Relevance Engineering: SEO is dead, long live relevance

The click on which 25 years of SEO were built is disappearing right now. Not slowly, not ‘one day’. Right now.

That’s a bold claim, so here’s the evidence straight away. A field study reported on by Search Engine Journal reported on in autumn 2025 found that a Google AI Overview reduces organic clicks on the relevant search queries by around 38 per cent. The proportion of searches ending without a single click rose from 54 to 72 per cent in the same study. Ahrefs reported at the end of 2025 that the click-through rate for the number-one result plummets by up to 58 per cent as soon as an AI response appears above it. And in Google’s new AI Mode – a pure chat interface – several analyses suggest that over 90 per cent of queries end without anyone visiting a website.

Anyone who continues to set ‘ranking number 1’ as their goal in this environment is optimising for a stage from which the audience is currently walking away. The good news is that a new discipline is taking the place of the old one, and it even has a name already. Relevance Engineering.

First, the uncomfortable truth: The click is dying, but visibility isn’t

Before this sounds like the end of the world, there’s an important distinction to make. What’s dying is the click as the sole currency. Visibility isn’t dying. It’s shifting.

It used to be a simple equation: you rank for a search query, the user clicks, you get traffic, and that traffic turns into enquiries and revenue. This chain worked for 25 years, and an entire profession was built upon it. AIsearch breaks the chain at the second stage. Google, ChatGPT and Perplexity answer the question themselves, directly within the interface, and at best cite you as a source.

This is no longer a marginal phenomenon. An analysis of over 21 million search queries showed that in the first quarter of 2026, around a quarter of all Google searches triggered an AI Overview. For B2B technology queries, precisely the terrain in which SMEs and their buyers operate, the rate stood at over 80 per cent. In Germany, Austria and Switzerland, AI Overviews have been rolled out since 26 March 2025, and AI Mode since 7 October 2025. This isn’t some futuristic US scenario; it’s the SERP your customers are seeing today.

The crucial question has therefore shifted. It is no longer “How do I get to number 1?”, but “How do I become part of the answer?”. And that is a fundamentally different challenge.

What Relevance Engineering actually is

The term was coined by Mike King, founder of the US agency iPullRank and named “Search Marketer of the Year” in 2025. King was one of the people who brought the major Google Search leak to light, so he has an unusually detailed understanding of how the engine works from the inside. His thesis: SEO in the traditional sense is over, and what is coming is no longer a marketing tactic, but a technical discipline.

Relevance Engineering is the intersection of five fields that used to be separate departments: machine learning, information retrieval, content strategy, digital PR and user experience. The essence in a single sentence: you no longer structure your content so that it ranks for a keyword, but so that an AI uses it as relevant, trustworthy building blocks for its answers .

In short: traditional SEO optimised pages for a ranking list. Relevance Engineering optimises passages for a synthesis.

The difference may sound like semantics, but it is technical in nature, and this is precisely where it becomes interesting for anyone who understands how these systems work. Let’s take a look at the machine.

How AI search really works: fan-out, vectors, passages

When you type a question into AI Mode, Google doesn’t just search once. It searches eight to twelve times, in parallel. The technique is called Query Fan-Out: A specially trained Gemini model breaks your question down into several sub-questions, each covering a different aspect of your intent, sends them off simultaneously and combines the results into a single answer. For complex searches, in Google’s ‘Deep Search’, this can amount to hundreds of sub-queries. Elizabeth Reid, Google’s Head of Search, has described precisely this process as the mechanism that makes AI Mode possible in the first place.

An example from everyday B2B life brings this into focus. If a buyer types in “which shop system for B2B wholesale with ERP integration”, the system breaks this single question down into a set of sub-questions in the background: B2B features of shop systems, ERP interfaces in e-commerce, platform comparisons for wholesale, typical costs, migration effort. It answers each of these sub-questions separately and constructs a response from them. So your page doesn’t need to match the original question, but rather one of these sub-questions – of which the user never realises they were even asked.

For this to work, there is no longer any keyword matching running in the background, but rather a dense vector search. Every query, every sub-query, every document and every passage is translated into a vector – a long sequence of numbers that encodes its meaning. The system then searches not for matching words, but for spatial proximity between these sequences of numbers. Two texts that mean the same thing, without using a single word in common, lie close together in this space.

In simple terms: the machine has stopped searching for terms and started searching for meaning. Your keyword density tool is optimising for a game that is no longer being played.

And it doesn’t search for pages. It searches for passages. For years now, Google has been ranking individual sections of a page independently of the rest. AI systems take this to the extreme: they extract individual chunks of text, group their vectors into thematic clusters, and, for each aspect of the answer, select the chunk that most closely matches the respective sub-query. Your carefully crafted 2,000-word page is broken down into dozens of individual parts, and each part competes on its own merits.

This has a drastic consequence for the way we write.

From keyword to passage: your page is no longer the unit

If AI thinks in chunks, you must write in chunks. A long, meandering paragraph that only gets to the point in the fifth sentence and relies on context from the previous paragraph is worthless for passage extraction. If the system takes it out of context, it makes no sense, and a chunk that makes no sense will not be cited.

What works instead are self-contained sections. A question-oriented heading, followed by a block of around 130 to 160 words that fully answers that very question, with its own evidence, without relying on knowledge from the previous paragraph. It is no coincidence that AI Overviews usually structure their answers to this exact length. This is the length at which a thought is both complete and portable.

This is the tricky bit for good writers: narrative, interlinked texts – the craft of which many are proud – perform less well in the ‘chunk’ world than modular, almost lexical structure. It’s not about writing worse. It’s about structuring each section so that it can stand alone whilst still fitting into the whole. For many editorial teams, this is a bigger shift than any technical measure.

Information Gain: Why ‘also quite good’ is now invisible

Here comes the lever that hurts the most. AI systems favour content that contributes something new. Google even holds a patent for this: “Contextual Estimation of Link Information Gain”, filed in 2018, granted in June 2024. It describes a mechanism that measures how much additional information a document provides, compared to what the AI already knows from other sources. Google does not confirm whether it actively uses this particular method. However, the principle aligns exactly with what the Helpful Content logic and AI Overviews reward.

The consequence is stark. The tenth blog post listing “the 7 benefits of headless commerce” in exactly the same order as the nine before it has an Information Gain of practically zero. The machine already has this information. It doesn’t need you as an eleventh confirmation. It cites the source that adds something: its own measurement, a counter-argument, a detail from real-world project experience, a figure that no one else has.

In the old days of SEO, you could reach page 1 with a solid, well-optimised copy of the consensus content. In Relevance Engineering, copying the consensus is the opposite of relevant. Originality is no longer just a stylistic device, but a ranking factor.

Relevance is not just text: entities, authority, structure

A common misconception is to reduce Relevance Engineering to ‘writing better’. In fact, text is just one of three pillars.

The second is the entity. AI systems work with knowledge graphs; they understand the world as a network of things and their relationships, not as a list of strings. For a machine to recognise your company, your product or your area of expertise as a distinct entity at all, the signals must be consistent and available everywhere. Structured data via Schema.org is not just a nice extra here, but the language in which you tell the machine what the things on your page are. For an online shop, this means, in concrete terms: clean product data, consistent markup and a well-maintained feed. Anyone who has their product data is technically well-structured, provides the AI with the building blocks it needs to cite it.

The third pillar is authority, and this is where the new is catching up with the old. AI systems assess whether a source is trustworthy, and they do so based on signals that SEO has always recognised: Who links to you, who mentions you, and in what context does your brand appear. Digital PR and the targeted generation of brand mentions are not a side issue in relevance engineering, but a core component. A brilliant, unique passage from a domain that the machine does not trust is cited less frequently than an average one from an authoritative source.

What this means in practice: Relevance Engineering in practice

Enough theory. What does the work look like when you take it seriously? The most honest way to show the difference is through a direct comparison.

This table sets out a working method, not a list of tricks. In concrete terms, Relevance Engineering means: you build every important page around a clear main question and several precise sub-questions, each with its own, independently readable answer section. You mark up entities and relationships with structured data so that the machine doesn’t have to guess. You ensure that every passage conveys at least one idea that hasn’t already been written by ten other people. And you no longer just measure your rankings, but whether you appear at all in the answers from Google AI Mode, ChatGPT and Perplexity. Checking whether AI can even find your website is the logical first step in assessing your current position.

A small, concrete example of the ‘entities’ pillar. An FAQ block marked up in this way can be directly adopted by an AI as a building block for an answer:

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "What is Relevance Engineering?",
    “acceptedAnswer”: {
 "@type": "Answer",
 "text": "Relevance engineering is the optimisation of content for AI search systems. Rather than optimising pages for ranking positions, individual passages are structured in such a way that AI models can use them as relevant building blocks for their generated answers."
    }
  }]
}

This is not magic, but discipline. This is precisely what the word ‘engineering’ means in relevance engineering: a repeatable, technically sound process rather than a collection of tricks.

What you are measuring now: The KPIs are shifting with

When the unit of optimisation changes, the dashboard changes too. And for many marketing managers, this is the most painful part, because the figures they’ve grown fond of become unreliable. If 90 per cent of AI-generated fashion queries end without a click, your sessions curve no longer measures your impact, but only the dwindling remainder that still clicks through. In future, a drop in traffic may mean that you have become more successful, because the AI displays your answer directly.

The new key question is: Share of Voice in AI responses. How often are you mentioned in the generated responses on your relevant topics, and what is your share compared to the competition? Alongside this come metrics that were virtually non-existent in traditional reporting: the frequency with which your brand is cited as a source; the recognition of your company as a distinct entity in the Knowledge Graph; and brand mentions without links, which still serve as a signal of authority for AI systems. A separate, rapidly growing category is referral traffic from the AI tools themselves: visitors who come not via Google, but directly from ChatGPT or Perplexity, because an answer containing your link appeared there.

In practical terms, this means: you need a monitoring system that regularly asks the major AI systems the same purchase-relevant questions and records whether and how you appear. This is not a ready-made, off-the-shelf tool, but a deliberate monitoring routine. Those who set this up early will see the shift in their own figures before it manifests as a dip in turnover. Those who continue to focus solely on sessions and rankings are flying blind through the very phase in which the rules of the game are being rewritten.

The counter-argument: Is SEO really dead?

Now for the honest part, because a provocative claim deserves honest scrutiny. ‘SEO is dead’ is an exaggerated headline. Anyone who takes it literally is mistaken, and for one simple reason: AI search systems do not possess their own truth. They synthesise information from the web. Without crawlable, indexed, structured content, the models would have nothing from which to build their answers. The technical foundation of traditional SEO – that is, clean crawlability, fast delivery, server-side rendering, and valid structured data – has not become any less important; rather, it is a prerequisite for your content to even be in the running.

It is therefore more realistic to say: SEO isn’t dying, it is being absorbed. Relevance Engineering encompasses traditional SEO as a technical subset and builds something new on top of it: namely, targeted work on meaning, passages, entities and authority. Anyone who hears ‘SEO is dead’ and neglects its technical foundations has completely misunderstood the argument. The old virtues remain essential. They’re just no longer enough.

And a second honest point: much of this is still in its infancy. The patents have been granted, the mechanisms documented, but exactly how the weightings work is a matter of observation and well-founded hypothesis – not set in stone. Anyone selling you an exact formula for ‘citation ranking’ today is selling you a certainty that doesn’t exist. The direction, on the other hand, is clear, and for the past two years it has consistently pointed in the same direction.

What you, as a decision-maker, should do now

It’s very tempting to treat all this as a tool problem and buy the next “AI SEO” subscription. That’s the wrong first step. The right one is a sober assessment of where you stand.

First, find out where you stand today in terms of AI responses, not rankings. Ask Google AI Mode, ChatGPT and Perplexity the questions your customers should be using to find you, and check whether and how you appear. Then take your most important pages and read them as a machine would: does each section contain a standalone, quotable answer, or is it all a narrative flow from which no clear chunk can be extracted? Check whether your entities are tagged with structured data. And be brutally honest when it comes to the ‘information gain’ question: what on your page does the machine not yet know?

For most medium-sized businesses, this isn’t a side project, but a strategic overhaul that brings together editorial, technical and PR teams – precisely the combination that King refers to as ‘engineering’. If you need support at this stage, this is one of the key areas of our work on SEO and GEO.

The sentence I started with is true: The click on which SEO was built is disappearing. But visibility isn’t disappearing. It’s simply being allocated in a new place, according to new rules. Anyone who understands these rules early on and structures their content accordingly won’t be a victim of this shift. They’ll be its winner. That is precisely what Relevance Engineering is.