How AI Search Picks Sources: From Query Rewriting to Citation

Every AI answer passes through up to four stages. Most confusion about AI source selection comes from mixing them up.

A page can only be cited if it clears access and retrieval first. Most published GEO experiments only test the last stage, on pages already handed to the model. Keep that in mind when you read any claim about what gets cited.

Retrieval is not automatic. As of October 2026, every major engine decides per query whether to search the web.

• ChatGPT searches at the model's discretion.
• Google AI Overviews often do not trigger at all, according to Google's own documentation.
• Claude.ai search is a user toggle, and Anthropic publishes no default.
• Perplexity's Agent API only searches when the developer turns web search on.

When no search happens, the answer comes from training data frozen at a knowledge cutoff. Nothing you publish this quarter reaches that answer. This is why the same question can return a cited answer one day and a sourceless one the next. If you see answers about your brand with no links, read why some ChatGPT answers have no sources.

The practical point: any visibility number blends answers given with and without live search. Treat a single score as an average, not a verdict.

When an engine does search, it rarely sends your words to an index unchanged.

What each engine documents

• OpenAI's help page says ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to search providers.
• Google's AI features documentation says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics.
• Google's Gemini API documentation says grounding "automatically generates one or multiple search queries" and runs them.

OpenAI also says saved memories can shape the rewrite when memory is on. So two people asking the same thing can trigger different searches.

Why this matters for you

The queries that actually retrieve pages are generated by the model. You cannot see them for ChatGPT, Gemini, Claude or Perplexity. The one exception is Bing Webmaster Tools. Its AI Performance report, in public preview since February 2026, lists grounding queries: the phrases the AI used when retrieving content it later cited in Copilot and Bing answers. Microsoft notes this data is a sample, and the report has no API.

Every prompt-tracking tool, including any that tests the questions your buyers type, measures a proxy for an input it cannot observe. That is still useful. Just do not mistake it for the real query log.

Each rewritten query goes to an index. The engine collects a set of candidate pages and passes some of them into the model's context.

Which index matters. Google's AI Overviews and AI Mode draw on Google's own index. Google states a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link, and that there are no additional technical requirements. Copilot grounds on Bing. ChatGPT search sometimes partners with other providers, per OpenAI. The platform guides cover each one: how Google picks AI Overview supporting links, where ChatGPT citations come from and how Perplexity chooses sources.

Organic rank is a loose proxy for entering the set, not a guarantee. Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the top 10 results, in a study published in March 2026. The rest came from lower positions or from pages outside the top 100.

Retrieval also depends on reach. A crawler that cannot fetch your page, or fetches it before JavaScript runs and finds nothing, never adds it to the index. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but do not execute them.

Once pages are in the model's context, the model writes the answer and attaches citations. This is where most GEO research happens.

Topical relevance and position lead

• A 252,000-trial factorial study (arXiv 2605.25517) found that "topical relevance and list position are the biggest drivers of being cited first."
• C-SEO Bench (NeurIPS 2025) found that improving a source's ranking in the model's context was "significantly more effective" than rewriting the content, and that most rewriting methods were ineffective or harmful.
• A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) named topical relevance and context position as the most reproducible levers, and warned that citation-oriented rewrites can impair retrieval.

Why position is not your lever

In C-SEO Bench, the "SEO" arm worked by moving the target document to position 1 in the model's context. No index was queried and no ranking was earned. The factorial study used only two candidate pages per trial on a synthetic corpus. Both show position matters. Neither shows a publisher can set it. Only the engine's retrieval system orders the candidates.

Topical relevance is different. It means your page actually covers what the generated query asks, not that it repeats keywords. The original GEO paper found keyword stuffing lowered its own visibility metric. More on this in whether staying on topic helps AI choose your page.

Cited is not the same as used

OpenAI's API returns the full list of retrieved sources separately from citations, and notes the retrieved list is often longer. Anthropic cites short text spans, while Perplexity cites whole sources by number. Mechanistic research (arXiv 2606.28358) repaired over 90% of missed citations by changing the model's internals, without changing the sources. Citation counts are a noisy proxy for influence.

Be plain with yourself and your team about the split.

In your control

• Crawler permission. robots.txt rules are per bot. OpenAI says each setting is independent: OAI-SearchBot governs ChatGPT search, while GPTBot governs training only.
• Network reachability. A firewall or CDN can block a crawler that robots.txt allows. OpenAI's help page says eligibility needs both crawl permission and a host that allows its search bot's IP addresses.
• Rendering. Content that only appears after JavaScript runs may be invisible to most AI crawlers. Googlebot is the main exception.
• Index and snippet status. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers, sometimes by accident.
• Specific facts on fetchable pages. A stated price, number or one-line definition gives an engine a sentence it can quote.

Not in your control

• Whether the engine searches for a given query.
• How it rewrites the query.
• The order of candidates in the model's context.
• Which retrieved page the model chooses to cite.

These access levers work by removing blocks. They make you eligible. They do not promise retrieval or citation.

Work from the first gate outward. A fix at a later stage is wasted if an earlier gate is closed.

1. Check crawler access per bot. Confirm robots.txt lets OAI-SearchBot, PerplexityBot and Claude-SearchBot in. Then request a page as each crawler to see whether your firewall returns a 403. The AI Crawler Checker does this request for you.
2. Check rendering. Compare your raw HTML with the rendered page. Move prices, product names and key claims into the server-rendered HTML.
3. Check index status per engine. Look at Google Search Console and Bing Webmaster Tools separately. One being healthy tells you nothing about the other.
4. Audit suppression tags. Search templates for stray noindex, nosnippet or data-nosnippet.
5. Read your grounding queries. If you use Bing Webmaster Tools, export its grounding queries and compare them with the questions you assumed buyers ask.
6. Find the third-party pages engines cite. When engines cite other sites instead of yours, being named on those sites is associated with appearing in the answer. Start with finding the pages AI engines cite.
7. Measure per engine, and log retrieved versus cited. The retrieved vs cited logging playbook shows how to capture both through the APIs. For a quick snapshot, the free checker on this site shows whether AI engines name you for up to three buyer questions.

• Treating your prompt list as the real query. Engines rewrite and fan out. Your tracked prompts are a proxy.
• Allowing GPTBot to get into ChatGPT search. GPTBot governs training. Search uses OAI-SearchBot.
• Adding schema or FAQ schema to get cited. Google says structured data isn't required for generative AI search, and an Ahrefs matched test found AI Overview citations fell 4.6% after pages added it. Existing markup does no harm, but it is not a lever.
• Rewriting for an AI-friendly style or chunking pages. Google's optimization guide says you don't need to write in a specific way for generative AI search. The survey above warns citation-oriented rewrites can hurt retrieval.
• Tracking an AI rank. Google exposes AI impressions only, with no position data to check a rank against.
• Trusting one blended score. Engines differ in index, model and citation format. Score each one on its own.
• Planting hidden instructions for AI crawlers. Prompt injection and cloaking exist, and research shows some agents are vulnerable to them. They are manipulation, they put your reputation at risk, and you should never use them.

Can I see the search queries ChatGPT actually ran for my brand? Not for ChatGPT itself. OpenAI says ChatGPT rewrites prompts into targeted queries but does not show them to site owners. As of October 2026, the only first-party source of real grounding queries is the AI Performance report in Bing Webmaster Tools, which covers Copilot, Bing AI summaries and select partners. It is a sample and has no API, so export it manually.

Does ranking in Google's top 10 mean AI Overviews will cite me? No. Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the top 10 organic results, in its March 2026 study. Most citations came from elsewhere. Google requires a page to be indexed and snippet-eligible to appear as a supporting link, but says eligibility does not guarantee inclusion. Rank is a loose proxy, not a route.

Is being retrieved the same as being cited? No. OpenAI's API returns retrieved sources and citations as separate lists, and notes the retrieved list is often longer. A page can enter the model's context, shape the answer and still go uncited. Logging both through the APIs gives a fuller picture than counting citations alone, though neither count measures influence directly.

Can I improve my position in the model's context window? Not directly. Studies like C-SEO Bench show position strongly affects citation, but they test it by placing a document first in the context by hand. In live engines, the retrieval system orders candidates, and no publisher setting changes that. What you can do is make sure crawlers reach your page, index it, and find content that genuinely matches the topic.

Why do AI engines sometimes cite pages with no Google ranking? Engines draw on different indexes. Copilot grounds on Bing, and Perplexity and Claude use their own retrieval paths. In Otterly's 2026 experiment, Copilot kept citing two sites after Google deindexed them, because Bing still held them. Google's own AI surfaces almost stopped citing them. Check your index status in each engine separately.