Disclaimer and shout out:
All the following is based on the research of Metehan Yesilyurt who has published a far more complex article on this on his blog.
This is just me breaking it down for myself and others into simpler words and actionable insights.
Why am I interested in this?
For years, everyone tried to “reverse-engineer” Google’s AI systems through vague statements, half-baked patents, and random experiments.
But the funny thing is this:
Google is already showing us how their AI search stack works — just not in the place most SEOs look.
It’s called Discovery Engine (part of Google Cloud -), and while it’s meant for enterprise search, it exposes many of the same architectural ideas that power AI Overviews and AI Mode in consumer search.
No conspiracy theories.
No guesswork.
Just documentation, UI panels, and config screens that reveal how Google thinks about AI search.
I’ve gone through it in depth, and here’s the part SEOs and C-level leaders should care about.
The Big Shift: From Ranking Pages → Evaluating Information Units
Discovery Engine follows a clear four-stage pipeline:
- Prepare – understand and interpret the user’s query
- Retrieve – pull relevant content chunks
- Signal – score those chunks with seven distinct ranking signals
- Serve – pass the top content into Gemini and generate an AI answer
Once you see this pipeline, one truth becomes obvious:
Google no longer ranks pages. It ranks chunks of information that best answer the user’s intent.
And that changes everything about SEO.
Stage 1: Query Understanding (And Why Your Semantics Matter)
Before your content is even considered, Google transforms the query:
- Synonym mappings that change over time (very helpful for seasonal or political topics)
- Autocomplete predictions
- Natural language interpretation
- Prefix or substring matching depending on the scenario
All of this happens before retrieval.
What this means for SEO:
Your website must cover the whole semantic field around a topic — not just today’s keywords, but related concepts, historical wording, and adjacent questions.
If a concept evolves, your content must evolve with it.
Stage 2: Retrieval (The 500-Token Rule SEOs Aren’t Talking About)
This is where things get interesting.
Google processes your pages into chunks of up to ~500 tokens (~375 words).
These chunks can also include their H1 → H2 → H3 headings, so the structural context stays intact.
And Google even LLM-parses:
- tables
- images
- complex layouts
using Gemini enhancements.
Why this matters:
You can have the best 2,000-word article in the world —
but Google is not retrieving it as one unit.
If your important point is split across 3–4 sections, it will never surface as one coherent answer.
Practical takeaway:
- Keep sections short and self-contained
- Use a strong heading hierarchy
- Make key information extractable in <500 tokens
- Don’t bury insights in long paragraphs
Google rewards clarity and structure, not text walls.
Stage 3: The Seven Ranking Signals (Yes, Google Shows Them)
Discovery Engine exposes seven signals Google uses to rank content before an AI answer is generated:
- Base Ranking
- Embeddings (Gecko) – semantic similarity
- Semantic Relevance (Jetstream) – cross-attention, understands negation
- Keyword Matching – classic BM25
- Engagement Signals – popularity → PCTR → personalised PCTR
- Freshness
- Boost/Bury Rules – manual business logic
Let’s break down why these matter.
Keyword matching is alive and well
BM25 is explicitly listed.
But it’s one of seven signals — not the star.
Semantic relevance is king
Gecko (embeddings) + Jetstream (cross-attention) = a powerful semantic layer.
Jetstream especially understands distinctions like:
- “with X vs without X”
- “best A except B”
- “alternatives to Y”
This is why vague content fails in AI answers.
Engagement is a 3-tier quality system
Google uses:
Tier 1: Popularity (simple event counts)
Tier 2: PCTR (predicted CTR, model-based)
Tier 3: Personalised PCTR (only after 100k+ queries)
Your site may literally be locked out of higher-tier ranking signals if you don’t generate enough user engagement.
Freshness is dynamic
It matters more for news and time-sensitive topics.
Less for evergreen but it still matters.
Boost/Bury is the smoking gun
It proves Google can systematically push whole categories up or down based on rule-sets — basically the machinery behind how E-E-A-T influence actually works in practice.
Stage 4: Serving the Answer (Where Content Can Still Be Rejected)
Even after ranking well, your content can be blocked at the answer stage:
- low relevance
- adversarial patterns
- unclear grounding
- insufficient factual density
Google’s LLM will not generate an answer using content it deems low-quality or insufficiently grounded.
Translation:
Thin content destroys AI Search visibility —
even if the page “ranks”.
You’re optimising for answerability, not just rankings.
The Big Architectural Reveal: One Pipeline for All Search Products
Discovery Engine exposes three modes:
- Search → like traditional SERPs
- Search + Answer → AI Overviews
- Search + Follow-ups → AI Mode (conversational)
These map directly to consumer search.
Same pipeline.
Different output layer.
This means:
If you optimise for AI Search, you get better visibility everywhere.
What This Means for SEO (The Actionable Bit)
1. Structure is now a ranking factor
If your content isn’t chunk-ready and heading-aligned, it loses.
2. Semantics beat keywords
Cover the topic universe, not keyword clusters.
3. Engagement determines how many ranking signals you even unlock
Traffic → engagement → stronger ranking signals → more AI citations.
This compounds over time.
4. Content must be extractable
Short, factual, declarative sentences make it into AI answers.
Fluffy “SEO prose” does not.
5. UX and behavioural data matter more than ever
PCTR models actively influence ranking.
Final Thoughts
Discovery Engine doesn’t show the full secret recipe.
But it shows enough to understand the direction:
- Chunks, not pages
- Semantics, not keywords
- Engagement tiers, not generic ranking
- Unified AI pipeline, not two separate systems
If you want to win in AI search, the strategy is simple:Produce the best structured, semantically rich, fact-dense content in your market, and do it consistently.
Everything else is noise.