Old search picked a winner. AI search builds a case.
Take a question people actually argue about — watch how the two approaches answer it, then what that means for your content.
Follow one signal and stop. "Arsenal won the league — so their keeper must be the best."
Breaks the question into smaller searches — each one a real query it runs, routed to wherever that evidence lives. Your page only enters the answer if it's what one of those searches pulls in:
Collate & compare. Pool every source, then weigh the leading keepers head-to-head. Strong across many measures beats a single freak stat.
The verdict — quoting its sources. "He leads on goals prevented in the data, the highlights back it up, and the fans agree" — one answer stitched from several sources, each named.
None of this is hand-waving: under the hood it's a documented pipeline — fan-out, routing, rank fusion, pairwise ranking, synthesis — each backed by a Google patent. The "Explore the mechanics ▸" tabs show every step.
Search stopped being a ranking. It became a research assistant.
The abstract version of the story above — kept here while we decide which framing lands better.
You type a search. Google shows ten blue links.
Being #1 won the click.
AI splits your question into smaller ones, searches a different place for each, compares what it finds, and writes a single answer — quoting a handful of sources.
Getting quoted in the answer is the new win.
So your content no longer just needs to rank — it needs to survive long enough to get quoted. And because one answer pulls from many searches, a single citation hides most of the story: appear in 4 of 12 sub-searches but get cited once, and classic rank tracking misses around 75% of your real footprint.
To be quoted, a page clears five gates in a row
Miss any one and you're not in the answer. This is a map of where content can fall out — not a literal look inside Google.
Walk one question through the five gates
A searcher asks "—". Here's the path to the answer — frozen, nothing to click.
Five priorities that decide whether you get cited
No guarantee — but content that scores well across all five tends to get quoted. This is the working checklist the rest of the report turns into real actions.
Coverage
Be present across the sub-questions, not just the headline term.
Format
Cover the right surface for each intent — guide, video, forum, product.
Evidence
Make passages specific enough to quote — names, numbers, real steps.
Comparison
Beat the competing page when an AI reads both side by side.
Action
Decide what to create or fix — and give it an owner.
See the machinery — or see where you stand
The tabs below are optional depth: each steps through one stage at your own pace. When you're ready for your real data, jump to the report.
AI search works more like a research assistant than a ranking engine.
It breaks a question into smaller jobs, searches different places for each job, compares the evidence it finds, and builds an answer from the strongest passages. Your content doesn't just need to rank. It needs to be discoverable across the right subtopics, available in the right formats, specific enough to beat competing passages, and useful enough to be cited in the final answer.
We can't see inside Google. This is the best working model based on Google's patents and observed behaviour — useful as a "where can content fail to be cited" map, not a literal architecture.
In practice, your page has to clear five gates in a row to be cited:
The sub-queries
- Google's smaller questions will list here.
What each search returned
- Top-10 results from each Google search will list here.
The AI's verdict
Live Execution Pipeline
The AI picks 3–5 pages by comparing them, not by ranking
After Google ranks the top 10 for a search, the AI doesn't just pick #1. It reads two pages at a time and asks itself: "Which one better answers this question?" Across many pairs, it adds up the wins. The pages with the most wins get cited in the AI Overview.
Below, we walk through this for one smaller question: "—" from the scenario "—". Four pages came back from Google — watch them fight it out.
Google ranks the top 10. Then the AI reads them.
Google's normal ranking picks the top 10 for the sub-query. The AI then reads the actual page text of those 10 and compares them two at a time. If your page doesn't make Google's top 10, the AI never reads it.
ranks it
compares
The 4 candidates · with live W-L tally
Ready to judge
Round 0 of 4Press play below — one round per click.
The AI reads both, then picks.
The AI keeps the top 2
After 4 rounds, the AI reads the W-L tally and keeps the pages it needs. Sometimes more than one — answers often cite multiple complementary pages.
Ready to play
Referee viewNo active match.
No active match.
One question becomes many — then gets stitched back into one answer
A single user question gets split into 5 smaller questions. Each one runs as its own Google search — some on the regular web, some on YouTube, some on Reddit. The results get combined into one list, the AI compares the top pages, and picks 3–5 to cite in the final answer.
Google doesn't publish exactly how it does this — what's below is our best read, grounded in the patents Google has filed. See docs/research/agentic-rag.md for what's confirmed vs inferred.
Question we're simulating: "—"
Google generates 5 smaller questions
Each runs as a Google search · top 10 results
Watch for pages that appear in more than one search — that's what gets rewarded next.
Pages in many searches rise to the top
Consistency beats peak rank. A page in 3 of 5 searches beats a page ranked #1 in just one.
The AI picks 3–5 pages to actually quote
The top-scoring pages go through one more . The AI picks the ones that best support each claim — not the top 3 by rank.
Grand Consensus Output
Waiting for winnersQualified evidence
- No bracket winners yet.
Run the consensus simulation to see how winning passages become a synthetic AI Overview answer.
Three generations of how AI search works. The framework — sub-query coverage, survival rate, citation share — holds across all three. What changes is whether the stages are visible as separate steps or fused into one model.
One search, one answer
The model runs one search, grabs a handful of nearby pages, and writes from those. No splitting, no comparing, no second pass.
- Weak on questions that have several parts.
- Doesn't route to forums, video, or specialised data.
- No checking that the evidence is good enough.
- Generic copy slips through because pages aren't compared.
Many searches, then judged
The system splits the question, runs many searches, compares the top pages two at a time, then writes the answer. This is what the rest of this report visualises.
- Splits the question into 5–20 smaller related questions.
- Routes each to its best source — blog, forum, video, shop.
- Goes back for more if the evidence is thin.
- Picks pages by comparing them, not by ranking signals alone.
One model does it all
A single model trained specifically for search — splitting, retrieving, comparing, and writing all happen inside it. Domain-tuned versions are emerging (Charcoal, SID-1, Waldo).
- No separate splitter / router / critic — the model owns the whole loop.
- Tunable to a specific domain or vocabulary.
- Better at the "last 20%" — your jargon, your aliases.
- Risk: citations may fragment across many search systems.
Five priorities that decide whether your content gets cited
AI search doesn't reward one #1 page anymore — it rewards content that meets a higher bar across these five dimensions. Treat this as a working checklist that emerges from the model we just walked through.
No guarantee of citation — but pages that score well across all five tend to. Use it to prioritise where to invest content effort.
Coverage
Be present across the sub-questions Google is generating, not just the headline query. Map your topic out to the 5–10 related angles a user would ask next.
Format
Cover the right surface for each intent — guide, video, forum, product page, FAQ. Google routes different sub-queries to different surfaces; if you only exist in one, you only reach one.
Evidence
Make passages specific enough to quote — named entities, concrete steps, real numbers, real comparisons. Generic framing loses the LLM-as-judge comparison every time.
Comparison
Beat the competing passage when an LLM reads both side-by-side. Specificity and direct-answer phrasing tend to win. Audit your top-cited competitors — what would beat them?
Action
Decide what to create or improve to lift the previous four — and assign an owner. The framework only works if it becomes a roadmap, not a poster.