# How to Find the Sources Behind a Competitor's AI Mentions

Author: Quentin Megevand
Published: 2026-10-06
Source: https://getnamedrop.ai/en/blog/how-to-find-which-sources-feed-a-competitor-s-ai-mentions

To find what feeds a competitor's AI mention, run the prompt with web search visible, open every cited page and confirm it actually contains the claim. When no citation appears, treat the mention as likely coming from training data or uncited retrieval, then test it. Repeat the prompt several times before trusting any source pattern.

## What can sit behind a competitor mention in an AI answer

The source behind a competitor mention can be the competitor's own site, a third-party or editorial page such as a review or comparison, or user-generated content such as a Reddit thread or forum post.

Knowing that the competitor is named tells you little. You need the competitor and third-party domains cited next to the name, recorded engine by engine. If mentions and citations get mixed up in your reporting, start with [how an AI mention differs from an AI citation](/en/blog/mention-vs-citation-what-s-the-difference-in-ai-search).

### Live retrieval, training data and citations added after the answer

- **Live retrieval.** The engine searches the web, reads pages and writes the answer from them. The pages it cites are the first candidates for the source, and they still need checking.
- **Training data.** The model learned about the competitor before the conversation and names it without retrieving anything. No page is shown because none was fetched for this answer.
- **Attached links.** An answer can carry links that do not match the sentence you care about. A link shows the engine associated the page with the answer, not that the page produced the mention.

Anthropic's help page on web search, checked on 6 October 2026, says: "When you ask about topics that benefit from current information, Claude invokes a search tool to inform and ground its generated responses with content from the live web."

Google's page on AI features, dated 10 December 2025 and checked on 6 October 2026, says 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. Our inference, not Google's statement: a page cited in a Google AI answer may have been found through a related search rather than the query you typed.

### How each engine shows its sources

Each engine exposes sources differently, so every check has to be run and recorded per platform. The table separates what we verified in vendor documentation on 6 October 2026 from what you need to check in the interface yourself.

| Engine | When it searches the web | How sources appear | What you can and cannot see | Where to look when none are shown |
| --- | --- | --- | --- | --- |
| ChatGPT | No OpenAI documentation verified for this article | Check for inline links or a sources list | Links shown, not which passage produced a sentence | Follow-up request; search on versus off |
| Perplexity | No Perplexity help documentation verified for this article | Check for numbered citations or a source list | Links shown, not how they were weighted | Open each source; rerun with a rephrased prompt |
| Gemini app | Not covered by the Google page cited here | Check for links attached to the answer | Links shown, not whether training data shaped the sentence | Follow-up request; treat links as candidates |
| Google AI Overviews and AI Mode | Documented: both may use query fan-out across related searches | Check the links displayed with the response | Links, not the related searches that surfaced them | Related phrasings of the query in classic search |
| Claude | Documented: invokes a search tool for topics that benefit from current information, once web search is enabled | Check for links attached to the response | Links shown, not what came from training data | Web search off versus on |

A blank source column is not proof of training data; the next sections show how to test it.

## When the answer shows citations, check that the page supports the claim

A citation is a lead to verify, not proof. In a study by the Tow Center for Digital Journalism, published in Columbia Journalism Review on 6 March 2025, the authors write of the AI search tools they tested: "Collectively, they provided incorrect answers to more than 60 percent of queries." That figure describes the study's own queries, not brand prompts, and it measures incorrect answers rather than citations alone. It is still a reason to open every link before you record it.

Check each cited URL in the same order:

1. Open the page and note whether it loads, redirects to another URL or returns an error page.
2. Search the page for the competitor's name. If it is absent, the citation is misattributed for your purpose.
3. Search for the specific claim the answer gives the competitor: a feature, use case, price position or ranking.
4. Copy the passage that supports the claim, with its publication or update date and its author.
5. Compare dates. An old page cited for a current claim is a weak source, even when the wording matches.

Log unsupported citations anyway, with the reason, because the same broken link may come back in later runs.

Record the passage, not the domain. One page can feed several claims, and each needs its own verified passage. The domain that matters is also often not the competitor's own. A competitor named on a third-party review site, a directory or an industry blog points you to a page where your brand could also appear, and that gap is often part of [why competitors get cited by ChatGPT when you aren't](/en/blog/why-competitors-get-cited-by-chatgpt-and-you-don-t).

## When the answer names a competitor without any citation

An uncited mention has several possible explanations: the model drew on training data, the engine retrieved pages without displaying them, or no search ran at all. You cannot read the cause from the answer, but you can test it.

Control personalisation first. Use a logged-out session or a temporary chat where the product offers one, turn memory off, and fix the region and language. Memory and history settings differ by product, so check them before each test series.

Then run the comparison:

- **Force or disable search** where the interface allows it. Claude's help page describes web search as a feature you enable; other products offer their own controls.
- **Compare a search-on answer with a search-off answer.** If the competitor appears only with search on, retrieval is the likelier route. If it appears in both, training data is a plausible contributor.
- **Ask for sources in a follow-up.** The links that come back are candidates, not proof of origin: an engine can find pages after the fact that agree with what it already wrote.
- **Check the same prompt on each platform.** One engine may cite where another does not, and that cited page is a candidate for the uncited answer too.

On what ChatGPT draws on: in general, an answer can come from training, from a live web search, or both, and which applied is visible only when the interface shows search activity or sources. We could not verify OpenAI's help pages on 6 October 2026, so read this as general model behaviour, not OpenAI's documentation.

## How to find the Reddit threads and forum posts behind a mention

Reddit, review platforms and forums are a distinct source type. Engines can cite them alongside editorial and brand pages, and they hold comparisons written by users in their own words, the kind of language a recommendation can echo.

Start from the distinctive wording of the answer: an unusual adjective, a specific complaint, a use case, a pairing of brands. Search it with a site restriction, for example `site:reddit.com "competitor name" "distinctive phrase"`, then repeat on the review and forum platforms that matter in your category. Match phrasing and claims, not only the brand name.

A matching thread is a probable source, not a confirmed one. Many people may have posted the same opinion, and the model may have met it somewhere else. Label it accordingly in your log: confirmed when the engine cited the page and you verified the passage, probable when you found it yourself and the wording matches.

Some platforms have announced content licensing agreements with AI companies, a direct route into model data. We could not verify a primary announcement for this article, so we name no deal, and nothing public we found explains how such content is weighted.

## How many runs before a source pattern means something

One answer is a sample, not a pattern. SparkToro's research write-up, dated 28 January 2026, reports that if ChatGPT or Google's AI is asked the same question 100 times, there is fewer than a 1 in 100 chance that any two responses give the same list of brands. The study measured brand lists, not cited sources. We infer, without the study testing it, that the sources behind the names move as well.

Names and sources shift between runs, sessions, regions and logged-in state. A practical rule, not a measured threshold: repeat each prompt several times, on different days and with a few phrasings, before calling a source recurrent. Record sources per platform and per run, because a pattern on Perplexity says nothing about Claude.

For each run, log the date, engine, mode (search on or off, logged in or out), exact prompt, region, language, competitor named, cited URLs and the verified passage.

Public data does not establish how stable sources stay over weeks. Re-check on a fixed schedule instead of assuming that a source found once still holds. The same log also supports [benchmarking your AI visibility against competitors](/en/blog/how-to-compare-your-ai-visibility-against-competitors).

## Which sources to pursue, and what to do when a mention is wrong

A list of competitor and third-party domains is only useful once ranked. Score each source on these criteria:

- **Recurrence**: how often it appears across runs and engines.
- **Topical relevance**: whether it covers your category or only mentions it.
- **Editorial independence**: whether the publisher chooses what to include or sells placement.
- **Legitimate fit**: whether your brand could honestly appear there, as a reviewed product, a contributor or a cited expert.
- **Cost and lead time**: a directory listing can be quick, an independent review can take months.

When an answer says something wrong about a competitor or your brand, contact the publisher of the page carrying the error with a sourced correction, update your own facts page so a correct, dated statement exists, use the engine's feedback control on the answer, and re-run the prompt later. None of these guarantees that the answer will change.

Do not plant fake reviews, seeded forum posts or undisclosed sponsored threads. They mislead users, they can break platform rules, and they turn your brand into the kind of claim worth correcting.

## Trace a competitor mention to its source this week

This routine covers one prompt and needs no paid tool.

1. Pick one high-value prompt where a competitor is named and a buyer would act on the answer.
2. Set a clean session on each engine, as described above.
3. Run the prompt on every engine you track and record results per platform.
4. Open and verify every cited page.
5. Run the search-on and search-off test on uncited mentions.
6. Search forums and review sites for the answer's wording and label matches as probable.
7. Classify each source as competitor-owned, third-party editorial or user-generated.
8. Repeat across days and phrasings, then rank the sources with the criteria above.

Keep one spreadsheet with these columns: date, engine, mode, exact prompt, region, language, competitor named, cited URL, source type, verified passage, passage date, author, status (confirmed or probable) and notes.

To see which sources each engine cites for your competitors on your own prompts, [run a free Namedrop scan](/en?src=blog-article#scanner) as your next step.

## Sources

- [Google Search Central, AI features and your website, consulted 2026-10-06](https://developers.google.com/search/docs/appearance/ai-features)
- [Anthropic (Claude Help Center), Enable and use web search, consulted 2026-10-06](https://support.claude.com/en/articles/10684626-enable-and-use-web-search)
- [Columbia Journalism Review (Tow Center for Digital Journalism), AI Search Has a Citation Problem, consulted 2026-10-06](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php)
- [SparkToro, New Research: AIs are highly inconsistent when recommending brands, consulted 2026-10-06](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands/)

## Frequently asked questions

### What data sources does ChatGPT use to answer questions about brands?

In general, a ChatGPT answer about a brand can come from what the model learned during training, from a live web search, or from both. Which one applied is visible only when the interface shows search activity or source links. We could not verify OpenAI's own help pages when this article was written, so treat this as general model behaviour and test each answer with search on and off.

### Why does an AI answer mention a competitor without showing any source?

The model may be drawing on training data, the engine may have retrieved pages without displaying them, or no search ran at all. The answer alone cannot tell you which. Use a clean session, compare a search-on answer with a search-off answer, and ask for sources in a follow-up, treating any links returned as candidates rather than proof of where the mention came from.

### Can I find the Reddit threads ChatGPT used to recommend a competitor?

You can find likely threads, not confirmed ones, unless the engine cites them. Take the distinctive wording of the answer, search it with a Reddit site restriction and on the review and forum platforms in your category, and match phrasing and claims. Log a match as probable, because many users may have written the same opinion and the model may have learned it elsewhere.

### How many times should I rerun a prompt before trusting its sources?

We know of no public threshold for sources. SparkToro reported in January 2026 that brand lists almost never repeat across runs of the same prompt, so a single answer is not a pattern. As a practical rule, repeat each prompt several times, on different days and with a few phrasings, on every engine you track, and re-check on a fixed schedule.

### What can I do when an AI answer says something wrong about a competitor or my brand?

Find the page that carries the error and contact its publisher with a sourced correction. Update your own facts page so a correct, dated statement exists, use the engine's feedback control on the answer, then re-run the prompt later and log the result. None of these routes guarantees a change, and planting fake reviews or posts is not a legitimate fix.