Who gets the credit? Audit the original source behind an AI citation

Separate being mentioned from being credited

An assistant can name your business while linking to somebody else’s description. A publisher can supply the original reporting while a copied version receives the citation. A link can also work perfectly and still lead to the wrong article. Count those outcomes separately before deciding that visibility improved.

The useful record is a chain: buyer question → exact answer claim → linked page → original evidence. A mention count only covers one part of that chain. This article focuses on the ownership and identity of the source, alongside the evidence and review checklist at /resources/ai-visibility-proof-and-review-signals.

What the Tow Center actually tested

Published March 6, 2025, the Tow Center study tested eight search tools using 200 news excerpts each: ten articles from each of 20 publishers, for 1,600 queries. Researchers selected excerpts whose original source appeared within Google’s first three results and checked article, publisher and URL identification. The tests took place in February 2025.

The report explicitly gives incorrect-identification rates of 37% for Perplexity and 94% for Grok 3; ChatGPT misidentified 134 of 200 articles, or 67%. These historical cases are not a current product ranking or a measurement of general answer accuracy.

The chart is a diagnostic example, not a buying guide

The three bars are the numeric cases stated directly in the report’s prose, not all eight tested tools. The task was to identify an existing news article from an excerpt. It was not to recommend a software product, answer every kind of question, or estimate a buyer’s chance of discovering your brand.

This narrow task is useful because it makes an attribution failure concrete. It does not justify telling a customer that a present-day assistant is wrong at that rate. Product versions, retrieval behavior and the material being searched can change. Keep the test date and task in the same sentence as any percentage you reuse.

Build a provenance ledger you can act on

For every important citation, record the question, collection time and exact answer wording. Then save the linked URL, final destination after any redirect, page title, publisher, publication or update date when stated, and a short note locating the supporting material. Preserve an original copy or permitted evidence excerpt for your internal review.

Assign one provenance verdict: original source; authorized or clearly labeled republication; unaffiliated copy; different article; unavailable destination; or unclear. Keep “unclear” when evidence is insufficient. Do not turn a failed fetch, a paywall or an inaccessible archive into an accusation of copying.

A fictional example: the report exists, but the credit moves

Imagine that Northstar Research publishes an original comparison of appointment-reminder methods. A trade publication summarizes it, and an assistant cites the trade publication while describing the research. The summary may be accurate, but the original publisher has not received a direct citation. Mark the attribution and the factual support separately.

If the linked summary names the original report and supplies a working source link, the discovery path is different from an unattributed copy with no traceable origin. If the answer instead links to the publisher’s homepage, do not award an article-level citation merely because the domain is correct. This example is fictional; it is not a finding about any customer or publisher.

Choose a fix that matches the broken link in the chain

When your original page is hard to identify, give it a stable, specific title, a clear author or organization, an honest date, and a visible description of what was measured. Link your own explanatory pages to that source. If the source contains data, state its unit, population and collection period so a reader can recognize what a summary has borrowed.

When a third-party page contains an outdated description of your product, first document the difference against your current public evidence. A correction request can then name the exact claim and source. Do not manufacture independent endorsements, publish copies to crowd out the original, or assume that a technical markup change will force a particular assistant to cite you.

Recheck provenance without rewarding a new mistake

Repeat the same unaided question in the same product and search mode, then inspect both the answer and destination. A switch from an accurate summary to your own page is one observation. A switch to your domain with an unsupported claim is not an unqualified win.

Report a compact result: “Original-source credit present, absent or unclear; claim supported, contradicted or unresolved; destination available or unavailable.” Keep these dimensions readable instead of compressing them into one percentage that hides the failure the team needs to fix.

Use the evidence before buying more monitoring

Start with a handful of citations tied to consequential buyer decisions. If the problem is a missing original source or an inaccurate product fact, fix and document that first. If you need to check many questions, markets or source changes repeatedly, compare tools by whether they preserve this evidence chain. /methodology and /demo show the kind of record to ask for; neither a citation nor a paid subscription guarantees discovery or referral traffic.

Sources

• AI Search Has a Citation Problem — March 6, 2025 — Tow Center for Digital Journalism, Columbia Journalism Review (accessed 2026-10-17)