A month ago, closing the manuscript for Augmented Expert in the Agentic AI Era, I wrote this sentence.

If your brand is not mentioned when AI generates an answer, you do not exist to the customer.

I still think the direction is right.

But this month, while turning AEO and GEO into an actual business, I reached a different conclusion about the execution. Much of it has to be rewritten.

This project update started with a video. Kim Hyungtae, CEO of Daily Growth, appeared on Builder Josh’s podcast and opened up an agentic marketing workflow: twelve numbered folders for running marketing with Claude Code. Analyze Company, Competitor, and Customer first, produce creative, read the ad data, optimize.

What stayed with me was not the tooling.

It was the way brand context accumulates physically in folders, and the two places where human judgment was deliberately left in. At step 04 a person decides which message goes out. At step 12 a person decides what to keep and what to kill.

I decided to apply that structure to an AI Discovery (AEO and GEO) tool inside a strategic marketing product.

The plan I first designed was a monthly subscription AEO and GEO service for Korean local businesses.

Generate the core questions for each client, ask them repeatedly across several AI engines, calculate brand mention and citation rates in the answers, and deliver monthly site improvements. To make that repeatable, I built a skill and five scripts modeled on the twelve-step structure from the video.

Judged on the prototype alone, it looked viable. Site diagnostics ran. Customer questions were generated. Brand mentions and sources were collected from AI answers, scored, and turned into a report.

Then I killed the plan at the business audit stage.

The technology did not fail. It was more dangerous precisely because the features worked and produced plausible numbers. Behind a working prototype sat four unanswered problems: real customer demand, terms of service, measurement credibility, and the unit cost of running it repeatedly.

A product you can build and a business that sustains itself are not the same thing.

The AEO I wrote about and the AEO I met in the field were different

AEO is commonly framed as producing answers an answer engine can lift, and GEO as raising the odds of being selected or cited as a source in a generative engine’s response.

In the book I explained this along three axes.

  • Trust Signals: build trust through consistent information across multiple sources
  • Information Advantage: build an edge with proprietary data and expert insight
  • Semantic Structure: make it machine legible with clear headings, FAQs, and structured data

Those principles still hold.

The problem is that they do not automatically add up to a new ranking industry.

Google’s 2026 official guide does not treat AEO and GEO as separate magic. Generative search runs on the existing search index and quality systems, so the SEO fundamentals of unique, useful content and clear technical structure remain the core.

It also states plainly that hacks like llms.txt, AI-specific markup, and finely chopped content do not improve Google Search visibility.

Translated into working language:

Owning an original worth finding and verifying comes before writing sentences an AI will like.

This connects to the CPR system I described in an earlier post.

Starting from a company name that returned two or three search results, over eighteen months we produced 51 press articles, 76 major media exposures, and 20 to 100 pickups per article, more than 1,000 pickups in a year. At the time I thought of it as building search visibility. Looking again now, it was also accumulating the evidence assets about the company that an AI can retrieve and cross-check.

AI will not invent facts about a company that do not exist in the world.

The first thing to watch was not technology. It was demand.

The first market I examined was Korean local businesses.

OpenSurvey’s second-half 2026 study shows generative AI growing fast for work and knowledge discovery, while Naver still holds roughly 60 percent primary usage for news and everyday information.

The fact that people use ChatGPT heavily and the fact that they look for restaurants, lodging, and clinics on ChatGPT are not the same fact.

App user counts should not be read as market demand.

A search grounding experiment I ran on 15 Korean local queries surfaced an interesting gap. Naver Blog and Naver Place never appeared as a cited source. Yeogi appeared near the top in all three lodging queries, and Daangn’s local profile ranked high for hair salon queries.

Part of the reason was visible in published crawling policy.

The sample is small. It was an approximation run against a US locale search index, so it needs to be measured again in the Korean locale. And robots.txt does not directly produce search rankings.

Even so, one thing was clear.

Crawler policy is not a technical setting. It is a distribution strategy. Surfaces that decide to be visible to AI can take the open space.

But an open space in supply and a customer willing to pay for that space are entirely different problems.

The second gap was measurement

The initial product tracked mention and citation rates across engines, running 40 core questions per client five to seven times each.

Doing the arithmetic, a single measurement for one client required 600 to 1,120 queries.

Automating queries against consumer AI interfaces creates a terms problem. OpenAI’s Terms of Use prohibit automated or programmatic extraction of data or output from the service.

Using the API is legitimate, but you cannot claim it reproduces the same results a consumer sees in the search product. Doing it by hand costs 10 to 19 hours per client.

The early unit economics, which assumed 70 percent automation, died at this point.

A margin that exists only when the automation is not legitimate is not a business margin.

There was a more uncomfortable finding.

One script quietly summarized a failed query as “0 blocked bots”. Another calculator reported a 63 percent profit margin on a 50,000 won monthly product when the real number was a 12 percent loss.

The code ran. The numbers appeared.

But the product had already started lying.

It confirmed again that expertise in the AI era is not only the ability to build automation. It is the ability to notice when the automation is wrong and stop it.

The academic evidence did not point in one direction either

The early GEO study most often cited in the AEO and GEO industry reported that content transformation techniques could raise visibility in generative engines by up to about 40 percent in a controlled setting.

But C-SEO Bench, published at NeurIPS 2025, points out that under realistic conditions where multiple competitors apply the same methods, most content transformation techniques have little effect and can even lower rankings.

Which documents were retrieved at the search stage showed a far stronger influence than the sentence style of the final answer.

The two studies use different settings and metrics, so it would be wrong to write that one simply refutes the other.

Read together, though, they rule out any promise that applying a technique lifts exposure by some percentage.

What you can say instead is this:

Good GEO is not a technique for fooling a generative model. It is the practice of building evidence that deserves to be retrieved, and verifying whether that evidence turns into business results.

So I changed the definition of the product

At first I wanted the selling point to be how many times a brand was mentioned in AI answers.

Now I define it as building AI Discovery infrastructure the client owns and can verify.

1. Discovery Audit

Diagnose the site’s crawlability, indexation, entities, existing content, and measurement state.

2. Evidence and Claim Registry

Structure the facts, figures, sources, and expert review history first. Decide which sentences may be claimed before deciding which sentences to generate.

3. AI Discovery Optimized Owned Media

Build unique pages for practitioners, services, treatments, FAQs, and cases on the client’s own domain, not on a duplicate site.

4. Human Approval Gate

In industries where expertise and regulation matter, nothing publishes automatically without expert review and sign off.

5. Owned-Signal Measurement

Connect AI crawler logs, Search Console, GA4, and consultation or CRM data.

6. Optimization Gate

A person decides not whether to make more, but whether to keep, revise, or stop.

The two human gates from the video survived in this structure too.

Step 04’s message sign off became Evidence and Claim approval. Step 12’s optimization call became the keep, revise, or stop decision.

This structure made more sense to validate first in verticals with high expertise, high customer value, and a real need for review, rather than in general local business.

So I narrowed the first market to premium medical, and decided to expand into specialized legal only after the product and its costs are proven in healthcare.

The measurement target moved from scraped answers to owned signals

The measurement environment changed this year as well.

Google Analytics added an AI Assistant channel that classifies visits arriving from ChatGPT, Gemini, Claude, and others. Google Search Console began offering generative AI performance reports showing AI Overviews and AI Mode impressions for some sites.

This data is not complete either. Depending on the app or browser, AI referrals can land as Direct, and you cannot see cases where you were cited but not clicked.

So the numbers have to be reported as a floor, not as a complete share.

  • Did AI bots read our pages
  • Did those pages enter the search index
  • Did they appear on generative search surfaces
  • Did actual visits arrive from AI services
  • Did those visits turn into consultations and revenue

Instead of dressing up what cannot be measured as a precise score, connect the signals a client can verify in their own accounts, step by step.

If I rewrote the book’s AEO chapter today

I would not discard what I wrote.

I would add four sentences.

First, AEO and GEO do not replace SEO.
Content that was never found or retrieved is unlikely to be selected in a generative answer either.

Second, originality comes before structure.
FAQs and schema are tools for delivering good material clearly, not devices for making ordinary material special.

Third, AI Share of Voice is not a business result.
Mentions and citations are intermediate signals. Without a path to visits, consultations, and revenue, they are not customer value.

Fourth, define the stop criteria before the build.
If no meaningful AI referral traffic shows up in 30 days, the right move is to stop making more pages.

The biggest update from this round of fieldwork was not an AEO or GEO technique.

It was the standard for separating “can this be done” from “is this a business”.

AI search is clearly growing. But a growing market is not the same statement as every industry needing to spend money on it now.

Make the website readable by AI, accumulate unique evidence, keep a person in the approval loop, and confirm results with data the client owns.

I am choosing to call this AI Discovery, a wider frame than AEO or GEO.

It is not a service that guarantees rankings. It is the work of earning the right to be found and building plumbing you can verify.

AI Discovery is not a ranking hack. It is owned media infrastructure for the answer economy.

Sources

PROJECT INQUIRY

Ready to design the next project?

Tell me where the connection between technology and market is breaking down. I connect product strategy, brand narrative, IR, and global GTM into one execution system.

Start a project

NEWSLETTER

One field note a week, in your inbox

What I learn turning technology into market results, once a week. Product strategy, brand narrative, IR, and global GTM.

Subscribe