Analysis · SEO 2026 & Algorithms
SEO in 2026: Who (or What) Are You Optimizing For?
Google stopped looking for pages that answer keywords. It looks for sources that probabilistic language models can synthesize as authoritative. The SEO market hasn’t noticed the system has changed.
A site with Domain Authority 85, thirteen years of accumulated backlinks, a solid editorial record. It loses 34% of its organic traffic in six months. Next to it, a six-month-old blog with no significant link profile climbs the SERPs on the same queries.
This data circulated repeatedly through 2026 SEO reports, in different versions but with the same core: something in how Google evaluates relevance stopped following the rules the market internalized over the past two decades.
The problem isn’t technical. It’s epistemological.
What Changed in the Google Algorithm and for SEO 2026, with AI Overview and SERP?
In 2019 Google integrated BERT into its ranking system. Bidirectional Encoder Representations from Transformers: a language model trained on encyclopedic-scale text corpora, able to read each word in the context of every other word in a sentence, simultaneously in both directions. It wasn’t a search-algorithm update in the sense the market understood until then: no adjustment of the weights assigned to backlinks, no penalty for keyword stuffing. Something structurally different, rather — the introduction of a system that comprehends the intent behind a query, not just the words that make it up.
Then, between 2023 and 2025, the next layer arrived. Google AI Overviews, rolled out globally across more than a hundred countries starting March 2025, introduced a second system that doesn’t merely classify pages but synthesizes them, deconstructs them, recomposes them into generated answers. The site that once served as a destination becomes a potential component from which information is extracted, without the user ever arriving there.
March 2026 brought what some analysts have already called the most volatile core update in Google’s history: 79.5% movement in Top-3 results, 24.1% of Top-10 pages vanishing entirely. Three signals were reclassified: original data, unique perspectives, paraphrased content. The third recorded minus 71% in traffic. It isn’t a penalty for spam. It’s a penalty for genericity.
| Reclassified signal (March 2026) | Traffic change |
|---|---|
| Original data and primary research | +22% visibility |
| Unique perspectives and original analysis | Gains even for low-authority domains |
| Paraphrased / commodity content | −71% traffic |
What BERT Actually Understands, and What It Only Predicts
There’s a detail in the standard description of BERT the SEO market tends to slide past without stopping. It’s said that BERT “understands natural language.” That’s a formulation Wittgenstein would have found problematic, and not for academic reasons. In the Philosophical Investigations, he argues that the meaning of a word doesn’t reside in the word itself but in the use made of it within a shared language game. BERT has access to none of these things: it has no experience of the world, participates in no social practices, belongs to no community.
What it does, with extraordinary sophistication, is model the probability that a sequence of words appears in certain contexts, produced by certain types of sources, in certain textual configurations. It doesn’t understand. It predicts. The distinction is practically crucial: optimizing for a system that predicts contextual plausibility requires resembling, in textual structure, the sources the system has learned to recognize as authoritative. Which is a completely different operation from logical clarity.
Paraphrased content has low semantic entropy. For a probabilistic system, the predictable is redundant. Redundancy contributes nothing to the estimation of authority.
Here enters information theory in the technical sense Shannon formalized in 1948. Content is informative to the degree it’s surprising: it adds something the system couldn’t already predict. The March 2026 core update, with its minus 71% on paraphrased content, isn’t a Google editorial choice. It’s the algorithm formalizing what Shannon said seventy-eight years ago: information is what surprises.
How Topical Authority Works in SEO 2026, and Why It Replaces Keyword Density
A domain doesn’t accumulate topical authority through how often it mentions a term. It builds it through the density and coherence of the semantic graph its content forms over time: how many related entities are covered, with what depth, with what internal interconnection, with what alignment to the expected response patterns of a specific query cluster.
It’s the shift from the document-keyword logic to the logic of the node in a network of meaning. A site with domain authority 85 producing generic content across twenty different topics has topical authority on none of them. A six-month-old blog covering a single domain with vertical semantic coherence, citing primary sources, building interlaced topic clusters, can climb the SERPs on that domain because the algorithm recognizes it as an epistemically dense node, not because it accumulated links. Classic SEO optimized one document for one term. SEO in 2026 builds an epistemic identity for a semantic domain.
What E-E-A-T Is and Why It’s a System of Distributed Epistemology
E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) is the framework Google’s human Quality Raters use to evaluate content, and those evaluations train the algorithmic systems over time. Google added the first “E” for Experience in December 2022, recognizing that direct first-hand experience of a topic is an epistemic signal distinct from theoretical expertise.
What makes E-E-A-T interesting isn’t the list of its components. It’s the underlying logic: Google is building a computational proxy for something human institutions have always done in slower, more expensive ways — evaluating a source’s credibility before using it as a reference. Universities do it through peer review. Journalism through source verification. Law through chains of evidence. Google does it with signals distributed across the web at scale: who cites whom, who gets cited by sources already recognized as authoritative, who produces content others find worth referencing. Trust, the fourth component, is explicitly defined by Google as the center of the framework. Not expertise, not authority: trust. Because trust epistemologically precedes every other judgment.
What GEO and AI Overview Are and How They Change SEO 2026 Ranking in the SERPs
Google AI Overviews appear today in a variable share of queries — between 4.5% and 12.5% depending on the study and market — concentrated sharply on informational queries. When an AI Overview is present, CTR at organic position one drops 58-61%. Around 60% of all Google searches now end without a click. 99.5% of the sources cited in AI Overviews come from the top ten organic results: traditional SEO remains the entry prerequisite.
GEO — Generative Engine Optimization — is the discipline that answers this scenario. It doesn’t replace traditional SEO: it layers on top of it. The goal isn’t only to rank at position one; it’s to be cited as a source within answers generated by AI systems, Google AI Overviews but also ChatGPT, Perplexity, Gemini. Optimization for citation requires information structure readable by retrieval systems, verifiable and independently citable data points, direct answers in the first 100 words of each section. 44.2% of AI Overview citations come from the first 30% of the content.
How to Do SEO in 2026: Strategies and Practices to Rank in the SERPs with AI
The practical structure that emerges from the convergence of these signals isn’t complicated to describe. It’s complicated to execute, because it requires editorial discipline over time, not spot technical optimization. The first element is vertical topical authority: choosing one semantic domain and covering it with real depth, not breadth. A topic cluster isn’t a series of articles on the same keyword; it’s a network of content that presupposes itself mutually, covering the entities related to the central theme, building the graph of meaning algorithmic systems use to recognize a domain as authoritative.
The second element is originality in the technical Shannonian sense: primary data, non-replicable perspectives, analysis that wouldn’t exist without the specific expertise of whoever produces it. The algorithmic discriminant is replaceability: content that could be written by anyone with the same sources and the same conclusions contributes nothing to a domain’s epistemic identity. The third element is structure for retrieval systems: a direct answer in the first 100 words of each section, citable data as standalone units, semantic blocks of 200-300 words that LLM systems can extract without losing coherence. The fourth element is the author as signal: verifiable identity, documented credentials, primary sources cited, content that demonstrates direct experience through proprietary data or case studies.
Trust as Infrastructure — and as a Question of Power
Back to the starting point: the site with domain authority 85 losing traffic while the six-month-old blog climbs the SERPs. The data seems counterintuitive only if you think of domain authority as an accumulated asset that generates automatic returns. Thirteen years of backlinks built across heterogeneous topics, content produced for volume rather than density, an editorial record that signals breadth but not depth: none of this builds the epistemic trust the system now seeks. What builds it instead is a blog that in six months has produced twenty vertical pieces on a single domain, with original data, primary sources cited, coherent semantic structure, high semantic entropy in the sense Shannon meant the term.
There is, however, a question the SEO market prefers not to face directly. If the system rewards epistemic trust, who decides which sources deserve to be recognized as authoritative? Google is a private system that constructs, through its algorithmic updates, a hierarchy of credibility at global scale. E-E-A-T isn’t a neutral system of quality evaluation: it’s a set of proxy signals reflecting the expectations of a specific system, built in a specific context, with specific biases. The circularity is structural: you’re authoritative if you’re cited by authoritative sources, you’re cited by authoritative sources if you’re authoritative. Optimizing for Google in 2026 means understanding how institutional trust works — but also understanding you’re optimizing for a system that decides, with not entirely transparent criteria, who deserves to be found.
Post scriptum
The open questionWhat remains invisible in the system that decides who deserves to be found, and what does that mean for those producing content in good faith?
SEO in 2026 didn’t change because Google decided to punish keyword density. It changed because the system that manages information — the ensemble of language models, retrieval systems, human and algorithmic evaluation layers — became sophisticated enough to approximate the question human institutions have always asked before trusting a source: does this entity really know what it’s talking about? Does it have something to say that hasn’t already been said? Is it worth citing? The answer to these questions isn’t optimized with a technical checklist. It’s built over time, with the quality of the output, with the coherence of the editorial identity, with the capacity to produce something that couldn’t exist without the specific expertise of whoever produced it. And that’s not a technical question. It’s a question of power.
