ChatGPT UTM Parameters Removed: What AEO Teams Should Do

Yes, it makes ChatGPT traffic harder to measure, and it probably makes your AI search numbers look smaller than they are. Search Engine Roundtable reports that OpenAI appears to have removed or replaced the UTM parameters it used to add to links in ChatGPT answers, starting around the GPT-6 release a week ago. The fix is not a new tag. It is a reporting setup that does not depend on any single parameter.

Key takeaways

• ChatGPT clicks are now harder to separate from direct and unattributed traffic, so reported AI search numbers may drop even if real visits do not.
• Treat this as a measurement problem, not a visibility problem. Nothing reported so far suggests ChatGPT changed how it picks or cites sources.
• Rebuild attribution from several signals: referrer data, landing-page patterns, server logs, branded search and self-reported source.
• Write your channel rules into a tracking specification so a vendor change cannot silently rewrite your reports.
• Annotate the break date in every dashboard and stop comparing pre- and post-GPT-6 AI traffic as one continuous series.

What changed in ChatGPT links after GPT-6

Until recently, links that ChatGPT placed in its answers carried a UTM parameter that named ChatGPT as the source. That was a gift to analysts. A visit with that tag could be grouped into a clean AI channel with one filter, and nobody had to guess.

Search Engine Roundtable now reports that those parameters seem to be gone or replaced since GPT-6 shipped. The write-up frames the consequence simply: it is harder to tell whether a click came from ChatGPT or from some other platform.

Two cautions belong here. First, this was reported as something that seems to be happening, not as an announced product change. OpenAI may restore, rename or reshape the tagging, and you should expect the behaviour to move again. Second, a missing UTM is not the same as a missing referrer. Whether a given click still arrives with a referring domain depends on the browser, the app, the link type and the visitor's settings. Look at your own raw landing-page data before you decide how bad it is.

If you want a working definition of what to check, open your analytics and compare the last two weeks of sessions that used to be tagged. Look at where they went: direct, referral, or an unassigned bucket.

Why a missing UTM matters for AEO and GEO reporting

The damage is to your story about AI search, not to AI search itself.

Consider a content lead at a B2B software company who has spent a year making the case for answer-engine work. The monthly report shows a line called AI referrals, built from the ChatGPT UTM plus a few referrer rules. It has gone up every quarter. This month it falls off a cliff. Nobody changed the content. Nobody lost a citation. The tag disappeared.

Without context, finance reads that as a failed programme. That is the real risk: a measurement break gets interpreted as a performance break.

There is a second, quieter problem. Traffic that used to be labelled ChatGPT now lands in direct or in generic referral. Your direct channel grows for no visible reason, and anyone using direct as a proxy for brand strength will read the growth as brand strength. Some of it is AI traffic that lost its label.

And a third: attribution models that give credit to the last tagged touch will quietly reassign value. A visitor who found you in ChatGPT, then returned later through a branded search, was already partly invisible. Now the first half of that journey is fully invisible.

None of this is unique to OpenAI. Any answer engine can change its link handling, and several already differ from each other. The lesson is to stop treating one vendor's parameter as the foundation of a channel.

How answer engines choose and cite sources

It helps to separate two questions that get blurred: whether you are being chosen, and whether you can see the click.

Choosing is the visibility side. An answer engine takes a prompt, often expands it into several related searches, pulls candidate pages, and assembles an answer that cites some of them. Ahrefs published a study of how ChatGPT does this kind of query fan-out, and the useful takeaway for practitioners is that a single prompt can touch many sub-queries you never typed. Your page may be retrieved for a question adjacent to the one the user asked.

What tends to make a page citable is mundane. The passage answers one question completely. The claim is specific and attributable. The page is crawlable and fast enough to fetch. The entity, whether that is your company, product or author, is described consistently across the web. None of that is affected by a link parameter.

Seeing the click is the measurement side, and this is where the UTM change bites. A citation can be earned and never produce a trackable visit, because many users read the answer and leave, or copy a name and search for it later. Those are real effects of your visibility that never appeared in a UTM report even before this change.

So keep the two layers apart in your reporting:

• Presence: are you cited, for which prompts, in which engines, and how often?
• Pull: do visits, searches and conversions follow?

The UTM change damages the second layer. It should not push you to stop working on the first.

What to change in your analytics this week

You do not need a new platform to respond. You need a few deliberate changes, in this order.

1. Mark the date. Add an annotation in every dashboard where AI referral traffic appears. Use the date the tag stopped showing up in your own data, not the date of any article.
2. Audit the referrer. Pull a list of sessions whose referrer domain belongs to ChatGPT, then compare it against sessions that carried the old UTM. The gap tells you how much you were relying on the parameter alone.
3. Write the channel rule down. Define in one document which referrers, parameters and landing-page patterns count as AI search, who owns the rule and when it was last checked. A tracking specification is the right format for that, and OmniSpec exists for teams that want that governance kept in one place.
4. Create an unattributed-likely bucket. Do not pour uncertain traffic into direct and hope. Give it its own name so the uncertainty is visible to readers.
5. Add a source question. On demo forms and signup flows, add an optional how did you hear about us field with an AI assistant option. It is blunt, and it catches visits that no tag ever would.
6. Check server logs. If your CDN or origin logs record the referrer and user agent, they are not dependent on client-side tagging and can corroborate what your analytics tool shows.

One point of judgement: resist the urge to add your own UTM to every link you can. You do not control the links ChatGPT builds. You do control links you place in profiles, newsletters and partner pages, and tagging those is still worth doing, but it will not recover the lost signal.

A measurement plan that does not rely on one parameter

The sturdier approach is triangulation. No single signal is trustworthy on its own, but several imperfect ones that agree are good enough to run a programme on.

Signal	What it tells you	Weakness	Owner
Referrer domain	Visits that arrive with a ChatGPT or other engine referrer	Dropped by some browsers and apps	Analytics lead
Landing-page pattern	Rising entries on pages that answer common prompts	Overlaps with organic search	SEO lead
Server or CDN logs	Fetches by AI crawlers and referred human visits	Needs engineering time to query	Web engineering
Branded search volume	Demand created by being named in answers	Slow, and affected by other campaigns	SEO lead
Self-reported source	What the buyer says	Small sample, recall bias	Demand generation
Citation tracking	Whether and where you are cited in answers	Prompts sampled, not exhaustive	Content or AEO lead

For the citation row, MediaPilot is the Cresia product aimed at media and AI-search visibility, and the product page describes what it covers.

A practical way to use this table is to pick three signals, agree how often you review them, and report them side by side. If referrer visits fall but citations hold and branded search is steady, you have a tagging story. If all three fall together, you have a visibility story. That distinction is exactly what a missing UTM used to make impossible to see.

What not to do after the change

A few reactions look sensible and cause harm.

• Do not rebuild the old number with guesses. Taking this month's direct traffic and assigning a fixed share to ChatGPT produces a figure that looks precise and has no basis. If you must estimate, label it as an estimate and show the range.
• Do not rewrite content because a report dipped. Pages that were earning citations last month are still the same pages. Changing them in response to a tagging break can remove what was working.
• Do not chase the new parameter. If OpenAI replaced the tag with something else, build your rules to survive another replacement rather than swapping one dependency for another.
• Do not report a single blended AI channel as if it were stable. Engines differ in how they pass referrers. Split them where you can.
• Do not stop reporting presence. Citation share and prompt coverage are the leading indicators. Clicks lag them and always will.

The common thread is restraint. Measurement breaks invite overreaction, and the overreaction is usually more expensive than the break.

Who needs to be told, and what to say

The fastest way to lose a budget conversation is to let someone else find the dip first. Brief the people who read these numbers before the next reporting cycle.

Keep the message short. Say what changed, the date it showed up in your data, which reports are affected, and what you are doing about it. Then say what has not changed: your citation work, your content, and the other channels.

Growth teams will care about the funnel numbers. Analytics teams will care about the channel definition and whether history needs a break marker. If you are sorting out which team should own which part, the analytics teams solution page shows how Cresia frames that work.

One more position worth taking: do not wait for OpenAI to confirm anything before acting. The cost of putting an annotation and a tracking rule in place is small, and if the tag returns you simply move the annotation.

Frequently asked questions

Did OpenAI officially announce that UTM parameters were removed?

The report from Search Engine Roundtable describes the parameters as seemingly removed or replaced, not as an announced change. Check your own landing-page and referrer data to see what your site actually receives, since that is the only evidence that matters for your reporting.

Will this reduce how often my content is cited in ChatGPT?

Nothing in the reported change relates to source selection. Citations depend on whether your pages answer the prompt and its related searches clearly and can be fetched. What you lose is some of your ability to see the resulting clicks.

Can I still identify ChatGPT visits in GA4 or another analytics tool?

Often partly, through the referrer domain and landing-page patterns, but not as cleanly as before. Expect some visits to fall into direct or unassigned, and set up a separate bucket so that uncertainty is visible.

Should I stop reporting AI search traffic until things settle?

No. Keep reporting it, add a break annotation, and pair it with citation and branded-search signals. A number that is flagged as incomplete is more useful than a gap in the report.

Is this likely to happen again with other answer engines?

It is reasonable to plan for it. Engines change how they build and tag links without much notice, so the safer assumption is that any single parameter can disappear. Documented channel rules and several corroborating signals make the next change a small annotation instead of a crisis.

Sources

• https://www.seroundtable.com/utm-parameters-missing-chatgpt-42299.html
• https://ahrefs.com/blog/query-fan-out-chatgpt/