How AI Engines Pick Their Sources: What Our Measurements Show

In short

An AI answer's citations show the sources presented to the reader, not the internal recipe used to select them. Our studies reveal observed frequencies, differences between response modes and limits to measurement. They do not expose engines' hidden selection criteria. Understanding an answer means separating the cited page, the claim it supports and whether the brand actually appears. Those are different results, and they need different evidence.

How do AI engines pick their sources? Our measurements cannot reconstruct that decision from beginning to end. What we can inspect is its visible output. That distinction matters when reading an audit. A LinkedIn link does not establish that an engine always favors LinkedIn, or that the page determined everything in the answer. We can examine what was displayed. We cannot see every document considered and then left out.

Two datasets answer different questions

The AI visibility barometer covers 6,820 responses, 104 complete audits and 65 brands between September 2025 and June 2026. It includes 44 distinct models. Its subject is brand visibility, with different subsets used for particular measurements. It does not provide a complete breakdown of the sources used to produce every answer.

The study of sources cited about brands examines 123 audits and 7,264 responses. It excludes the audited brands' own websites and counts a given source only once per audit. Its percentages describe the share of audits in which an external source appeared at least once. They are not percentages of the barometer's 6,820 responses.

Why dwell on the denominator? Because getting it wrong changes the decision you make. LinkedIn appearing in 37% of audits in the source study does not mean that 37% of all AI answers rely on LinkedIn. It does not mean that 37% of brands gain a recommendation there, either. Each published result needs to stay attached to the population it actually measures.

A citation does not expose the whole selection chain

It helps to separate several possibilities: information may be accessible, a document may be consulted, its content may contribute to an answer, and a link may be shown. This is an analytical framework, not a verified account of any particular engine's architecture. The final list of references cannot, by itself, tell us how each of those stages unfolded.

The missing evidence includes the candidates that were not retained. If your page does not appear, that observation alone cannot tell us whether it was inaccessible, absent from a retrieved set, considered unsuitable or simply not displayed. Saying the engine rejected it for lacking authority would jump ahead of the evidence. A missing link does not disclose the reason it is missing.

The reverse deserves equal care. A displayed URL is something we can record and check. Saying it explains the entire response requires more evidence. An answer may connect several claims to several references. The most prominent link is not necessarily the reference that supports the detail the reader needs to verify.

What observed source frequencies actually tell us

A few findings from the 123-audit source study illustrate the range of surfaces presented to readers:

• LinkedIn appears in 37% of audits and Wikipedia in 35%.
• Trustpilot appears in 32% of audits.
• Pappers, a French business information service, appears in 24% of audits.
• YouTube appears in 23% of audits and Reddit in 21%.

This is not a reliability ranking. A business register can document legal identity. A review records someone's experience. A video may demonstrate a use case. Their appearance in the same dataset does not make their evidence interchangeable. To understand the answer, ask what kind of claim is attached to each source, rather than treating every reference as an equivalent endorsement.

The sample's French component matters as well. Pappers' frequency does not predict the importance of business registers in another country. Publishing this article in English does not turn the data into a study of the US or UK market. The figures still describe the original published corpus, with its original limits.

Native mode and web search measure different conditions

In the barometer, 63 audits support a comparison of both modes. Average visibility scores are 63.4 out of 100 in native mode, without web search, and 70.5 out of 100 with web search enabled. These are visibility scores. They are not domain citation rates, and they are certainly not click-through rates.

The averages show different measurements under the two conditions in that subset. They do not prove that adding search improves every response or makes a recently published page likely to be cited. Nor can we attribute the difference to LinkedIn, Wikipedia or another service. The published comparison does not measure that attribution.

AI Labs Audit documents a dual mode that compares native responses with responses using web search. For this question, its value is to keep the observation conditions separate. Comparing a screenshot without search to another with search, then calling the difference a content effect, confuses the diagnosis. The mode belongs beside the result, not in a forgotten footnote.

Being cited and being recommended are different outcomes

A brand can be mentioned without a link to its website. A website can be referenced for a fact without its owner being recommended. An answer can also display several sources while never naming the company being measured. Calling all of these outcomes visibility, without qualification, hides distinctions that a decision-maker needs.

The barometer's spontaneous-presence result comes from a subset of 657 responses concerning 13 brands with entity extraction. It records 106 mentions, or approximately 16%. That figure does not describe all 6,820 responses. It also does not measure the share of cited links belonging to the audited brands. Neither interpretation would use the right denominator.

Our view is straightforward: a narrow metric with a clear denominator is more useful than a large score that mixes unlike outcomes. A team might want to correct a company fact, earn a citation for expertise or appear in a recommendation. Those are different objectives. The report should say which one was actually tested and what evidence would count as a change.

Read the relationship between a claim and its link

Consider a teaching example, not a documented client case. An answer says that a company operates in a city, offers a service and receives positive reviews. It displays a business register and a review platform. The register might support legal identity without confirming the current service. The reviews might document opinions without proving that the service remains available.

The check therefore happens claim by claim. Open the exact page. Find the passage that supports the statement. Check that it concerns the right entity, that its date fits the question and that the answer has not expanded its scope. An individual opinion turned into a general verdict is an interpretation problem even if the link works perfectly.

If you cannot find supporting text, record support not established. That is not proof that the engine consulted no other information. If the page cannot be reached, record the failed check and its conditions. A temporary access failure, on its own, does not establish that the engine invented the URL. Preserve the uncertainty instead of giving it a more dramatic label.

Test a hypothesis without inventing a causal story

Suppose you correct an old description on a profile you administer. Your next test cites that profile. Have you demonstrated that the edit caused the citation? No. The result is consistent with that explanation, but the model, its tools, the prompt or the available search results may also have changed.

For a comparison people can interpret, retain the exact question, language, requested market, reported model, response mode and date. Archive the answer and its URLs. Repeat observations under conditions that are as similar as you can make them. These are measurement precautions proposed here. They are not a description of a causal experiment already carried out in our published studies.

Define success before interpreting the result. For a factual correction, check whether the wrong fact disappears. For a citation objective, check the actual page presented. For a recommendation objective, check spontaneous mention and context. Movement in one of these outcomes does not demonstrate an improvement in the others.

What to do when the reason for selection remains unknown

Start with errors you can prove and have a legitimate way to correct. An outdated address on your website or an obsolete description on your own profile calls for a specific action. A general source-frequency ranking does not justify fake reviews, invented profiles or manufactured community mentions. Accuracy is a defensible objective without pretending to control an engine's decisions.

Keep negative observations too. A corrected page that still does not appear is a result, not something to hide. It reminds the team that making accurate information available does not guarantee its inclusion in an answer. Recording that limit makes the next decision less arbitrary and prevents a convenient success story from replacing the actual evidence.

For the data and their scope, return to the two linked studies. Our conclusion is deliberately narrower than a universal recipe: the measurements show sources presented to readers and differences in visibility, not the engines' complete selection rules. The better question is not which platform AI prefers. It is which part of this particular answer the cited evidence actually lets you verify.