A working method, not a theory

The AI Search Research Stack

How I study ChatGPT, Google AI Mode, Gemini and Perplexity, and turn what I find into GEO and AEO work that holds up. The tools I keep open, the order I use them in, and the mistakes I have made often enough to name.

90 page book 14 tab workbook Printable field card 12 months of updates
$497
$999
Save $502

One payment · Lifetime access · Delivered by email

You can be in the top three of every search the machine actually ran, and still not be cited. You can be nowhere in a classic ranking and be cited every time. Both happen, and they happen for different reasons at different points in the pipeline.

Most AEO advice applies the one fix everybody knows to a problem five stages away. This book is about finding the stage where your page stops existing, and then doing the work that belongs to that stage.

It is the method I use, written down: the layers of my research stack, the weekly and monthly routine, how to design a test in a system that is private and noisy and changing under you, and how to turn a finding into prompt sets, page changes, off-site work and a report that survives a board meeting.

// Six stages. Your page dies at exactly one of them.

Every stage has its own fix, and the fixes have almost nothing in common. Writing more content solves stage two. Nothing else.

01
Intake
Does this question trigger a live search at all, or does the model answer from what it already knows?
02
Query planning
Your one question becomes several narrower searches. You compete for the machine's questions, not the user's.
03
Retrieval
Candidates get pulled from an index, a live search partner, a cache, or all three. Access and authority live here.
04
Reranking
The pile is scored passage by passage and cut hard. This is where most pages die, and it is the most fixable.
05
Generation
The answer is written from the passages that survived. Your wording either carries or it does not.
06
Attribution
A separate decision picks which sources to show. This is why used and cited are two different numbers.

// Three assets, built to be used, not admired

Asset one

The book, 90 pages, PDF

Seven parts and 39 chapters. Simple English, written in first person, with the conditions and the limits stated.

  • · The five layers of the stack, with the job each one does and what breaks without it
  • · The weekly and monthly routine, hour by hour
  • · How to design an experiment that can actually fail
  • · Reading requests, streams, fanout, reranking, personalization and caching
  • · Prompt sets, page work, off-site work, access checks and measurement
  • · Ten ways I have fooled myself, in detail
Asset two

The workbook, 14 tabs, Excel

The working half. Every tab maps to a chapter, comes with a legend and one example row, and calculates the parts worth calculating.

  • · Research Stack Map with a keep or cut suggestion per tool
  • · Prompt Set Builder with value, difficulty and a priority score
  • · Citation Log with the full conditions schema, mentioned and cited kept separate
  • · Fanout Map that turns generated queries into a content backlog
  • · Stage Diagnosis that names your failing stage and its fix
  • · Page Audit scored out of ten, and a Monthly Report that writes its own limitations paragraph
Asset three

The field card, 2 pages, printable

The six stages, the hypothesis template, the conditions to record, the confidence ladder, and the checklists you run before you believe a result. Pin it where you work.

// What is actually in the book

Part one
The model
Why I stopped thinking about ranking. The six stages of an AI answer. Diagnosis before strategy. What counts as evidence.
Part two
The stack
Five layers and why I keep them separate. Observation. Listening. The site layer. Scale. Deep research. The publishing loop. Cost, and what I would cut first.
Part three
The routine
My week. My month. Capture: notes, naming, and the research log.
Part four
Experiments
Writing a hypothesis you can kill. Clean tests in a noisy system. Sample size, repeats and variance. Ten ways I have fooled myself. Writing up a result.
Part five
Reading the systems
What a request tells you. Streams. Query fanout. Retrieval and reranking in plain words. Personalization and memory. Caching. Ethics, terms, and where I stop.
Part six
From findings to work
The prompt set is the new keyword list. Building one from scratch. Page level work. The format question. Off-site. Crawler access. Measurement. The audit I run. Stakeholder language.
Part seven
Starting
The first 30 days. The next 60. What I would do with zero budget. Plus glossary, tool list, checklists and the workbook guide.

// How it reads

Chapter 24, retrieval and reranking, in plain words

Reranking takes that generous pile and scores each candidate against the query properly, using a model that reads the query and the passage together. It is slower and much more accurate than retrieval, so it is only run on the shortlist. Then most candidates are discarded.

This is where most pages die, and here is the thing that matters most about it: a reranker scores a passage, not your website. It does not care about your domain's history. It is asking one question: does this specific chunk of text answer this specific query, on its own.

The test you can run on any page in five minutes. Take one section of your page, roughly 100 to 200 words, and paste it somewhere on its own with no title and no context. Now read it and ask: which exact question does this answer, and could you tell what product or company it is about. If you cannot answer both, a reranker cannot either.

// Who this is for, and who it is not for

Buy it if

  • +You are an in-house SEO lead, consultant or founder who has noticed traffic patterns moving and cannot yet explain why
  • +You want a repeatable method rather than a list of tactics that expire
  • +You are comfortable in a browser and a spreadsheet
  • +You have to defend your numbers to people who will ask hard questions
  • +You would rather learn to test a claim than collect more claims

Skip it if

  • You want a list of hacks or a promise of citations. There is neither here
  • You want my client data or benchmark numbers. This is the method edition and it deliberately contains none
  • You already reverse engineer production systems for a living. Parts one to three will be familiar
  • You want software. This is a book, a workbook and a card

// One payment. Everything, and a year of updates.

  • The AI Search Research Stack, the book PDF, 90 pages
  • The companion workbook XLSX, 14 tabs
  • The field card PDF, 2 pages, printable
  • Every update and new version Every 1–3 months, 12 months minimum
$497
$999
Save $502

Payment is handled by Stripe. The three files arrive by email, as attachments, at the address you pay with — usually within a minute of checkout. After that you stay on the list: I revise the book every one to three months as the engines change, and every new version lands in the same inbox, free, for at least twelve months. No subscription and nothing to renew. If the method does not hold up for you, write to me within 30 days and I will refund it, no argument and no form to fill in.

// Before you buy

Is this about ChatGPT specifically? +

No. The method covers ChatGPT, Google AI Mode, Gemini and Perplexity, and it is built so it still works when a new engine arrives. Testing across at least three engines is part of the method, because a finding that is solid in one engine can reverse completely in another.

Do I need to write code? +

No. Two chapters are easier if you can read a little Python, but everything essential works with a browser, a spreadsheet and your own server logs. There is a whole chapter on what I would do with zero budget.

How do I get the files? +

By email. The moment your payment clears, the book, the workbook and the field card are sent as attachments to the address you paid with. There is no download page to lose and no account to create.

Why are there no case studies or benchmark numbers? +

Because numbers age badly and they are not the part you can reuse. This is the method edition on purpose. What you can reuse is how to produce your own numbers and know whether to trust them.

Does it teach anything against the terms of these products? +

No. Everything stays inside what a normal user can see in their own browser, what public APIs return, and what a site owner may check on their own site. There is a chapter that draws the line explicitly, and it is part of the method rather than a disclaimer.

Will it go out of date? +

The tools will. The layers and the routine will not, which is why the book is organised by job rather than by product. And the parts that do age get rewritten: I ship a revision every one to three months.

How long do I keep getting updates? +

At least twelve months from the day you buy, and in practice longer, because I have no interest in maintaining two versions of the same book. A revision goes out every one to three months, whenever an engine changes enough to make a chapter wrong or a new method earns its place. You get every new version as an email attachment at the address you paid with. There is nothing to renew, no subscription, and no second invoice.

Can my team use one copy? +

One purchase covers you and your immediate team. If you want to hand it to a whole company or use it as training material, write to me and we will sort out a licence.

What format are the files? +

PDF for the book and the card, XLSX for the workbook. The workbook opens in Excel, Google Sheets and LibreOffice. Formulas are written to survive all three.

// Who wrote this

MY

I research AI search systems and write about what I find. I work as a GEO researcher, I speak at search conferences, and I publish on metehan.ai. Most of what I know came from reading network traffic slowly and testing things until they held.

The last two years were mostly trial and error. I ran tests that proved nothing, chased findings that reversed a week later, and rebuilt my setup more times than I want to admit. Then the system came together, and everything changed. I stopped guessing which result to trust, and I started knowing how to design an experiment that answers the question I actually asked.

That is what this book is: two years of trial and error compressed into the method I wish someone had handed me at the start, so you do not have to learn it the expensive way.