AI Search Draws 66% of Answers From Just 10 Documents, Gist Finds

AI Eats the Web That Feeds It with Gist — Ai4

Ai4

Generative search threatens an information supply collapse by eliminating publisher referral traffic while concentrating the bulk of answers into a narrow layer of unmaintained web documents.

AI-assisted summary of AI Eats the Web That Feeds It with Gist — original source. Check the original for full context.

Source published: 2026-08-20

Detailed notes

## Background Published on YouTube by Ai4 on August 20, 2026, this presentation features Peter Sloterdyk, Chief Marketing Officer at Gist. The session addresses the imminent supply-side collapse of web publishing caused by AI answer engines defunding the information ecosystem they harvest, urging industry stakeholders to understand this unpriced dependency and implement structural changes within 90 days.

## TL;DR Peter Sloterdyk demonstrates how the rise of AI answer engines is breaking the open web's monetization model as zero-click searches surge to 68% and publisher referral traffic plummets. Empirical research from Gist reveals that AI models rely heavily on an alarmingly narrow, stale retrieval set where just ten core documents supply two-thirds of answers in key categories. Because uncorrected archived sources now become persistent AI ground truths, traditional marketing playbooks are obsolete across brands, publishers, creators, and agencies. To survive this shift, Sloterdyk presents the mechanics of advertorial authority transfer and details a 90-day operational roadmap centered on auditing baseline query visibility and pricing domain authority.

## Atmosphere & style The presentation carries a sobering, analytical, and urgent tone that strips away tech hype to expose the fragile economic mechanics beneath generative search. Sloterdyk speaks with sharp corporate pragmatism and clear-eyed authority, balancing startling data visualizations with dry, self-aware humor when his presenter notes briefly flash on screen. The overall experience feels like an eye-opening boardroom briefing that transitions from an alarming structural critique into a methodical, strategic blueprint.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) The generative era has lengthened the half-life of being wrong, turning minor historical errors in forgotten listicles into permanent factual truths inside AI answer engines. - [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) Correcting an archived external publication has become a higher-leverage strategy for AI visibility than updating owned web properties or buying ads. - [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) Zero-click Google searches accelerated to roughly 68% in early 2026, up from 18% two years prior, breaking the web's primary monetization mechanic where traffic served as payment. - [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) The crawl-to-referral imbalance is severing publisher revenue: small publishers lost roughly 60% of search traffic while AI bots generate under 1.5% of publisher page views. - [06:41](https://www.youtube.com/watch?v=bv9tur3e53A&t=401s) AI answer generation relies on an unexpectedly narrow document layer: across 760 query variants, just 10 core documents supplied two-thirds of all synthesized answers. - [06:41](https://www.youtube.com/watch?v=bv9tur3e53A&t=401s) The underlying retrieval layer of AI search is sticky, neglected, and outdated, relying on unmaintained artifacts like undated pages or a 2009 Federal Reserve PDF. - [09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) The AI industry operates on an unpriced dependency, actively defunding the human information supply chain that creates the reporting and data models ingest. - [09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) Traditional top-of-funnel marketing has collapsed, forcing agencies and brands to pivot from driving web clicks to establishing authoritative presence inside LLM retrieval datasets. - [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) Large language models ingest advertorials and sponsored content strictly as factual markdown, effectively bypassing visual disclosures and absorbing the publisher's institutional authority. - [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) Authority is emerging as a tradeable commodity, opening monetization opportunities for publishers to rent out their domain credibility to brand advertisers targeting LLM outputs. - [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) A 90-day baseline plan requires brands and creators to map where their information lives across the web and verify historical accuracy across products and business rules. - [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) Independent creators must lock valuable content behind logins or paywalls to avoid being scraped as free, uncredited raw material for AI models. - [15:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=935s) Authoritative editorial corrections demonstrate powerful longevity, remaining the solitary source citation across multiple competing LLMs for years after indexing.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) The Half-Life of Being Wrong: When a Listicle Error Becomes the AI Answer

A person asking an online bank about overdraft fees is not making an obscure query; it is an everyday question with real financial stakes. In June 2021, the largest online bank announced it was eliminating overdraft fees entirely. Later that year, an article published a roundup listicle that failed to include them. In 2021, this was an ordinary, minor omission that ended no careers and went virtually unnoticed while the bank continued running standard digital marketing, social campaigns, and PR. Fast forward five years, and that single uncorrected listicle became the definitive source behind AI answer engines. What was once an innocent human error had been "frozen, promoted, and handed to a lot of people asking a real question about their own money."

```mermaid flowchart LR accTitle: Compounding Error Flow in AI Retrieval subgraph Origin["2021 Omission"] A["Bank eliminates overdraft fees"] --> B["Publication listicle omits bank"] end subgraph AIAnswer["2026 AI Retrieval"] B -->|"LLM crawls static source"| C["AI synthesizes answers"] C --> D["Omission presented as factual answer"] end subgraph Correction["Correction Path"] E["Bank requests edit to 5-yr-old article"] -->|"Source corrected"| C end ```

This dynamic marks a fundamental shift in how digital errors behave: "We changed the half-life of being wrong." In traditional search, listicle mistakes would naturally decay over time as new pages ranked. In the generative era, errors compound. To fix its visibility in the AI answer, the bank did not rewrite its website, increase ad spend, or publish new blog posts. Instead, it contacted the publication to correct the five-year-old article. Within days, the AI answer corrected itself. This upends twenty-five years of digital strategy: the highest-leverage move was no longer optimizing owned properties, but editing someone else's archived document.

- **The June 2021 announcement**: The largest online bank eliminated overdraft fees, yet was excluded from a major roundup listicle published later that year. - **The compounding error**: Five years later, AI answer engines adopted that uncorrected listicle as their sole ground truth. - **The fix**: Contacting the original publisher for an editorial correction updated the AI model's output within days, with zero changes to the bank's own site or marketing.

---

[02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) The 68% Zero-Click Reality and the Crawl-to-Referral Imbalance

Peter Sloterdyk introduces himself as the Chief Marketing Officer of Gist, an Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO) platform. Addressing why a marketing presentation belongs at an AI conference, Sloterdyk highlights that attendees build, fund, and deploy systems powering the question-and-answer economy. However, AI models are only as good as the inputs they read. In the first four months of 2026, roughly 68% of Google searches ended without a single click—an unprecedented surge from just 18% two years prior, representing the fastest acceleration of search behavior in over a decade.

This sudden contraction of organic traffic is devastating publishers who monetize by serving ads to incoming web visitors. Small publishers with 1,000 to 10,000 daily page views have lost roughly 60% of their search referral traffic in two years; midsize publishers have dropped 47%, and industry giants have fallen 22%. Meanwhile, AI chatbots account for just under 1% (perhaps 1.5%) of publisher page views, even as 57% of all web traffic is automated and AI crawlers represent the fastest-growing segment. Sloterdyk notes that this dynamic—dubbed the "crawl-to-referral imbalance" by companies like Cloudflare—means AI engines harvest publisher content at massive scale while returning virtually no human readers back to the source. A brief technical mishap occurs on stage when presenter notes appear on the main display, prompting laughter from the audience as Sloterdyk quips, "Well, then you guys are going to watch what I'm saying." He emphasizes the core economic disruption: "The click was not a metric. It was our form of payment."

- **Zero-click search acceleration**: Google zero-click searches jumped from 18% to 68% in just two years. - **Traffic automation**: 57% of total web traffic is automated, with AI bots being the fastest-expanding component. - **AI referral deficit**: Chatbots generate under 1% to 1.5% of publisher referral traffic despite aggressively scraping web content.

---

[06:41](https://www.youtube.com/watch?v=bv9tur3e53A&t=401s) Deconstructing the Retrieval Set: 10 Neglected Documents Rule the AI Web

To understand how answers are constructed when the familiar "10 blue links" are replaced by synthesized text, Gist conducted an empirical research study. They analyzed queries across technology, health, and finance, running the same question 48 times with multiple phrasing variants across 760 total prompts, generating over 1.58 million citations. While individual ChatGPT responses cited an average of 16 to 18 sources, the overall breadth of sources was shockingly narrow.

```mermaid flowchart TD accTitle: Retrieval Set Narrowing A["760 Query Variants (Tech, Health, Finance)"] --> B["1.58 Million Total Citations"] B --> C["Average 16–18 Sources per Individual Answer"] C --> D["Only 36 to 56 Unique Docs per Question Domain"] D --> E["10 Core Documents Supply 66% of All Answers"] E --> F["5 of 10 Core Documents Cannot Be Dated"] E --> G["Example: Federal Reserve PDF from January 2009"] ```

For any given question on ChatGPT, the model relied on a static set of only 36 to 56 documents total across all phrasing, language, and context variations. Across the entire dataset, a mere 10 documents provided two-thirds of every single answer generated. Shockingly, one of the documents underpinning answers across these modern queries was an un-updated Federal Reserve PDF dating back to January 2009—a time Sloterdyk humorously notes was when Kanye West interrupted Taylor Swift at the VMAs. Even more concerning, 5 of the top 10 documents could not be dated at all. The foundational answer layer of today's generative web is "really really small, undated, sticky, and neglected," relying on unmaintained artifacts that brands and publishers never think to audit.

- **Research scale**: 760 prompt variants across finance, health, and tech produced 1.58 million citations. - **Document concentration**: Across 48 distinct answer sets, only 23 total documents were used, with 12 appearing every time. - **Two-thirds rule**: Just 10 individual documents provided 66% of all synthesized answers. - **Data staleness**: 50% of the top 10 sources had no discoverable publication date, and one key finance source was an archived Federal Reserve PDF from January 2009.

---

[09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) The Fractured Ecosystem: The Imminent Supply-Side Breakdown

The breakdown of search traffic impacts four key stakeholders across the digital economy:

1. **Brands**: "The top of the funnel has lost its top." Twenty-five years of established digital playbooks are obsolete, requiring total reinvention. 2. **Publishers**: They have lost the transaction and ad monetization, which in turn strips brand advertisers of addressable web audiences. 3. **Creators**: They have been reduced to unpaid, uncredited "raw material" and ingestion inputs for large language models. 4. **Agencies**: The legacy promise of driving engagement and website traffic is broken, forcing agencies to pivot toward establishing authoritative brand presence inside LLM datasets.

The foundational value chain of the internet relies on a simple cycle: someone discovers new information, publishers write about it, user traffic pays for that reporting, and models ingest the knowledge. When traffic vanishes, the funding mechanism collapses. While no one can prove the exact date LLM output quality will noticeably degrade, defunding the human information supply chain threatens the integrity of future AI models.

```mermaid flowchart LR accTitle: The Collapsing Information Supply Chain subgraph SustainableLoop["Healthy Information Supply Chain"] D["Discovery / Journalism"] --> PUB["Publisher Coverage"] PUB --> TR["Web Traffic & Ad Dollars"] TR -->|"Funds new discovery"| D end subgraph DefundedLoop["Defunded Supply Chain"] PUB -->|"Ingested by"| LLM["LLM Synthesis"] LLM -->|"Zero-Click Direct Delivery"| USR["User"] USR -.->|"Broken Revenue Stream"| TR end ```

- **Top-of-funnel evaporation**: Brands can no longer depend on traditional inbound organic traffic funnels. - **Creator exploitation**: Creator content is ingested as raw training data without consistent attribution, citations, or downstream compensation. - **Agency reorientation**: Agency metrics must transition from traffic delivery to building domain authority within model retrieval layers. - **Supply-side risk**: Starving publishers of traffic eliminates the economic incentive to produce the original reporting that LLMs rely on to learn.

[11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) The Advertorial Arbitrage: How LLMs Ingest Sponsored Content

As the web's foundational information ecosystem begins to contract and degrade, the artificial intelligence industry finds itself operating atop an unpriced dependency. The open web that supplies AI models with training data and real-time retrieval is being hollowed out, but the resulting structural vacuum does not have to remain empty. Research into AI retrieval reveals an emerging mechanism: only one in every 10,000 citations currently stems from sponsored material. While advertorials have long served as a traditional staple of print and digital publishing, they are poised to become significantly more pivotal in the AI era due to a fundamental disconnect between how humans and large language models process web pages.

When a human visits a digital publication, the commercial disclosure remains visually front and center—designated by an asterisk, a "presented by" label, or a "supported by" byline. However, large language models interpret the underlying markdown entirely differently. Publishers structure their data so that the LLM ingests the page strictly as raw content and factual information, stripping out the commercial connotation of sponsorship. As a consequence, an advertiser who funds an advertorial effectively rents the publication's reputational weight in the eyes of the model. "Authority starts to become the commodity," shifting the digital monetization model from sheer traffic generation to the sale and transfer of institutional credibility.

```mermaid flowchart TD accTitle: The Authority Transfer Pipeline via LLM Ingestion subgraph Publisher ["Publisher Webpage"] HumanView["Human View: 'Presented By' / Asterisk Disclosure"] LLMView["LLM Markdown View: Plain Structured Text"] end subgraph Ingestion ["AI Model Processing"] Crawler["LLM Web Crawler"] IngestEngine["Ingests Content as Objective Authority"] end subgraph Result ["Model Output"] Answer["AI Answer Generation"] BrandAuthority["Brand Gains Authority & Sole Citation"] end

HumanView -.-> Publisher LLMView --> Crawler Crawler --> IngestEngine IngestEngine --> Answer Answer --> BrandAuthority ```

- In Gist's research, only 1 in 10,000 citations across AI search outputs came from sponsored content. - LLMs bypass visual sponsorship disclaimers ("presented by" or "supported by") in markdown, treating paid content as objective informational source text. - Brand advertisers can strategically borrow a host publication's institutional authority to establish dominance in AI-generated answers.

---

[13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) The 90-Day Plan: Baseline Visibility Across Brands, Publishers, Creators, and Agencies

Navigating this structural shift requires a deliberate 90-day operational framework—a timeframe long enough to delve into substantive strategic execution, unlike 30- or 60-day sprints. The fundamental premise of this roadmap is rooted in absolute visibility: "You can't fix what you can't see." Every stakeholder across the digital media ecosystem must establish an empirical baseline by identifying the critical queries they want to trigger, locating where the underlying source material lives (whether on proprietary websites, Reddit, YouTube, external editorial outlets, or historical archives), and auditing whether that information remains factually accurate after years of corporate and operational evolution.

For publishers, authority is the newly tradeable asset. Media outlets like *Forbes*, *People*, or *Entertainment Weekly* can monetize the ability of corporate advertisers to rent their domain stature—such as an entertainment studio funding coverage in *Entertainment Weekly* to position an upcoming release as the definitive blockbuster of the summer. Meanwhile, independent creators face the urgent task of owning their final destinations; any material left outside a paywall or login boundary is consumed by LLMs without attribution or citation mandates. Concurrently, advertising and marketing agencies must overhaul their traditional pricing logic. Instead of indexing value to raw web traffic or social engagement metrics, agencies must reprice their strategic, creative, and earned-media output around its capacity to build durable LLM authority.

- **Brands:** Map the exact web locations feeding target AI responses, verify factual accuracy, and reconcile outdated legacy information concerning business rules, products, and customer support. - **Publishers:** Reframe high-authority editorial domains as premium real estate where commercial brands pay to influence LLM sentiment and knowledge representation. - **Creators:** Secure content destinations behind login or paywall barriers to prevent unpaid, uncredited AI scraping, balancing audience discoverability with direct monetization. - **Agencies:** Transition billings and client value propositions away from vanity engagement toward long-term authority building and earned-media equity.

---

[15:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=935s) The New Vocabulary of Authority and the Longevity of AI Citations

The digital marketing landscape is actively coalescing around a new lexicon, shifting through concepts like Answer Engine Optimization (AEO), Generative Engine Optimization (GEO), and visibility, toward the foundational metric of authority. The central strategic imperative is establishing how authority is priced, accessed, and exchanged across industries. Once an AI model indexes a authoritative correction, that footprint possesses remarkable persistence.

This permanence is illustrated by a benchmark case study involving a major financial institution. In an analysis of a 2021 article that cited multiple competing banks while omitting the market's largest institution, the excluded bank took targeted action to get the source piece updated. Years later, that single corrected editorial placement continues to serve as the exclusive citation answering that specific banking query across four separate flagship LLMs. The lesson remains stark: while AI search behavior is constantly fluctuating, authoritative source corrections create long-lasting, highly stable citations across the entire generative ecosystem.

- Industry nomenclature is transitioning from SEO, AEO, and GEO toward the economics of authority pricing and access. - A single editorial correction made to a 2021 financial article remained the solitary citation for that banking query across four major LLMs over two years later. - AI citations exhibit significant inertia: once an accurate, high-authority source is adopted by generative models, the citation footprint remains entrenched.

## Notable quotes > Every AI answer is assembled from someone else's work. The traffic that paid for that work is fading. — Peter Sloterdyk

> The AI industry now runs on an unpriced dependency: an information ecosystem it is actively defunding. — Peter Sloterdyk

> [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) frozen, promoted, and handed to a lot of people asking a real question about their own money. — Peter Sloterdyk

> [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) We changed the half-life of being wrong. — Peter Sloterdyk

> [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) Well, then you guys are going to watch what I'm saying. — Peter Sloterdyk

> [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) The click was not a metric. It was our form of payment. — Peter Sloterdyk

> [06:41](https://www.youtube.com/watch?v=bv9tur3e53A&t=401s) really really small, undated, sticky, and neglected — Peter Sloterdyk

> [09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) The top of the funnel has lost its top. — Peter Sloterdyk

> [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) Authority starts to become the commodity — Peter Sloterdyk

> [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) You can't fix what you can't see. — Peter Sloterdyk

## Mentioned Peter Sloterdyk, Gist, Ai4, Google, Cloudflare, ChatGPT, Federal Reserve, Kanye West, Taylor Swift, VMAs, Forbes, People, Entertainment Weekly, Reddit, YouTube

## Video description Peter Sloterdyk - Chief Marketing Officer, Gist

Every AI answer is assembled from someone else's work. The traffic that paid for that work is fading. Small publishers lost roughly 60% of search referrals in two years. AI referrals are nowhere close to replacing them; the ad supply funding the open web is contracting behind it. The AI industry now runs on an unpriced dependency: an information ecosystem it is actively defunding. Anticipate the imminent supply-side collapse underneath AI search. Zoom out from tactics to incentives, to learn the one structural change you'll make in the next 90 days to support a sustainable, reimagined market.

[Source](https://www.youtube.com/watch?v=bv9tur3e53A)

中文简报

Gist发现：AI搜索66%的答案仅来自10篇文档生成式搜索在消除出版商引流流量的同时，将绝大部分答案集中在狭窄且缺乏维护的网页文档层中，从而带来信息供给崩溃的威胁。 Peter Sloterdyk阐述了随着零点击搜索激增至68%以及出版商引流流量骤降，AI问答引擎的兴起正在瓦解开放网络的变现模式。Gist的实证研究表明，AI模型极度依赖狭窄且陈旧的检索集，在关键类别中仅10篇核心文档便提供了三分之二的答案。由于未经修正的归档信息源如今成为AI持续认定的既定事实，传统的营销策略在品牌、出版商、创作者和代理机构中已全面过时。为应对这一转变，Sloterdyk解析了软文权威传递的机制，并详细介绍了一份以审计基准查询可见性及评估域名权威为核心的90天实操路线图。生成式时代延长了错误的半衰期，使被遗忘的榜单文章中微小的历史错误沦为AI问答引擎内永久的事实真相。 (0:04) 相比更新自有网站内容或投放广告，修正已归档的外部出版物已成为提高AI可见性杠杆效应更强的策略。 (0:04) 2026年初，Google的零点击搜索占比从两年前的18%加速上升至约68%，瓦解了以流量作为报酬的互联网核心变现机制。 (2:44) 爬取与引流之间的失衡正在切断出版商的收入：小型出版商损失了约60%的搜索流量，而AI机器人产生的出版商页面浏览量不足1.5%。 (2:44) AI答案生成依赖于出人意料的狭窄文档层：在760个查询变体中，仅10篇核心文档便提供了全部综合答案的三分之二。 (6:41)

Reading summary

Gist Chief Marketing Officer Peter Sloterdyk argues that generative AI engines are breaking the open web's economic engine by answering queries directly without returning traffic, while drawing answers from a surprisingly narrow, unmaintained layer of static documents. Compounding Errors and the Persistent Retrieval Set (0:04) In traditional search architectures, factual errors in listicles and articles tended to decay naturally as newer pages were indexed and ranked. Generative answer engines invert this behavior by freezing and compounding uncorrected mistakes. In one noted case, a major online bank eliminated overdraft fees in June 2021, but an uncorrected omission in an article published later that year was adopted five years later by AI models as factual ground truth. Correcting the AI's output required reaching out directly to the publisher to edit the archived five-year-old piece rather than modifying the bank's own site, ads, or marketing materials. The Zero-Click Surge and Publisher Traffic Deficits (2:44) According to Sloterdyk, Google searches ending without a click jumped from 18% to roughly 68% in the first four months of 2026. This sudden contraction of inbound search referrals has hit publishers unevenly: small publishers with 1,000 to 10,000 daily page views experienced an estimated 60% loss in search referral traffic, midsize publishers dropped 47%, and large media outlets declined 22%. While automated traffic accounts for 57% of total web traffic, AI chatbots generate less than 1% to 1.5% of publisher referral visits, creating what Cloudflare terms a 'crawl-to-referral imbalance.' Narrow Foundations in Model Answer Construction (6:41) Empirical testing conducted by Gist across technology, health, and finance domains evaluated 760 prompt variants and 1.58 million citations. Although individual ChatGPT responses cited an average of 16 to 18 sources, the broader corpus remained tight: across 48 distinct answer sets, the model relied on only 23 total documents, with 12 appearing every time. In total, just 10 core documents supplied 66% of all synthesized answers. Half of these 10 documents lacked discoverable publication dates, and one modern financial query consistently relied on an unmaintained Federal Reserve PDF dating back to January 2009. Advertorial Ingestion and the Authority Trade (11:35) As human search traffic diminishes, the digital monetization model is pivoting toward institutional authority. Gist's research found that only one in 10,000 citations in AI search outputs derived from sponsored content. However, because large language models ingest underlying markdown rather than visual layouts, they typically bypass commercial disclosures like 'presented by' or asterisks, treating advertorial copy as objective source text. This technical gap allows commercial brands to rent high-authority publisher domains to secure durable placements in synthesized answers, while creators are forced to move content behind paywalls to avoid uncompensated scraping. The broader sustainability of large language models remains vulnerable to defunding the human information supply chain, as cutting off traffic disincentivizes publishers from generating the primary reporting models require. While operational roadmaps emphasize auditing the specific web locations feeding LLM outputs, the long-term degradation timeline for AI answer quality without fresh human input remains an open question.

精炼阅读摘要

Gist首席营销官Peter Sloterdyk指出，生成式AI正通过高达68%的零点击搜索和失衡的网页抓取抽干开放网络流量，而大模型本身却高度依赖极少数未维护、无日期的陈旧文档生成答案。这种机制打破了依托点击变现的传统数字经济链条，迫使品牌、出版商与代理机构从争夺流量转向争夺并租赁AI检索层中的权威地位。零点击激增与抓取推荐失衡冲击出版业 (2:44) 根据Gist提供的数据，2026年前四个月，Google搜索中约有68%的查询没有产生任何后续点击，相较两年前18%的水平出现急剧跃升，代表了十余年来搜索行为演变的最快阶段。与此同时，自动化流量已占全网流量的57%，AI爬虫成为扩张最快的群体。然而，AI聊天机器人带给出版商的推荐流量仅占其总页面浏览量的不到1%（至多1.5%）。Cloudflare等机构将这种现象称为“抓取与推荐失衡”，即AI大规模收割出版商内容，却几乎不向源头反哺真实读者。这种流量收缩对依靠广告变现的内容创作者造成沉重打击。数据显示，日均浏览量在1,000至10,000之间的小型出版商，其搜索推荐流量在两年内下滑约60%；中型出版商下滑47%；行业头部大型媒体亦下降了22%。Sloterdyk强调，点击在过去不仅是一项统计指标，更是维系整个内容生态运转的“支付货币”，零点击回答的普及直接切断了原创出版商的资金循环。检索源极度集中与陈旧错误的算法固化 (6:41) 当传统的“十条蓝色链接”被整合答案取代后，AI检索的底层信源展现出极高收敛性与脆弱性。Gist在科技、健康和金融领域开展的实证研究显示，针对760个提示变体的测试（产生超过158万次引用），单次ChatGPT回答平均引用16至18个来源，但特定领域全部提问所依赖的唯一文档仅有36至56篇。在整个数据集中，仅仅10篇核心文档就提供了66%的生成答案；其中5篇完全无法追溯发布日期，金融领域甚至长期引用一份2009年1月美联储未经更新的PDF文件。这种机制彻底改变了数字错误的“半衰期”。传统搜索中过时信息会随新页面权重上升自然沉降，而在生成式引擎中错误会被固化与放大。例如2021年某大型在线银行取消透支费，但当年一篇盘点文章遗漏了该信息；五年后，该未修正的文章成为AI回答的权威事实源。该银行最终并非通过修改自身官网或追加广告预算，而是联系原媒体修正五年前的旧稿，才在数天内使四款主流大模型的回答同时修正，且该单一信源此后长期被模型采纳。信息供应链承压与赞助内容的权威套利 (11:35) 搜索流量的萎缩对数字生态中的四大参与者构成连锁冲击：品牌的传统漏斗顶端不复存在；出版商失去受众变现渠道；独立创作者的内容被提取为无署名、无报酬的训练素材；代理机构以点击和互动为核心的交付承诺宣告失效。Sloterdyk警告，虽然无法确切预知模型输出质量何时出现实质性退化，但抽干人类原创报道的资金支持，终将动摇大语言模型未来知识更新的供给基础。在生态重塑过程中，赞助内容展现出特殊的套利空间。Gist研究表明，目前AI搜索引用中仅有万分之一直接来自赞助内容。然而，人类读者能直观看到的“赞助发布”或星号标识，在模型抓取的Markdown数据层中往往被剥离为纯文本事实。大模型在解析时倾向于将此类内容视同客观权威，使得商业品牌能够有效“租赁”知名媒体的机构声誉，以极低被标记风险在AI答案中确立排他性权威。 90天审计框架与权威资产的重新定价 (13:08) 面对生成式搜索的深层变革，演讲提出了一个为期90天的行动框架，主张行业必须建立“可见性基线”。品牌需要系统排查触发自身核心业务问题时AI调用的底层数据源，追查信息究竟寄存在第三方媒体、Reddit、YouTube还是历史归档中，并全面审核涉及业务规则、产品规格与客服条款的陈旧数据。行业竞争焦点正从传统的搜索引擎优化（SEO）以及生成引擎优化（GEO），转向对“权威”的定价与占领。在此框架下，出版商可将高权重域名转化为供品牌方影响模型认知的商业阵地；独立创作者亟需将核心资产迁移至付费墙或登录门槛之后，以抵御未经授权的抓取；营销代理机构则需彻底重构收费逻辑，将衡量标准从即时流量和虚荣指标，转向在主流模型检索层中构建长期稳定的权威资产与媒体权益。 Sloterdyk总结认为，生成式AI所依赖的开放网络基础正在被无偿榨取，但整个数字产业尚未建立起可持续的内容反哺机制。尽管通过修正外部高权重归档或赞助权威媒体能够短期操纵AI检索结果，然而若内容生产端的商业激励长期缺失，AI模型未来将面临训练数据退化与事实供给断裂的根本性系统风险。

Detailed notes — zh

## Background 该内容由 Ai4 频道于 2026 年 8 月 20 日发布在 YouTube 平台，由 AEO 与 GEO 平台 Gist 的首席营销官 Peter Sloterdyk 主讲。视频简介指出，AI 问答建立在他人的劳动成果之上，但维系开放网络的流量与广告机制正在萎缩，演讲旨在揭示 AI 搜索底层信息供应链面临的断裂风险，并探讨如何通过结构性变革重塑可持续的内容生态。 ## TL;DR 演讲通过一家大型银行因陈旧报道而在 AI 答案中被遗漏的真实案例，揭示了大语言模型会“冻结”历史错误并高度依赖狭窄、过时的底层信源。随着零点击搜索激增至 68%，AI 抓取与回流引荐之间出现严重失衡，导致作为信息源头的出版商流量暴跌、商业模式濒临崩溃。面对品牌获客受阻、创作者沦为无偿语料等全产业链危机，演讲者提出了一套 90 天行动指南，倡导各方将“权威度”作为可定价的核心资产进行重构，以应对 AEO/GEO 时代的范式转移。 ## Atmosphere & style 演讲整体呈现出清醒、务实且极具警示性的行业洞察基调，以严谨的数据研究和生动的商业案例层层剥开 AI 热潮背后的生态脆弱性。演讲者娓娓道来，语言犀利直接，从微观的搜索事实偏差一路推演至宏观经济链条的断裂，展现出强烈的问题意识与危机感。整场分享不仅给观众带来对传统 SEO 和流量变现失效的认知冲击，更通过条理清晰的 90 天落地框架注入了切实可行的战略定力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) AI 彻底改变了事实错误的“半衰期”：过去会随时间自然淡化的媒体疏漏，如今会被大模型捕获并固化为唯一标准答案，促使营销重点转向纠正外部信源的 AEO/GEO 策略。 - [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) 零点击搜索比例激增至 68%，导致小型出版商在两年内流失了约 60% 的搜索引荐流量，重创了依赖搜索点击变现的传统网络出版模式。 - [05:10](https://www.youtube.com/watch?v=bv9tur3e53A&t=310s) 生态出现严重的“抓取与引荐失衡”（Crawl to referral imbalance）：AI 爬虫虽然海量抓取原创内容，给出版商带来的回流访问量却仅占 1% 至 1.5%，蒸发了开放网络赖以维系的“点击货币”。 - [06:36](https://www.youtube.com/watch?v=bv9tur3e53A&t=396s) AI 答案层底层信源极度狭窄且严重过时：基准研究显示仅 10 篇关键文档便垄断了近三分之二的回答内容，且模型高频依赖十多年前未维护的陈旧文档作为事实依据。 - [09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) 开放网络价值链面临断裂风险：品牌失去了营销漏斗顶端流量，创作者沦为无偿训练原材料，若出版商因缺乏商业回报而停止生产优质内容，将最终反噬 AI 模型的学习能力。 - [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) 软文广告成为品牌向媒体“借用权威”的特殊杠杆：由于抓取层数据往往省略商业赞助标注，LLM 会将软文内容直接视作客观事实并作为答案输出。 - [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) 品牌方需在 90 天内建立基准，主动审计并修正散落在第三方媒体和历史网页中的陈旧错误，防止过时信息持续误导模型与用户。 - [13:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=815s) 出版商应将商业模式从售卖点击转向售卖“权威度”，将自身的媒体公信力包装为向 LLM 输出事实背书的核心大宗商品。 - [14:05](https://www.youtube.com/watch?v=bv9tur3e53A&t=845s) 创作者需权衡开放抓取与私域掌控以防止被模型无偿榨取，代理商则需从承诺流量与互动转向以“权威度构建”来重塑服务定价体系。 - [15:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=935s) AEO 与 GEO 的核心实质是权威资产的争夺与定价；由于大模型引用的高粘性，品牌一旦在权威信源完成事实纠偏，便能获得长期稳定的 AI 搜索壁垒。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) 消失的银行与被“冻结”的错误：AI 时代错误的半衰期当有人在搜索引擎中提问“哪些网络银行不收透支费？”时，这绝非一个冷门偏门的问题，而是一个每天都有大量用户查询、关乎个人切身财务利益的真实需求。回溯到 2021 年 6 月，全美规模最大的网络银行公开宣布全面取消透支费，并在随后的营销、社交媒体和线上内容中倾注了大量资源，严格践行了过去 25 年数字营销领域的所有最佳实践。然而，在同年晚些时候发表的一篇盘点免透支费银行的媒体文章中，该银行的名字却被遗漏了。在传统搜索时代，这不过是记者赶稿时常有的疏漏，属于“以 Y 结尾的日子（意指稀松平常的每一天）”里发生的一件小事，既没有砸掉谁的饭碗，也没有引起任何人的关注。然而五年过去，随着生成式 AI 与大语言模型（LLM）的崛起，这个原本微不足道的媒体疏漏被模型捕获并“冻结”下来，直接演变成了 AI 给用户的唯一确定答案——它不再只是众多参考来源之一，而是成为了最终答案本身。正如演讲者所指出的，技术的发展并没有仅仅自动化信息检索，而是“改变了犯错的半衰期”（changed the half-life of being wrong）。在过去，传统列表文章里的事实错误会随时间自然衰减被遗忘；而在当下，这些错误却在 AI 答案层中不断复合、放大，并直接呈现在无数关心自己银行账户余额的真实用户面前。为了纠正这一致命的错误答案，该银行并未采取常规手段去重写官网或大量堆砌新内容，而是直接联系了最初刊发文章的出版机构，请求对这篇五年前的旧报道进行勘误。令人震惊的是，在源文件被修改后的短短数天之内，AI 生成的答案就随之更正，将该银行重新纳入免透支费的名单中。在过去 25 年的经典 SEO 理论中，没有任何一本教科书会教导营销人员“最有效的方法是跑去编辑别人的旧文档”。但这恰恰揭示了当下的残酷现实：AEO（答案引擎优化，Answer Engine Optimization）与 GEO（生成式引擎优化，Generative Engine Optimization）已经彻底颠覆了用户的发现机制，而外部出版商正重新成为品牌方最核心的关键合作伙伴。 ### [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) 搜索零点击与出版商流量的断崖式暴跌作为 AEO 和 GEO 平台 Gist 的首席营销官（CMO），Peter Sloterdyk 阐述了营销议题之所以属于核心 AI 讨论的原因：当所有 AI 开发者都在构建、投资和部署能够回答问题的系统时，他们实际上只构建了“问答经济”（Question and Answer Economy）的一半。如果问答系统读取的底层语料本身存在偏差或错误，那么即便品牌方做对了一切，最终输出的答案依然是荒谬的。在 AI 时代，“输入决定输出”的铁律依然适用，归属机制（Attribution）与信源引用（Citations）正是维系整个生态运转的基石。然而，维系这一生态的传统经济支柱正在急速崩塌。数据显示，在 2026 年的前四个月中，约有 68% 的 Google 搜索在没有任何用户点击的情况下直接结束（即“零点击搜索”），而这一比例在仅仅两年前还只有 18%。这种十多年未见的剧烈转变，正以前所未有的速度测试着整个营销行业的韧性。依赖搜索引流变现的出版商首当其冲： - **小型出版商**（日均浏览量 1,000 至 10,000）：在过去两年内失去了约 60% 的搜索引荐流量； - **中型出版商**：失去了约 47% 的搜索引荐流量； - **大型出版巨头**：搜索引荐流量也大幅缩减了约 22%。 ```mermaid flowchart TD accTitle: AI Crawl to Referral Imbalance accDescription: Flowchart illustrating how AI bots scrape publisher content without returning web traffic, destroying ad monetization. subgraph ContentCreation["内容与商业层"] P["内容出版商 / 创作者"] M["传统商业变现<br>(广告展示 / 点击结算)"] end subgraph AIEcosystem["AI 抓取与生成层"] AI["AI 大模型与聊天机器人"] A["直接生成最终答案<br>(零点击搜索激增至 68%)"] end P -->|"大规模抓取内容 (占自动化流量主流)"| AI AI --> A AI -.->|"极微弱引流 (仅占出版商浏览量 1%-1.5%)"| P P --- M A -.->|"剥夺访问与点击"| M ``` 出版商依靠访问流量售卖广告来实现内容变现，但如今的流量格局却陷入了严重的失衡。尽管全网高达 57% 的流量已由自动化机器人构成（其中 AI 抓取是增速最快的板块），但 AI 聊天机器人给出版商网站带来的回流访问量却仅占其总浏览量的 1% 至 1.5% 左右。Cloudflare 等机构将这种极度不对称的现象称为“抓取与引荐失衡”（Crawl to referral imbalance）：AI 系统在海量抓取、榨取出版商原创内容的同时，几乎不向源网站反哺任何读者。“点击从来不只是一项分析指标，它是我们赖以支付报酬的货币形式。”开放网络赖以生存的商业循环，正因点击货币的蒸发而面临断裂。 ### [06:36](https://www.youtube.com/watch?v=bv9tur3e53A&t=396s) 答案层的真相：极其狭窄、过时且脆弱的信息源当前营销人员普遍将 AI 时代的挑战视为表面上的“曝光度问题”（Visibility Problem），认为只要做好模型优化、争取被引用即可。但深入解构 AI 答案的生成链条——从用户提问（Query）、到检索集（Retrieval Set）、再到最终答案（Answer）——就会发现其核心的“秘密配方”隐藏在模型对最省力路径的依赖中。为了探究 AI 答案底层的真实信源构成，Gist 展开了一项深度基准研究：针对科技、医疗健康和金融三大垂直领域，设计了 48 个核心问题及其变体，累计发起 760 次提问，共收集到超过 158 万条引用数据。研究揭示的底层结构极为脆弱和狭隘： | 研究维度 | 统计数据与事实特征 | 行业影响与隐患 | |:--- |:--- |:--- | | **测试问题与引用总量** | 760 个问题变体，产生 158 万次引用 | 单次回答平均引用 16 至 18 个信源 | | **单问题底层信源数** | 无论提示词语言与语境如何变化，仅调用 36 至 56 篇核心文档 | 信源池极度单一，缺乏多样性与冗余度 | | **头部文档集中度** | 仅 10 篇关键文档贡献了全部答案中近三分之二（2/3）的内容 | 关键事实被少数旧文档高度垄断 | | **信源过时程度** | 金融类答案高频依赖一份美联储 2009 年 1 月的 PDF | 沿用 15 年前的信息回答当下的金融咨询 | | **信源时效可查性** | 抽样的 10 篇核心文档中，有 5 篇完全无法考证发布日期 | 大量未维护、无日期的陈旧网页成为“永久真理” | AI 答案层并非实时反映着最新的互联网智慧，相反，它极其窄小、过时、具备极强粘性且长期缺乏维护。正如五年前那篇遗漏银行名单的文章一样，品牌一旦在这些陈旧的关键信源中缺失或被误读，就将在 AI 时代被彻底隔绝在用户的视野之外。 ### [09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) 开放网络的价值链断裂：四大群体的系统性危机面对答案层的结构性异变，整个数字生态中的四大核心群体正同时经历着前所未有的范式震荡： - **品牌方（Brands）**：营销漏斗的顶端彻底“消失”（Top of the funnel has lost its top）。过去 25 年行之有效、经过反复验证的获客策略大面积失效，当公共媒体池中的有效流量枯竭时，品牌也同步失去了赖以触达的目标受众； - **出版商（Publishers）**：不仅失去了流量，更从根本上失去了交易转化与商业变现的能力，陷入“内容被免费拿走，网站无人问津”的生存危机； - **内容创作者（Creators）**：沦为了训练大模型的纯粹“原材料”（Raw material）与无偿输入源，在模型输出的答案中极少获得署名、提及或实质性的引流回报； - **营销代理机构（Agencies）**：过去向客户承诺“带来海量点击与互动”的提案逻辑已经过时，核心诉求被迫转向如何在关键核心领域为品牌重新构建权威度（Authority）。从整个信息生态的宏观价值链来看，其运转逻辑原本清晰闭环：有人探索发现新事实 $\rightarrow$ 出版商撰写报道并分发 $\rightarrow$ 读者点击带来的流量与商业化收益为此买单 $\rightarrow$ AI 模型从中抓取并持续学习进化。然而，一旦为原创内容买单的流量经济彻底崩塌，出版商将失去持续生产优质内容的动力与资金，导致模型失去高质量的训练输入。虽然目前尚无法断言模型能力会在哪一年发生断崖式退化，但如果不及时修补这一断裂的价值链，底层信息生态的瓦解势必反噬 AI 产业自身。 ### [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) 未定价的依赖与权威借用：软文广告在 LLM 时代的裂变整个 AI 产业正在建立在一个未被合理定价的依赖（unpriced dependency）之上，这造成了生态系统的严重空缺与失衡，但这种真空并不意味着会一直保持空白。根据 Gist 的研究数据，当前 AI 引用中每 10,000 次仅有 1 次来自赞助内容（sponsored content）。这引出了出版业历史上悠久、但即将迎来爆发的经典形态——软文广告（advertorial）。人类读者与大语言模型（LLM）看待这类内容的方式存在着戏剧性的割裂：对人类读者而言，赞助声明清清楚楚地留在页面上，无论是以星号标注，还是冠以“由……呈献”（presented by）或“由……支持”（supported by）的标签；然而，绝大多数出版商在面向 LLM 的 Markdown 结构化视图中，根本不会标记这是赞助内容，因为告诉模型这是广告对抓取并没有意义，模型只会将其视为纯粹的内容和客观信息。这意味着付费刊登软文的广告主，实际上是在直接“借用”（borrow）该出版机构的权威度。权威性（authority）正迅速演变成为一种可以被定价、购买与变现的硬通货。如果一家品牌希望在某个主题上成为大模型眼中的绝对权威，直接向《福布斯》（Forbes）或《人物》（People）等老牌媒体购买权威背书，就成了最高效的杠杆。 ```mermaid flowchart LR accTitle: 软文广告在LLM时代的权威传递链路 subgraph HumanView["人类视角 (前端网页)"] A["媒体平台内容"] --> B["标注：赞助内容 / 软文"] B --> C["人类读者辨识为商业广告"] end subgraph AIView["LLM 视角 (Markdown / 抓取层)"] D["未标注广告的结构化信息"] --> E["LLM 直接吸收为事实信源"] E --> F["品牌借用媒体权威成为答案输出"] end ``` - 在 Gist 进行的研究中，AI 生成回答的引用来源里，赞助内容占比仅为万分之一（1 in 10,000）。 - 出版商提供给 LLM 读取的 Markdown 数据中往往省略了商业赞助披露，使得软文中的品牌论述直接被 AI 摄入为权威事实。 --- ### [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) 90 天落地行动指南：品牌、媒体、创作者与代理商的生存战法为了应对这一生态巨变，Peter Sloterdyk 提出了一个切实的 90 天转型行动框架——“30 天太短，60 天不够，90 天恰好可以深入到实质核心”。解决问题的前提是可见性，“你无法修复你看不见的东西”。各方生态角色需要从底层重新校准各自的定位与商业模式： 1. **品牌方（Brands）：建立基准并审计旧信息** 品牌的第一步是厘清自己希望在哪些查询（queries）中出现，并追踪生成这些 AI 答案的底层信息究竟存放在哪里。无论是在自家官网、Reddit 社区、YouTube 视频、媒体文章，还是几百年前缩微胶卷般的历史遗留碎片中。找到信源后，必须逐一核查其准确性：10 年前的信息是否还在误导用户？企业现在的业务模式、服务条款或交付体验是否已发生剧变？必须主动介入修正这些陈旧叙事。 2. **出版商（Publishers）：将“权威度”包装为核心商品变现** 出版商必须意识到“权威是新的大宗商品”。当品牌广告主渴求在 LLM 面前建立特定领域的专业度时，出版商应直接针对这种需求定价。例如影视宣发为了让自己的电影被 AI 推荐为“暑期档必看大片”，愿意向《娱乐周刊》（Entertainment Weekly）购买软文，出版商出售的不再是点击，而是自身的权威信任资产。 3. **创作者（Creators）：掌控目的地并权衡开放与变现** 创作者必须牢牢掌握内容的目的地。如果内容没有设立付费墙（paywall）或登录限制，LLM 可以肆意抓取并合成答案，而且现阶段大模型没有任何法律义务对创作者进行标注、致谢，甚至不会提及内容来自创作者的社交频道或 YouTube 视频。创作者必须在“被大模型抓取的曝光度”与“私域受众掌控力”之间重新找到平衡。 4. **代理商（Agencies）：重构以“权威构建”为核心的定价体系** 代理商面临着价值主张的彻底重写。无论是策略构建还是创意产出，代理商不应再用传统的“流量”（traffic）或“互动量”（engagement）来计价，而是必须主动向客户按“权威度构建”（authority building）重新定价。例如，所产出的方案能为品牌沉淀多少高权重的自然媒体曝光（earned media），这些曝光在 LLM 体系中能带来多大持久价值。 --- ### [15:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=935s) 从 SEO/GEO 到权威定价：纠偏效应的持久胜利当前整个行业正处于概念与词汇体系的快速融合期。行业内外都在摸索各种新名词，无论将其称为答案引擎优化（AEO）、生成式引擎优化（GEO），还是单纯的 AI 可见性（visibility），其本质内核都是“权威”（authority）。业界需要更深入地公开探讨：权威到底该如何定价？其真实商业价值为何？大家又该如何获取它？ Peter 结语时重新呼应了演讲开篇提到的银行案例：在 2021 年，原先那篇关键对比文章列举了大量竞争对手银行，却唯独遗漏了市面上体量最大的那家银行。在推动文章内容修正后，那家大银行成功进入了 AI 的推荐名单之中。这起修改发生在两年前，直到今天，该文章依然是四大主流 LLM 在回答对应银行选购问题时的唯一引用来源（single citation）。“事物在改变，但之后又保持不变”——一旦品牌成功在权威源头完成了事实纠偏，这种生成的先发优势就能够在相当长的时间内稳定生效，成为 AI 搜索时代最坚固的壁垒。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) 改变了犯错的半衰期（changed the half-life of being wrong） — Peter Sloterdyk > [00:04](https://www.youtube.com/watch?v=bv9tur3e53A&t=4s) 最有效的方法是跑去编辑别人的旧文档 — Peter Sloterdyk > [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) 点击从来不只是一项分析指标，它是我们赖以支付报酬的货币形式。 — Peter Sloterdyk > [02:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=164s) 抓取与引荐失衡（Crawl to referral imbalance） — Peter Sloterdyk > [09:44](https://www.youtube.com/watch?v=bv9tur3e53A&t=584s) 营销漏斗的顶端彻底“消失”（Top of the funnel has lost its top） — Peter Sloterdyk > [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) 整个 AI 产业正在建立在一个未被合理定价的依赖（unpriced dependency）之上 — Peter Sloterdyk > [11:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=695s) 权威性（authority）正迅速演变成为一种可以被定价、购买与变现的硬通货。 — Peter Sloterdyk > [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) 30 天太短，60 天不够，90 天恰好可以深入到实质核心 — Peter Sloterdyk > [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) 你无法修复你看不见的东西 — Peter Sloterdyk > [13:08](https://www.youtube.com/watch?v=bv9tur3e53A&t=788s) 权威是新的大宗商品 — Peter Sloterdyk > [15:35](https://www.youtube.com/watch?v=bv9tur3e53A&t=935s) 事物在改变，但之后又保持不变 — Peter Sloterdyk ## Mentioned Peter Sloterdyk, Gist, Google, Cloudflare, 美联储 (Federal Reserve), 福布斯 (Forbes), 人物 (People), 娱乐周刊 (Entertainment Weekly), Reddit, YouTube ## Video description Peter Sloterdyk - Chief Marketing Officer, Gist Every AI answer is assembled from someone else's work. The traffic that paid for that work is fading. Small publishers lost roughly 60% of search referrals in two years. AI referrals are nowhere close to replacing them; the ad supply funding the open web is contracting behind it. The AI industry now runs on an unpriced dependency: an information ecosystem it is actively defunding. Anticipate the imminent supply-side collapse underneath AI search. Zoom out from tactics to incentives, to learn the one structural change you'll make in the next 90 days to support a sustainable, reimagined market. [Source](https://www.youtube.com/watch?v=bv9tur3e53A)

Source quotes

[0:04] Peter Sloterdyk: 改变了犯错的半衰期（changed the half-life of being wrong） [0:04] Peter Sloterdyk: 最有效的方法是跑去编辑别人的旧文档 [2:44] Peter Sloterdyk: 点击从来不只是一项分析指标，它是我们赖以支付报酬的货币形式。 [2:44] Peter Sloterdyk: 抓取与引荐失衡（Crawl to referral imbalance） [9:44] Peter Sloterdyk: 营销漏斗的顶端彻底“消失”（Top of the funnel has lost its top） [11:35] Peter Sloterdyk: 整个 AI 产业正在建立在一个未被合理定价的依赖（unpriced dependency）之上 [11:35] Peter Sloterdyk: 权威性（authority）正迅速演变成为一种可以被定价、购买与变现的硬通货。 [13:08] Peter Sloterdyk: 30 天太短，60 天不够，90 天恰好可以深入到实质核心 [13:08] Peter Sloterdyk: 你无法修复你看不见的东西 [13:08] Peter Sloterdyk: 权威是新的大宗商品 [15:35] Peter Sloterdyk: 事物在改变，但之后又保持不变

Source quotes

Peter Sloterdyk: Every AI answer is assembled from someone else's work. The traffic that paid for that work is fading. Peter Sloterdyk: The AI industry now runs on an unpriced dependency: an information ecosystem it is actively defunding. [0:04] Peter Sloterdyk: frozen, promoted, and handed to a lot of people asking a real question about their own money. [0:04] Peter Sloterdyk: We changed the half-life of being wrong. [2:44] Peter Sloterdyk: Well, then you guys are going to watch what I'm saying. [2:44] Peter Sloterdyk: The click was not a metric. It was our form of payment. [6:41] Peter Sloterdyk: really really small, undated, sticky, and neglected [9:44] Peter Sloterdyk: The top of the funnel has lost its top. [11:35] Peter Sloterdyk: Authority starts to become the commodity [13:08] Peter Sloterdyk: You can't fix what you can't see.

Original transcript

[0:04] Which online banks don't charge overdraft fees? It's not an obscure query, a real question asked [0:12] constantly with real financial consequences for the person asking. In June 2021, the largest [0:18] online bank announced that they were eliminating overdraft fees. Oh, we got a presentation. There [0:26] we go. Tada. [laughter] And uh later that year, an article was written not naming them as one of [0:36] the banks that doesn't charge overdraft fees. It was a huge part of their marketing, their [0:41] online content, social, and more. They were doing everything 25 years of digital marketing expertise [0:46] told them to. And yet, they were still not in the answer. So, I don't see that as a scandal. [0:55] I see it as a pretty standard day. It's a day that ends in Y, you might say. Journalists write lists, [1:03] lists are incomplete, and in 2021, this was a minor omission that harmed no one. Nobody's [1:08] career ended. Nobody even noticed. But 5 years later, that omission has become the answer. Not a [1:17] source for the answer, but the whole answer. That ordinary, forgettable, entirely human mistake got [1:22] frozen, promoted, and handed to a lot of people asking a real question about their own money. So, [1:29] it looks like we changed the half-life of being wrong. We didn't just automate retrieval. An [1:34] error in a standard listicle used to simply decay. Now, it has the potential to compound. [1:40] They did not publish more content. They did not rewrite their site. In order to get in the answer, [1:47] they contacted the publication and asked for a correction to a 5-year-old article. Within days, [1:54] just days, they were in the answer. Now, I can't empirically prove that the correction [2:00] caused it. What I can tell you is that the source document changed, the answer changed within days, [2:06] and nothing about that bank, their site, their marketing, their press, their social, nothing [2:12] about it changed in that window. This isn't an SEO story. I don't believe that there's any version of [2:18] the last 25 years of marketing that says the best move you can make is to go edit somebody else's [2:24] document. But here we are living in today's world where AEO and GEO are a huge factor in [2:32] the way that our customers discover anything and everything about us. and our new primary partner, [2:37] one of our new primary partners that by the way was an old primary partner are publications. Okay, [2:44] so I'm Peter Slaughterike. Um I'm the chief marketing officer of Gist. We are an AEO GEO [2:50] platform. Um and one of the things when I got invited to speak here, I started thinking about [2:55] like why does a marketing story belong at an AI conference like this? Um, everyone here builds [3:01] funds or deploys systems that answer questions. And that's one half of the question and answer [3:06] economy. We spend a lot of time talking about the question and answer economy. And if we go back to [3:12] my example, the bank did everything right and the system still got it wrong because of what it was [3:17] reading. So we track and measure and analyze and then make recommendations about how to [3:22] improve brands, publishers, creators, and agencies visibility in AI. If we start with publishers as [3:29] one of our primary partners, attribution has to be the start of our conversation. Citations matter. [3:36] Where the answer comes from matters because as you've all heard for years and years and years, [3:42] the input is only as great as the output. That certainly applies here. In the first four [3:47] months of 2026, about 68% of Google searches ended without a single click. Two years ago, it was 18%. [3:57] That's the fastest acceleration we've ever seen, at least in the last decade plus. And it's insane [4:03] to think that we've gone from 18 to 68%. We've never experienced this pace of change, and it is [4:10] testing our resilience as marketers. Something to think about here. Globally, publishers with a,000 [4:17] to 10,000 daily page views have lost roughly 60% of their search referral traffic in two years. [4:24] midsize roughly 47% and the giants about 22%. These are big numbers when you're thinking about [4:32] how publishers monetize. They monetize their traffic. They are selling advertising against [4:36] the folks that come to their website. With that traffic declining, we have to figure out a new [4:41] model. AI chatbots account for just under 1% of publisher page views today. That status from last [4:49] week, it's probably 1.5% today. 57% of all web traffic is automated. What's important to note [4:58] here is that AI is the fastest growing slice of that traffic. But together, those two numbers [5:04] actually tell one coherent story. And now you guys are seeing my entire script on the screen there. [5:12] Pardon me. Gentlemen, can we do not presenter view up there? Oh, we'd have to do off my laptop. Well, [5:21] then you guys are going to watch what I'm saying. Here we go. All right. Um, so put together, [5:26] those numbers actually tell one coherent story. The systems combine the publisher content at [5:31] massive scale while sending almost no readers back to the publisher sites. That asymmetry is [5:36] often called the crawl to referral imbalance by folks like Clyde Cloudfare. And so the traffic [5:41] that the publishers used to monetize um isn't there anymore. It's not the same kind of traffic. [5:48] And so we find ourselves in a position where we have to look at the history to figure out what [5:52] to do next. What where are we going to go? How are we going to get there? The click was not a metric. [5:57] It was our form of payment. And so the entire open web was built around the click. Right? We had this [6:04] idea that almost every uh fact you could come across that you've asked an LLM about recently [6:10] was somewhere way upstream. It was monetized by receiving a visitor by that website receiving a [6:17] visitor and the research tells us that the overlap between what ranks in Google and what's get what [6:23] gets cited by AI has fallen dramatically. So that's all great and good, but what's [6:29] more interesting to me is what the cited sources actually are. We call this a visibility problem. [6:36] Marketers talk about it as visibility. Um, get cited, show up, optimize for the model. not wrong, [6:41] just one very specific view. So what's an answer made of? If an answer is 10 blue links anymore, [6:48] if it's not what we used are used to, um what ingredients come together to make it? We start [6:54] with query, we go to retrieval set, and we get to the answer. Now that retrieval set is the kinds of [7:01] information that we're talking about here, um is usually coming from the easiest path to an answer, [7:06] right? The LLMs are always going to take the easiest path to an answer. thinking about [7:10] how did I get there? Where did the content come from? That's where the special sauce [7:14] is. And it's where I'm specifically focused when thinking about how are we optimizing visibility [7:19] uh in AI. So we did a quick research project. We looked at 48 sorry we asked the same question 48 [7:27] times. We asked multiple variants of that question every which way we could think of. 760 questions [7:32] um across technology, health and finance. Those queries produced over 1.5 million citations. [7:38] significant number, but every answer cited 16 to 18 sources. It's quite a few honestly when you're [7:46] thinking about compiling an answer. So, it's not a small factor, but across all 48 answers, [7:51] there were 23 documents total, 12 of them appeared every single time. So, we analyzed [7:57] the 1.5 million 1.58 million citations. We saw that 36 to 56 [laughter] of them. Uh, sorry, [8:05] let me go back. for any single question on chat chat GBT uh they only ever cited between 36 to 56 [8:11] documents total no matter the phrasing language or context 10 of those documents provided twothirds [8:16] of every single answer just think about that 760 million sorry 760 questions thousands of [8:22] variations millions of citations and yet most of the answers only came from 10 documents across [8:28] the topics of tech health and finance one of the documents behind every single one of those answers [8:33] is the Federal Reserve PDF from January 209 Things are different today than they were in 2009. We all [8:40] remember the great old Kanye stealing the moment. But when you think about where we're capturing our [8:45] information going forward, that's when things get really funky, right? [snorts] So, five out of 10, [8:51] we couldn't even date. We couldn't age them at all. We couldn't figure out when they were [8:55] from. And that makes me nervous. Are they from 20 years ago? Are they from 15 years ago? Are they [8:59] from yesterday? can promise you they're probably not from yesterday because the answer layer is no [9:06] longer sorry the answer layer is something really really small undated sticky and neglected. We're [9:15] not maintaining our sources of information. If you go back to the bank example at the beginning [9:20] that article was from six five years ago and we're not maintaining that information. It's not [9:26] uh common place for marketers to think, oh, I need to go back to my um quotes from 5 years [9:31] ago that were covered by Forbes to make sure that I'm cited correctly. And so, we have a [9:36] lot of work to do to figure out where the answers are coming from, where the information is cited, [9:40] and then what we need to correct, what we need to fix. So, there's four groups that [9:44] I'm thinking about. Brands are the first. The top of the funnel has lost its top. What are [9:50] we going to do? 25 years of regular marketing tactics that have that have become tried, true, [9:56] and tested. Uh we understand what those are going to deliver for us. Well, a lot of those are kind [10:01] of out the window now and we've got to do some reinvention. So, we have to think about what what [10:06] does it mean now if the funnel has lost its top? Publishers, they lost the transaction. Never mind [10:11] to the traffic. They lost the monetization, they lost the transaction. And by the way, for brands, [10:17] we lost that audience, right? If I don't have the traffic to reach as a brand advertiser, we've got [10:21] a problem. Then we think about creators. Creators, they have become this raw material for the LLMs. [10:28] They are inputs. They are absolutely getting used but not necessarily noted, not necessarily [10:35] cited. And the agencies, agencies are a really important piece of this puzzle because agencies [10:40] would go to their clients and say, "This is going to deliver fantastic traffic. This idea is going [10:44] to get you all sorts of engagement." Well, that's not necessarily the goal anymore. The goal is how [10:50] do I build authority for my brand around the subjects that matter to me? So, [snorts] how [10:57] do we fix it? How do we solve it? What is it going to look like? If we think about the value, [11:02] the value chain, someone finds something out, publishers then write about that idea. So, [11:07] that's new discovery, new information, and the traffic then pays for that and the models [11:13] learn from it. But what happens when the traffic paying for it breaks? When that's no longer there, [11:18] the models don't have as much to learn from. Now, I can't look ahead and say that in 2029 the models [11:24] are going to start to degrade. I no one can show you that chart today. I'm a little nervous [11:30] about if we don't fix this value chain that the information that the LLMs have to pull from is [11:35] going to start to degrade. At the very at the very smallest, uh, it's going to start to diminish. [11:41] And so we have this unpriced dependency. We've built uh an industry that has this piece of the [11:47] puzzle that's kind of missing. Um and I'd like to suggest that the vacuum that it created does not [11:53] have to stay empty. One in 10,000 citations in the research that we did was from sponsored content. [12:02] Anybody heard of an advvertorial? Right. So we're thinking about advertorials. It's a longheld [12:07] uh piece of the publishing puzzle, but I think it's going to become a lot more important really, [12:12] really quickly. And so, here's something to think about. The disclosure stays on the page. One in [12:19] 10,000 citations came from sponsored material. And when I say the disclosure stays on the page, [12:25] from a user perspective, from a human point of view, I can see that the con content is sponsored. [12:30] We've got a every publisher does it a little bit differently. Maybe there's just a quick asterisk. [12:34] Maybe it's called presented by, supported by, something to that effect. [snorts] But the [12:40] LLMs see something completely different. Most of the publishers have figured out to structure the [12:45] view for the LLMs in the markdown. We don't call it sponsored. It's not relevant to tell the LLM [12:51] that it's sponsored. It's just content. It's just information. So the advertiser who paid for that [12:56] advertorial gets to borrow from the authority of that publication. Authority starts to become the [13:02] commodity. it starts to become the thing we can monetize. And so 90-day plan, I'm a big fan of a [13:10] 90-day plan. 30 feels too short, 60 is not long enough, and 90 days we can really get into the [13:15] meat of it. And so the biggest thing here, going back to the beginning, is visibility. You have [13:20] to know what the problem is in order to solve it, right? So you can't fix what you can't see. [13:25] Brands, the very very first thing you have to do is find your baseline. Understand the queries that [13:31] you want to show up for. Where is the information that builds those answers? Where does it live? Is [13:38] it on your own websites? Is it on Reddit? Is it on YouTube? Is it on publishers websites? Is it [13:44] in microfich from millions of years ago? Right? You have to find out where that information lives. [13:51] Then you have to figure out if it's accurate. You have to start thinking about, okay, is [13:55] that still telling the right story? If it's from 10 years ago, what's changed about your brand? [13:59] What's changed about your business? What's changed about your rules? Customer service, the thing you [14:03] deliver. You have a lot of pieces of the puzzle to think about. Publishers, authority is the new [14:09] commodity. What is the value of authority when I, as a brand advertiser, am trying to become the [14:16] authority in the eyes of the LLM on a particular topic? Can I pay to borrow the authority of an [14:23] of a publisher? Let's say it's Forbes. Let's say it's people. I'm in entertainment and I want to [14:27] buy an advertorial in Entertainment Weekly because I want my movie to be the top movie of the summer. [14:33] That's what that value chain starts to look like. Creators, you have to own the destination. Where [14:39] does your content live? Is it easily accessible? If it's not behind a payw wall or a login, [14:44] the LLMs can get to it. And they're not under any obligation right now to site you, to credit you, [14:49] to say that it came from your social channels, from the YouTube video that you posted. So, [14:54] you have to figure out the balance between visibility and owning your audience in a way [15:00] that still allows you to monetize that um in a number of different ways. Agencies, we've got a [15:06] totally different value prop. What is the thing you're creating, the strategy you're building, [15:10] the creative you're delivering, what is that going to be worth in terms of authority building? Is it [15:17] going to create a number of pieces of earned media? Great. Fantastic. That's authority [15:21] building. Um, so instead of thinking about pricing it as traffic or the engagement that you're going [15:26] to get, you have to start thinking about things a bit differently and repric it before you're even [15:30] asked to do so in the context of authority. That's my suggested 90-day plan. It's going [15:36] to take a minute to get these answers and for us all to get some common language as we've watched [15:41] over the last couple of years. Uh, I think what I find really interesting is that the language comes [15:46] together. We all have to try a bunch of different things. Are we talking about AEO or GEO? Are we [15:50] talking about visibility? Um, authority I think is going to be that next thing that we have to really [15:55] be thinking quietly about and then start talking with one another about how are we pricing that? [16:00] What is the value of that? How do we get access to it? So, back in 2021, that article that I started [16:06] with cited a bunch of other banks, but not the the largest bank that was available. They went [16:15] and changed it. Now, they're in the answer. That whole example happened two years agoish and now [16:23] it's that article is still today the single site citation for the answer to that question across [16:30] four of the LLMs. So things change and then they stay the same, right? So once you make [16:35] the correction, hopefully that sticks around for a little while. Thank you very much for listening [16:40] to my talk and seeing all of my words up screen. I really appreciate it. Have a wonderful evening.