Traditional search returns a ranked list of web pages for the words a person typed, and the person opens those pages and reads them. AI search does the reading on the person's behalf: a language model sends queries to a search index, pulls passages from the pages it gets back and writes one answer that cites a few of them. Both rely on crawling and indexing the web, and they differ in what they rank, what they show and how stable the result is.
How does traditional search work?
A traditional search engine has three stages. A crawler, which is a program that fetches pages by following links, downloads pages from the web. An indexer stores the text of each page in a search index, a database organized so pages can be looked up by the words they contain. When someone types a query, a ranking system scores the matching pages and returns them in order.
The ranking uses signals such as how closely the page text matches the query, links from other sites, page quality and freshness. The result is a list of links, each with a title and a short snippet. From there the person chooses which links to open, reads the pages and assembles the answer in their own head.
How does AI search produce an answer?
AI search keeps the index and puts a language model on top of it. A large language model is a neural network trained on large amounts of text to predict the next word, which also lets it read a question and write a reply. The model reads the question, rewrites it as one or more search queries and sends them to an index.
After the index returns candidate pages, the system works through a fixed sequence:
- Fetch the pages and remove navigation, ads and other boilerplate.
- Split the remaining text into passages of a few sentences each.
- Score each passage against the question and keep the highest scoring ones.
- Give those passages to the model as context.
- Generate the answer and attach citations to the sentences that came from a source.
This pattern of fetching documents first and writing from them second is called retrieval-augmented generation. It is how a model can answer questions about recent events, prices or specific products that were missing from its training data.
What are the main differences?
The two systems share an index, and almost everything after the index works differently.
| Traditional search | AI search | |
|---|---|---|
| Input | A short keyword query | A full question, often with earlier conversation |
| Queries sent to the index | The one the user typed | One or more written by the model |
| What gets ranked | Whole pages | Pages, then passages inside them |
| Who reads the pages | The user | The model |
| Output | A list of links with snippets | A written answer with a handful of citations |
| Repeatability | The same query gives a similar list | The same question can give a different answer |
The input row affects how people ask. In a chat, a follow-up such as "which one is cheaper?" depends on the previous turn, so the model has to carry the context into the queries it writes. A keyword engine treats every search as new.
Ranking passages instead of whole pages changes what competes. A page can rank well as a whole and still contribute nothing to an AI answer if none of its paragraphs answers the question directly. How AI search chooses which sources to cite describes the passage scoring step in more detail.
Why are AI search results less stable?
A traditional results page changes slowly. Rankings move when pages are updated, new pages appear or the ranking system changes, so two people typing the same query on the same day usually see much the same list.
AI answers vary for several reasons that stack on top of each other:
- The model chooses its words with some randomness, so the wording and the order of items differ between runs.
- The queries the model writes can differ from one run to the next, which brings back a different set of pages.
- Earlier messages in the conversation change how the question is interpreted.
- Some assistants decide per question whether to search the web at all, and answer from training data when they do not.
Because of this, there is no fixed position to track the way there is with a ranked list. A brand can be named in one answer and missing from the next for the identical question. The useful measure is how often a page or brand appears across many repeated runs, which is why a single check says little.
What stays the same?
AI search still depends on the web being crawled and indexed. Each AI product uses a search index, either one it builds itself or one it licenses from a search engine. If a page is missing from that index, the model never sees it.
AI companies also run their own crawlers. AI crawlers have separate user agents for training, for live search and for fetching a page a user asked about, and a site's robots.txt file can allow or block each one. A page blocked for the search crawler of a given product cannot be cited by that product, however well it is written.
What does the difference mean for a website?
The first change is in how traffic arrives. In traditional search, a visit happens when someone clicks a result. In AI search, the answer is already on the screen, so many questions end without a click, and the visits that do come from citation links are counted as AI referral traffic.
The model also reads content in a different way. Because it picks passages, a paragraph that answers one question in its first sentence is easier to select than one that only makes sense after the three paragraphs above it. Definitions, short lists and tables with real values are easy to lift into an answer.
Measurement changes too. Rank tracking records the position of a page for a keyword. AI visibility records how often AI answers mention or cite a brand across a fixed set of questions and platforms, run repeatedly over time, and it also counts mentions on third-party pages such as reviews and comparisons that the model read while writing.
