What Economics Taught Me About AI Search Strategy
| Game Theory, Signaling, and Adverse Selection Applied to LLMO
Hi, I'm Tsubasa.
I watched a six-part economics lecture series on YouTube — a collaboration between journalist Hiroki Takahashi and Professor Fuhito Kojima of the University of Tokyo's Graduate School of Economics. The series covered market mechanisms, game theory, information asymmetry, and more. The content was dense, so I fed each video into Google's NotebookLM to create structured summaries and stocked the insights for later use. Links to all six videos (in Japanese) are at the end of this article.
This article is for EC operators and web managers who wonder: why doesn't AI search cite my site? Why do competitors keep showing up? What should I actually do for LLMO (Large Language Model Optimization)?
While reviewing the summaries, I realized that the mechanism by which AI search (AI Overviews) selects and recommends specific sites can be explained through three economics concepts: "coordination games," "adverse selection," and "signaling."
The strategy follows from the same three concepts. Aggregate information in your area of expertise to become AI's "consensus reference point." Build structured data and first-party data — signals that are expensive to fake. Write with facts instead of unsubstantiated superlatives. Looking back, I'd been doing some of this without realizing why it worked. I also found frameworks I want to apply more deliberately going forward.
Contents
- How Economics YouTube Videos Led to an AI Search Framework
- Why AI Overviews Cited My Article — The Coordination Game Structure
- Unsubstantiated "No. 1" Claims Still Work in AI Search — The Adverse Selection Problem
- Implementing Signaling That AI Can Distinguish from Noise — JSON-LD and First-Party Data
- Three Practical LLMO Guidelines for EC Operators
- FAQ
- Summary
- Videos That Inspired This Article
How Economics YouTube Videos Led to an AI Search Framework
The catalyst was a series called "Takahashi Hiroki vs. Economics" on ReHacQ, a Japanese YouTube channel. Professor Fuhito Kojima of the University of Tokyo covered market mechanisms and the invisible hand, monopolies and externalities ("market failures"), information asymmetry leading to moral hazard and adverse selection, behavioral economics (loss aversion, sunk costs), game theory fundamentals and the prisoner's dilemma, and his specialty — matching theory and coordination games. Six videos, several hours total.
Watching them one by one was engaging, but stepping back to ask "which concepts are actually applicable?" required a different approach. I fed each YouTube URL into Google's NotebookLM, which can ingest YouTube videos as sources. After processing all six, the accumulated summaries made cross-video patterns visible.
Three concepts stood out as directly applicable to AI search: coordination games, adverse selection, and signaling.
Why AI Overviews Cited My Article — The Coordination Game Structure
When you search "ささげ代行 おすすめ" (best product listing outsourcing services) on Google in Japan, AI Overviews names specific companies — Sasageya, fleston, Sasage.tokyo, BAXS — as recommendations. The citation source? My article: "Best Product Listing Outsourcing Services — 20 Companies Compared" (Japanese). That article also ranks No. 1 in organic search for the same query. A personal blog post being used as an AI search source while simultaneously holding the top organic position.
Why was my article chosen? The economics concept of "coordination games" explains it.
A coordination game is a situation where "if everyone else chooses A, it's best for me to choose A too." The Windows/Mac dynamic is an intuitive example. If everyone around you uses Windows, file compatibility and available support make Windows the better choice for you as well. A feedback loop forms: what everyone uses gets reinforced further.
AI search operates on the same structure. When multiple articles across the web name the same company as a recommended provider, AI treats that as "consensus among information sources" and adopts it. Conversely, if only one site recommends a particular company, AI assigns it lower confidence.
My article compared 16 companies side by side. It aggregated company names that multiple independent sources mentioned individually. For AI, this made it an efficient single reference point — "cite this one page and get information on multiple providers at once." Without intending to, I had become a participant in a coordination game, contributing to AI search's consensus formation.
I've written similar articles in other categories too — "Photo Retouch Pricing — 15 Companies Compared" (Japanese) also appears in search results for "retouch company recommendations." Same structure. Articles that line up multiple options on a specific topic become natural "consensus formation sources" for AI search.
Bing Webmaster Tools' AI citation data (as of June 18, 2026) confirms this: the retouch pricing comparison article was cited by AI 170 times. The product listing article had 9 AI citations but was selected as a source by Google AI Overviews and held the No. 1 organic position. Bing's Copilot and Google's AI Overviews don't always cite the same articles — looking at only one platform's data gives an incomplete picture. Monitoring citation status across multiple AI platforms is necessary.
Unsubstantiated "No. 1" Claims Still Work in AI Search — The Adverse Selection Problem
AI search has a weakness it hasn't resolved: it cannot verify whether text is true. If a company writes "industry-leading" or "recommended No. 1" on its own website, AI may incorporate that claim directly into its recommendation.
In economics, this structure is called "adverse selection." Originally proposed in the context of used car markets, it's also known as the "market for lemons." Only the seller knows the car's true quality; the buyer can't tell from the outside. This means sellers of low-quality cars can charge high prices, while sellers of genuinely good cars get undercut. The result: bad drives out good.
The same dynamic can occur in AI search. As long as AI builds recommendations from website text alone, "companies with weak services but strong self-promotion" and "companies with strong services but modest self-promotion" look the same to AI. The former may capture AI's recommendation slots.
Does this mean you should also claim "No. 1" to gain an advantage? Two concepts from the lecture series argue against it.
First, the logic of "tit-for-tat." In game theory simulations, the most successful long-term strategy was "cooperate by default, retaliate when the other side defects." Business is not a one-shot game — it's a repeated game played over years. An unsubstantiated No. 1 claim is equivalent to "defecting." Short-term gains are possible, but platforms (Google) and regulators exist as rule-enforcers. When the defection is detected, retaliation follows. In Japan, this means cease-and-desist orders, surcharges (3% of revenue), and public naming under advertising law. Other jurisdictions have equivalent enforcement mechanisms — the FTC in the U.S., the ASA in the UK, the EU's Unfair Commercial Practices Directive.
Second, the adverse selection spiral. If one company benefits from claiming No. 1, others follow. Once everyone claims No. 1, the informational value of the claim drops to zero. Procurement managers start thinking "no company's website can be trusted," and the entire market's credibility collapses. Even companies providing genuinely excellent services get caught in the damage. Participating in adverse selection ultimately destroys the market you operate in.
Adverse selection isn't limited to No. 1 claims. Pricing information has the same structure. When you search for product listing service pricing, AI Overviews generates a price range — but the source data is limited. If one company floods the web with pricing information favorable to itself, AI's response may be skewed accordingly.
I covered the legal risks of No. 1 claims and AI search dynamics in detail in "Is Google Cracking Down on 'Best Of' Comparison Articles?" — including the Lily Ray study showing that self-promotional brands get excluded from AI Overviews' recommendations 69% of the time.
What I focus on is not joining the adverse selection race. Writing "industry's lowest price" or "No. 1" might give a short-term AI search advantage. But it accelerates adverse selection and degrades the entire market's information quality. Write facts as facts. Compare pricing using actual data collected from multiple official sources. That accumulation is the counter-strategy against adverse selection.
Implementing Signaling That AI Can Distinguish from Noise — JSON-LD and First-Party Data
"Recommended No. 1" is something anyone can write in plain text. Because the cost is zero, its informational value is also low.
In economics, "signaling theory" holds that information which is expensive to produce functions as a credible trust indicator. The classic example is education in the job market. Completing a university degree requires four years and significant tuition — the cost itself signals that the person has the capability and persistence to bear it. The harder it is to fake a signal, the more reliably it functions as a trust marker.
Looking back, my product listing comparison article contained several elements that qualify as signals.
Structured data (JSON-LD) was implemented. I had embedded FAQPage, Article, and BreadcrumbList JSON-LD in the article. Writing schema-compliant code by editing HTML requires effort far beyond typing "recommended No. 1" in plain text. For AI, the presence or absence of structured data can serve as a proxy for "how carefully was this information source built?" I documented the implementation process in "How to Write Structured Data (JSON-LD) — FAQPage and Article Implementation" (Japanese).
Pricing data from 16 companies was researched and published as first-party data. Checking each company's official site one by one and compiling pricing structures into a comparison table takes considerable time. That research cost is itself a signal. "Information that required effort to produce" differs in quality from "information that was just written."
FAQ structured data covered anticipated questions. The article includes five or more FAQs matched in both the body text and JSON-LD. Pre-answering questions that readers are likely to ask AI makes it easier for AI to reference the article when generating responses to those questions.
These weren't implemented as a deliberate signaling strategy — they accumulated as basic SEO and LLMO practices. But applying signaling theory retroactively explains why they worked.
Bing Webmaster Tools' AI citation data (as of June 18, 2026) provides supporting evidence. Across my site, 52 pages have been cited by AI a total of 3,342 times. The top-cited pages — a glossary entry on image resizing (666 citations), an Instagram image dimensions guide (477), and an Exif metadata article (415) — share a common trait: they comprehensively compile specific numerical data and specifications on a focused topic. Pages with "high-cost-to-fake information" get cited more by AI — consistent with signaling theory.
Three Practical LLMO Guidelines for EC Operators
From the three economics concepts, actionable LLMO (AI search optimization) guidelines emerge. I'll separate what I was already doing unconsciously from what I plan to do deliberately going forward.
Apply the coordination game: aggregate information in your area of expertise. My product listing article became an AI Overviews source by compiling 16 companies on a single page. EC operators can do the same by writing comprehensive "best X" or "X compared" articles in their specific industry niche, lining up multiple options side by side. I was already doing this. The next step is writing multiple related articles and clustering them, reinforcing AI's reference point through internal linking.
Practice signaling: maintain structured data and first-party data. JSON-LD implementation and first-party data compilation are already in place. The next focus is periodic pricing data updates — re-surveying each company's pricing page at least every six months and noting the update date in the article. Stale data degrades signal freshness. Another priority is earning backlinks through article quality. Backlinks are difficult to manufacture. The fact that external sites reference your content is one of the hardest signals for AI to dismiss.
Counter adverse selection: refuse to join the superlative race. Not using unsubstantiated "industry No. 1" or "lowest price" is already practiced. Additionally, consistently publishing fact-based articles at a regular cadence is necessary. AI's accuracy improves continuously. When AI becomes capable of distinguishing unsubstantiated claims, fact-based articles will be the ones that survive.
For an introduction to LLMO fundamentals, see "What Is LLMO? Strategic Differences from SEO and How to Start" (Japanese).
FAQ
Q. What determines whether AI Overviews cites my site?
A. No official criteria have been published, but AI tends to adopt content where multiple information sources agree. If your site comprehensively covers a specific topic and has structured data (JSON-LD) implemented, it's more likely to be selected as a reference source.
Q. Can a small personal site get cited by AI Overviews?
A. Yes. My site is a personal blog on GitHub Pages, and it was selected as an AI Overviews source for "best product listing services" while also holding the No. 1 organic position. Site scale matters less than whether you provide comprehensive information on a specific topic.
Q. I wrote a comparison article on my company blog, but AI Overviews doesn't cite it. What can I improve structurally?
A. First, check whether structured data (JSON-LD) is present. FAQPage and Article structured data help AI understand the article's content. Next, review the number of comparison targets and depth of information. My article compared 16 companies with pricing and characteristics side by side — that comprehensiveness contributed to being selected. An article covering only three or four companies may be judged by AI as having insufficient information density.
Q. A competitor claims "industry No. 1" and seems to be winning in AI search. We write factually but seem to be losing. What should we do?
A. In the short term, No. 1 claims may get picked up by AI. But unsubstantiated No. 1 claims carry legal risk under advertising regulations in many jurisdictions. AI accuracy is continuously improving, so writing factually pays off long-term. Focus on enriching information in the areas where your company is strong, and outcompete on information volume as a coordination game participant. Also consider this: while AI may not catch these violations yet, human decision-makers evaluating service providers are not easily fooled by companies with poor compliance awareness. For humans, unsubstantiated claims erode trust rather than build it.
Q. What is the difference between LLMO and SEO?
A. SEO aims for high rankings on Google's search results pages. LLMO aims for your site's information to be cited when AI generates direct responses — in ChatGPT, AI Overviews, Copilot, and similar services. The two are not opposing strategies. Structured data and first-party data improvements benefit both.
Q. What are coordination games, adverse selection, and signaling in simple terms?
A. A coordination game is when "everyone benefits from choosing the same thing" — like how everyone using Windows makes Windows more useful. In AI search, when multiple sources agree on a recommendation, AI adopts it as consensus. Adverse selection is when bad products drive out good ones because buyers can't tell the difference — like a used car market where sellers know quality but buyers don't. In AI search, companies making unsubstantiated claims can capture recommendations while honest companies get overlooked. Signaling is providing information that's expensive to fake, proving your quality — like a university degree in the job market. In AI search, structured data and researched first-party data serve as signals that distinguish serious content from surface-level claims.
Summary
Watching six economics lecture videos and summarizing them with NotebookLM revealed that AI search can be explained through three concepts: coordination games, adverse selection, and signaling.
From the coordination game perspective, aggregate information in your area of expertise and become AI's "consensus reference point." My 16-company comparison article was cited by AI Overviews and ranked No. 1 in organic search. From the adverse selection perspective, refuse unsubstantiated superlatives and maintain market information quality through fact-based writing. From the signaling perspective, accumulate "high-cost-to-fake information" — structured data, first-party data, and comprehensive FAQ coverage.
Going forward, I'll incorporate biannual pricing data re-surveys, article cluster reinforcement through internal linking, and quality-driven backlink acquisition as deliberate LLMO strategy components.
Videos That Inspired This Article
The "Takahashi Hiroki vs. Economics" series on ReHacQ (Japanese YouTube channel). Professor Fuhito Kojima of the University of Tokyo's Graduate School of Economics covers economics fundamentals through game theory and matching theory. Six videos total. All videos are in Japanese.
- Episode 1 — Purpose of economics, the invisible hand, market failure, introduction to matching theory
- Episode 2 — Monopoly, externalities, public goods, auction theory
- Episode 3 — Information asymmetry, moral hazard, adverse selection (used in this article)
- Episode 4 — Rationality, procrastination bias, nudge theory, loss aversion, sunk costs
- Episode 5 — Game theory fundamentals, Nash equilibrium, prisoner's dilemma, tit-for-tat strategy (used in this article)
- Episode 6 — Coordination games (used in this article), matching theory, designing serendipity
🇯🇵 この記事の日本語版を読む
よく読まれている記事
