit, re-rank, keep-or-discard loop instead of assumed model preferences

Projects

ML ModeladvancedPublished April 2025

Query Deserves Grounding (QDG) Classifier

DeBERTa-v3-largeGemini 2.5 Pro API (search grounding)Hugging Face TransformersWeights & BiasesSynthetic data

A binary classifier that predicts whether a prompt will trigger a web-search grounding step or be answered from the model's own training data. He ran 10,000 prompts through Gemini 2.5 Pro with grounding enabled and labeled each one by whether Google chose to ground it. He padded the minority class with Gemini-generated examples, then fine-tuned DeBERTa-v3-large to replicate Google's decision.

It separates prompts where search visibility matters from prompts where only training data matters, which is a real GEO targeting decision. He answered it with measured data and an open model rather than an opinion.

  • 10,000 Gemini 2.5 Pro prompts labeled by Google's own grounding decision

  • Fine-tuned DeBERTa-v3-large, 90:10 train/validation split, best checkpoint picked by F1

  • Open weights on Hugging Face (dejanseo/query-grounding)

Ready to build your portfolio?

Follow the GEO / AEO Specialist roadmap — skills, projects, and timeline to get hired.

GEO / AEO Specialist Roadmap