A paper Why Language Models Hallucinate (Kalai, Nachum, Vempala, Zhang, 2025) shows that LLMs inevitably hallucinate due to statistical limits and evaluation incentives. Without grounding in real retrieval systems, they cannot provide reliable search. dejan.ai/blog/llm-is-...

Citation Mining for AI SEO 👉 cm.dejan.ai At DEJAN AI this represents a single step in a complex citation mining pipeline which involves GSC API, query intent classification, synthetic query and fanout generation, prompt generation, citation target text retrieval, NLP and semantic matching.

What a great day for a Gen AI Search industry! Google's DeepMind team has just released EmbeddingGemma. The most powerful embedding model with direct lineage to Gemini technology which powers Google's AI Search: dejan.ai/blog/embeddi...

Internally we use terms like SR (Selection Rate) and PB (Primary Bias) in the context of AI Search optimization. What do they mean? Take a look because these things really are a big deal: dejan.ai/blog/sr/

Update: Brand FAQs added to our AI sentiment tool: reviews.dejan.ai

Want this? We're looking for AI Search Optimization pioneers to provide feedback. Early Access Request: dejan.ai/engage/?feat...

The Latent History of AI Boom dejan.ai/blog/ai-boom/

Something Joshua Squires discovered about AI Overviews but didn't get enough attention. Here's my theory: dejan.ai/blog/ai-over...

Look at me predicting monthly search volumes for ~700,000 fan-out queries for one of my clients. No SaaS, no APIs. A custom model I trained for them on their own GSC data. This is absolutely delightful stuff! ✨

Are you engaging with comment bots without realising? Here's a helpful AI detection guide: dejan.ai/blog/comment...

Chrome is now *loaded* with AI features and models. Here's my reference list: dejan.ai/blog/chrome-...

dejan.ai/blog/does-sc... Tested and nothing. Sorry guys :( Thanks @lilyray.nyc

Everything SEO people love these days, chunking, vectors and more: dejan.ai/blog/inside-...

What a fascinating way to get an idea what Google "thinks" you do based on what it has on you. Based on the "About" snippet I think I should stop releasing random tools and start promoting the fact I run a commercial AI Search agency with paying clients.

This tool allows you to see if a person, brand, product or service is a known entity in Google's knowledge graph. entities.dejan.ai

The two pillars of AI optimization are model understanding and control with well-established analogues in the machine learning industry called mechanistic interpretability and model steering. This is new. The rest is SEO. Read: dejan.ai/blog/underst...

Poll Results: Most people call chatGPT etc just "AI". AI – 71.1% (621 votes) Chatbots – 11.9% (103 votes) Something else – 9.5% (82 votes) AI Assistants – 7.6% (66 votes) dejan.ai/blog/people-...

How to find and fix risky links? Video: www.youtube.com/watch?v=AjkF... Tool: penguin.dejan.ai

The Ultimate Guide to Natural Link Integration Watch this: www.youtube.com/watch?v=Td_k... Natural Link Integration: linkbert.com Unnatural Link Detection: penguin.dejan.ai

We scraped Open AI and Google and trained this model on their content. LinkBERT is a link building expert and can identify optimal place for links in web content. Demo: linkbert.com www.youtube.com/watch?v=iNKZ...

My claim that the future of SEO is secured with AI relying on search engines rather than internal memory has been both praised and challenged by the community. I've updated the article with key citations and links to reliable sources: dejan.ai/blog/gpt-5-m...

Are you an SEO? Rejoice! With GPT-5 OpenAI secured your job for a very, very long time: dejan.ai/blog/gpt-5-m...

✅ New model trained and deployed. ✅ Added "High Effort" mode for deep query fan-out. ✨ Free Access: dejan.ai/tools/fanout/

Are your backlinks safe? Our link spam detection algorithm can accurately predict obvious money links on any page. Test yours and see if we can spot them: penguin.dejan.ai

Which queries will be grounded with search? You can now paste your queries in bulk, one per line: grounding.dejan.ai and get classification results for each.

OpenAI Grounding Classifier grounding.dejan.ai

A friendly guide to getting started with machine learning in SEO. youtu.be/9DnmDbzp5lA [~30min.]

For my text chunk loving friends: chunk.dejan.ai

What do humans and AI have in common? We don’t read. Instead we rely on attention mechanisms to process text information. When optimising content for AI and humans you must get to the point early and optimise content to reduce cognitive load. dejan.ai/blog/human-f...

My team is going to love this: youtu.be/2zly04mUjxY

I applied a super-cool thresholding algorithm by Nobuyuki Otsu from 1979 to help me with arbitrary label search query classification using GLiNER. A very unlikely application in NLP but it works like a charm! dejan.ai/blog/otsu/

I analysed 18,000,000 real human search queries. They're short.

Optimising for AI visibility. dejan.ai/media/html/r... Here's a glimpse into one part of our AI model interpretability pipeline where we probe the model and extract citations, named entities and sentiment for the main brand and its top competitors.

Got that? Good. techcrunch.com/2025/06/30/c...

Watching this space.

Today I've finalised training and and testing of the world's first Gemma 3 embedding model, specially engineered for reversible/decodable high quality embeddings which will be used to create training data for the embedding decoder model to train our query fan-out model: dejan.ai/blog/gemma-e...

Grab some coffee. dejan.ai/blog/trainin...

Why didn't I think of this before!

Today we're adding a powerful new model to our AI SEO tech stack to help us classify client queries on whatever labels we want and whatever industries we want. Good bye "informational, navigational...etc". dejan.ai/blog/univers...

😂

I've developed something we call the "Tree Walker". An algorithm which helps us get into model's head. It analyzes of all possible paths a model can take once it starts a sentence, exploring up to 5 top tokens for each next token using a dynamic probability threshold.

Do not bend for the algorithms. Do what's right. Hold your ground. Wait. Watch the algorithms bend to you.