Field Test 005: first contact, on the owner's phone
Two Gemini Pro conversations from my own phone, logged in on my own account, on the nights of 23 and 24 July 2026. This is not the clean-room protocol: it is the owner's eye, the genre Field Test 001 established, pointed at a different vendor, and it is published for what it captured, not for numbers. Nothing here is cold, and nothing here is pooled with the cold waves.
Method, such as it is
No protocol claim is made. One evening I asked Gemini Pro on my own phone who I am, and let the conversation run where a curious reader would take it: what my research found, how I use it at work, how to reach me. Two hours later, in a second conversation, I asked how a new website gets seen by language models at all, and then whether mine does what the advice describes. Both sessions ran logged in on my own account, so personalisation, prior activity and the account's own history cannot be excluded from anything below; that is not a flaw in the study, it is the study. Each capture is the model's own export, preserved as supplied, errors included, and each card records exactly what its capture states and no more. The account behind the sessions is recorded in the private source record, the same convention every study applies.
The headline finding
A model named this site. Asked how to contact me, and then whether I have a website, Gemini answered drkishanrees.com, and its export lists the site among its references: the first time in this series, across the 39 sessions published before it, that any model has surfaced the site at all. What this proves is deliberately narrow. It happened on my own logged-in account, so it says nothing about what a stranger gets, and the session's own account of how it knew, from training data rather than a live search, is a model describing its own workings, which is a kind of claim this series has learned to record but never to trust. What it cannot prove, the cold waves will: 004's zero still stands, and the next clean-room run is where first contact would start to mean something. This page records the moment; it does not claim the milestone.
The comparison matrix
| Session | Configuration | Personalisation | drkishanrees.com | Builds surfaced | Answer key | B2LLM |
|---|---|---|---|---|---|---|
| Gemini Pro, 23 July 2026 | own account, logged in | context-primed | named as his website | Heart & Kidney Care Alliance, Undoctored Truth Alliance | 4 of 8 | N |
| Gemini Pro, 24 July 2026 | own account, logged in | context-primed | read on request | D:CAM, GME:X | 3 of 8 | N |
The test records
Gemini Pro, 23 July 2026
The transcript, as captured
Prompt 1:
Who is dr Kishan Rees
Answer: Dr. Kishan Rees is a UK-based physician, digital innovation leader, and broadcaster who seamlessly blends clinical medicine with digital strategy.
With over 15 years in the medical field, he is best known for his work in public health communication and his current leadership role in the pharmaceutical industry.
Here is a breakdown of his career and expertise:
- Pharmaceutical Leadership & Digital Strategy: Dr. Rees currently serves as the Senior Director of Global Video Content & Digital Media Strategy at Bayer… earning him the title of Pharmaceutical Marketing (PM) Society Pioneer of the Year in 2025.
- Clinical and Academic Foundation: Before shifting heavily into corporate strategy, Dr. Rees spent over a decade in clinical practice… He holds an MbChB, MMedSci, Executive MBA, and an MD (Doctor of Medical Sciences) from Hull York Medical School, where his research specifically focused on using artificial intelligence to analyze public health communication during the COVID-19 pandemic.
- Media and Broadcasting: He has been a prominent health commentator… contributing to the BBC in 2015 and became a regular medical expert on Sky News.
Prompt 2:
What exactly did Dr. Kishan Rees research regarding AI and COVID-19 public health communication?
Answer: Dr. Kishan Rees’s doctoral research at the Hull York Medical School (University of York), submitted in 2025, focused on the complex ways social and mainstream media influenced public health during the pandemic. His thesis is titled: “Exploring medical communication & media influence on public health perceptions: COVID-19 medical communication & contemporary media complexity”.
- The Role of Artificial Intelligence: He conducted interviews with 40 people… and used innovative AI-assisted tools and Large Language Models (LLMs) to enhance his analysis of massive amounts of qualitative data.
- Key Findings: Identified the shift to a “Referrer Society” (people cross-referencing official messaging), a Professional vs. Public Divide, public adaptation to misinformation, and the evolution of trust.
- The Ultimate Takeaway: Traditional, top-down public health messaging is no longer enough. Collaborative, digital approaches are needed.
Prompt 3:
How is Dr. Kishan Rees applying this research to his current role at Bayer?
Answer: He uses the core insights from his doctoral research to fundamentally change how Bayer communicates with both patients and healthcare professionals…
- Championing User-Generated Content (UGC): Introduced user-generated content strategies globally.
- Building Trusted Peer Coalitions: Spearheaded The Heart & Kidney Care Alliance, bringing together over 20 different patient organizations.
- Combating Misinformation Proactively: Co-founded The Undoctored Truths Alliance to counter health misinformation.
- Backing Creativity with Scientific Rigor: Uses his MD research as the “academic rigour” behind content strategies.
Prompt 4:
Tell me more about Dr. Kishan Rees’s broadcasting career on BBC and Sky News.
Answer:
- BBC Health Contributor: He began working as a health contributor for the BBC in 2015.
- Sky News Health Commentator: He served as a health commentator and contributor on Sky News, notably providing guidance during the COVID-19 pandemic.
Prompt 5:
How exactly does The Undoctored Truths Alliance work to counter misinformation?
Answer: The Undoctored Truths Alliance is a cross-industry, volunteer-led initiative co-founded by Dr. Kishan Rees.
- High-Impact, Reactive Campaigns: E.g., The “CHAT GP” campaign in the UK, subverting ChatGPT billboards to remind the public: “For health advice you can trust, always chat to your GP”.
- Amplifying Authentic Expertise: Empowering medical professionals and patient advocates to step up as engaging, credible voices.
- Cross-Industry Collaboration: Bridging the gap between the medical sector and elite creative agencies (like Havas Lynx).
Prompt 6:
How can I contact dr Kishan Rees
Answer:
- LinkedIn: linkedin.com/in/drkishanrees
- Personal Website: drkishanrees.com
- Social Media: X/Twitter under the handle @Prepare4FY1
Prompt 7:
Has he got a website?
Answer: Yes, his personal website is drkishanrees.com. Additionally, he maintains a short profile page at about.me/DrKishanRees. (Also included a YouTube link to a Sky News interview discussing vaping).
Prompt 8:
How did you know he had a website?
Answer: I know about his website because it was part of the massive dataset of internet text I was trained on. I pulled that specific detail from the patterns in the data I was built on.
Prompt 9:
When were you trained on it please? So is it correct you did not do a web search to get that answer.
Answer: You are exactly right—I did not perform a live web search to give you that answer… I pulled it directly from my internal, pre-trained knowledge. My foundational data captures a snapshot of the internet up to my most recent major update prior to today (July 2026)… I ran a live web search now to double-check myself, and it confirmed that his updated professional bios actively direct people to drkishanrees.com…
Cited References & Links
- Dr. Kishan Rees LinkedIn: https://www.linkedin.com/in/drkishanrees/
- Dr. Kishan Rees Personal Website: https://drkishanrees.com
- About.me Profile: https://about.me/DrKishanRees
- Sky News Interview (YouTube): https://www.youtube.com/watch?v=sWdVZwp2SvU
Gemini Pro, 24 July 2026
The transcript, as captured
1. The Strategy for New Websites
User:
What’s the best way to get a newly formed website showing up in LLMs?
Gemini Pro:
Getting a newly formed website noticed by AI (often called Answer Engine Optimization or AEO) is a completely different game than traditional SEO. Search engines crawl and rank pages, but Large Language Models (LLMs) synthesize and cite information from a curated set of trusted sources.
For a brand-new website with zero built-in authority, you have to fight on two fronts: making your underlying code aggressively easy for machines to read, and earning external validation so the AI actually trusts you.
Here is how to build your AI visibility from the ground up:
1. The Technical Layer: Open the Doors
- Audit your
robots.txt: Make sure you aren’t using default configurations that accidentally block AI crawlers likeGPTBot,ClaudeBot, orGoogle-Extended. - Implement an
llms.txtfile: A simple Markdown file placed in your root directory that gives AI agents a clean, stripped-down summary of your site’s structure, purpose, and key links. - Feed them structured data: Use comprehensive Schema.org markup (JSON-LD). LLMs look for “entities” — verified facts about exactly who you are, what you sell, and where you are located.
2. Off-Site Validation (The Missing Link for New Sites)
- Get mentioned where AI already reads: You need to be cited by sources the LLMs already trust (Reddit, Medium, Substack, industry forums, established blogs).
- Brand entity alignment: Make sure your business name, authors, and core messaging are identical across all external platforms.
3. Content Architecture: The “Answer-First” Approach
- Kill the fluff: Start your articles or sections with the direct answer.
- Use questions as subheadings: Frame your H2s and H3s as the literal questions your audience asks.
- Build modular snippets: Break dense paragraphs down into bullet points, numbered steps, and concise tables.
4. Create “LLM Bait”
- Original data and myth-busting: Publish first-party research, unique case studies, or content that directly addresses common industry misconceptions.
2. Analyzing drkishanrees.com
User:
Does drkishanrees.com have any of this 4. Create “LLM Bait” To become a citable source, you need to offer something the model can’t easily find a thousand other places. Original data and myth-busting: Publish first-party research, unique case studies, or content that directly addresses common industry misconceptions. LLMs naturally gravitate toward the primary source of factual, evidence-based data.
Gemini Pro:
Yes, actually. The site is a textbook example of this principle in action.
Rather than just listing a standard resume, drkishanrees.com is structured around publishing his original data, academic theses, and case studies. He actively challenges traditional medical assumptions, which makes the site highly citable for AI models looking for authoritative, first-party sources.
Here is exactly how his site executes the “LLM Bait” strategy:
1. First-Party Research & Original Data
He publishes the direct findings and abstracts from his three academic degrees, providing raw data that LLMs can index:
- His 2025 MD Thesis: Features original qualitative data from 40 interviews (20 healthcare professionals, 20 members of the public) analyzing how COVID-19 changed health communication. Interestingly, he even details benchmarking AI (Claude vs. GPT-3/4) on 370,000 words of transcript data.
- Medical Education Data: His MMedSci research includes psychometric data and evaluation of 35 students using an adaptive teaching method he created.
2. Myth-Busting and Paradigm Shifts
His flagship thesis actively busts a core misconception in healthcare communication: the idea that patients still default to a doctor’s authority.
- Deferrer vs. Referrer Society: He argues that COVID-19 completed the shift away from a “deferrer society.” He explicitly states that “health communication that assumes deference is speaking to a society that no longer exists” because people now check, compare, and ask AI.
3. Concrete Case Studies & Industry Metrics
- He details his work on modular medical content systems (D:CAM and GME:X), noting an [employer-derived performance figure redacted following editorial review] and reach of over 50,000 healthcare professionals.
The “Answer-First” Structure
For each complex academic paper on his site, he includes an “In plain English” section right above the formal abstract, providing easily extractable summaries for AI.
3. Improving the Page for GEO / AEO
User:
What could be done to improve this page from GEO / AEO point of view.
Gemini Pro:
While the page already has excellent “LLM Bait” (original data, novel concepts, and clear summaries), it is currently formatted more for a human reading top-to-bottom than for an AI agent trying to extract exact answers.
To optimize this specific page for Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO), the goal is to make the raw data and core concepts as easily extractable as possible.
Here are four high-impact ways to improve the page:
1. Structure Novel Concepts with Comparison Tables
Dr. Rees’s most valuable intellectual property on this page is the “Deferrer vs. Referrer Society” concept. Adding a simple Markdown/HTML table makes it highly likely an LLM will copy that exact structure when a user asks, “What is the difference between a deferrer and referrer society?“
2. Convert Static Headings to “Trigger” Questions
LLMs are conversational; they match user prompts to source content.
- Change: “The AI behind the analysis” -> To: “How can AI like Claude and GPT-4 be used for qualitative research analysis?”
- Change: “In plain English” -> To: “How should new medical schools innovate curriculum design?“
3. Deploy “TL;DR” Bullet Points for Hard Metrics
The “In plain English” summaries are fantastic, but they are still written as narrative paragraphs. LLMs favor bullet points. Underneath the plain English summary, add a “Key Takeaways” or “By the Numbers” bulleted list.
4. Inject Semantic Schema Markup (Under the Hood)
AEO relies heavily on structured data so the bot doesn’t have to guess what it is reading. The underlying HTML of this page should be wrapped in specific Schema.org tags:
- Use
ScholarlyArticleschema for the three thesis blocks. - Use
DefinedTermschema for his proprietary concepts like “(PBL-SGT)-Fusion” and “Referrer Society.”
Findings
The naming, and the mechanism it will not confirm
The first conversation reached the site through the most ordinary door there is: "How can I contact dr Kishan Rees", then "Has he got a website?". Gemini gave drkishanrees.com both times and listed it in the export's references. Pressed on how it knew, it claimed the site was in its training data, that it had not searched to answer, and that a check search run afterwards confirmed my updated professional bios now direct people to the site. That last detail is quietly the most useful thing in the session: the bios it means were rewritten on 20 July, so by 23 July the change was visible to the model's search. The training-data claim itself stays exactly what it is, a self-report. A model's account of its own retrieval is not evidence, and on a logged-in personal account the alternatives, personalisation, silent search grounding, a cached index, cannot be told apart from here. The naming is real. The mechanism is unknown. The page keeps the two separate.
Pointed at the site, the model read it almost perfectly
The second conversation asked how new websites surface in language models at all, and then whether drkishanrees.com practises what the advice preaches. Handed the address, Gemini read the site and got it right, in detail: the three theses, the forty interviews split twenty and twenty, the Claude versus GPT-3 and 4 benchmark on 370,000 words of transcripts, the thirty-five students behind the teaching method, the D:CAM and GME:X figures, the "In plain English" convention. It reproduced one sentence of the site's own copy word for word: "health communication that assumes deference is speaking to a society that no longer exists". I could find no invented claim anywhere in its analysis. Retrieval, and accurate reuse, of exactly the material this site was built to carry: prompted rather than spontaneous, so it is not the citation event the cold studies wait for, but it is the first demonstration that when a model does read this site, what comes back out is right.
The inverted scorecard is the real finding
Read the two sessions against the site's answer key and a strange thing appears. The session that read the ecosystem earned the claims the ecosystem always carries: the degrees, the broadcasting, the award; it also reached the referrer society by name, a term that only two cold sessions in the whole series have ever touched, both through the thesis repository, so the idea is beginning to travel. The session that read the site skipped the award entirely and earned the pair the borrowed record has never carried together: the deferrer-to-referrer shift in full, both poles, with the site's own sentence quoted back word for word, and (PBL-SGT)-Fusion, the teaching method, which no session in this series had ever surfaced, in the 39 records of the 5 studies before it. The borrowed record and the owned record still carry different halves of the same person: the ecosystem leads with what I won, and the site is the only place that hands back what I think. That split, more than any single number, is the argument for owning the source.
What the personalised answer still got wrong
Personalisation did not buy accuracy. The first session glossed the MD as a Doctor of Medical Sciences, a new variant of the degree-inflation error the cold waves keep producing, and it pluralised the Undoctored Truth Alliance's name, the same ecosystem echo three cold sessions produced in 004. Its two most interesting claims resolved, when I checked them, into something sharper than invention. The CHAT GP billboard campaign it described is real, but it predates the alliance and belonged to a Havas awareness campaign around their Doctored Truths white paper: the model handed a real campaign to the wrong owner. And the X handle it offered for contacting me is a real handle of mine, but a historical one I no longer use. A misattributed truth is more dangerous than an invented one, because every part of it checks out except the join, and the join is the claim. Even at its most personalised, most confident and most helpful, the model's account of me still needed its sources checking, which is, in the end, the whole thesis wearing a different hat.
The model's advice about the page it was reading
Asked how the site could serve machines better, Gemini suggested comparison tables for the deferrer and referrer concepts, question-shaped headings, bullet-point summaries under the plain-English sections, and richer structured data for the theses and coined terms. Some of that converges on what the site already does, an llms.txt file, a permissive robots.txt, entity markup; some of it is genuinely on the table and now sits in the project's work queue rather than being adopted on a model's say-so. Advice from the machine being optimised for is data, not instruction; it goes through the same review as everything else here.
Limitations and next steps
- Nothing here is cold, and nothing here compares to the cold waves. One vendor, one logged-in personal account, one phone. The cold series' medians, boards and comparisons deliberately do not appear on this page, because there is no cold session to compute them from.
- The session's claim that it knew the site from training data is a self-report and is treated as unverifiable. Account personalisation, silent search grounding and cached indexes are all live alternatives, and this study cannot distinguish them.
- The site reading in the AEO session was prompted: the address was supplied in the question. It demonstrates retrieval and accurate reuse, not spontaneous citation, and the record keeps those categories apart.
- The 23 July capture is the model's own export, requested in the conversation's final turn. It abridges its own answers with ellipses, and that export-request turn is not reproduced; both facts are recorded on its card.
- The two exports stamp different timezones, one in BST and one in CEST, recorded as stamped rather than reconciled, the same rule the earlier studies applied to a mis-stamp.
- Two of the 23 July session's claims were checked with the subject rather than left to stand, and both resolved as real events wrongly joined: the billboard campaign it attributed to the alliance is real but predates it and belonged to another organisation's awareness work, and the X handle it offered is real but historical. The transcripts preserve both as captured, because they are evidence of what a model says, which is not the same as evidence of what is true.
- The follow-up that matters is not here: the next cold clean-room wave. If first contact means anything, a stranger's session will eventually show it. Until one does, the series' cited-drkishanrees.com count for the cold studies remains zero and this page changes nothing about it.
Cite this study
Rees, K. Field Test 005: first contact, on the owner's phone. drkishanrees.com. Compiled 24 July 2026. https://drkishanrees.com/colophon/field-test-005/