How Can You Tell Whether LLMs Use Your Content or Only Index It?
You can tell whether LLMs use your content or only index it by checking two things separately. First, check whether your pages are in the places AI systems draw from: the search indexes behind each assistant and, for training, public crawl datasets. Second, check whether answers actually reflect your pages. That means tracking citations, asking questions that can only be answered with facts found on your site, and reading the first-party reports from Bing and Google. A page can be indexed for months and never shape an answer. It can also shape answers without being cited, when an assistant paraphrases your facts or cites a copy elsewhere. Asking a model "what do you know about my site?" mixes all of these up and often produces confident guesses.
What is the difference between indexed and used?
There are several stages between your page and an answer:
| Stage | What it means | How you can see it |
|---|---|---|
| In training data | The text was collected and may have shaped the model's knowledge | Partly, through public crawl datasets; never with certainty |
| In a search index | The page can be found by an assistant's search step | Webmaster tools, site searches, crawl logs |
| Retrieved | The page was fetched or read for a specific question | User-triggered fetches in logs |
| Used | Its facts or wording shaped the answer | Unique-fact probes, answer comparison |
| Cited | The answer links to or names the page | Citation tracking, first-party AI reports |
Most sites have far more pages indexed than used, and more used than cited. The difference between crawling and citation is covered in whether being crawled by an AI bot means your pages will be cited.
How can I find out what content AI systems have actually indexed about my website?
Each system has its own index, and only some offer tools:
- Google. Search Console's page indexing report and URL Inspection show which pages Google has indexed. AI Overviews and AI Mode draw on the same index.
- Bing. Bing Webmaster Tools has URL Inspection and index reports. Bing's index supports Microsoft Copilot and is used by other services that license Bing results.
- Brave. Claude's web search uses Brave. There is no webmaster console, but searching Brave for your domain or specific page titles shows whether pages are findable.
- OpenAI, Perplexity and Anthropic's own crawlers. None offers an index report. Server logs show whether OAI-SearchBot, PerplexityBot or Claude-SearchBot fetch your pages and receive a normal response, which is the best available sign that pages can enter their indexes.
- Training data. Common Crawl, a public web archive used in many language model training datasets, has a public index. You can look up which of your URLs it captured and when. That shows the page was available to model builders who use Common Crawl, not that any model retained it.
Retention in training is the least visible layer. A study of model knowledge cutoffs found that the effective cutoff often differs from the stated one and varies by topic. Older copies of pages in newer crawls are one reason. So even a page captured before a model's cutoff may not be reflected in what it knows.
How do you test whether your content is used?
Combine several methods, since each sees a different part:
- Track citations. Run a fixed set of buyer questions in each assistant regularly and record which of your pages are cited. This is the most direct evidence of use.
- Probe with unique facts. Choose facts that appear only on your site, such as a figure from your own data, a specific policy detail or a named method. Ask questions whose correct answer needs that fact. If the answer contains it, your page, or a copy of it, was used. If it cites you as well, it was used and credited.
- Test direct access. Ask an assistant with browsing to summarise a specific URL. A correct summary shows the page can be fetched and read, though not that it would be chosen for a general question.
- Read first-party reports. Bing Webmaster Tools' AI Performance report shows which pages Copilot and Bing's AI features cited, with the grounding queries used to find them. Google Search Console's generative AI performance report shows impressions from AI Overviews and AI Mode by page, though not the queries or answer text.
- Watch user-triggered fetches. Visits from ChatGPT-User, Claude-User or Perplexity-User show that an assistant opened a page to answer someone's question at that moment.
- Check AI referrals. Google Analytics 4's AI Assistant channel shows visits from assistant links, which confirms at least some citations are being clicked.
What does "used but not cited" look like?
Some use never shows up as a citation:
- Paraphrased facts without a link. The answer states your figure or your method in different words and cites someone else, or nobody.
- Your brand named without your site cited. Assistants often name companies without linking to them, and the reverse also happens.
- A copy cited instead of the original. Syndicated or scraped versions of your content can be cited in place of your page.
- Training-based answers. When an assistant answers from its training without searching, it may reflect your content with no citation at all.
Unique-fact probes are the main way to catch these. If your fact appears but your page does not, look for where the assistant found it.
Why is asking the model directly unreliable?
Questions such as "What pages from example.com do you know?" or "Have you been trained on my site?" produce unreliable answers. Models do not have a list of their training documents, and they will often invent plausible page titles or summaries. When search is on, the answer reflects a fresh search, not what was already known. Direct questions are useful for one thing: seeing how an assistant currently describes you. They cannot tell you what is indexed or used. The limits of this approach are covered in whether asking LLMs directly is a valid way to measure visibility.
What might an illustrative check look like?
Consider an illustrative cycling accessories retailer that publishes a detailed guide to fitting bike racks on e-bikes. It includes its own compatibility table covering 40 e-bike models. Search Console and Bing Webmaster Tools show the guide is indexed. Logs show OAI-SearchBot and PerplexityBot fetching it. Citation tracking for "which rear racks fit [e-bike model]?" finds the retailer cited in Perplexity and Copilot, and Bing's AI Performance report confirms Copilot citations. ChatGPT answers correctly for several models but cites a cycling forum that copied part of the table. The retailer asks the forum for attribution and adds a clear source line and date to its table. It also keeps tracking the same model-specific questions each month.
What do people ask about indexing and AI use?
Can I see exactly what ChatGPT learned from my site?
No. No model provider offers that. You can see whether your pages were in public crawls and test what answers reflect.
Is being in Google's index enough for AI Overviews?
It is required but not enough. Pages still need to answer the question better than other indexed pages.
Does Common Crawl inclusion mean a model trained on my page?
No. It means the page was available in a dataset many builders use. Each builder filters the data differently.
How often should I test for use?
Monthly for your main buyer questions, and after any significant content change.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources