How Search Engines Work: From Crawling to Ranking

When you type a question into a search bar, what happens next feels like magic. Billions of pages exist on the web, yet somehow the best ones appear instantly in front of you. To understand this, let’s step into the world of crawling, indexing, and ranking—not as dry steps, but as a story of how machines turn the chaos of the internet into order.

Crawling: The Endless Exploration

Think of the internet as an infinite graph:

G=(V,E)G=(V,E)

where V represents web pages (nodes) and E represents links (edges).

Search engines deploy “crawlers,” which move from node to node by following edges. Just like a random walker in graph theory, a crawler has a probability distribution of visiting a new page:

P(v∣u)=1/outdeg(u)

meaning the chance of visiting page v from page uu depends on how many outgoing links u has.

In practice, crawlers don’t move randomly—they prioritize freshness, authority, and importance of pages. But at its heart, crawling is the process of discovering nodes in this massive web graph.

Indexing: Building the Digital Library

Once discovered, each page must be stored and organized. Indexing transforms raw HTML into structured information. Imagine the index as a function:

I:word→{pages containing word}

For example:

I(“AI”)={p1,p2,p3,...}

This is called an inverted index—instead of mapping pages to words, it maps words to pages.

Just like a library catalog, it allows instant retrieval. When you search “how engines work,” the system intersects these sets:

I(“how”)∩I(“engines”)∩I(“work”)

and retrieves only the documents containing all three.

Ranking: Choosing the Best

Now comes the art and math of ranking. Not all pages are equal. Some are more authoritative, others are more relevant. Ranking blends multiple signals into a scoring function:

S(d,q)=w1⋅R(d,q)+w2⋅A(d)+w3⋅U(d)

where:

• R(d,q) = Relevance of document dd to query qq
• A(d) = Authority of document (based on backlinks, trust, etc.)
• U(d) = User experience signals (speed, mobile-friendliness, engagement)
• w1,w2,w3 = weights tuned by the search engine

One famous early formula is PageRank. In simplified form, the importance of a page PP is given by:

PR(P)=1−d​/N+d*∑​L(Q)PR(Q)​ (Q∈M(P))

where:

• d = damping factor (usually ~0.85)
• N = total number of pages
• M(P) = set of pages linking to P
• L(Q) = number of outbound links from page Q

This elegant formula captures the intuition that a page is important if other important pages link to it.

‍

The Evolution of Search

Early search engines relied heavily on link-based authority measures like PageRank, a mathematical model developed by Google’s founders in the late 1990s. Over time, these signals were enriched with more complex models that incorporated user behavior, semantic meaning, and real-time personalization.

Modern search engines use machine learning to estimate the probability that a document is relevant to a query:

P(relevant∣q,d)=fθ(q,d)

where fθ is a neural scoring function (often powered by deep learning models such as transformers). Unlike static formulas, these models learn from billions of past queries, clicks, and satisfaction signals.

In practice, this means the search engine no longer just asks: “Does the page contain the right words?” Instead, it asks: “Does this page truly answer what the user means?”

‍

‍

The Rise of Generative Engines

We are now entering a new era: the shift from retrieval engines (like Google Search, Bing) to generative engines(like ChatGPT, Perplexity, or Google’s Gemini-powered Search).

A traditional search engine follows:

Query    →    Retrieve documents    →    Rank results

A generative engine follows instead:

Query    →    Retrieve knowledge + Model inference    →    Generate synthesized answer

The difference is crucial:

• Search engines point you to documents.
• Generative engines produce answers directly, sometimes without showing the source unless explicitly linked.

From a mathematical lens, a generative engine is not computing:

arg⁡max⁡  P(relevant∣q,d)

but rather:

arg⁡max⁡  P(answer=a∣q,K)

where K is not just the index of documents, but also the latent knowledge encoded in the model’s parameters (Lewis et al., 2020 — Retrieval-Augmented Generation).

‍