Sources AIs cite most about brands: a study

AI answers about a brand do not rely on the brand’s website alone. Across 123 audits and 7,264 answers, we recorded the third-party sources models displayed. LinkedIn appeared in 37% of audits, ahead of Wikipedia and Trustpilot. That describes this sample. It does not rank source quality or predict what every model will cite next.

In short

We counted URLs that models presented as sources in answers about audited brands. A source counts once per audit after the brand’s own website is excluded. LinkedIn, Wikipedia, Trustpilot, business registries, review services and community spaces recur frequently. The useful response is to verify the information visible on those surfaces, not to assume that being present there guarantees a citation.

What we measured

We analysed 123 audits, representing 7,264 AI answers. For each audit, we extracted the URLs shown as sources when an answer discussed the audited brand. We removed the brands’ own sites so the analysis keeps third-party sources only. We then counted a domain only once within an audit. This prevents one brand or one unusually long answer from mechanically inflating a source.

The share in the table means the share of the 123 audits in which a source appeared at least once. It is not an overall citation rate and not an average number of citations per answer. The table does not measure website traffic, editorial reliability, or a model’s intent. It is a frequency marker for this specific corpus.

The third-party sources observed most often

The ranking below records observed frequency. Its categories help read the groups of sources; they are not a quality scale.

RankSourceShare of 123 auditsType
1LinkedIn37 %Professional network
2Wikipedia35 %Encyclopedia
3Trustpilot32 %Reviews
4Pappers24 %Business registry
5YouTube23 %Video
6Reddit21 %Community
7Societe.com19 %Business registry
8Facebook19 %Social network
9G214 %B2B reviews
10PagesJaunes10 %Business directory
11annuaire-entreprises.data.gouv.fr7 %Public registry
12Glassdoor7 %Employment reviews

How to read the table without overclaiming

LinkedIn appeared in 37% of audits, Wikipedia in 35%, and Trustpilot in 32%. In this corpus, those three services appeared more often than the other listed third-party sources. That does not show that every LinkedIn profile, Wikipedia page, or Trustpilot review is reused. An exposed URL may appear in an answer without being the sole cause of the wording that follows.

Pappers appeared in 24% of audits and Societe.com in 19%. These figures also reflect the French component of the sample. It would be wrong to turn them into a claim that French registries dominate every market or language. The same restraint applies to PagesJaunes and the public French company directory: they may be relevant in a given context, not universally decisive.

Four recurring source families

Professional networks, encyclopedias, and public profiles

LinkedIn, Wikipedia, and Facebook appear in the ranking. Such pages often carry identity details: name, activity, people associated with the organisation, history, or a public description. These are relatively easy to compare with primary records. Before trying to improve anything, check that the details are accurate, dated where appropriate, and consistent across surfaces the company actually controls.

Registries and directories

Pappers, Societe.com, PagesJaunes, and annuaire-entreprises.data.gouv.fr belong to this family. Answers may associate them with structured facts such as legal name, activity, or company status. The appropriate response is not to rewrite a company’s history. It is to distinguish official data, aggregated data, and descriptive copy, then correct an upstream source only where a verified correction is legitimate.

Reviews and first-hand accounts

Trustpilot, G2, and Glassdoor occur in the sample. They do not serve the same purpose: customer feedback, software review, and employee experience answer different questions. Folding them into one word, reputation, hides that difference. An AI answer displaying a review does not make the review a universal truth. It only shows that a third-party surface was visible in the answer we observed.

Communities and video

YouTube and Reddit sit among the first six sources. Their presence is a reminder that brands are also described through videos, comments, comparisons, and conversations. These formats are mixed by nature. Read the date, author, topic, and context before interpreting a source. A product demonstration, a support thread, and a personal opinion do not carry the same meaning.

A practical verification process

The starting point is not a universal checklist of platforms. It is the record of sources already shown in answers about your brand and your questions. From there, a short process reduces decisions based on vague impressions.

  • Keep the prompt, model, date, and answer mode linked to every observed URL.
  • Classify the URL: official page, registry, review, community, media page, or social profile.
  • Check identity facts, commercial claims, and experience-based opinions separately.
  • Correct inaccurate information only where a correction is legitimate and traceable.
  • Repeat the same panel later to observe change, without automatically assigning that change to one action.

This process separates an observed fact from an explanation. The fact is that a given URL appeared in a given answer under stated conditions. Explaining why requires more evidence. An updated profile may coincide with a different answer without necessarily causing it. Likewise, an unresolved URL or one absent from an answer is not proof that information was invented.

What the study changes for a brand

It shifts attention from the official website alone to the ecosystem of sources already describing the brand. Not to manipulate those sources, but to map them factually: where identity is described, where reviews are published, where legal or commercial data is repeated, and which pages actually recur in tested answers. Our article about Source Authority explains that URL-level approach.

Visibility alone does not qualify an answer. Accuracy, freshness, and context still matter. That is why the AI visibility barometer and our study of URLs invented by models complement this ranking. They answer different questions under their own scopes. Combining them as if they were one measure would make a stronger-sounding, less reliable conclusion.

Limits to keep in view

The brands in this corpus chose to be audited, so the sample does not represent every company. Its French component partly explains the visibility of local directories and registries. The ranking covers third-party sources displayed in recorded answers only; it does not predict future model behaviour. Models, prompts, language, and access to web search can all change what is observed.

The conclusion is deliberately modest. Your website remains important, but it does not contain everything an AI answer may expose about your brand. Measure the sources actually observed, verify the facts, document valid corrections, and keep frequency separate from causality.

What to record for each source

A useful source register is deliberately specific. Do not write merely that a model cited LinkedIn or a directory. Record the exact page, the question asked, the model, the date, whether web search was available, and the sentence in which the source was shown. This makes a later review possible. It also prevents a team from treating a domain-level observation as evidence about every page on that domain.

Then compare the information exposed by the source with a primary record. A company name, legal form, address, executive name, product description, review claim, or old price should not receive the same treatment. Some details may belong to a public register, some to a profile owner, and some to a reviewer. The relevant question is not “how do we make this page appear?” It is “is the information that appears here accurate, attributable, and current for the reader’s question?”

Keep the outcome of that check equally precise: confirmed, outdated, incomplete, opinion, unresolved, or not applicable. An unresolved URL should remain unresolved in the register. It should not be converted into a claim that a model fabricated a fact. Conversely, a factual discrepancy should include the primary evidence used to identify it and the legitimate route for correction, if one exists.

For published data that sits alongside this record, see the Observatory and the AGS methodology. They address different objects; they do not turn the frequency recorded here into an average, causation, or a forecast.

This audit trail has another advantage. If a later answer changes, the team can compare like with like: the same prompt family, model, language, and response mode. That is more informative than comparing two screenshots taken under unknown conditions. It does not prove causation, but it makes the observation reproducible and useful for editorial, legal, and brand teams.

Questions and answers

Which sources do AIs cite most about brands?

Across 123 audits, the most-cited third-party sources are LinkedIn (37% of audits), Wikipedia (35%), Trustpilot (32%), Pappers (24%), YouTube (23%) and Reddit (21%). Business registries and review platforms weigh especially heavily.

Why isn't my website enough to be visible in AI?

Because AIs build a brand's image from third-party sources — professional networks, encyclopedias, reviews, business registries — not only from the official website. If those sources are inaccurate or incomplete, the AI repeats the error.

Comparison: do LinkedIn or Wikipedia matter more than my website?

In our sample, LinkedIn is the third-party source most frequently cited by AIs about brands, ahead of Wikipedia and Trustpilot. An up-to-date LinkedIn page and an accurate encyclopedic presence therefore directly affect what the AI says about you.

How was this study conducted?

We extracted the URLs models cite as sources across 123 real audits (7,264 answers), removed the brands' own sites, then counted each source once per audit to measure cross-cutting authority rather than a single brand's weight.

About the author

Davy Abderrahman

Founder of

As a specialist in visibility in AI answers (AEO/GEO/LLMO), I help agencies and consultants measure and analyze their clients’ presence on ChatGPT, Claude, Gemini, Perplexity and other AI answer engines.

AEO GEO LLMO AI visibility AI audits
How do your brands appear in AI answers?

Agencies, consultants and in-house teams: compare the answers generated for your own scenarios, languages and selected models. Our infrastructure, your scenarios, your expertise.

Was this article helpful?

- (0 votes)