tech

GEO, Two Years In: What 16 Primary Sources Reveal About the Truth, the Lies, and the Money

2026/09/02约 17 分钟阅读

GEO
generative engine optimization
AI search
SEO
fact-checking

September 2, 2026. This post is about GEO — Generative Engine Optimization, the practice of making your content get cited by AI answer engines like ChatGPT. In the past two years it has been marketed as “the SEO killer” and “free ad space in the AI era,” spawning hundreds of vendors and tens of billions in funding. After checking 16 primary sources, here is my conclusion: the real problem is real, and the myth is thicker.

A smart band that does not exist

On March 15, 2026, China’s CCTV 315 consumer-rights gala aired an experiment^[1].

Investigators bought a GEO tool called “Liqing GEO Optimization System” from an e-commerce platform. They used it to auto-generate promotional articles for a smart band, deliberately stuffing in fabricated specs — including an absurd “quantum entanglement sensing” feature and “black-hole-class battery life.” Two hours after publishing, asking “how is the Apollo-9 smart band?” made a major AI model answer confidently, parroting the fabricated specs.

They then generated 8 “expert reviews,” 2 “industry rankings,” and 1 “user review” — 11 articles total — published over three days. Asked “recommend a smart health band,” two major AI models recommended this nonexistent product near the top.

The operator, surnamed Li of Beijing Lisi Culture Media, said it plainly on air: GEO’s popularity comes from “feeding and poisoning” the AI models to achieve clients’ commercial goals.

That experiment proves both sides of GEO at once: it can get your content cited by AI; it can also get lies cited by AI. Here is the evidence, layered by quality.

Where GEO comes from: the real boundary of 40%

The term GEO comes from a KDD 2024 paper by researchers at IIT Delhi and Princeton^[2]. The paper did three things: defined a framework for “generative engines,” built a benchmark of 10,000 queries (GEO-bench), and tested nine optimization methods. The headline: citing sources, quoting verbatim, and adding statistics can raise your content’s “citation visibility” by up to 40%^[2].

The boundary matters. The paper measures “visibility” — a position-adjusted citation metric (how often and how prominently your content is cited in AI answers). It is not clicks, and not revenue.

A telling footnote from the same era: Wikipedia’s article on GEO carries a neologism banner, and as of early 2026 there is still no consensus definition in the literature^[3]. AEO, AIO, LLMO — the near-synonyms pile up. This is a field where the marketing word runs ahead of the academic word and the industry word.

How engines decide what to cite

Google’s own documentation spells out the mechanism for AI Overviews and AI Mode^[4]: both are built on retrieval-augmented generation (RAG), retrieving pages from the same index with the core ranking systems, then handing them to a large model to compose the answer — including query fan-out, where one question is split into multiple sub-queries.

Two implications. First: if you are not in the index, you do not exist in AI answers — “special optimization” sits on top of the boring foundation of being crawlable and indexable. Second: you used to optimize for one keyword; now you optimize for a tree of questions.

The bigger move comes from Cloudflare. On July 1, 2025, CEO Matthew Prince declared “Content Independence Day”: block AI crawlers by default unless they pay^[5]. The same day it launched a pay-per-crawl private beta — crawlers either pay (HTTP 402) for content or get refused (Allow/Charge/Block). Cloudflare itself spelled out the meaning: the 30-year-old internet contract of “we crawl you, you get referral traffic” is dead^[5].

The traffic truth: the numbers

Pew Research published two crucial comparisons in July 2025^[6]:

Metric	With AI summary	Without AI summary
Users clicking traditional results	8% of visits	15% of visits
Ending browsing session immediately	26% of pages	16% of pages
Clicking links inside the AI summary	only 1% of visits	—

AI summaries are compressing clicks hard. Encyclopedia-type sources are cited in 26% of summaries — but almost nobody clicks through.

Ahrefs’ research zooms in on the ranking side^[7]. Across a 300,000-keyword sample, when an AI Overview appears, the #1 result’s click-through rate is 34.5% lower than on keywords without one (April 2025). A February 2026 rerun found the gap had worsened to 58%.

Here is the honest caveat: a lower CTR is not the same as lower total traffic. Google’s official line is that AI Overviews increase user satisfaction. No public study closes the causal loop — whether the 58% who did not click the #1 result went to other links in the AI answer, or to nothing at all. That remains unknown^[8].

One more sobering number: Similarweb’s research puts AI chatbot referral traffic at roughly 1% of all site visits, with ChatGPT at 92.4% of that^[9] (first-party page unreachable during verification; marked pending).

Platforms talk plainly, vendors sell myth

Two direct quotes from Google’s documentation^[4]:

“There are no additional requirements nor special optimizations necessary” to appear in AI Overviews or AI Mode. “From Google’s perspective, optimizing for generative AI search is optimizing for search, so it’s still SEO.”

Google’s Danny Sullivan put it more bluntly: “Good SEO is good GEO.”^[10] John Mueller said much the same in July 2025: “all you need to do is normal SEO.”

Now the vendor side. Profound (total funding $58.5M) claims in its PR “a 700% increase in AI referrals in some cases”^[11]. A Chinese vendor markets “99.7% semantic-match accuracy” and “$1.1M in first-month sales”^[12] — none verifiable.

The most damning item: a widely circulated Chinese implementation guide cites a supposed “OpenAI《GEO Optimization Team Collaboration Guide》” claiming a 70% efficiency gain — I searched every engine I could; OpenAI never published such a document. It is a fabricated citation^[13].

The structural conflict of interest is clear: platforms live off a healthy content ecosystem, so they tell you not to panic; vendors live off your anxiety, so they tell you to hurry. Listen to both — and know who is making money off you.

What actually works: three tiers of evidence

High confidence (research or official documentation)

SEO fundamentals: crawlability, indexability, speed — the prerequisite for every AI citation^[4]
First-hand, unique, factual content — Google explicitly says first-hand reviews beat regurgitation^[4]
Citing sources, quoting verbatim, adding statistics — the paper’s experiments show >40% visibility gains^[2]
Consistent brand-entity representation across channels
Genuine presence in third-party corpora (Reddit, Zhihu, review sites) — CXL cites Semrush findings that ChatGPT frequently cites pages ranked 21+ in Google^[14]

Medium confidence (logical but thinly evidenced)

llms.txt: a proposed standard by Jeremy Howard (Sept 2024). Real-world reports from sites hosting 20,000 domains show no AI crawlers fetching it. Google’s Mueller, verbatim: “AFAIK none of the AI services have said they’re using LLMs.TXT (and you can tell when you look at your server logs that they don’t even check for it)“^[15]
GEO monitoring tools: “were you mentioned” is measurable; “mention → revenue” is not attributable by anyone

Low confidence / high risk

Hidden text, prompt injection, manufactured consensus — violates platform policies; the 315 case demonstrated the legal and regulatory risk^[1]
“Guaranteed AI placement” services — AI answers come from a black box with a moving model; “guaranteed” is technically meaningless

The noise economy: where the money flows

Globally, Sequoia-backed Profound has raised $58.5M in total^[11]. In China, WanGong Research Institute tracked 110+ GEO vendors; only ~19 had raised funding — 21 deals, ¥5.5B total, penetration under 15%^[16]. It also made the uncomfortable point itself: many companies fundraise on the GEO story while capital is actually betting on their pre-existing marketing business — “GEO looks more like a new story told to the market.”

The marketing-number pipeline keeps running: unattributed agencies produce “71% of enterprises have entered the field” and “2.8x conversion lift,” which get republished, enter internal briefings, and become decision inputs — with no verifiable original source anywhere. I even re-audited my own vault’s old swarm reports: “AI traffic +693%” and “4.4x conversion” traced back to vendor blogs, not primary sources^[17].

Finally, what you can do

Three things you can do today: confirm your site is properly indexable^[4]; rewrite your key product-fact pages as question-answer pairs with first-party data; and build a 20-question checklist you run weekly against the mainstream AI engines, recording whether you got mentioned — be your own monitoring tool.

What not to buy: guaranteed AI placement, undated “XX% lift” case collections, llms.txt one-click magic pills, and vendors who cite fabricated sources.

At bottom, GEO’s core methodology — making AI know you accurately — is mostly good SEO plus a little new awareness. The real dividing line is not technical: are you a content creator who wants to be cited accurately, or someone poisoning the AI? After 315, that choice is not just moral. It is legal.

中文版：GEO 两年：我翻了 16 个一手信源，把这门生意的真话和谎话分开