AI-Powered Search Engines: A Complete Guide

An AI-powered search engine answers a question in prose instead of returning a list of links, and attaches citations to the pages it read. ChatGPT, Perplexity, Google's AI Overviews and AI Mode, Gemini and Microsoft Copilot all work the same way underneath: a search index in front of a language model. The model does not know your site. It reads whatever the search step returns, and it cites from that. So the whole question of visibility in AI search is whether your page gets retrieved, and retrieval still runs on an index that links help you enter. Google states the floor plainly: a page has to be indexed before it can be cited at all.

Where does the answer come from, training or retrieval?

Two different things get called AI search, and the distinction decides what you can influence.

Training is the first. A model learns from a large body of text scraped years before you asked your question, and reference sites, news and documentation are heavily represented in it. Nothing in that process produces a link. A model answering from training alone has no source to point at, and any URL it prints is a guess.

Retrieval is the second, and it is where citations come from. The assistant decides the question needs current information, runs one or more searches, reads the pages that come back, and links what it used. Every platform below does this. The consequence is worth stating flatly: a page that is not in the index cannot be retrieved, and a page that is not retrieved cannot be cited, however good it is.

How do AI search engines work?

Traditional search indexes pages and ranks them by relevance. AI search adds a generation step on top of the same retrieval problem.

Retrieval-augmented generation

Most AI search engines use retrieval-augmented generation, usually shortened to RAG. Five steps:

1. Query understanding. The system parses the question to extract intent and key concepts.
2. Retrieval. It searches an index, often the live web, for relevant documents.
3. Ranking. Documents are converted to numerical representations and ranked by similarity to the question.
4. Generation. The model writes one answer from the top-ranked sources.
5. Citation. Sources are linked so the reader can check them.

Google calls its version of step two query fan-out: one question becomes several related searches across subtopics, and its AI features documentation says the models "identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search."

Embeddings and vector search

At the core is the embedding: a numerical representation of a passage that captures its meaning rather than its words. A search for "best coffee shops in Brooklyn" can match a page that never uses the word "best". Vector databases store those embeddings and compare them across billions of documents quickly.

The practical upshot for a writer is that exact-match keywords matter less than they used to, and a passage that answers one question completely matters more. Retrieval pulls passages, not pages.

Which AI search engines matter?

Perplexity

Perplexity calls itself an answer engine rather than a search engine, and citation is the interface rather than a footnote. Its help centre says it gathers information "from authoritative sources like articles, websites, and journals" and that "Each answer includes numbered citations linking to the original sources".

Features that affect what it cites: real-time web search with inline citations, a Pro mode that runs multi-step research, and focus modes that restrict retrieval to a source family such as academic papers, Reddit or YouTube.

ChatGPT with search

ChatGPT searches the web when the question needs it. OpenAI's help page on searching the web with ChatGPT says it "may search the web automatically when your question would benefit from current information" and that responses using web search "may include citations".

Measured citation behaviour differs from the other platforms. Profound's study of 680 million citations (published 5 June 2025, covering August 2024 to June 2025) found Wikipedia was 7.8% of all ChatGPT citations, and nearly half of the citations that went to its ten most-cited sources.

Google AI Overviews and AI Mode

AI Overviews sit above the blue links in ordinary Google results. AI Mode is a separate conversational surface. Both matter because they sit inside the search product people already use.

Google's eligibility rule is the most useful sentence any vendor has published on this: "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet." The same document adds that there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." There is no submission form and no opt-in.

How often they appear is now measured rather than guessed. A Washington University measurement study of 55,393 trending queries over 40 days (13 March to 21 April 2026) found AI Overviews on 13.7% of queries overall and 64.7% of question-form ones. Why isn't my site showing up in AI Overviews works through the causes in the order they are usually the answer.

Microsoft Copilot

Copilot is Microsoft's assistant with live web access, built into Windows, the Edge browser and Microsoft 365. Its retrieval runs on Bing's index, which makes Bing Webmaster Tools the place to check whether a page is indexed for it. Bing is a separate index from Google's, and a site can be in one and not the other.

Gemini

Google's developer documentation for grounding with Google Search says the feature "connects the Gemini model to real-time web content" and returns annotations that "link a text segment" to a source URL. The consumer Gemini app cites web pages when it draws on Search, and Google's indexing rule above is the one to work from.

Claude

Anthropic's web search tool documentation states "Citations are always enabled for web search", with each citation carrying the source URL, its title and up to 150 characters of the cited content. Developers can pass an allowed_domains or blocked_domains list, so a product built on Claude can restrict citations to a chosen set of sites before the model sees a result.

Smaller engines worth knowing

You.com lets a user pick between AI modes and has moved towards enterprise search infrastructure. Phind is tuned for programming questions. Kagi is a paid, ad-free search engine with AI features on top. NotebookLM answers only from documents you upload, which makes it a useful control: it shows what an assistant does when retrieval is restricted to a known corpus.

How do AI models decide what to cite?

Nobody outside the vendors can decompose the ranking, and we say so rather than imply we have read it. What the published measurement studies agree on is the shape.

Being referenced by other sites correlates with being cited. SE Ranking's study of ChatGPT citation factors (24 November 2025), across 129,000 domains and 216,524 pages in 20 niches, found "Sites with over 32K referring domains are 3.5x more likely to be cited by ChatGPT than those with up to 200 referring domains." That is correlation. No published study shows the mechanism.

Citation is not the same contest as ranking. The Washington University study found that of the domains an AI Overview cites, "nearly 30% do not appear in those results at all", meaning the first-page results shown beside the overview. The authors read that as "a source selection mechanism distinct from Google's ranking algorithm."

The platforms lean on different source families. Ahrefs' analysis of Google AI Mode (published 2 September 2026) put reddit.com at 17.9% mention share and youtube.com at 17.8%, where mention share is "a domain's citations as a percentage of the summed citations of the top sources". ChatGPT leans on Wikipedia. The mix differs; the set of domain types is stable across studies. LLM citations compares five of those studies and explains why their percentages disagree with each other.

Beyond the link profile, the content signals that recur are unremarkable: the page answers the question directly, the claim can be checked against another source, the page is current where currency matters, and the headings match the question asked.

How do you get cited by an AI search engine?

Four things, in the order they bind.

Get indexed. Google's rule is the floor, and the equivalent applies to Copilot through Bing. Check the property in Search Console and Bing Webmaster Tools before optimising anything else.

Answer the question in the first paragraph. Retrieval pulls passages. A paragraph that answers the question without the rest of the page is a passage an assistant can lift; a paragraph of preamble is not.

Earn references from sites the engines already read. Links are one of the things that put a page in the index, and the studies above show the correlation with citation. Do backlinks matter for AI search states the honest limit of that claim, and how to get cited by ChatGPT covers the page-level work.

Let the crawlers in. Each assistant fetches with its own user agent, and a robots.txt written for Googlebot can block the rest by accident. AI crawlers and robots.txt lists the agents and what blocking each one costs.

The part that takes years is the third. A new domain has been referenced by nobody, and no amount of formatting substitutes for that.

Using the Revised directory for this

Start from the citing side rather than the domain side.

1. Open the source hubs. Each one groups expired domains by a trusted site that still links to them, so you can start from a source family the studies show the assistants read.
2. Narrow to the families that recur in those studies: Wikipedia for encyclopaedic references, .edu and .gov for institutional pages, Reddit for long-lived community threads.
3. Set the Agent Citability filter in the left rail. Agent Citability is a 0 to 100 measure of how deeply the trusted layer of the web references a domain. Its methodology is published, the weights are not, and a score can fall as well as rise between releases.
4. Read Linked by before Revised Score. The named sources tell you whether the links came from the kind of page an assistant retrieves.
5. Check the archived history and the spam screen on the listing page. As of 16 September 2026 the directory holds roughly 11,000 expired domains, 84% of which pass a low spam screen, and every listing is spam-screened before it goes up.

Searching, filtering and sorting need no account, and free listings publish their name outright. Run a shortlisted domain through the agent citability checker and the backlink checker before you commit to anything.

FAQ

What is an AI-powered search engine? A search product that answers a question in prose and cites the pages it read, instead of returning a ranked list. Underneath it is an ordinary search index feeding a language model, which is why being indexed is the precondition for everything else.

Do AI search engines use Google's index? Some do. AI Overviews, AI Mode and Gemini's grounding run on Google's. Copilot runs on Bing's. Perplexity and ChatGPT use their own crawling plus third-party search, and the details change. Being indexed by both Google and Bing covers most of it.

Can I submit my site to ChatGPT or Perplexity? No. Google's documentation says there are no additional requirements and no special optimisations for AI Overviews, and no vendor publishes a submission route. What you control is whether the page is indexed, crawlable by each assistant's user agent, and referenced by sites the engines already read.

Which websites do AI search engines cite most? It depends on the engine. Ahrefs measured reddit.com at 17.9% mention share in Google AI Mode in September 2026; Profound measured Wikipedia at 7.8% of ChatGPT citations over 2024 to 2025. Community sites, reference works, video and institutional pages recur across every study.

Does AI search replace SEO? Not so far. Retrieval sits on a search index, so indexing, crawlability and references from other sites still decide whether you are a candidate. What changes is the unit of work: a passage that answers one question rather than a page that ranks for a phrase.