Prompt
How do I find out which sources LLMs are using to describe my company?
Latest observation
To find out which sources LLMs are using to describe your company, think in terms of “what the model likely learned from” and “what it cites at answer time.” Those are related, but not the same.
1) Check whether the model is using live web sources
Some LLMs answer from their training data only; others use retrieval or browsing and can cite sources directly.
Try prompts like:
- “What sources are you using to describe [Company]?”
- “Cite the web pages you used for this answer.”
- “List the sources that support each claim about [Company].”
If the product supports citations, it may show:
- page titles
- URLs
- snippets
- publication dates
That’s the easiest path.
2) Search where the company is mentioned on the public web
If there are no citations, the answer is usually influenced by public web content. Audit likely inputs:
- your website pages
- press releases
- news articles
- Wikipedia/Wikidata
- LinkedIn company page
- Crunchbase / PitchBook / similar directories
- review sites
- forums and Reddit
- job postings
- government filings
- app store listings
- developer docs / GitHub repos
Search:
- your company name + “about”
- your company name + leadership names
- your company name + competitors
- your products + key features
- your brand name + “scam”, “lawsuit”, “review”, etc.
This helps identify what the model may have absorbed or retrieved.
3) Ask multiple LLMs the same question and compare outputs
Use several systems, for example:
- ChatGPT
- Claude
- Gemini
- Perplexity
- Bing/Copilot
Ask:
- “Describe [Company] and include sources.”
- “What are the top sources you used?”
- “What facts about [Company] are you confident about, and which are uncertain?”
Compare:
- which facts repeat
- which sources recur
- which claims appear without sources
Repeated mentions often indicate influential sources.
4) Test for source influence with targeted prompts
If you suspect a source is shaping the answer, ask narrow questions:
- “What do you know about [Product]?”
- “Where did you get the claim that [Product] does X?”
- “Do you see this in the company website, news, or third-party sites?”
Then compare responses after changing the prompt or context. If the answer shifts, the model may be relying on retrieved context or strong surface associations.
5) Use citation-oriented search tools
For public-source discovery, tools like:
- Bing
- Perplexity
- GDELT
- Common Crawl-based search
- SEO tools / media monitoring platforms
can help identify which pages are most visible and likely influential.
If you want a more systematic approach, track:
- page rank / visibility
- backlinks
- recency
- frequency of mentions
- how often the page is mirrored or syndicated
6) Consider that training data is not directly inspectable
For most closed models, you generally cannot see the exact training sources used to describe your company. You can only infer likely sources from:
- public content
- citations from retrieval systems
- repeated phrasing
- model outputs across platforms
If you need stronger certainty, ask the vendor about:
- training data policies
- opt-out mechanisms
- crawl exclusions
- retrieval source logs
- citation support
7) Practical workflow
A good workflow is:
- Compile a list of likely public sources about your company.
- Ask several LLMs to describe the company with citations.
- Record repeated claims and sources.
- Compare those against your public pages and media coverage.
- Update the content you control if needed.
- Monitor monthly for changes.
8) If your goal is reputation management
If you want LLMs to describe your company accurately:
- improve the clarity of your website “About” and product pages
- publish authoritative pages on key topics
- use consistent naming and descriptions
- add structured data/schema where relevant
- maintain current press, FAQ, and docs
- correct inaccurate third-party listings
- ensure major indexes can crawl your site
If you want, I can also help you with one of these:
- a step-by-step audit plan,
- a prompt set to test LLMs for source attribution, or
- a template spreadsheet for tracking sources and claims.