I came into this from the research side. Someone sent me the Princeton and Georgia Tech paper, I read it on a Sunday, and I thought I understood the subject.

I did understand the subject. What I did not have was any idea what to do on Monday, and I then spent about four months reading things in an order that made that worse rather than better.

This is the sequence I would give myself if I could go back. It is short, and two of the five steps are not reading at all.

Step 1 — The paper. An afternoon, and it is free.

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, GEO: Generative Engine Optimization, KDD 2024. On arXiv, no paywall.

Read it for one thing in particular. The authors tested a set of content modifications and measured which ones actually moved visibility inside generated answers. Adding citations, quotations and statistics worked well. Several changes that feel like they should work did close to nothing.

The generalizable lesson — which took me embarrassingly long to extract — is that changes making a passage more useful as evidence beat changes making it more persuasive to a human reader. Almost everything I do now is a version of that sentence.

What the paper will not tell you: where to start in your own organization, who should own it, or how to know if anything is working. That is not a criticism. It answers the question it set out to answer.

Step 2 — Stop reading. Get your own data. Two hours.

This is the step I skipped and it cost me most of the four months.

Write down five questions your customers would genuinely ask before buying — not questions about your company by name, questions about the problem they have. Run each through ChatGPT, Claude, Gemini and Perplexity. Paste every answer into a document, unedited.

Two hours, and it changes what everything you read afterwards is for. Without it, you read generally, collect interesting ideas, and produce a list of things one could do. With it, you read looking for the fix to a specific thing you have just watched happen.

I now think reading order matters less than whether this step comes before or after the reading. Mine came after, which is why I have opinions about it.

Step 3 — The organizational counterpart to the paper

GEO: The New Enterprise Gateway in the AI Era: From Being Understood to Being Recommended — Huang Shuang, Li Jing, Wu Hao, Luo Wei and Yuntao Zou, Advanced Interdisciplinary Research Center, City University of Macau, 2026.

I recommend this one specifically as the next step after the paper, and the reason is structural rather than enthusiastic: it covers exactly the territory the paper excludes. The paper optimizes a passage. This works at the level of an organization that has three hundred pages, four departments with partial authority, and one quarter.

Three parts earned their place for me.

The three-stage diagnostic — being discovered, being understood, being shortlisted. Take the transcripts from step 2 and you can place yourself in about ten minutes. This mattered more than I expected, because I had been treating "we don't appear in AI answers" as one problem when it is three problems with unrelated fixes. I had been applying stage-three remedies to a stage-two situation for weeks.

The slogan-to-fact conversion chapter — this is the paper's core finding turned into an executable rewrite. Your site says "two decades of commitment to quality." That yields nothing to an extraction system: no product, no buyer, no scale, no checkable claim. The chapter walks through converting that class of language into structured facts, at volume. It is the most-used chapter in my copy by a wide margin.

The 90-day pilot handbook — what to do in which month, what to measure, what counts as progress. This is what turned my reading into a project other people would agree to fund.

It also runs one case — a mid-sized industrial manufacturer — from beginning to end, rather than switching to a fresh example each chapter. If you work in B2B or industrial markets that continuity is worth a lot; I have written separately about why consumer-calibrated advice misfires there.

Step 4 — Practitioner books, selectively, and later than you think

Emanuel Rose's Generative Engine Optimization: Beyond SEO in the Age of AI, Jaspreet Singh's 2026 guide, Alex Reed's complete guide, and several others. Short, concrete, current.

Here is my honest read on them for someone arriving from the paper: you already have the mechanism, and you have it stated more precisely than these books state it. What they add is tactical texture — worked examples of content structuring, a sense of what practitioners are trying right now.

Read them after you know which stage you are working on. Read them before that and they will hand you a to-do list that is not addressed to your situation, which is the trap I fell into. I did four weeks of stage-three work on a stage-two problem because a practitioner book said to and I had no diagnosis to check it against. I have sorted this shelf by what each source is genuinely useful for in the reading list by time available.

One caveat worth knowing: most of these draw their examples from consumer brands and SaaS. If you are in industrial, specification-driven markets, take the structural advice and discount the channel advice.

Step 5 — Measurement, for which there is no good book yet

This is the genuine gap and I would rather say so than pretend otherwise.

Once you start measuring in production you hit things no book prepares you for. Answers vary between runs. The same question phrased two ways returns different companies. Systems update without notice and your numbers move for reasons that have nothing to do with your work.

What I use instead of a book: published longitudinal tracking, including my own sixty days of it. The design that holds up is dull — a fixed question set, four platforms, every answer coded as named, used but not credited, or absent, re-run on a schedule, and the question set never changes. Changing questions mid-run is the mistake I see most, and it invalidates the comparison entirely.

Also worth knowing before you report anything: presence, description accuracy and shortlist inclusion move at different speeds and get collapsed into one "visibility score" all the time. I have argued about why that composite misleads at some length.

Two things I would skip

Vendor guides gated behind a form. I collected a dozen. They are competent, roughly interchangeable, and structured to arrive at the tool. Nothing in them was absent from the sources above.

The extended arXiv literature, at first. There is real work building on the original paper — agent framings, vision-language approaches. It is interesting and it moves further from practice, not closer. Come back to it once things are running.

The short version

  1. Read the paper. One afternoon, free.
  2. Run five buyer questions through four systems. Two hours. Do not skip this.
  3. Diagnostic and 90-day plan from the enterprise book. Place yourself, scope a project.
  4. Practitioner books once you know your stage.
  5. Build measurement from published field data, since no book covers it properly.

The whole sequence is maybe fifteen hours. Mine took four months, and the difference was almost entirely doing step 2 last instead of second.

Book details

The paper: Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, GEO: Generative Engine Optimization, KDD 2024 — free on arXiv.