How to Tell if an AI Assistant is Citing You From Training Data or Live Retrieval

A prospect tells your sales team they "saw us mentioned in ChatGPT." Your CMO wants to know what worked. Before anyone touches a content calendar, there's a more basic question to answer: was that mention pulled from a live web page your team can influence this quarter, or from a frozen snapshot of the internet that won't update until the next model release?

Those are two different systems with two different playbooks. Confusing them wastes budget: teams chase backlinks to fix a retrieval problem, or rebuild a page's technical SEO to fix a training-data gap that only closes when a model retrains. For a B2B SaaS marketer trying to get cited alongside (or instead of) competitors, the diagnostic comes first. Here's how to run it, with the actual mechanics named platform by platform.

Two Completely Different Ways an AI Assistant "Knows" Your Brand

Every large language model carries two separate sources of information, and they behave nothing alike.

The first is parametric memory: the patterns baked into the model's weights during training. Nothing is looked up when the model answers from this layer. It's prediction based on what the model absorbed before its knowledge cutoff, and it doesn't change again until the next training run.

The second is live retrieval, sometimes called grounding or RAG (retrieval-augmented generation). The model recognizes that a question needs current or specific information, issues one or more search queries against a live index, reads what comes back, and writes an answer from that fetched content. This is the layer that produces the clickable citations you actually see.

A brand's presence in a model's trained knowledge shows up without a citation shown or a source URL displayed, while live retrieval triggers a web search and citations show up as clickable footnote-style links. Same brand, same model, two entirely different mechanisms depending on which layer answered the question. Understanding AI search visibility means treating these as two separate tracking problems, not one.

The Clearest Tell: Is There an Actual Clickable Citation?

Start here before anything more technical. If the assistant's answer includes a numbered footnote, an inline bracketed link, or any clickable source attached to the specific claim about you, that's retrieval. If the model states something about your company with no link, no footnote, and no visible source, it's answering from training data, or in ChatGPT's case, possibly a licensed publisher deal running in the background.

Two things complicate this simple check, and both matter for a marketer trying to be precise:

• Memory and custom instructions can shape an answer without any web page being fetched. ChatGPT can carry information across conversations, both as explicit "saved memories" and as a synthesis of past chats, so earlier context can quietly inform new answers, and this is private to that user's account, not a lever for public brand visibility.
• A citation being present doesn't guarantee the underlying claim is accurate. Click through. Confirm the URL is actually yours and that the cited passage supports what the model said, not a page that happens to mention your category.

With that groundwork set, the mechanics differ enough by platform that a single test won't cover all three.

ChatGPT: A Frozen Model Layered With Bing as the Named Partner, Google in the Mix Too

The underlying model has a fixed knowledge cutoff, and for everything after that the base model either doesn't know or confabulates based on patterns, which is where hallucinations about recent events come from. ChatGPT Search gives the model real-time web access. OpenAI's own documentation names Bing and Shopify as its third-party search providers and adds that ChatGPT search "sometimes partners with other search providers" without saying who. That caveat is doing real work: independent testers, including a former Google engineer and Backlinko, got ChatGPT Plus to cite pages that were indexed only in Google, not Bing. Nick Turley's court testimony confirms ChatGPT's search draws primarily on Bing, that OpenAI approached Google directly for a licensing deal and was turned down, and that OpenAI is building its own first-party search index, a project Turley said is now years behind its original goal of covering 80% of searches by the end of 2025. Treat Bing as the backend you can audit through your own server logs, and treat a Google-sourced citation as real but harder to trace to a specific mechanism.

The decision to search isn't a fixed rule; it's a probabilistic judgment the model makes per query, and the rate at which it fires has been moving. A Semrush analysis of more than a billion lines of U.S. clickstream data found that ChatGPT enabled its search feature on just 34.5% of queries as of February 2026, down from 46% in late 2024, meaning most responses still rely on training data alone. A separate Profound analysis of roughly 700,000 U.S. conversations found that about 18% trigger at least one web search, a rate that held steady across the three months of the study. The exact figure moves depending on how each firm samples its data and when, but both point the same direction: most ChatGPT turns never touch the live web.

A UK-based research team went further and instrumented this directly. By logging ChatGPT's hidden turn-level metadata through a custom Chrome plugin, researchers at SALT.agency captured two fields attached to every message: turn_use_case, which records how ChatGPT classified the turn, and result_source, which records which retrieval backends supplied results. A third of the classified turns in their session never touched the web at all, typically definitional, conversational, or evergreen questions, where no amount of crawling, structured data, or ranking strength puts a brand in the answer, because the answer comes straight from training data.

You can force the split yourself. Click the globe icon before sending a prompt, or add phrasing like "search the web for" or "as of today" to push ChatGPT into search mode deliberately. Run the same question both ways and compare. If the answers diverge, or the non-search version is outdated, you've confirmed which layer is doing the work.

On the crawler side, OpenAI's crawler documentation describes three distinct bots, each controllable independently via robots.txt:

• GPTBot. Used for AI training.
• OAI-SearchBot. Used for ChatGPT search.
• ChatGPT-User. Used for direct user requests. Because these fetches are triggered by a specific user action, OpenAI states that robots.txt rules may not always apply the same way to that agent as they do to the other two.

Checking your own server logs for OAI-SearchBot and ChatGPT-User activity around the time you ran a test query is the closest thing to hard proof that a live fetch actually happened.

Claude: Training Corpus First, Web Search as a Deliberate Tool Call

Anthropic's web search tool documentation is unambiguous about the citation mechanism: Claude decides when to search based on the prompt, the API executes the searches, and provides Claude with the results. Claude cites sources when its web-search tool has been invoked, meaning the model has retrieved live web content during the response. Outside web-search mode, Claude generates from training weights without retrieving external content, and there are no formal citations even though the response may mention brands or sources from training.

The behavioral difference from ChatGPT is the part worth internalizing. Where ChatGPT's default disposition is to reach for its browse tool on many query types, Claude leans more heavily on its training corpus and invokes its web-search tool more selectively, so brand presence inside Claude's answers depends substantially on whether the brand's content was absorbed during training, and only secondarily on tactical browsing-driven citation tactics.

Anthropic hasn't officially named its search provider, but the technical trail is unusually well documented for a "we don't confirm" situation. TechCrunch reported that Anthropic added Brave Search to its subprocessor list on March 19, 2025, one day before Claude's web search launched. Developer Simon Willison independently verified the connection the same week, finding a parameter named BraveSearchParams inside Claude's search tool definition and citations that matched Brave's own results for the same query. That evidence chain is strong, but it isn't the whole picture: Anthropic's subprocessor list also added a second "Web Search" entry, TurboPuffer, on May 6, 2026, a vector and full-text search database whose specific role in the pipeline Anthropic has not explained. That second entry is a sign the retrieval layer behind Claude isn't fully mapped by outside researchers, so treat the Brave connection as a strong, verified inference about part of the pipeline rather than the complete mechanism.

Claude's crawler structure mirrors OpenAI's split. Anthropic's crawler documentation describes three:

• ClaudeBot. Collects web content that could contribute to model training.
• Claude-User. Supports Claude users by accessing websites when individuals ask questions.
• Claude-SearchBot. Navigates the web to improve search result quality. Disabling it can reduce a site's visibility and accuracy in search results.

Unlike OpenAI's and Perplexity's user-triggered fetchers, Anthropic says all three of its bots, including Claude-User, honor robots.txt. If your logs show Claude-SearchBot or Claude-User activity tied to a specific test prompt, that's your confirmation the answer came from the live layer, not the training corpus.

One practical detail from Anthropic's API worth knowing: Claude's web search citations include a cited_text field alongside the URL and title, so in products built on the API you can see the exact excerpt the model pulled from your page, not just a link to it.

Gemini: The One Platform That Shows Its Work

Google exposes the training-versus-retrieval distinction as structured, inspectable data rather than just a visual cue, which makes Gemini the easiest of the three to verify with certainty.

When grounding is enabled, Gemini's API documentation explains that the model analyzes the prompt and determines if a Google Search can improve the answer, and if needed, automatically generates one or multiple search queries and executes them, then returns a response grounded in the search results with inline citations tied to specific spans of text. On this API surface, that grounding shows up as annotations containing url_citation entries, each one linking a start and end position in the response text to a source URL, alongside the underlying search queries actually run.

Vertex AI's legacy grounding documentation exposes the same mechanism through a groundingMetadata object instead:

• webSearchQueries. The searches the model actually performed.
• groundingChunks. The web sources it read, listed by URL and title.
• groundingSupports. The connection between specific text segments and those source chunks.

The trigger is a scored decision, not a toggle a user has to find. Dynamic grounding lets a developer set a threshold, so the model only searches when it determines the query needs external data, and questions answerable from training data alone can be left ungrounded. Google demonstrated why this matters with a concrete example: asked who won the 2024 Emmy for outstanding comedy series, Gemini without grounding named the 2023 winner, while the grounded version returned the correct answer with source links, because the model's training data had a cutoff before the ceremony took place.

For a marketer, the practical version of this test is simple:

• No search entry point, no source list. Treat it as training data.
• Citations and a "sources" element. That's grounding, and it means your page was actually indexed and pulled in real time.

Gemini and Google's AI Overviews also use a related technique worth knowing by name. Google's AI features documentation states that both AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources to develop a response, allowing the systems to identify more supporting web pages and display a wider and more diverse set of helpful links than a classic web search. Visibility now depends on being retrievable for the handful of sub-queries the system generates behind the scenes. A different, broader target than ranking for the literal question someone typed.

Five Diagnostic Tests You Can Run

1. Run the same prompt twice. Once as the model would answer by default, once forced into search mode (globe icon in ChatGPT, "search the web for..." in Claude, a query specific enough to trigger Gemini's grounding threshold). A material difference between the two answers tells you the default answer was training data, possibly stale.
2. Ask something that postdates the model's cutoff. A correct, current answer with no citation attached is a red flag worth double-checking against a licensed-content partnership rather than assuming retrieval happened silently.
3. Click through every citation you find. Confirm the URL is actually yours, and that the cited passage supports the specific claim. A citation on an unrelated page is a mismatch worth flagging, not a win.
4. Test in a clean, logged-out or fresh session. Ask ChatGPT directly on a fixed buyer-question panel, and note which competitors are cited alongside you or instead of you, doing this in a fresh session, since chat history skews results. The same caution applies to Claude and Gemini if you're signed in with a personalized account.
5. Check your own server logs for the specific bot, not just "AI traffic." A hit from OAI-SearchBot, ChatGPT-User, Claude-SearchBot, or Claude-User around the time you ran a test is closer to proof than inference. A hit from GPTBot or ClaudeBot only confirms training-data eligibility for a future model version, not today's citation.

Why the Distinction Changes What You Actually Optimize

The distinction changes where a marketing budget should go, and it isn't academic.

Training-corpus presence is slow and indirect. The work to get into a model's knowledge is similar across frontier LLMs, and it's built over the training cycle rather than the crawl cycle:

• Entity signals
• Named coverage
• Wikipedia and structured reference data
• Consistent brand mentions and backlinks across the open web

That's the case for earned coverage at scale, the kind of work covered in Backlinks vs Brand Mentions vs Citations, and it's the reason earning citations and brand mentions at scale matters even when no single mention produces an immediate, trackable citation.

Retrieval and grounding presence is fast and controllable by comparison. It's a function of four things you control:

• Crawlability
• Structured content
• Freshness
• Fan-out coverage: whether your pages answer the sub-queries a system generates around a topic, not just the head keyword

That's a technical and editorial checklist your team can act on this quarter, and it's worth understanding how ChatGPT and Claude actually pick their sources, a question covered in Not a Ranking Factor, before assuming a ranking win on Google translates automatically into an AI citation.

For B2B SaaS marketers specifically, both layers matter because buyer research spans both modes. A prospect asking "what is field service management software" gets an evergreen, training-data answer. A prospect asking how two competing platforms compare on pricing in 2026 is far more likely to trigger live retrieval, because the model recognizes it needs current, specific information. If your category work targets B2B SaaS marketing, planning for both query types, not just one, is the difference between showing up in the definitional answer and showing up in the comparison that actually drives a demo request.

Run the Test Before You Change the Strategy

The most common mistake is skipping the diagnostic and assuming every AI mention behaves the same way. A brand mention with no citation attached tells you something was absorbed months ago and won't move until the next training run. A citation with a clickable link tells you something on your site, right now, is retrievable, readable, and trusted enough to be pulled into an answer today.

Run the five tests above on your own brand name and your two or three closest competitors, across ChatGPT, Claude, and Gemini, before you brief anyone on a content plan. The pattern you find, mostly training data, mostly retrieval, or a split that changes by query type, should decide whether the next quarter's work goes toward earning broader mentions across the web or tightening up the technical and editorial signals that get a specific page fetched and cited. If you want a structured read on where your brand currently stands across both layers, reach out to map the growth system your market needs.