The fall of blue links and the rise of GEO.

The blue link is not dying because users got bored with it. It is dying because the page underneath it kept growing. Universal Search arrived in 2007 and pulled images, news, and video onto the results page. Knowledge Graph arrived in 2012 and pulled facts onto the results page. Featured Snippets, People Also Ask, Local Packs, Shopping Carousels, and AI Overviews each pulled something else onto the results page. By the time generative AI showed up, the answer had already moved out of the link and into the SERP itself.

What changed in the last two years is the speed of that shift, and what it means for any business whose pipeline depends on being found.

The data behind the shift

Two numbers do most of the work in this story.

50%

Of Google searches ended without a single click by 2019 — users got their answer inside the SERP itself.

SparkToro, 2019

−25%

Forecast drop in traditional search volume by 2026 as AI chatbots absorb informational queries.

Gartner

The SparkToro figure is now seven years old, and the trend it captured has only sharpened. The Gartner forecast came out before generative AI was a consumer product. Neither number assumes the existence of ChatGPT, Perplexity, Claude, or Gemini. They describe the gravitational pull of the SERP itself before AI accelerated it.

The third number is the one that matters most to operators. When an AI Overview cites a source, it usually pulls from a deep page, not a homepage. That has direct implications for how content is structured and where authority needs to sit. A polished homepage that ranks for your brand name does not protect you if your supporting pages are thin, vague, or missing.

>80%

Of AI Overview citations point to deep pages, not homepages — supporting content does the load-bearing work.

iPullRank

Three eras of search, compressed

It is useful to mark the three eras the search system passed through, because each one changed what optimization meant.

Each era kept some of the mechanics of the era before it. Crawlability still matters. Structured data still matters. Backlinks still matter, but their role is narrower. What is new in era three is that the surface of the answer is generated, not selected, and the rules of inclusion are different.

What GEO actually is

Generative Engine Optimization is the practice of getting your brand named, cited, or recommended inside answers produced by AI engines. The framework, as articulated by Mike King of iPullRank, names three core requirements:

• Semantic clarity. Content has to be unambiguous at the passage level. A model deciding which sentence to lift cannot guess.
• Authority signals. Third-party validation and trustworthy citation chains, not domain authority alone.
• Multimodal accessibility. Models pull from text, images, video transcripts, tables, PDFs, and code, in any combination.

That last requirement is the one most businesses underestimate. A YouTube transcript can carry your brand into a Perplexity answer. A diagram with a clear caption can be parsed into a structured fact. A PDF on a partner site can outrank your own marketing copy as a source.

Relevance Engineering: the broader frame

King uses a second term that we think is worth adopting: Relevance Engineering, sometimes shortened to r19g. He defines it as a channel-agnostic discipline that borrows from information retrieval, machine learning, UX, content strategy, and digital PR.

The reason the term is useful is that it lifts the work above any single engine. Optimizing for ChatGPT today and Perplexity tomorrow, with each engine quietly adjusting its retrieval pipeline, is a losing game if it is the whole strategy. Relevance engineering treats those engines as instances of a more general system: a retrieval-augmented pipeline that selects passages, ranks them by vector similarity, evaluates them for trustworthiness, and composes an answer.

If you build for the underlying system, the brand-specific tactics fall out of it. If you build for one engine, you are constantly retooling.

Why the blue link broke down

The intellectual case for the shift was made earlier than most people realize. In 2023, Andrei Broder and Preston McAfee of Google Research published Delphic Costs and Benefits in Web Search, which framed search as a transaction with three hidden costs: access cost, cognitive cost, and time cost. Every blue link a user has to click and read is a cost. Every page they have to interpret and discard is a cost.

The paper made the case that search engines compete on lowering those costs. AI Overviews and chatbot answers are the natural endpoint of that competition. They collapse access cost (one answer, no clicking), cognitive cost (synthesized prose, not ten conflicting blogs), and time cost (under five seconds for most queries) at once.

That is why AI search is not a fad. It is the cost-minimizing endpoint of a trajectory that has been running for fifteen years.

The infrastructure layer underneath all of it

The next layer is starting to surface, and it is worth naming because it changes how content gets selected. The Model Context Protocol, introduced by Anthropic and now adopted across the major AI vendors, lets agents delegate tasks to other agents and connect to external data sources during a single request. Instead of one prompt producing one answer, an agent can decompose the question, fan it out across specialized retrieval services, and assemble the result.

Mike King has written extensively about query fan-out as the mechanism behind Google's AI Mode. The idea is the same in MCP-style architectures: a user-facing query is rewritten into many synthetic subqueries, each retrieved against a different corpus, with results merged at the top.

The implication for content owners is that being mentioned once on your homepage is no longer enough. The retrieval system is going to ask for evidence at the passage level, multiple times, from multiple angles, before it composes the answer.

What this means for measurement

This is the part most businesses are getting wrong. The instinct, when AI search shows up, is to bolt an "AI rank tracker" onto the existing dashboard and call the problem solved. That misses the structural change.

You cannot measure GEO with the same instruments you used for SEO, for three reasons.

• The unit of analysis is different. SEO tracks rank position for a query. GEO tracks whether your brand was mentioned, recommended, or cited inside an answer. A page that ranks #1 on Google can be absent from the AI answer for the same query.
• The output is non-deterministic. Two users asking the same question can get different answers, because the model personalizes retrieval. The right metric is presence rate across many captures, not position on a single SERP.
• The signals are distributed. Your performance depends on which third-party sources each engine trusts in your category, and on how your passages survive vector retrieval. Backlink counts and keyword volume do not capture either of those.

A defensible GEO measurement program tracks engine-level visibility (presence rate by engine), recommendation rate (how often you are actively recommended, not just mentioned), share of model (your slice of mentions vs. competitors), and source tier coverage (whether you appear on the editorial sites each engine pulls from). Position weighting matters too: being the first brand named in a list is worth more than being the fifth.

What to do this quarter

None of this means tearing up the SEO program. The page that ranks well on Google is also the page an AI engine is more likely to find, parse, and quote. Crawlability, structured data, canonical URLs, fast load times. Keep all of it.

What changes is the work you add on top.

1. Audit your passage-level structure. Are your most important claims self-contained inside a single paragraph? Or are they distributed across three sections that only make sense together? Models extract passages, not pages.
2. Map the gatekeeper sources in your category. Which third-party editorial sites does ChatGPT cite when answering "best [your category] for [use case]"? Which subreddits does Perplexity rely on? You probably know your top SEO referring domains. You almost certainly do not know your top AI source pool.
3. Measure presence, not position. Run a structured capture across multiple engines and prompts on a recurring schedule. Look at recommendation rate, share of model, and the gap between mentioned and recommended. This is what BeCited audits do, and it is what every operator we talk to says they wish they had built sooner.
4. Treat each engine as a separate market. ChatGPT and Perplexity share roughly 11% of cited domains. Gemini favors sources that already rank in Google Search. The strategy that wins one engine does not automatically win the others.

The fall of the blue link is not an event that will finish. It is a slow rebalancing of attention, citation, and trust away from the ranked list and toward the synthesized answer. The brands that read the shift early and instrument for it are the ones who get cited. The rest get described by their competitors.

Frequently asked questions