---
title: "When buyers don’t know your name, AI cites a wider range of sources"
description: "Across 81,778 API-reported citation links, discovery questions drew on a much wider range of domains than questions naming a brand. Here’s what PR teams can do with that finding."
author: RunPR
status: approved-for-publication
proposed_slug: ai-citation-concentration-brand-vs-discovery
data_through: 2026-09-27
---

# When buyers don’t know your name, AI cites a wider range of sources

The 100 most-cited domains accounted for **81% of citation links in questions about a named brand**. For questions asking which companies or products to consider, that share fell to **35.6%**.

That is the clearest finding from our analysis of 81,778 API-reported citation links. When a question started with a need rather than a company name, the links were spread across a much wider range of domains.

For a PR team deciding where to spend its time, that wider spread is worth investigating. The sources behind a buyer’s question may extend well beyond the websites that dominate answers about an established brand. Start with the question your client wants to answer, then inspect the pages the engines cite.

This is an observational study of RunPR API scans from June 30 through Sept. 27, 2026, organized around 19 monitored websites. Those websites were the subjects of the scans, not the limit on sources the engines could cite. Across the included responses, citation links pointed to 7,060 distinct website domains. Four engines met our citation-data criteria. The sample is small and selected, includes repeated scans and does not represent consumer chat usage. We did not test whether a placement causes an AI mention or whether one kind of publisher performs better than another.

## The question changes the source picture

“What does RunPR do?” gives the assistant a company to describe. “What PR software should a small agency consider?” asks it to find options without supplying our name. These are illustrative questions, not customer prompts from the study.

We ranked cited domains separately within each of five question categories:

| Question category | Citation links | Share going to its top 100 domains |
|---|---:|---:|
| Founder | 10,112 | 83.6% |
| Named brand | 20,613 | 81.0% |
| Competitor comparison | 10,347 | 55.6% |
| Subject matter | 10,605 | 38.8% |
| Category discovery | 30,101 | 35.6% |

Each row has its own top 100, ranked by citation frequency within that category. A domain can rank highly in one category and rarely appear in another. These are not prestige rankings or a list of recommended media targets.

The discovery group cited 4,001 domains. The brand group cited 1,078. Across all five categories, the top 100 domains received 46.1% of links, a combined figure that obscures the differences a client team needs to see.

![Share of API-reported citation links going to each question category’s top 100 domains: founder 83.6%, brand 81.0%, competitor 55.6%, topical 38.8% and discovery 35.6%.](/studies/2026-09-citation-concentration/citation-by-question.png)

*Source: RunPR API scans, June 30 to Sept. 27, 2026. Rankings are calculated separately within each category.*

## Some concentration is built into the question

A question about a named company or founder naturally points toward a smaller set of relevant sources: the company website, a LinkedIn profile or a database listing. Our scans were organized around 19 monitored websites and included questions about their brands, founders, competitors and broader categories. Across those scans, AI answers cited 7,060 distinct website domains. It would be surprising if questions about those entities produced the same source distribution as broad questions about an industry.

The 19 monitored websites are a small, selected sample. Repeated scans generated many citation links, but they do not make the sample representative of all businesses. That is a real limitation, not something the analysis can explain away. The finding is useful because those two situations reflect different research tasks: checking a company someone already knows and looking for companies they have not named.

The difference is also not explained by citation volume alone. Founder questions contributed 10,112 links and subject-matter questions contributed 10,605, yet their top-100 shares were 83.6% and 38.8%. Nearly equal-sized groups still produced very different concentration. That comparison does not remove differences in subject matter or company mix, but it shows why the result cannot be dismissed as simply a larger group spreading links across more domains.

Brand citations were more concentrated than discovery citations within each of the four included engines. The gap also remained when we counted each domain only once per answer: 69% for brand questions versus 29.8% for discovery questions. Full sensitivity checks are in the [methodology](/methodology/ai-citation-concentration).

## What should a PR team do for a client today?

### Start with one question the buyer would actually ask

Keep questions that name the client separate from questions about a need. For a staffing client, “What does this staffing company specialize in?” checks the company’s description. “Which healthcare staffing firms help hospitals hire permanent nurses?” asks the assistant to identify options before a vendor has been chosen.

Read the answers separately. An accurate description when the company is named does not establish that it appears when the buyer describes a problem.

### Open the cited page before choosing an outreach target

Suppose the hospital-staffing answer cites an article about nurse retention. Read it. Is it about the problem your client can address? Does the reporter cover that subject regularly? Could your client contribute original data or a specific explanation the article lacks?

Those questions are more useful for deciding today’s pitch than a domain’s position in an overall citation table. Our study did not measure page quality against domain rank. It gives teams a reason to investigate the broader set of pages appearing in discovery answers, not a reason to disregard major publications.

The next move might be to research that reporter, prepare a supported angle or improve an inaccurate explanation on the client’s own site. It might also be to leave a source alone because there is no credible contribution to make.

### Keep coverage, citations and mentions separate

An answer can name a company while citing an independent review. It can also cite the company’s website without naming the company in its text. A mention is the name appearing in the answer; a citation is a source link reported with it.

If a later answer names your client and cites a new placement, record both observations. Neither establishes a visit or sale. Recheck comparable questions to see what changed rather than treating a single answer as proof that the pitch worked.

The dashboard calculations use different denominators. We explain those in [Understanding AI visibility rates](/resources/understanding-ai-visibility-rates), outside this study.

## Engines, models and the sample

RunPR supports six engine families. This study includes four: **Exa, ChatGPT, Claude and Grok**. The 81,778 links came from 13,294 responses across 182 runs and pointed to 7,060 registered domains.

We excluded historical Gemini records with unresolved destination domains and Perplexity’s mixed search-and-citation feed. The [methodology](/methodology/ai-citation-concentration) details all exclusions, which removed 53.7% of the starting records.

These were programmatic API scans, not tests of consumer apps or their default models. Recorded routes included Exa Answer, OpenAI’s GPT-4o, GPT-5.5, GPT-5.6 Terra and GPT-6 Sol, Claude Sonnet 4.6 and Sonnet 5 and Grok 4.5. The [methodology](/methodology/ai-citation-concentration) lists their contributions and explains the limits of backfilled historical model labels. Exa supplied 49.5% of included links, so the combined result is not an equal-weight average of the four engines.

We counted each recorded URL citation, including different pages on one domain and repeated citations across scans. Ordinary subdomains were grouped under their registered domain. A reported citation was not independently checked for support of every statement in its answer.

Our product-wide source-link totals use a broader population than this study, including feeds excluded here. They should not be read as the study’s sample size or compared as growth in the same dataset. Customer identities, prompt text and per-customer results are not part of this publication package.

For client work, the distinction is simple: what AI says about a company it has been asked to describe is different from whether it brings that company into a conversation about a need.

That is why RunPR separates those questions. Send us a client’s website and we’ll show you how to review both, inspect the sources and choose a next PR move worth making.