UTM tagging for AI channels: what works and what breaks

Why a tag survives when a referrer does not

Referrer stripping works by instructing the browser not to send the referring page in the request header. A campaign tag is part of the URL itself, so it arrives with the request no matter what the referring page does.

This is the single most useful fact in AI measurement, and it has a hard limit: it only helps on links whose URL you wrote. If an assistant cites your page because it found it, the URL it cites is your plain URL and there is nothing in it to tag.

Where tagging actually applies

Links you control are a real but modest set. Your entries in directories and marketplaces. Links in your own documentation and knowledge base that assistants read and reproduce. Links inside structured data where you specify a URL. Profiles on third-party sites you maintain.

Everything else — the citations assistants generate by finding your pages — cannot be tagged, because you never write those URLs. Any tagging plan that promises full AI attribution is either misunderstanding this or hoping you do not notice.

A convention worth using

Keep it simple enough that it survives staff changes. Set utm_source to the assistant or surface where the link lives. Set utm_medium to something stable across the whole programme, such as ai_assistant, so a single filter catches everything. Use utm_campaign for the specific placement.

Do not put a campaign tag on links inside your own site. Internal tagging restarts the session in GA4, which breaks the attribution of the original source and produces the strange result of your own site appearing as one of your traffic sources.

Write the convention down in a shared place. An undocumented tagging scheme decays within a year and leaves you with a channel whose history is uncomparable.

What ChatGPT already does for you

OpenAI appends utm_source=chatgpt.com to outbound links in some contexts. That tag arrives in your analytics as a source without you doing anything, and it survives even when the referring page is absent.

This means some of your ChatGPT attribution comes from tagging rather than from the referrer, which is worth knowing when you see the same host appear through two mechanisms. It is not double counting; it is one visit identified by whichever signal survived.

Do not rely on it continuing. It is a behaviour of a third-party product, not a contract, and it has changed before.

What tagging will not fix

Tagging cannot recover a visit from a citation you did not write. That is most AI traffic for most sites, and no configuration changes it.

Tagging also cannot tell you about a mention that produced no click. If an assistant names you and the reader simply remembers it, nothing in your URL was ever loaded.

The honest position is that tagging raises the identifiable share and does not close the gap. It is worth doing because a modest improvement in a figure you report monthly compounds, not because it makes the measurement complete.

Auditing the tags you already have

Before adding tags, look at what is already arriving. In GA4, an exploration with Session campaign as the dimension shows every tag currently in use, including ones nobody remembers creating.

Look for three problems. Inconsistent casing, because utm_medium values differ between AI_Assistant and ai_assistant and will report as two channels. Internal links carrying tags, which restart sessions and corrupt attribution. And tags on links in email, which are legitimate but should not land in an AI channel by accident.

Then check whether your AI filter matches on source, on medium, or on both. Matching on medium alone is more robust as the host list grows, provided every tagged link uses the same medium value.

Fix the inconsistencies before adding anything new. A tagging scheme with three spellings of the same value is worse than no scheme, because it produces confident numbers that are quietly split across categories.

Deciding whether it is worth the effort

Tagging pays off in proportion to how many links you actually control. A business with a large directory presence and extensive documentation has real ground to gain; one whose AI traffic comes entirely from organic citations has almost none.

Count your controllable links before building a scheme. If the answer is a handful, the honest conclusion is that tagging will not move your measurement much and the effort belongs elsewhere.

Where it does apply, the gain compounds quietly. A tagged link keeps identifying visits long after the person who added the tag has forgotten it, which is the rare kind of work that keeps paying.

It is also worth being clear about what a good outcome looks like here, because expectations set wrongly turn a reasonable improvement into a perceived failure. Tagging will not take you from partial attribution to complete attribution. It will move the identifiable share up by whatever proportion of your AI traffic arrives through links you authored, and for most businesses that is a modest but permanent gain.

Treat it the way you would treat any other small compounding improvement. It costs an afternoon once, it keeps working without maintenance provided the convention is documented, and it makes every subsequent monthly figure slightly more complete than it would otherwise have been. That is a good trade, and it is a different claim from the one usually made for it.

The last consideration is who maintains it. A tagging convention survives exactly as long as the person who wrote it remains reachable, unless it is documented somewhere the next person will actually look. That place is usually not a personal document, and it is usually not a message thread either. Put it wherever your team keeps decisions that outlive individuals.