How Google Search Actually Works
Most SEO advice is folklore: guesses dressed up as rules. But between 2023 and 2024 the real thing came out in public: sworn testimony from Google's antitrust trial, a leak of Google's internal search documentation, and decades of patents. Put together, they show how the ranking machine actually fits together, and once you see it, the right way to do SEO stops being a matter of opinion.
Two planes, joined at your page
Google Search is really two systems bolted together. An offline plane constantly crawls, renders, and indexes the web, and precomputes every expensive quality signal it can, attaching them to each page ahead of time. An online plane runs the moment you search: a fast, staged funnel that starts cheap and narrows to a handful of pages before the expensive models ever run. The whole design follows one pattern: precompute everything you possibly can, then spend real computing power only on ever-smaller sets of documents.
The offline plane: Google grades your page before anyone searches
Long before a customer types anything, Google has already crawled your site, rendered your JavaScript, pulled your links, and decided how much to trust you. According to the leaked documentation, the index sorts pages into storage tiers: a hot, frequently-served slice at the top, down to a cold tier for stale, long-tail pages that barely get retrieved (the leak nicknames the bottom tier the “landfill”). Links coming from higher tiers carry more weight than links from the bottom. We break down what that tiering means for your pages in what SEO is.
Then Google attaches a bundle of precomputed signals to each page: its PageRank (computed offline over the whole link graph, then simply looked up at search time), site-level authority and quality scores, and spam scores. This is the part most owners get wrong: a signal like “Panda”, Google's site-quality system, isn't a model that judges you live during a search. The leak indicates it is a precomputed input, baked into your page's record in advance. The expensive grading happens up front; serving your page is then a cheap lookup.
The takeaway for a business owner: a large part of your fate is decided before the query even runs. You have to be cleanly crawlable and indexed, you have to earn enough real retrieval to sit on a hot tier instead of the landfill, and you have to build genuine site-level quality and authority, the inputs Google bakes in ahead of time.
The online plane: a cheap-first cascade
When someone searches, Google runs a serial funnel. Each stage only reorders what the previous stage handed it, and each stage is more expensive than the last, so Google keeps shrinking the candidate set before it spends the big compute. From the trial testimony and the leak, it looks roughly like this:
| Stage | What happens | Pages still in the running |
|---|---|---|
| Understand the query | Rewrite and expand it: synonyms, entities, what you really mean | — |
| Lexical retrieval | Old-school inverted index and postings lists, plus a semantic model to catch matches the keywords missed | tens of thousands |
| ~100 hand-built signals | Score on anchors, body, links, clicks, and page quality (PageRank feeds in here) | a few hundred |
| Click re-ranking | Reorder using how people actually clicked and stayed (the “Navboost” system) | a few hundred, re-sorted |
| Neural re-rank | The heavy AI models (RankBrain / DeepRank), too expensive to run on everything | the top ~20–30 only |
| Assemble the page | Write snippets, lay out the features, merge ads at the page level (ads never touch organic scoring) | the results you see |
Here is the rule that should reshape how you do SEO: nothing downstream can save a page that fails an earlier gate. The famous AI models only ever see the top couple dozen pages. If you never made it through lexical retrieval because your page wasn't clearly relevant, no amount of “great content” or clever tricks reaches the stage that could reward it. You win by clearing the gates in order: be retrievable and relevant first, be trusted by the hand-built signals second, and be genuinely chosen by real users third.
Where your clicks actually count (and where they don't)
Clicks matter, but not the way people fear or hope. The click re-ranking system works on aggregated behavior across a long window, reported in the leak as roughly 13 months of click data. Your click today doesn't personalize your own results; it becomes one vote in a 13-month slice that nudges results for everyone searching like you. The patents describe correcting for the fact that the number-one result gets clicked more regardless of quality, so raw clicks alone don't win: it's clicks and genuine engagement relative to what a position should earn.
Two things follow. First, you can't fake this at scale: you earn it with titles and pages that get chosen and then actually satisfy the visitor. Second, it is a big reason SEO takes months, not days: a 13-month click signal genuinely takes time to accumulate. Personal data enters in only three places, and personalization is the smallest of them: your context (location, language, device) shapes results from the very start, aggregate behavior is the heaviest input in the middle, and true per-user tailoring is a light final touch, with genuinely personalized answers living in the separate Gemini / AI Mode layer that sits on top of this stack.
The famous “updates” aren't the system: they map onto it
Once you can see the architecture, Google's named updates stop being mysterious weather and become changes to specific stages:
- Hummingbird: a rewrite of the matching core (how queries are understood and pages retrieved and scored).
- Panda: the offline site-quality signal baked into your page's record.
- Penguin: link-spam handling that feeds the PageRank / authority side.
- RankBrain / DeepRank: the neural re-rankers bolted onto the head of the funnel.
This is the whole reason the architecture is worth understanding. When a core update lands and rankings move, it changed one of these stages. A team that knows the system can reason about which one (was it a site-quality signal, a link-weighting change, a click-model shift?) and adjust deliberately. A team that only knows tactics just guesses, panics, and hopes. That difference is what makes a business more update-agnostic: not immune to change, but able to diagnose and pivot instead of scramble.
Why we do SEO this way
We optimize for this system, not for exploits. Exploits target a single stage, work until Google patches it, and then collapse, taking your rankings with them. Optimizing the system compounds: be cleanly retrievable, build the real site-quality and authority Google precomputes, earn the clicks that survive presentation-bias correction, and you grow in a way that holds through updates. It also gives us the best possible diagnostics (when a client's rankings move, we locate the stage that changed rather than guessing), which is exactly how we keep our clients steady when the algorithm shifts. It is the same discipline behind our SEO service.
Key takeaways
- Google is two planes: it grades pages offline, then serves them through a fast, staged funnel.
- The funnel is cheap-first, and nothing downstream saves a page that fails an earlier gate.
- Site-quality and PageRank are precomputed inputs, not live judgments. You build them ahead of time.
- Clicks count as aggregated votes over ~13 months, corrected for position bias. You can't fake them at scale.
- Named updates map onto specific stages, so knowing the system lets you diagnose and pivot instead of guess.
Frequently asked questions
Is this really how Google works, or is it speculation?
It is a synthesis of the public record, not an official spec. The structure comes from sworn testimony in the 2023 antitrust trial, the 2024 leak of Google's internal Content Warehouse documentation, and Google's own patents. The system names are internal Google identifiers as they surfaced in that testimony and leak. Google hasn't published an architecture diagram, so this is the best-supported picture available, assembled from primary sources rather than folklore.
What is the single most important thing this tells me about SEO?
Clear the gates in order. A page has to be retrievable and clearly relevant before it can be scored, trusted before it's re-ranked, and genuinely chosen by users before the AI models ever see it. No downstream trick rescues a page that failed an earlier stage, which is why fundamentals beat exploits.
Why does SEO take months to work?
Because the machine is built on precomputed and accumulated signals. Google has to re-crawl and re-index your changes, your site-quality and authority inputs are computed ahead of time, and the click-based re-ranking system works off roughly a 13-month window of behavior. Those signals genuinely take time to build, which is why durable rankings compound over months rather than appearing overnight.
Do clicks affect rankings?
Yes, but as aggregated, population-level votes over roughly 13 months, not as personal or instant signals. The patents also describe correcting for the fact that the top result gets clicked more regardless of quality, so raw clicks don't win. Earning clicks and genuine engagement relative to your position does. It can't be faked at scale; it's earned with pages that get chosen and satisfy.
Does this apply to Google's AI search / AI Mode?
AI Mode and Gemini sit as a separate system on top of this stack, and they still draw on the same underlying signals (structure, relevance, authority, and reviews) that classic ranking rewards. So the fundamentals here are the foundation for getting surfaced by AI too, which we cover in how to get recommended by AI.
How does knowing all this actually help my business?
Two ways. Diagnostics: when your rankings move, we can reason about which stage of the system changed instead of guessing, so fixes are targeted. And resilience: because we optimize the system rather than an exploit, a core update is something we diagnose and adjust to, not a disaster, which keeps your growth steadier through the changes that wipe out trick-based competitors.