How an AI search engine picks and cites its sources

An AI search engine no longer returns just a list of links: it writes an answer and cites a few sources. This page is for B2B marketing and SEO leads who want to understand how ChatGPT, Perplexity, Gemini, Copilot, Le Chat, AI Overviews and AI Mode choose those sources, so they can decide where to focus their effort.

It complements our guide to Generative Engine Optimization: a map of the engines, how passages are selected, ChatGPT in detail, and a method to get cited by several engines at once.

AI search engine vs traditional search engine: what changes

A traditional search engine crawls the web, indexes pages, then ranks them to present links. An AI search engine relies on the same kind of search, then a language model writes an answer in natural language and cites the sources it has kept. Users read the answer before clicking, if they click at all.

From SERP to answers: what you are now aiming for

You are no longer aiming only for the best page, but for the clearest passage, the one easiest to reuse. A brand is "cited" when it is named or linked in the answer. Four practical differences follow:

• The unit being judged: a self-contained passage rather than a whole page.
• The target outcome: being named or cited in the answer, on top of being ranked.
• Stability: the answer varies from one model to another and from one session to the next.
• Value: sometimes few clicks, but a direct influence on a B2B buyer's shortlist.

Usage is already established in France

This is no longer a forward-looking issue: 39% of French people use AI engines for their searches (IPSOS, 2026). A B2B brand missing from these answers therefore misses part of the moments when its buyers frame a need, compare solutions or build a shortlist.

Usage figures specific to ChatGPT are gathered in our ChatGPT statistics.

Mapping the AI search engines and AI assistants to watch

Talking about a "new search engine" does not necessarily mean a brand-new web index: most of these services combine a language model with an information retrieval layer. The question therefore becomes: where am I retrieved and cited as a source? The table compares the six environments to track.

Claude belongs to the same family and is tracked with the same question set. Differences between engines mostly come from the index being queried: with identical content, your odds of appearing vary from one service to another.

How generative search works: RAG, citations and sources

An AI search engine is not a "brain" that knows everything. A brand appears in an answer through two paths: knowledge learned during training, which changes slowly, and real-time retrieval, which fetches pages when the question is asked. The second path is the one your content can influence in the short term.

RAG: retrieve, select, generate, cite

RAG (Retrieval-Augmented Generation) follows a stable sequence. Each step filters, and each filter is a lever for you:

What weighs in source selection

From one engine to another, the same criteria come back when two passages need to be separated:

• Indexing: a page missing from the index being queried does not exist for the engine.
• Extractability: a passage that keeps its meaning once isolated is reused without distortion.
• Authority: mentions and links from sites that AI engines already cite count.
• Freshness: dated, maintained information beats an outdated version.
• Coherence: what the web says about the brand must match your pages.

Coherence is largely decided outside your site: 48% of AI citations come from community platforms (State of AI Search, 2025). Your presence on public sources that engines retrieve therefore matters as much as your own pages.

Citations, biases and answers without sources

In reality, generative search remains inconsistent. Depending on the query, context or product policy, you may see links, partial citations or no explicit sources at all. Three biases explain most of the gaps:

• Selection bias: already popular sources are over-represented versus specialised sources.
• Format bias: structured content (lists, tables, short definitions) is easier to extract.
• No sourcing: the system answers "from memory" or considers the question too broad, and the risk of approximation increases.

Reliability: hallucinations and vague content

A model can confidently state a wrong answer, especially when its data is outdated: this is a hallucination. Retrieving external documents reduces this risk without eliminating it, and a vague passage can become a certainty in your brand's name. So publish content that is dated, maintained and traceable: the more cross-checkable a piece of information is, the better it is reused.

ChatGPT: web search, OpenAI crawlers and the link with Bing

ChatGPT has built-in web search: when a question calls for it, it queries the web, filters the results and displays clickable citations in its answer. In a conversation, users chain follow-up questions: cover the neighbouring angles of a topic, not a single wording.

How ChatGPT builds an answer

The journey follows the mechanics described above, with steps of its own:

• Interpret the request, which is conversational and tied to previous exchanges.
• Retrieve sources through web search.
• Discard duplicates, spam and obvious contradictions.
• Keep convincing passages rather than whole pages.
• Write the synthesis and display the citations.

GPTBot, OAI-SearchBot, ChatGPT-User: three crawlers, three decisions

Access is the first thing you control. OpenAI distinguishes three crawlers, and each calls for a different decision in your robots.txt file:

• OAI-SearchBot: it crawls pages for ChatGPT search; blocking it removes you from the possible sources.
• ChatGPT-User: it visits a page in real time at a user's request; allow it if you are targeting ChatGPT.
• GPTBot: it collects content to train models; blocking it does not prevent visibility in search.

The GPTBot decision is therefore about training, not citation. Syntax and common mistakes are covered in our guide to the robots.txt file. You do not, however, control the final decision: ChatGPT may prefer another source if it judges your passage less clear or less corroborated.

The link with Bing and what it changes

Part of ChatGPT's selection happens upstream, in the index: 87% of ChatGPT citations match the top Bing results (Squid Impact, 2025). SEO on Bing becomes an entry condition:

• Check that your key pages are indexed in Bing, just as in Google.
• Apply SEO basics to Bing: useful and unique content, optimised media, accessibility.
• Track sessions from chatgpt.com in Google Analytics 4, separately from other engines.

Perplexity: an engine that shows its sources

Perplexity presents itself as a hybrid between a search engine and a conversational assistant. It runs a web search for every question and displays its sources with the answer, which makes the clarity of each passage decisive. Its crawler, PerplexityBot, must be able to access your pages. The formats reused, the page types that surface and the tracking method are detailed in our guide to SEO for Perplexity AI.

Google: AI Overviews and AI Mode

AI Overviews (formerly SGE, renamed in May 2024) and AI Mode have been available in France since summer 2026. An AI Overview places a synthesis citing links above the results; AI Mode answers in a conversation and also cites sources. Search Console counts their impressions and clicks in the Performance report, without separating them. The effect on CTR and the reuse criteria are covered in our article on Google's AI Overviews and AI Mode.

Getting cited by several engines without duplicating your content

Each engine has its own habits, but the fundamentals remain stable: an indexed, accessible page, strong SEO, extractable passages, evidence and external brand credibility. One page per engine therefore makes no sense. Start from one pillar page per topic, then build pages by intent (definition, comparison, implementation, mistakes), each with a distinct angle: this is the best answer to the sub-queries generated by assistants.

What engines reuse most

Generative engines favour what they can isolate and reuse without distorting meaning:

• Entities: concepts, categories, acronyms and market players, defined and contextualised.
• Data: dated figures, with their public source.
• Procedures: numbered steps, each with a validation criterion.
• Evidence: examples, operational definitions, limitations and conditions of application.

Open each section with a sentence that can stand on its own, then expand, without unverifiable superlatives.

Measure engine by engine, never as an average

Presence varies from one model to another and from one session to the next. An average mixing ChatGPT, Perplexity and AI Overviews therefore hides the gaps that matter. The protocol rests on four rules:

• Ask each engine the same fixed question set, at regular intervals.
• Record for each question whether the brand is named, the domain cited, the domain used as a source, or absent.
• Read the results engine by engine, without adding them up.
• Compare each page with itself before / after, as correlation and never as causation.

The full set of indicators (presence, citations per question, AI share of voice, traffic, conversions) is defined in our article on GEO KPIs.

Reading Search Console signals

On the Google side, the first signals show up in Search Console, read by cohorts of pages (guides, comparisons, product pages) rather than overall:

Google Analytics 4 then shows whether the click drop actually affects conversions, and which ones.

Prioritising the pages that deserve to be cited

A page deserves the effort if it meets at least two criteria: strong demand, high business value, and the ability to provide evidence. Ask yourself a simple question: if an AI had to summarise this topic, which parts of my page would be worth citing? Three page types respond best:

• Definition and criteria (TOFU): to capture framing citations.
• Comparison and limitations (MOFU): to influence the shortlist.
• Method and steps (BOFU): to become the procedural source.

Depending on your situation

Your starting point determines which page to read next:

• You are new to the topic: the GEO definition sets out the principles and what the term does not cover.
• You want to take action: the GEO referencing method starts with the presence audit and the choice of questions.
• You are weighing SEO against GEO: our GEO vs SEO comparison shows what they share and where they differ.
• Your competitors are cited and you are not: the article on backlinks in the AI era explains how external mentions weigh in answers.

To choose where to focus effort, the Incremys opportunity analysis module distinguishes, at the writing stage, keywords, which drive a SERP and an AI Overview, from questions, which drive an answer requested from ChatGPT. Presence on those questions is then measured after publication, and each page is compared with itself before / after.

FAQ on AI-powered search engines

What is an AI search engine?

An AI search engine is a service that answers a question with a written synthesis, based on sources it cites. It combines a web search, the selection of relevant passages and a language model that writes the answer. ChatGPT, Perplexity, Gemini, Copilot, Le Chat, AI Overviews and AI Mode are the main examples.

How is an AI search engine different from a traditional search engine?

A traditional engine ranks pages and presents links, whereas an AI search engine writes an answer and cites a few sources. The unit being judged becomes the passage rather than the page. The target outcome changes too: being named or cited in the answer, on top of ranking well in the results.

Does ChatGPT search the web?

Yes, ChatGPT has built-in web search and uses it when a question calls for it. It retrieves pages, keeps passages from them, writes its synthesis and displays clickable citations. For a page to be selected, it must be indexed and accessible to its crawlers, in particular OAI-SearchBot and ChatGPT-User.

Which sources does ChatGPT cite?

ChatGPT cites pages it has retrieved through its web search, and these pages largely match the top Bing results. It favours passages that are clear, verifiable and attributable: definitions, lists, tables, direct answers, with an identifiable author and date. Checking that your pages are indexed in Bing is therefore part of the work.

Do Perplexity and ChatGPT cite the same sites?

No, Perplexity and ChatGPT do not always cite the same sites, because they do not query the same index and do not select passages in the same way. A brand can be named in one and absent from the other. Presence is therefore measured engine by engine, on a fixed question set, without mixing the results.

Is AI Mode an AI search engine?

Yes, AI Mode is Google's AI search engine: it answers in a conversation and cites sources. Available in France since summer 2026, like AI Overviews, it relies on the Google index. Search Console counts its impressions and clicks in the Performance report, without separating them from those of AI Overviews.

Should you allow GPTBot?

Allowing GPTBot is not necessary to be cited in ChatGPT search, because this crawler collects content to train models. Blocking it is a decision about training. For search, OAI-SearchBot and ChatGPT-User are the crawlers that must be able to access your pages through your robots.txt file.

How can you measure visibility inside AI answers (not just on Google)?

Visibility in AI answers is measured on a fixed question set, asked to each engine at regular intervals. For each question, you record whether the brand is named, whether the domain is cited, whether it appears as a source or whether it is absent. Search Console and Google Analytics 4 complete this tracking on the traffic and conversion side.