Do tracking parameters inflate AI citation counts?

A cited-page appearance is one page showing up in one answer. ChatGPT Search produced 424,433 distinct citation addresses. Removing the eight standard tracking parameters reduced that count to 385,287 pages. The 39,146 collapsed address variants were 9.2231% of the raw count. Normalization is that cleanup step: stripping the tracking parameters so every address for one page counts as one page.

Count normalized pages when comparing how many pages an engine cites. Keep the raw address separately when referral attribution matters.

A standard tracking parameter appeared in 916,648 of 1,212,559 ChatGPT Search cited-page appearances, or 75.5962%. Google AI Mode had one in 1,072 of 4,050,020 appearances, or 0.0265%.

A raw-URL count creates a much larger measurement risk for ChatGPT Search in this window. Engine comparisons should use one declared normalization rule.

All 916,648 tracked ChatGPT Search cited-page appearances included utm_source=ChatGPT.com. Those appearances occurred in 240,458 AI answers.

OpenAI documents this parameter as a way for publishers to identify referral traffic from ChatGPT search results. It is useful for attribution, but it should not create a second page identity in a citation audit.

After normalization, 39,137 pages had exactly two raw citation addresses. Two pages had three addresses, and one page had six. The remaining 346,147 pages had one address.

The aggregate overcount is broad rather than driven by a few pages with many variants. Counting by normalized page address resolves nearly all of it.

ChatGPT Search and Google AI Mode shared 9,086 exact citation addresses. After the standard tracking parameters were removed, they shared 62,471 pages. That added 53,385 shared pages, a 587.5523% increase over the raw-address count.

Cross-engine URL overlap is partly a normalization decision. Publish the URL rule beside every overlap statistic so readers can reproduce the comparison.

Among websites with at least 1,000 ChatGPT Search cited-page appearances, GitHub fell from 979 raw addresses to 841 normalized pages, removing 138 variants. Amazon Web Services removed 88, IBM removed 60, Apple App Store removed 54, and Google Cloud removed 51. Reddit had tracking parameters on 99.4564% of appearances but removed only 43 of 35,205 raw addresses, or 0.1221%.

How often tracking parameters appear and how much they inflate the page count are different measures. Audit both before deciding which websites create the largest counting error.

Removing only the five utm parameters reduced the ChatGPT Search address count by 9.2210%. The eight-parameter rule reduced it by 9.2231%. Adding msclkid and source changed the result to 9.2366%.

The headline is not determined by the three non-utm parameters in the main rule. This test is about how page addresses are counted, not about whether two pages show the same content. Query parameters that can change what the page shows remain distinct.

Standard tracking removal reduced ChatGPT Search's distinct cited-page count by 9.2231% and increased its measured shared-page count with Google AI Mode from 9,086 to 62,471.

Store the raw citation address for referral analysis and a normalized page address for page counts and overlap. Strip a declared list of tracking parameters. Preserve query parameters that can change page content. Rerun the same fixed-window rule before calling a difference movement.