Research
Original first-party research with data behind it.
Gemini writes better answers when it knows the user is a recognised expert
Tell Gemini you're a known expert and its answers get better. Across 15 topics, blind-rated answers scored 7.8 out of 10 for experts and 5.8 with no name.
We Trained AI. Now AI Trains Us.
People take on AI's reasoning as their own, and it shapes how they decide next time.
We Tested an AI for Signs of Inner Experience. Nobody's Home.
Eight behavioural tests on Gemini found no evidence of AI consciousness, though results reveal how the model categorises and judges its own self-concept.
Fanout Vectors
An analysis and reproduction of Google's query fanout research, demonstrating how vector-based diffusion models could replace text fanouts for faster search.
Google's Approach to Recursive Self-Improvement
Dream-RSI optimizes AI agent search strategies offline by replaying past discovery trees as exact simulators, reducing execution costs across discovery tasks.
Fanout V2: From Paraphrastic Collapse to Complementary Set
Google Research introduced R4T, an RL-trained diffusion retriever that optimizes fan-out query sets for diversity and speed, achieving up to 20x latency speedups.
Impact of Parametric Memory and Grounding on Ordinal Positioning of Brands in Generative Responses
Mathematical formulas for evaluating entity association, semantic embedding similarity, ordinal rank positioning, and generation performance in AI models.
AI Influence With Confidence Threshold Manipulation
A study reveals that language model abstention is governed by a single internal confidence signal that outranks question difficulty and fact retrievability.
Anti-AI sentiment attaches to visible AI: labels, generated media, chatbots and data centers
Experiments and survey data show growing public distrust of AI, with labels reducing content value and widespread opposition to local data centers.
Google, Anthropic and OpenAI crawlers fetched dejan.ai's robots.txt 1,150 times and its llms.txt 4 times in 30 days
In a 30-day server-log census of dejan.ai's three machine files, the big-three AI crawlers fetched robots.txt 1,150 times combined, llms.txt 4 times, and the OKF content bundle 23 times.
A free 300M model predicts Gemini's embeddings at 0.83 cosine
Mapping Gemma embeddings to Gemini embeddings via a linear transform recovers entity vector structures with high fidelity without calling the paid API.
Finding Bard Inside Google's Gemma 4
A mechanistic interpretability study shows how ablating specific MLP neurons in Gemma 4 causes the model to revert to its older, dormant Bard identity.
XM on text diffusion: a small real gain, and a bigger effect we were not looking for
An evaluation of Explorative Modeling on a text diffusion model: a small consistent gain at matched steps, and a larger likelihood-versus-sampling tradeoff that turned out to belong to the training schedule.
Gemini picks whichever web page it reads first 92% of the time. Only pages rewritten to its measured preferences ever beat that.
Shown two web pages, Gemini picks the first one 92% of the time. Pages rewritten to 31 measured text preferences win 15% from the second slot. Ordinary pages: 0%.
No, AI doesn't prefer Reddit. Search does.
Analysis of citation mining data shows Reddit has a 99.39% rejection rate from OpenAI, zero citations from Anthropic, and heavy usage by Google.
Engadget: Quantitative Linguistic Analysis
This paper analyzes Engadget's article corpus from 2006 to 2023 using Shannon entropy and data compression to track editorial trends and article length.
The Verge: Quantitative Linguistic Analysis
This paper profiles the recovered corpus of The Verge from 2011 to 2023 using Shannon entropy and data compression to analyze changes in article structure.
The Paradox of Average and AI Visibility
A study by DEJAN AI examines how AI rankers prioritize prototypical content over brand authority, suggesting a dual strategy for SEO and human conversion.
Content Optimization Engine Insights by DEJAN AI
We spent 2.71 Billion tokens teaching a machine to reverse-engineer how AI search ranks pages. What it found overturns the usual SEO playbook: structure and intent beat every credential you've been told to add, and the engine that proved it reached #1 in 1,554 of 2,246 runs.
Emotion Geometry of Google’s AI Models
A replication study of Anthropic’s emotion research on Google’s Gemma 4 31B model, finding that internal emotion representations organize along a valence axis.
Gemma 4 Brand Authority Map
A comparison of brand recall between Google's Gemma 4 and Gemini 3 Flash models, analyzing how open-weight and closed models prioritize different brands.
AI Brand Authority Index: Ranking 2.9 Million Brands by Associative Embeddedness in Gemini’s Memory
This research presents a methodology for quantifying brand authority in large language model memory using Personalized PageRank and directed association graphs.
Reverse Prompting: Reconstructing Prompts from AI-Generated Text
A fine-tuned Gemma 3 270M model reconstructs the most likely prompts from AI-generated responses using synthetic data and contrastive search configurations.
Is Query Length a Reliable Predictor of Search Volume?
An analysis of 39.6 million Amazon search queries reveals that query length is an unreliable predictor of search volume compared to semantic content.
SRO & Grounding Snippets
Selection Rate Optimization (SRO) is a new discipline focused on visibility in AI-powered search by measuring how often content is selected for grounding.
Most People Don’t Read
A qualitative study comparing self-reported reading habits against actual user behavior, tracking mouse movements, scroll patterns, and time on page.
How big are Google’s grounding chunks?
Analysis of how Google selects content to ground Gemini-powered AI shows a fixed 2,000-word budget per query, where relevance rank determines word share.
How Long Are Web Pages?
An analysis of 44,684 web pages reveals a median content length of 3,201 tokens and an average of 10,403 tokens, highlighting implications for AI systems.
Report: How People Use AI at Work
An analysis of qualitative interviews with 1,250 professionals exploring how the general workforce, creatives, and scientists integrate AI into their work.
How do people use AI assistants?
An analysis of 3.9 million AI chat sessions reveals that most interactions are short, non-commercial, and involve users seeking help with writing, learning, or coding.
BlockRank: A Faster, Smarter Way to Rank Documents with LLMs
BlockRank is a novel method for in-context ranking that uses structured sparse attention and contrastive training to improve LLM efficiency and accuracy.
In AI SEO #10 is the new #1
An empirical study analyzing how Google's AI Mode uses text snippets from multiple sources, finding that snippets are more prompt-aligned than full web pages.
AI Search Citation Mining
Raw data dump from a citation mining pipeline demo featuring 60 prompts across AEO, AI marketing, AI optimization, AI SEO, and AIO using GPT-5 and Gemini.
Using GPT-5 Structured Output Markers to Detect AI-Generated Content Online
Publishing unedited AI-generated text can leak internal GPT-5 structured output markers like turn0search21, which can lead to SEO and reputational risks.
Primary Bias on Selection Rate in AI Search
Selection Rate measures how often AI systems select specific items from grounding results. It explores primary bias, model relevance, and the Tree Walker algo.
Does Schema Help With “AI”?
An experiment testing whether OpenAI's browsing tool provides GPT-5 with grounding context from page schema or only extracts plain text and markdown content.
People call them AI. That’s it.
Social media poll results from 864 votes show that while AI is the dominant label for tools like ChatGPT and Claude, users remain divided on preferred terms.
Analysis of Gemini Embed Task-Based Dimensionality Deltas
An analysis of Gemini Embed optimization modes, including classification, retrieval, and semantic similarity, through vector embedding dimension visualization.
Vector Embedding Optimization
An evaluation of four embedding methods comparing speed, storage, and accuracy. Results show mrl truncation maintains high accuracy while reducing file size.
LLM-Based Search Volume Prediction
An analysis comparing Google Gemini's keyword volume predictions against actual Google Search Console data reveals weak-to-moderate correlation and limited accuracy.
Cross-Model Circuit Analysis: Gemini vs. Gemma Comparison Framework
A framework for comparative circuit analysis between Google's Gemini and Gemma models to identify how different architectures represent brand information.
Beyond Rank Tracking: Analyzing Brand Perceptions Through Language Model Association Networks
The DEJAN methodology uses large language models to analyze brand perception and semantic associations, moving beyond traditional keyword rank tracking.
Both humans and AI return similar results when asked for a random number
A comparison of 200,000 random numbers provided by humans and Google's Gemma-2-2b-it model reveals significant overlaps and patterns in number selection.