How each AI engine sources and cites content

Every AI answer engine you want to be cited by reads from a different library. ChatGPT pulls its live web results from Bing. Microsoft Copilot does too. Google’s AI Overviews, AI Mode, and Gemini all use Google’s own index. Perplexity runs its own crawler on top of search APIs. And Claude’s web search reads from Brave’s independent index. So “I rank on Google” tells you almost nothing about whether ChatGPT or Claude can quote you.

This matters more every month: ChatGPT alone reached roughly 900 million weekly users in early 2026 and crossed a billion monthly app users by June (ALM Corp). Here’s how each engine finds and cites content — and, because this is a site about rigour, a confidence rating on every claim based on how well independent sources agree.

The short version

All of them cite and link sources. The catch is that they read from different indexes — so Bing, Google, and Brave all matter, and the third one is the one almost nobody optimises for.

ChatGPT: it runs on Bing

ChatGPT’s live web search is built on Bing’s index, so if Bing hasn’t indexed your page, ChatGPT can’t cite it — no matter how good the page is. It’s technically a hybrid: Bing’s index is foundational, and OpenAI’s own crawler (OAI-SearchBot) supplements it with fresh fetches of specific pages (The Keyword, Stackmatix). High — this is consistently reported across independent sources.

The crawler names matter. GPTBot gathers training data; OAI-SearchBot fetches pages for live search citations; ChatGPT-User is the agent that visits when a user clicks through. Blocking GPTBot removes you from future training but does not block live ChatGPT search — that depends on OAI-SearchBot and Bing (Capston.ai). High

The move: verify in Bing Webmaster Tools, submit your sitemap, and switch on IndexNow so new pages reach Bing in minutes rather than weeks.

Google AI Overviews, AI Mode, and Gemini: Google’s index

All of Google’s AI surfaces — AI Overviews, the AI Mode tab, and Gemini — draw on Google’s existing search index. High What’s not settled is how much your organic ranking actually drives AI citations.

Here the data genuinely conflicts. BrightEdge measured AI Overview overlap with organic rankings growing to about 54%. But more recent 2026 analyses report the overlap with top-10 results falling — to somewhere between 17% and 38%, with one study claiming a collapse from ~70% to under 20% — a shift several analysts tie to Gemini 3 becoming the default for AI Overviews in late January 2026 (BrightEdge, ALM Corp). Low — sources conflict The honest read: ranking well still helps, but in 2026 it’s no longer a reliable proxy for being cited, and a growing share of citations comes from pages ranking outside the top 10.

Perplexity: its own crawler, in real time

Perplexity retrieves fresh results on essentially every query and cites more sources than most engines — around 8 per answer, typically in the 5–12 range (MarGen). (An earlier figure I’d seen put this near 22; the better-supported 2026 number is roughly 8, so treat the exact count as Medium — it varies with query complexity.) It runs its own crawler (PerplexityBot) layered on search APIs, so new content can surface within hours of being indexed — a faster loop than training-based systems. High on the mechanism, Medium on the exact source count.

Because Perplexity cites widely and leans on community platforms, it rewards two things: genuine freshness, and a presence on the sources it trusts (Reddit, YouTube, Wikipedia, official docs).

Claude: the Brave index nobody optimises for

Claude’s web search reads from Brave’s independent index — not Google’s and not Bing’s — so strong Google rankings don’t guarantee Claude can see your page. This is now well-evidenced: Anthropic lists Brave Search on its official subprocessor list, and an independent test found an 86.7% overlap between Claude’s cited results and Brave’s top organic results (TechCrunch, RivalHound). Brave’s index covers 30+ billion pages with its own coverage gaps. Med-high — strongly evidenced, though Anthropic hasn’t made it a headline announcement.

Anthropic separates training crawls (ClaudeBot) from live-search fetches (Claude-SearchBot); allow both. The takeaway: if Claude citations matter, check whether Brave can find your pages — a separate question from Google and Bing.

What this means for you

You can’t “optimise for AI” as one thing, but the differences collapse into a short checklist:

• Get indexed in all three: Google, Bing, and Brave. The first two cover ChatGPT, Copilot, and Google’s AI; Brave covers Claude. Most people do two and forget the third.
• Allow the live-search bots, not just the training bots: OAI-SearchBot, PerplexityBot, Claude-SearchBot, Google-Extended.
• Publish fresh, keep it fresh. Real-time engines lean hard on recency.
• Earn mentions on the sources each engine trusts — Reddit, Wikipedia, YouTube, G2. Being cited by a third party tends to drive more AI visibility than your own pages do.

How fast, and how stable?

A well-structured page indexed in Bing can be cited by ChatGPT within roughly a week — one benchmark puts the median time-to-first-citation near 7 days, with a long tail past a month (AuthorityTech). Medium — it’s a single benchmark, though it lines up with the wider 30–90-day window most sources cite. Training-data citations are slower — months, because they only update when a new model trains.

And citations are volatile: the set an engine cites for a question churns 40–60% month to month, even though week-to-week it’s surprisingly stable (Similarweb). Med-high You don’t win a position and hold it; you re-earn it each cycle — which is exactly why freshness matters.

Common questions

The takeaway

There is no single “AI search.” ChatGPT and Copilot read from Bing; Google’s AI reads from Google; Perplexity reads from its own crawler; Claude reads from Brave. Get indexed in all three underlying systems, allow the live-search bots, publish fresh and genuinely useful content, and get mentioned where these engines already look. That makes you eligible everywhere — the precondition for being cited anywhere.

I’m testing how much each lever actually moves the needle. Results land on the experiments log as they come in.