Microsoft Copilot UTM parameters

Microsoft Copilot is the assistant layer Microsoft has stitched across its consumer surfaces — the standalone app at copilot.microsoft.com, the Copilot button in Edge, and the conversational answers inside Bing. Ask it a question and it answers in prose, often citing sources with outbound links. Those clicks are real visits, and your analytics already identifies them.

Copilot is a place people read answers, not a console where you configure campaigns. Every other “sender” in this catalog — Meta Ads, Google Ads, the rest — gives you a field for a UTM template and a click ID it appends automatically. Copilot gives you neither, and needs to give you neither: the referrer does the job.

How Copilot traffic appears in analytics

GA4 ships a native AI Assistants default channel. Google’s documented rule is that the medium “is set to ai-assistant and the campaign is set to (ai-assistant) if the referrer matches a list of AI Assistants” — and Copilot is one of the sources Google names, alongside ChatGPT, Gemini, Grok, and Deepseek. So ai-assistant is a value GA4 writes once it recognizes the referrer, not something you supply.

AnyTrack does the equivalent, assigning every session a channel from the referrer and URL parameters, with conversions attached in the Campaign Report.

Two wrinkles specific to Copilot:

• The Bing and Edge blur. Copilot’s tight coupling to Bing and Edge means that depending on the surface and the session, the same kind of click can surface under a Bing referrer or a generic browser referral rather than copilot.microsoft.com. Read the split by session source and accept that it isn’t perfectly clean. No tag on your pages would make it cleaner — it would only relabel every visitor on that link.
• Referrer-less clicks. A share arrive with no header at all and land in (direct)/(none). Auto-detection can’t fire on a signal that never came.

Why you shouldn’t apply a UTM convention here

The common advice — tag “links you control that Copilot may cite” with utm_source=copilot&utm_medium=ai-assistant — doesn’t survive inspection.

Copilot cites your canonical URL. The only way to get parameters onto it is to publish your canonical URL with them attached, so every visitor from organic search, email, referrals, and bookmarks lands on the tagged link and reports as Copilot traffic. UTMs also take precedence over referrer detection in GA4, so you’d overwrite a classification that was already correct — and writing ai-assistant by hand asserts a referrer match GA4 never made.

For the receiver-side mechanics, see Google Analytics .

Limitations and gotchas

• No click ID, no per-citation token. Copilot exposes nothing like gclid. You can confirm that a click came from Copilot, never which prompt or answer produced it.
• You can’t tag a link the model writes. If Copilot cites your homepage, that link carries whatever is on your canonical URL. There’s no way to reach into the model’s output and append a parameter.
• The Bing/Edge blur is structural. Don’t expect one clean referrer, and don’t try to force one by tagging your pages.
• Referrer-less clicks are a hard floor. They land in Direct and nothing recovers them. Report the AI channel as a known undercount.
• Server-side capture is the durability fix. It preserves sessions that did carry a referrer but would otherwise be lost to ad blockers, consent gates, and fast bounces.

Verification

1. Check the AI Assistants channel. In GA4, open Reports → Acquisition → Traffic acquisition by default channel group. AI Assistants should be a row with no configuration. Switch the dimension to Session source to see copilot.microsoft.com inside it.
2. Ask Copilot a question it should answer with your page, click the citation, and inspect the landing URL and the referrer your analytics records — that tells you how Copilot’s referrer behaves for your domain specifically, including whether it surfaces under Bing.
3. Confirm your canonical URLs are bare. Check the pages Copilot is most likely to cite for stray utm_ parameters.

For the bigger picture — reading AI traffic across all five assistants without inventing a convention for it — see the related guides below.