Seeing a publisher mentioned in an AI answer does not tell you whether that publisher earned a commission, lost one or caused a sale. Start with the journey you can actually observe. Which page was cited? Did a person visit it? Was a qualifying affiliate referral recorded before the purchase? Those questions lead to a more useful review than assuming the entire attribution system has broken.
A citation is evidence of a displayed source
OpenAI's current search guidance says ChatGPT can search the web and include links and citations, while warning that results can be incomplete or incorrect. A displayed citation is not, by itself, proof that the page was used in model training or that a reader acted on it. The earlier article blurred those ideas and used an unsupported figure of nearly 70% for affiliate content's share of cited sites. That figure and the unverified shopping-query growth claim have been removed.
Compare two possible journeys
Imagine a fictional backpack publisher whose comparison page appears as a source in an AI answer. In one possible journey, a shopper visits that page, chooses a product and follows the publisher's approved affiliate link to the merchant. In another, the shopper sees a brand name in the answer and later visits the merchant directly. These are illustrative paths, not observed customer behavior. The publisher's citation alone cannot establish which path occurred or what would have happened without the comparison.
For the first path, the manager can investigate whether the referral and conversion were recorded according to the integration and agreement. For the second, the manager should not invent a missing click identifier or award causal credit from a guess about influence. If the available evidence cannot join the events, label the gap. That preserves the difference between a possible contribution and a payable transaction.
Server tracking still needs a way to match the referral
impact.com's tracking documentation describes API conversion reporting that compares conversion data with tracking-link click data, and an FTP method that can use ClickID and other data points. That illustrates a matching requirement, not a promise that sending server events reconstructs every missing journey. Review the actual integration with its owner. Moving an event from browser to server does not establish a referral that was never observed or remove the need to handle permissions and matching correctly.
A useful tracking-review brief names where the identifier is created, how it reaches the merchant, how the order is matched and how duplicates or adjustments are handled. Keep website visits, network clicks, approved actions and paid commissions separate. If one recorded referral has no corresponding action, inspect the specific evidence and applicable terms before declaring a platform-wide failure. Any controlled implementation tests belong in an approved test process, not an unreviewed production change.
Keep visibility reporting useful without overclaiming
For a fictional manual review, suppose a publisher appears in four of ten saved answers to a defined set of questions. That is four appearances in that small review. It is not a 40% share of all AI recommendations or evidence of four purchases. Record the question, service, date, answer and cited URL so someone can understand what was checked. Use available referral reporting for visits rather than converting citation counts into traffic estimates.
Use the evidence to choose the next job. An inaccurate cited product description needs an editorial correction. A broken recorded referral needs a tracking investigation. A claim about incremental sales needs a credible comparison addressing what would have happened otherwise. These are different questions, and answering one does not automatically answer the others.