AI Search Explained: How AI Overviews, ChatGPT Search and Perplexity Build Their Answers

Search used to end with homework: ten blue links, and you did the reading. Increasingly it ends with an answer — a paragraph written by AI at the top of Google (AI Overviews), inside a chatbot (ChatGPT search), or from a dedicated answer engine (Perplexity). The convenience is real, and so is the confusion: where did that paragraph come from, which pages is it based on, and how much should you trust it? This pillar guide explains the machinery in plain English — how classic search worked, what AI added, how the three big systems differ, what the measured evidence says about their accuracy and their effect on the web, and how to use them without being misled. Five companion guides go deeper: how AI Overviews work, a three-way comparison, why AI answers go wrong, how to ask better questions, and an FAQ.

The old model: crawl, index, rank

Classic search engines do three things. They crawl the web with automated programs that follow links and copy pages. They store what they find in an index — a giant, searchable catalogue. And when you type a query, they rank pages from that index by predicted usefulness and hand you the list. The crucial property of the old model is what it does not do: it never claims anything. A ranked list can be badly ordered, gamed or incomplete, but every link is a real page that exists, and the judgement — clicking, reading, believing — stays with you. Everything AI search adds builds on this foundation; none of the three systems below replaced crawling and indexing, which is why Google’s documentation for its AI features still starts from the same requirement: to appear as a supporting link, a page must be indexed and eligible to be shown in Search with a snippet.

The new model: retrieve, then generate

AI search bolts a writer onto the retriever. Your question is first turned into one or more searches — Google calls its version of this a “query fan-out,” issuing “multiple related searches across subtopics and data sources” to build the response. Pages come back; their text is fed to a large language model (the same technology as a chatbot — see our plain-English LLM explainer); the model composes an answer and attaches links to some of the pages it used. Two different kinds of error now become possible. Retrieval errors: the searches found the wrong pages, missed the right ones, or read a syndicated copy instead of the original. Generation errors: the model misreads a correct page, blends two pages into a claim neither makes, or writes a citation that does not support — or does not even match — the sentence it hangs from. When an AI answer is wrong, it is almost always one of those two failures, and the fixes differ, which is why the “why it goes wrong” companion guide treats them separately.

The three systems, side by side

Google AI Overviews live inside ordinary Google Search: a generated summary appears above the classic results for some queries — Pew’s March 2025 browsing data found summaries on about 18% of US searches — alongside links to supporting pages. Google says Overviews are shown when its systems judge the summary “additive to classic Search,” and its separate AI Mode goes further, built for “further exploration, reasoning, or complex comparisons” with a fuller generated response. ChatGPT search lives inside the chatbot: OpenAI’s help centre explains that ChatGPT may decide on its own that a question needs current information, or you can choose Search yourself; it then turns your request into one or more search queries, retrieves results (drawing on search partners), and answers with links, with inline citations you can open and a Sources list at the end. Perplexity is the purpose-built answer engine: its help centre defines it in one line — it “answers your questions with cited sources,” searching the web, summarising what it finds, and showing “where each part of the answer came from” with numbered citations; its Pro Search runs a deeper, multi-source pass for research questions. Same skeleton, three personalities: Google’s is an addition to the results page, ChatGPT’s is a mode of a general assistant, Perplexity’s is the whole product.

What changes for you, the reader

The answer-first design changes behaviour measurably. Pew Research Center tracked 900 US adults’ browsing through March 2025 (68,879 Google searches) and found that when an AI summary appeared, users clicked a traditional result link in 8% of visits, versus 15% when no summary appeared; they clicked a link inside the summary in just 1% of visits; and they ended their browsing session entirely in 26% of summary visits versus 16% without. The summaries themselves are usually sourced — 88% cited three or more sources — and the most-cited sources overlap heavily with classic results: Wikipedia, YouTube and Reddit together made up 15% of summary sources, with government sites appearing more often in summaries (6%) than in standard results (2%). The pattern is consistent: the AI layer absorbs the click. Whether that is good or bad depends entirely on whether the summary deserved the trust — which brings us to the evidence nobody in this industry enjoys quoting.

The accuracy picture, without the marketing

In March 2025, Columbia Journalism Review’s Tow Center for Digital Journalism ran the most careful public test of AI search accuracy: 1,600 queries built from excerpts of 200 real news articles across 20 publishers, given to eight AI search tools, each asked to identify the article’s headline, publisher, date and URL. Collectively, the tools answered more than 60% of queries incorrectly. Perplexity did best of the named tools at 37% incorrect; ChatGPT Search got 134 of its 200 wrong (67%); Grok 3 reached 94%. Worse than the errors was the manner: the tools “presented inaccurate answers with alarming confidence,” rarely hedging — ChatGPT signalled a lack of confidence just fifteen times in 200 responses and never declined to answer. More than half of Gemini’s and Grok 3’s responses cited fabricated or broken URLs (154 of Grok 3’s citations led to error pages), and several tools cited syndicated copies instead of original articles. The test was deliberately hard — tracing a quote to its source is retrieval’s sternest exam, and the tools have been updated since — but it remains the benchmark demonstration of the two failure modes above, and no comparable study has overturned its shape. Our companion guide on why AI search answers go wrong walks through it in full.

What this means

AI search is best understood as a transfer of effort and of trust. The old bargain made you do the work — opening pages, comparing sources, synthesising — and in exchange the system’s claims were minimal and its sources inspectable. The new bargain does the synthesis for you and asks, in exchange, for trust in a composed paragraph whose evidence is one click further away than it used to be. Pew’s data shows the public accepting that bargain at scale: the answer is read, the links mostly are not, the session ends. That makes the quality of the generated paragraph a civic matter in a way a ranked list never was — a wrong link in position seven misled almost nobody; a wrong sentence in the summary misleads everyone who stops there, which is most people. It also quietly rewires the economics of the web the summaries feed on: publishers whose work grounds the answers receive a fraction of the visits, a tension already visible in licensing deals, crawler-blocking and lawsuits that the Tow Center documented alongside its accuracy test. For the individual US reader, the practical settlement is the one this series teaches: use AI search for orientation, overviews and first drafts of understanding; follow the citation when a fact will be repeated, spent or acted on; prefer the tool whose citations you can inspect fastest; and remember that a confident paragraph is a style, not a credential. The readers who thrive in the answer era are not the ones who avoid AI search or swallow it whole — they are the ones who treat it as a brilliant research assistant whose work, like any assistant’s, gets checked before it goes out under their name.

There is a second, quieter consequence for how knowledge itself gets maintained. The classic web ran on an exchange — sites published, search engines sent readers, and readers’ attention funded more publishing. AI answers keep more of that exchange’s value at the answer layer, which is why Pew’s finding that only 1% of summary visits click a citation matters beyond marketing: it describes a funding leak in the system that produces the sources. Whether that leak forces new models — licensing, subscriptions, public funding for journalism — or simply thins the open web is the open question of the next few years, and it is why the health of AI search and the health of the web it reads have become the same conversation.

Related Guides

• How Google AI Overviews Work: Where the Answers and Links Come From
• Perplexity vs ChatGPT Search vs Google AI Overviews: Which Should You Use?
• Why AI Search Answers Go Wrong: Wrong Citations, Made-Up Links and Confident Mistakes
• How to Ask AI Search Better Questions: Prompts That Get Specific, Sourced Answers
• AI Search FAQs: Is It Free, Is It Accurate, and Does It Replace Google?

Bottom line

The new bargain trades your reading effort for trust in a composed paragraph. Accept it for orientation, refuse it for evidence — and keep opening the sources, because the answer’s quality is capped by theirs.

Read Next

• How to Check AI Answers: A Simple Fact-Checking Checklist
• What Is a Large Language Model? LLMs Explained in Simple Words

Sources

• AI Features and Your Website — Google Search Central (accessed 7 November 2026)
• Introducing ChatGPT search — OpenAI (accessed 7 November 2026)
• AI Search Has a Citation Problem — Tow Center for Digital Journalism, Columbia Journalism Review (accessed 7 November 2026)
• Do People Click on Links in Google AI Summaries? — Pew Research Center (accessed 7 November 2026)