AI Search vs Traditional Search

By Bill Hartzer · Updated 19 August 2026

The short version

Traditional search returns a ranked list of pages and leaves the judging to the reader. AI search returns a written answer and does the judging itself, showing citations as evidence.

That is the whole of the interface difference, and it is smaller than it looks, because underneath the answer the machinery is unchanged. Something still has to crawl the web with a program, store what it finds in an index, and rank that index against a query. An AI search product adds a generation layer on top of that; it does not replace it. Every answer engine in existence is a retrieval system with a language model attached, and the retrieval system is ordinary search.

The consequential differences are therefore not about "AI versus not AI". They are about what the reader can inspect, what the reader can reproduce, who owns the index doing the retrieving, and who is paying for the whole arrangement.

What is identical underneath

Three things, and none of them were changed by generative models.

Crawling. A program still has to fetch pages. AI products brought new crawlers, not a new mechanism: OpenAI's OAI-SearchBot, Perplexity's PerplexityBot, alongside Googlebot and Bingbot. They obey — or in documented cases decline to obey — the same robots.txt convention that has governed crawling since the 1990s.

Indexing. Pages still have to be parsed, deduplicated, scored for quality and stored. Perplexity's September 2025 engineering write-up describing an index tracking over 200 billion unique URLs on tens of thousands of CPUs is a description of a search engine, not of a chatbot.

Ranking. Candidates still have to be scored and ordered before a model can read any of them. The pipeline Perplexity describes — lexical and embedding scorers for candidate generation, then cross-encoder rerankers, operating at document and passage level — is a ranking system tuned to feed a language model rather than to render a results page. The output format changed; the retrieval problem did not.

The practical test of this is what happens when a system has nothing to work with. If an index does not contain the page, no amount of conversational skill will produce the fact — the model will either say so or invent something.

What genuinely changed

The unit of interaction. A query was a bag of keywords; a prompt is a sentence, and follow-ups carry context. "And the cheaper one?" is a question no keyword engine can serve because it contains nothing to match. This is a real advance and it is the reason people stay.

The unit of output. A list of ten results is a set of options. A paragraph is a conclusion. Somebody has made the judgement either way; the difference is whether the reader can see the alternatives that were rejected.

Reproducibility. A ranked list is stable enough to record and compare next week. A generated answer varies between runs in wording and sometimes in substance. This is why a generated paragraph is a poor thing to cite and the page it points at is a good one.

Inspectability. Three specific opacities arrive together. The query is rewritten — Microsoft documents that Copilot writes its own Bing query from the prompt, and OpenAI says queries are rewritten before going to providers, so retrieval depends on words the reader did not choose. The candidate set is hidden. And the reason one source was preferred over another is not published by anyone.

A consistent view of the web is no longer guaranteed. Independent measurement of ChatGPT's retrieval stack published on 17 August 2026 found the free tier served largely from OpenAI's own index and the paid "thinking" tier served largely from purchased scraped Google results. That is not a quality tier — it is a different web, and the interface does not say so.

The ownership question, which almost nobody gets right

The most useful fact about any search engine is whether it crawls the web itself or resells somebody else's results. Applied to AI products, the picture is not the one the branding suggests.

Perplexity built its own. A documented crawler with published IP ranges and an index of its own — and it did not start that way. For roughly its first two years it was, in substance, a Bing API reseller with a summarisation layer on top, and it has never published a dated statement of when that stopped.
ChatGPT Search is a mixture. OpenAI's own index built by OAI-SearchBot; purchased scraped Google data on the expensive path; Bing named in OpenAI's own help documentation as a partner and as the sole third-party provider for Enterprise and Edu workspaces.
Microsoft Copilot has no crawler and no index at all. Microsoft's documentation describes it plainly: Copilot generates a search query, sends it to the Bing search service, and writes prose over the answer. The qualifier that matters is that Bing is Microsoft's own index rather than a rival's.
Yahoo Scout, launched 27 January 2026, is the clearest case in the whole category. Anthropic's Claude as the model, Microsoft Bing's grounding API for the web. Yahoo's own summary was that the underlying index is Bing's while the responses, ranking and experience are Yahoo's. Yahoo has not run a general web crawler since about 2010.
Google's AI Overviews and AI Mode run on Google's own index, via what Google calls a query fan-out. Same crawl, same index, different presentation.
Kagi crawls with Kagibot into its own indexes but supplements essentially every query with anonymised calls into third-party commercial indexes. Its independence is at the ranking layer, not the index layer.

The forcing event was commercial. Microsoft retired the Bing Search APIs entirely on 11 August 2025, decommissioning the cheap retrieval substrate much of the AI search field had been built on. Products with their own index carried on; the rest paid Azure prices, bought scraped data elsewhere, or negotiated a first-party arrangement.

So the honest comparison is not "AI search versus traditional search". It is a handful of organisations that crawl the open web — Google, Microsoft, Yandex, Baidu, Brave, Mojeek and a few others, joined recently by Perplexity and OpenAI — and a much larger number of products presenting one of those indexes with a different face. Consulting Copilot, then Yahoo Scout, then Bing is one index asked three times.

Who pays, and what that changed

AI search was widely framed as an escape from advertising-funded search. It was not, and the retreat happened quickly.

OpenAI announced advertising in ChatGPT on 17 January 2026 for free-tier adult users in the United States, with ads visible in production by March 2026; the appearance of an OAI-AdsBot agent in OpenAI's crawler documentation corroborates it independently. Microsoft Copilot has always sat inside the division that also owns Microsoft Advertising and Bing. Perplexity ran the experiment in the other direction: it launched ad formats in late 2024 and discontinued them in February 2026 in favour of subscriptions. Kagi takes no advertising at all and charges for search directly.

The structural point is about auditability rather than morality. In a ranked list, a paid placement is a labelled row that a reader can skip and a researcher can count. Inside a paragraph of generated prose there is no row, no position and no label, and the influence of a commercial relationship on a sentence is not observable from outside. Every critique that was ever made of ad-funded search now applies to AI answers, with one fewer instrument for checking it.

The publisher side of the ledger is unresolved. OpenAI signed content licences with a long list of news organisations; Perplexity has been sued by News Corp, Reddit and three Japanese newspaper publishers while running a revenue-share for cited publishers; Google faces a broad dispute over AI Overviews and traffic in which no independent, agreed measurement of the effect exists.

Has one replaced the other?

No, and as of August 2026 no engine has said it will.

Google's AI Mode is a surface alongside classic results rather than a replacement, and nothing Google has published states otherwise. Copilot Search in Bing, launched April 2025, deliberately shows conventional web results next to the generated answer. Yahoo Scout is a separate destination; classic search.yahoo.com still exists. Google's share of measured search traffic worldwide was 91.31% in July 2026 per StatCounter, on the classic surface that AI Overviews sit on top of, not in place of.

The two formats also fail differently, which argues for keeping both available rather than picking a winner.

A ranked list is better when the reader needs to see the range of what exists, choose their own source, compare positions, work with a primary document, or produce something reproducible.
A generated answer is better for multi-part or ambiguous questions, for synthesis across several sources, and for refinement by follow-up rather than by reformulation.
Neither is better at telling the reader when it is wrong. A list can be full of poor pages; an answer can be fluent and false. The list at least shows its working.

The option to have search without generated answers is unevenly distributed and worth knowing before changing a default. DuckDuckGo publishes a dedicated AI-free endpoint. Brave Search offers a setting to disable its Summarizer. Kagi treats AI as something the reader invokes. Google offers no global switch — only the Web tab and the &udm=14 parameter, applied one search at a time.

Frequently asked questions

What is the difference between AI search and traditional search?

Traditional search returns a ranked list of pages and leaves the judging to the reader. AI search retrieves pages and then generates a written answer with citations, doing the judging itself. The retrieval machinery underneath — crawling, indexing and ranking — is the same in both, and most AI products do not run it themselves.

Is AI search replacing traditional search?

Not as of August 2026. Google's AI Mode sits alongside classic results rather than replacing them, Copilot Search in Bing deliberately shows conventional results next to the answer, and Yahoo Scout is a separate destination from classic Yahoo search. Google's measured share of search traffic worldwide was 91.31% in July 2026 per StatCounter, on the classic surface.

Do AI search engines still crawl and index the web?

Some do; most do not. Perplexity and OpenAI both run crawlers and indexes of their own. Microsoft Copilot has no crawler at all and grounds on Bing. Yahoo Scout grounds on Bing's API with Anthropic's Claude as the model. Google's AI features run on Google's existing index. The crawling and indexing still happens — just not always by the company whose name is on the product.

Which is more accurate, AI search or a list of links?

Neither is reliably more accurate, and they fail differently. A ranked list can be full of poor pages, but it shows its working and lets the reader choose. A generated answer can be fluent and wrong with a citation attached, and gives no signal of which it is. The list is auditable; the paragraph is not.

Why do different AI search engines give the same answers?

Often because they are consulting the same index. Microsoft Copilot and Yahoo Scout both ground on Bing, so asking both is one index queried twice rather than a second opinion. Only a handful of organisations crawl the open web at scale, and most AI products present one of those indexes with a different interface and a different model.

Is AI search free of advertising?

No longer. OpenAI announced ads in ChatGPT on 17 January 2026 for free-tier United States adults, in production by March 2026. Microsoft Copilot sits inside Microsoft's advertising business. Perplexity launched ad formats in late 2024 and discontinued them in February 2026. Kagi takes no advertising and charges a subscription instead.

Can traditional search results be used without AI answers?

It depends on the engine, and the differences are documented rather than a matter of taste. DuckDuckGo publishes an AI-free endpoint. Brave Search has a setting to switch off its Summarizer. Kagi requires the reader to invoke AI answers. Google offers no global setting — only the Web tab and the udm=14 URL parameter, applied one search at a time.

Did AI change how search engines rank pages?

Not fundamentally. Candidates are still scored and ordered before a model reads any of them. Perplexity describes a pipeline of lexical and embedding scorers followed by cross-encoder rerankers, working at document and passage level — a ranking system tuned to feed a language model rather than to render a results page. The output format changed; the retrieval problem did not.

AI search vs traditional search