Google’s extended candidate set and selection problem

Google’s expanded candidate set reflects a profound change in the way search systems evaluate content. As AI systems process large pools of information, visibility increasingly relies on validation, relationships, and trust signals instead of traditional keyword targeting alone.

That change pushes SEO beyond retrieval and ranking mechanics to something closer to forensic architecture — systems designed to help machines verify and trust information at scale.

Search Engine Land recently published an article about Google’s expanded candidate set. When I read it, I felt a huge wave of relief and a shot of adrenaline. It confirmed that the rabbit hole I’ve been digging myself for the past five years is not just a personal obsession. This is where the digital ecosystem is headed.

For more than 30 years, I have worked to meet today’s requirements in ways that work for tomorrow. That experience teaches you to spot patterns early on and make decisions that are not just careers, but cornerstones of where the industry is headed.

Evolution: From librarian to forensic investigator

To understand why the “selection problem” occurs, you must first distinguish between a search engine and an AI agent.

In the early days, Googlebot was a download engine. It follows a strict, rules-based logic: find the link, download the page, and identify the words. It didn’t “think” about your content. Just record it. It was a librarian.

The evolution of intelligence

Over the past decade, that librarian successfully returned to school, earned a PhD in linguistics, and became a forensics investigator:

• Background thinking (2015): RankBrain allowed the system to understand the intent of queries it had never seen before.
• Change of status (2019): BERT allowed the search engine to understand the relationships between words, moving searches beyond keywords and into information acquisition (IG).
• Productive agent skips (2023–present): With Gemini’s overview and AI, the system now reads hundreds of pages at once to compile a single, unique answer.

The OpenAI catalyst and the selection problem

The arrival of ChatGPT in late 2022 accelerated the transition to responsive engines. Users stopped asking for recipes and began compulsively searching for meal plans.

This has created what I call the “problem of choice.” Because the AI ​​agent delivers a single, unified response, it must choose which facts to include and which to ignore. That leveled the playing field. The natural language interface allowed anyone to access high-quality information, regardless of their literacy search.

For those of us in the trenches, this has ensured that the acquisition of knowledge and atomic truths are the only currencies that matter. If an AI program can summarize your 2,000-word page in two sentences, the other 1,980 words are debt for context – unnecessary weight the machine will eventually ignore.

A 30-year journey towards atomic knowledge and truth

This conclusion did not come in a “magical” moment. It came from 30 years of identifying zombie facts, or outdated and inaccurate information that masquerades as truth, and extensive trial and error.

My path started in high-level industries: online pharmacies and regulated iGaming.

In these fields, trust is not a thing. It’s the only way to stay in business. Back in 2018, I started triple mining with semantic and knowledge graph. I realized that the browser just didn’t have to find it. It needed a logical map to understand us.

Property problem

Later, when I was managing eight ecommerce sites selling the same products at the same prices, I ran into an inventory problem. If everyone says the same thing, the search engine has no logical reason to choose you. You must provide atomic truth: the only unique, verified information you can provide.

I spent ten building tools to fix the gaps I found:

• EEAT engine: A 500-point system based on Google’s Search Quality Rater Guidelines.
• Atomic sandwich: A three-layer structure (atomic fact, information gain, structure layer) that handles content as a technical blueprint.
• IG forensic examiner: A tool for measuring whether your content really adds something new to the discussion.

Finally, the tool belt became too heavy. The issues – the core debt and the trust gap – require a more integrated approach.

That led me to create a framework designed to bring together a sophisticated understanding of engineering and the kitchen table.

Building trust in the response engine landscape

A recent audit I conducted of 28 digital companies confirmed that the selection problem has reached the web at large. As Search Engine Land reported, Google now evaluates dozens of pages for ranking.

In the domain of hundreds, the machine no longer asks who has the best keywords. It asks, “Who can I convince?” Standards alone are no longer enough. You need to be a source of AI programs that you can verify and trust.

To solve this, I use three pillars of forensic engineering:

• Pillar 1 – Cryptographic Authority: In the deep economy, I use the JSON Web Signature (JWS) standard (RFC 7515) to sign the entity manifest. Think of it as a quick pass for the candidate set because it enables quick verification.
• Pillar 2 – Semantic graph: AI thinks in relationships, not categories. Using W3C RDF-star standards, I send audits as structured information graphs. This reduces translation error when AI systems read your data.
• Pillar 3 – Regulatory alignment: I have mapped the structures in the EU AI Law (Law 2024/1689). This protects the digital GDP against legal shifts. If you want to be seen globally, you have to meet global needs.

The response engine changes that selection

The expansion of the candidate set indicates that search engines are becoming response engines. Visibility is increasingly dependent on whether AI systems can verify, connect, and trust the information associated with your business.

That change is changing the job of SEO. No more just retrieval and levels. It is increasingly about building systems that help machines understand relationships, verify information, and establish trust at scale.

The frameworks and standards needed to support that change already exist in the public domain. The challenge now is to learn how to assemble them into a reliable base so that they show up in AI-driven search.