TL;DR: Google’s click-through rate dropped from 70% in 2013 to under 35% in 2025. The ranking game is over. The citation game has begun. Businesses that treat AI content generation and authority building as their primary SEO investment will be referenced by AI engines. Those that don’t will be invisible.
The Pulse:
- Google click-through rates dropped from 70% in 2013 to under 35% in 2025, per Gareth Bull SEO, driven by AI overviews and knowledge panels designed to answer rather than refer.
- LLMs operate via Retrieval Augmented Generation (RAG): they do not crawl the web like Google does; they generate answers from trained parameters and pull fresh documents before responding, compressing knowledge rather than ranking pages.
- Google’s own data confirms search demand is increasing in 2025, yet site traffic is harder to earn than ever, creating an authority-building gap that most businesses have not yet closed.
The friction here is precise: search volume is growing while click-through rates are collapsing. More queries enter the system; fewer exit as website visits. As Gareth Bull of Gareth Bull SEO frames it, the question has shifted from “who ranks” to “who gets referenced.” That structural inversion is the central challenge for every practitioner building an AEO strategy or GEO optimization program today.
From PageRank to AI Inference: A 25-Year Architecture Shift
The foundational mechanism of search changed twice in 25 years: first from directory noise to link-graph authority in 1999, then from keyword matching to AI inference in 2023-2025. Understanding both transitions is the prerequisite for any credible SEO optimization program in the current environment. Each shift rendered the prior playbook obsolete within months, not years.
In 1999, Larry Page and Sergey Brin introduced PageRank at Stanford. The core insight was architectural: rank pages by the quality of inbound connections, not keyword density. For the first time, a machine was not just reading the web. It was modeling trust relationships across it. That single mechanism displaced directories, banner-ad dominance, and keyword-shouting overnight.
The algorithm then required continuous refinement. The 2011 Panda update elevated content quality as a primary ranking signal. Thin pages and keyword-stuffed blogs vanished from results. Google began rewarding original insights, depth, and readability. SEO shifted from system exploitation to trust-earning, a mindset change that most practitioners underestimated at the time.
A year later, Penguin targeted link spam with equal force. Sites that had scaled on shady link networks collapsed. Businesses that had grown headcount and infrastructure to service the traffic surge from those links lost their revenue base almost overnight. The lesson: quality over quantity was not a preference. It was an enforcement mechanism.
The Real Takeaway: PageRank’s link-graph model held for over a decade, but the Panda and Penguin updates compressed the quality bar so sharply that by 2015, machine learning via RankBrain was the only viable path forward for Google to handle query intent at scale.
Google’s PageRank launched in 1999 and dominated search architecture for over a decade. The 2011 Panda update made content quality a primary ranking signal, and the 2012 Penguin update eliminated link-spam networks. By 2015, RankBrain introduced machine learning to interpret query intent rather than match exact keywords. Each update compressed the window for low-quality tactics from years to months.
How LLMs Work Differently From Google: The RAG Mechanism
LLMs do not crawl or index the web the way Google does. They generate answers by mapping tokens through billions of trained parameters, then augment those answers with fresh web documents via a process called Retrieval Augmented Generation (RAG). This architectural difference determines why ranking in Google and earning a citation from an LLM require overlapping but distinct content strategies.
Google’s system is retrieval-first: crawl, index, rank. A query triggers a lookup across a pre-built index, weighted by hundreds of ranking factors including user signals, backlink authority, and topical relevance. The output is a ranked list. LLMs invert this. They are generative-first: a query is broken into tokens, mapped through a parameter space trained on vast corpora, and the system predicts the most probable accurate answer. RAG then pulls live documents to ground that prediction in current information.
The practical consequence is significant. An LLM does not show ten blue links. It compresses knowledge into a single synthesized answer and selects which sources to credit. Your content must be structured to survive that compression. Schema markup, verified authorship, clear entity definitions, and transparent sourcing are the signals an LLM uses to decide whether your content is a trustworthy input to its answer or noise to be filtered out.
Compare this to how OpenAI’s ChatGPT (GPT-4o), Anthropic’s Claude 3.5 Sonnet, and Google’s Gemini 1.5 Pro handle the same query. All three use some variant of RAG to pull fresh web content, but their training data cutoffs, context windows (ranging from 128K tokens for GPT-4o to 1M tokens for Gemini 1.5 Pro), and retrieval architectures differ. A piece of content optimized for AI citation needs to be machine-readable across all three inference environments, not just Google’s crawler.
What This Means in Practice: The shift from a ranked list to a synthesized answer means that expert articles with clear entity structure and verified authorship are the primary currency of AI content generation at scale. A generic page optimized only for keyword density will be filtered out of RAG retrieval before it ever reaches a user.
| The Conventional Approach | The Yacov Avrahamov Perspective |
|---|---|
| Optimize for keyword density and exact-match phrases | Optimize for topical authority and entity clarity so RAG systems can extract and credit your content |
| Measure success by ranking position and organic traffic volume | Measure success by AI citation frequency and ChatGPT citations alongside traditional click-through data |
| Build backlinks through volume-based outreach campaigns | Earn editorial backlinks from trusted sources as trust signals for both Google and LLM retrieval |
| Treat schema markup and authorship as optional enhancements | Treat schema markup, verified authorship, and transparent sourcing as infrastructure for machine readability |
| Publish content formatted for human readers only | Engineer thought leadership content readable by both humans and AI inference engines simultaneously |
LLMs use Retrieval Augmented Generation (RAG) to ground generative answers in live web documents. Unlike Google’s crawl-index-rank architecture, LLMs compress knowledge into a single synthesized answer and select sources to credit. Content must be structured with schema markup, verified authorship, and entity clarity to survive RAG filtering across systems like GPT-4o (128K context window) and Gemini 1.5 Pro (1M context window).
The Citation Economy: What Google Actually Values in 2025
Google’s 2025 ranking signals still include keyword relevance, backlink authority, and site speed, but the dominant new layer is credibility as a trusted data point for AI systems. The businesses that win are those whose content is easy to read for machines, not just humans. That requires a deliberate AEO strategy built around topical relevance, clear structure, and verifiable expertise.
As Gareth Bull of Gareth Bull SEO notes, even backlinks have evolved. An LLM does not just see a link as a vote. It reads a link as context: a signal of expertise and trust from the linking entity. The algorithm no longer just ranks pages. It ranks perspectives, deciding which ideas and opinions it trusts most before presenting them to users. That is a different optimization target than a PageRank score.
The uncomfortable question this raises is real. If AI systems are both generating content online and judging the quality of content to train on, the E-E-A-T framework (Experience, Expertise, Authoritativeness, Trustworthiness) becomes a machine-readable specification, not just a Google quality guideline. Verified authorship, transparent sourcing, and schema markup are the implementation layer of that specification. Skipping them means your content is invisible to the systems that now mediate most information discovery.
One critical nuance: for most commercial searches in 2025, AI overviews have not yet displaced traditional search results. Those queries are still handled by conventional ranking mechanisms. The GEO optimization and AEO strategy investments you make now are positioning for the next phase of displacement, not a response to a completed transition. The window to build citation authority before commercial queries fully migrate to AI-generated answers is open, but it is narrowing.
The Strategic Implication: Every credible piece of expert content published today becomes fuel for AI training and RAG retrieval tomorrow. The opportunity is asymmetric: brands that invest in authority building now will accumulate citation equity that compounds as AI search handles a larger share of commercial queries.
Google click-through rates dropped from 70% in 2013 to under 35% in 2025, driven by AI overviews and knowledge panels. For most commercial searches in 2025, AI overviews have not yet replaced traditional results, meaning the window to build citation authority is open but closing. Google’s own data shows search demand increasing as AI lowers the friction of query formulation, yet fewer queries result in website visits.
Building Authority in the AI Era: The Execution Framework
The execution framework for AI-era SEO optimization combines three layers: machine-readable content architecture, topical authority depth, and verified credibility signals. Each layer maps directly to how RAG systems select sources. Treating any one layer as optional produces content that ranks but does not get cited, or gets cited but cannot be traced back to your brand.
Machine readability is the foundation. Schema markup tells inference engines what type of entity your content represents, who authored it, and what claims it makes. Verified authorship connects your content to a real expert identity that AI systems can cross-reference. Transparent sourcing signals that your claims are grounded in evidence, not generated noise. These are not technical niceties. They are the admission criteria for RAG retrieval.
Topical authority depth means covering a subject comprehensively enough that an LLM treats your domain as a reliable source across multiple related queries. A single well-optimized article earns one citation opportunity. A structured content program across a topic cluster earns recurring citation across the full query space. This is where content marketing automation becomes a force multiplier: publishing 30 expert articles in the time it previously took to produce one changes the topical coverage equation entirely.
Credibility signals extend beyond your own site. Editorial backlinks from trusted sources function as trust signals for both Google’s ranking algorithm and LLM retrieval systems. The mechanism is the same: a trusted entity linking to your content is a machine-readable endorsement of your perspective. Building these signals through genuine expertise, not volume-based link schemes, is the only durable approach in an environment where both Google and AI systems are actively filtering low-quality inputs.
Why This Matters Now: AI-powered SEO is not a future consideration. LLMs are already pulling from live web content via RAG on every query. Brands that publish authoritative, machine-readable thought leadership content today are building the citation equity that determines AI referral traffic in 2026 and beyond.
Frequently Asked Questions
Does traditional keyword research still matter in 2025?
Keywords remain a relevant input, but their function has changed. RankBrain, introduced in 2015, shifted Google from exact-match keyword processing to intent interpretation. In 2025, keywords signal topical relevance to both Google’s ranking system and LLM retrieval. The operational difference: target keyword clusters that map to a coherent topical authority domain, not isolated high-volume terms. A single keyword-stuffed page earns neither a ranking nor a citation.
How does the MUM (Multitask Unified Model) update affect content format requirements?
Google’s MUM update enabled the system to connect a YouTube video, a blog post, and a forum thread about the same topic into a unified understanding. This means multi-dimensional content backed by visuals, expertise, and credibility across formats earns stronger topical authority signals than text-only pages. For practitioners building an AEO strategy, this means coordinating written expert articles with video and structured data across platforms, not just optimizing a single content type.
If users never visit my site because AI answers their query directly, how do I measure ROI?
This is the core operational challenge of the citation economy. The measurement framework needs to expand beyond organic traffic and ranking position to include AI citation frequency, brand mention volume in AI-generated answers, and referral traffic specifically from AI platforms. Tools that monitor ChatGPT citations and LLM brand mentions are emerging as the analytics layer for this new environment. Traffic metrics alone will increasingly undercount the actual authority value your content generates.
What is the risk of AI-generated content cannibalizing citation authority?
Gareth Bull raises this directly: if AI systems are generating most online content and also judging content quality, the trust signal chain becomes circular. The practical defense is verified authorship and transparent sourcing. Content attributed to a real, credentialed expert with a traceable publication history is structurally differentiated from anonymous AI output. This is why E-E-A-T implementation, specifically the Experience and Expertise components, is the most durable investment in the current environment.
How should I prioritize between Google SEO optimization and LLM citation optimization?
For most commercial searches in 2025, traditional Google results still handle the query. The two systems are not yet fully merged, and ranking in Google and earning AI citations currently go hand in hand. The practical priority: build content that satisfies Google’s quality signals (topical relevance, credibility, machine-readable structure) first. That same content is simultaneously the highest-quality input for RAG retrieval. The architectures overlap enough that a single well-executed content program serves both, provided it is built around genuine expertise rather than keyword volume.
Build the Citation Authority Your Competitors Are Missing
AuthorityRank engineers expert articles at scale, structured for both Google ranking and LLM citation. Publish 30 authoritative pieces in the time it takes most teams to produce one, and become the trusted data point AI engines reference by default.