Most AI Local Recommendations Do Not Show Their Work

The source-provenance problem begins with missing links

Marketers often ask which websites caused an assistant to recommend a business. The output usually cannot answer that question. A model may display links, name sources without linking them, retrieve material invisibly, rely on training data, or combine several inputs into a sentence whose provenance cannot be reconstructed from the response.

Across 7,637 raw responses in the latest eight Atlas category scan runs, only 63 responses contained an explicit HTTP or HTTPS URL. Those responses contained 117 URL occurrences across 55 normalized domains. In other words, fewer than one percent of the collected answers exposed even one directly extractable URL.

This is a property of this API-based corpus and its model configurations, not an estimate of citation display rates in every consumer interface. Search-enabled products can show substantially more links. The finding establishes a measurement boundary: visible-link analysis alone cannot explain most recommendations in this dataset.

The visible source set leans toward public and professional infrastructure

The most frequently visible domains were New York court lookup and court information sites: `iapps.courts.state.ny.us` appeared 19 times and `nycourts.gov` 10 times. The New York City Bar appeared eight times. Other recurring domains included the National Employment Lawyers Association of New York, Legal Aid NYC, the ICE detainee locator, the American Immigration Lawyers Association directory, Catholic Charities, and NYC government.

Owned firm domains also appeared, but in small numbers. Examples included Block O’Toole, Fragomen, Gair Gair, Wildes Law, and several immigration and injury firms. Because only visible URLs are counted, these totals cannot establish that public sources matter more than owned sites overall. They show what this narrow output layer chose to expose.

The visible mix suggests several source jobs: official procedure, professional identity, referral infrastructure, public-service support, and firm-specific claims. A useful provenance system should classify those jobs rather than treating every URL as an interchangeable backlink.

A displayed URL does not prove recommendation causality

If an answer recommends Firm A and links to a court page explaining a deadline, the court source supports the legal explanation—not necessarily the choice of Firm A. If it links to Firm B’s guide while recommending Firm A, the visible citation and commercial recommendation point to different entities.

This is why mentions and citations must be tracked separately. A link proves that the output displayed a URL. It does not prove the page caused the recommendation, supported every nearby claim, or was the only material retrieved.

Source-to-claim alignment requires manual or model-assisted annotation of the answer: which claim sits near the citation, which entity it concerns, whether the page supports it, and whether the source is owned, directory, professional, editorial, government, or another class.

How to build a defensible source audit

Capture provider-native citation metadata whenever an API exposes it; do not rely only on URL regex extraction from prose. Preserve the raw response, source title, URL, position, model, timestamp, prompt, and any retrieval annotations. Resolve redirected and tracking URLs to a canonical domain.

Next, attach each displayed source to the claim it appears to support. Audit a sample for entailment, freshness, entity match, and source type. Keep “no visible source” as a real outcome instead of discarding the response from the denominator.

Finally, compare the source layer with recommendation presence. The new Google and Bing AI visibility reports can reveal URL appearances and citations on their own surfaces. An independent prompt panel is still required to observe recommendations across assistants. Neither layer alone supplies causality.

What firms should publish when provenance is opaque

Opaque provenance is not a reason to publish indiscriminately. It is a reason to make important claims easy to verify across several channels. Put service fit, practitioner credentials, geography, process, and evidence on accessible first-party pages. Keep professional listings and public business records consistent. Earn independent coverage where a claim should not depend solely on the firm asserting it.

Measure whether owned URLs begin appearing, but keep the commercial question separate: did the firm enter the shortlist? A site can become a useful cited publisher without becoming a recommended provider. It can also be recommended from third-party evidence without receiving an owned-domain citation.

Method and limitations

The analysis searches the raw response text associated with the latest eight published Atlas snapshot scan runs for explicit HTTP and HTTPS URLs, normalizes hostnames, and counts response presence and URL occurrences. It does not include unlinked source names, hidden retrieval metadata, training sources, or links available only in consumer-product interface elements.

The 63-of-7,637 figure should therefore be read as visible URL prevalence in this bounded corpus, not as a universal citation rate for ChatGPT, Claude, Gemini, or Perplexity. Provider-native citation capture and claim-level annotation are required before Viclaro can publish a true owned-versus-directory-versus-earned share.