The practical consequence is a measurement gap. Visits that were briefly identifiable as Gemini-originated are again being absorbed into Direct traffic, and reports built on default analytics settings are likely to understate the contribution of content that AI systems cite. This article reviews what is known about the change, explains why AI referrals are prone to misattribution, and sets out a structured approach to measuring them with appropriate caution.

Why attribution accuracy matters

Attribution informs investment decisions. When leadership evaluates the return on content, or when budgets are allocated between channels, the underlying reports determine the outcome. If a growing source of visitors is recorded in a catch-all category, the work that earned those visits appears less effective than it is.

The issue is compounded by the changing shape of search. AI-generated answers now account for a larger share of research and comparison activity, and the click-through rate from those answers is generally low. Each identifiable visit therefore carries more weight, and each unidentifiable one represents a larger loss of information. Because platforms have offered no guidance on how AI assistants should pass visitors to websites, organisations must rely on their own measurement discipline.

What is known about the Gemini change

The reported sequence of events is as follows.

Around the start of September 2026, Gemini began adding the parameter utm_source=gemini to the source links in its answers. Google made no announcement. A Reddit user noticed the tags, and John Mueller of Google confirmed in the same thread that he could see them as well.

Server logs later published by search consultant Natzir Turrado show the parameter disappearing overnight around September 28 or 29, after approximately three weeks. Search Engine Journal reported the original addition on October 1. Turrado drew attention to the removal on October 27, and Search Engine Roundtable reported on October 30 that it could no longer find the parameter on Gemini links. Google has not commented on either the addition or the removal.

Why the parameter mattered

A UTM parameter is the most direct way to label the origin of a visit. In its absence, analytics platforms rely on the HTTP referrer, the information a browser passes to indicate where a visitor came from.

On desktop and the web, Gemini visits continue to arrive with a gemini.google.com referrer and can therefore be identified. The weakness lies on mobile. One analysis reports that Gemini’s iOS application opens links in an in-app browser that does not pass a referrer, causing those visits to be recorded as Direct in Google Analytics 4. The same analysis states that the parameter had previously exposed between 16 and 65 additional hidden clicks for every 100 sessions identified through a standard referrer. As this derives from a single source, the range should be treated as indicative rather than definitive. The underlying point is nonetheless clear: a portion of AI-originated traffic cannot be identified using standard methods.

The risk of artificial trends

Changes in tagging behaviour can produce patterns in the data that have no connection to audience behaviour. An analyst reviewing Gemini traffic across September would have observed it emerge, grow for three weeks and then fall sharply. That curve reflects a parameter being enabled and disabled, not a change in interest.

ChatGPT provides a comparable example. According to reporting, its links lost the utm_source=chatgpt.com parameter between approximately October 7 and 16 before it was restored, and the interruption was subsequently described as a bug. A comparison of October with September could suggest a decline in AI-referred traffic that did not occur.

The appropriate safeguard is to annotate analytics reports with the dates on which sources changed their tagging, and to treat any abrupt movement that coincides with those dates as a measurement artefact unless other evidence indicates otherwise.

Interpreting growth in Direct traffic

Direct traffic has always served as a residual category. It includes visitors who typed an address, used a bookmark, or arrived from an application that does not pass a referrer. AI assistants have added a further source of the last type, particularly within mobile apps.

This does not mean that all Direct traffic originates from AI, nor that Direct traffic should be reclassified wholesale. It does mean that sustained growth in Direct visits, especially to informational pages that users rarely enter manually, merits investigation. It is a lead to be tested, not a finding to be reported.

A multi-signal measurement approach

No single method resolves the problem. The following measures are most effective when used together.

1. Define an AI channel in GA4. A custom channel group allows known AI referrers, such as chatgpt.com and gemini.google.com, to be reported together instead of being dispersed across Referral and Direct. The group should be built from sources that actually appear in your own data.
2. Review server logs. Logs capture details that analytics tools omit, including the user-agent string. One analysis reports that Gemini’s iOS visits can be recognised by a distinctive token in that string, which makes it possible to estimate visits that analytics classifies as Direct. Feasibility depends on developer or hosting support.
3. Monitor landing-page patterns. Comparing Direct traffic to deep informational pages against a historical baseline can reveal unusual increases. This is supporting evidence, not proof.
4. Collect self-reported attribution. Adding a “How did you hear about us?” field to enquiry or checkout forms, with an option for AI assistants, captures discovery that no tag can record. Responses are imperfect but valuable as a cross-check.
5. Track brand demand. Where AI answers mention an organisation, branded search and direct visits to the home page often increase. Observing these indicators alongside content performance shows influence that click data does not.
6. Apply UTM parameters to controlled links. Links placed in newsletters, profiles and partner pages should carry the organisation’s own tags, because third-party platforms cannot be relied upon to add them.
7. Combine with Search Console data. Google Search Console’s AI reporting shows impressions in AI Overviews and AI Mode, though not clicks. Used alongside first-party analytics, it provides a second partial view of the same activity.

Avoiding two opposite errors

Two interpretive errors are common. The first is to conclude that AI traffic is negligible because the identifiable portion is small; a substantial share may be recorded as Direct, so the true volume is probably higher than reported. The second is to attribute every increase in Direct traffic to AI, which, without corroborating evidence, amounts to conjecture.

A defensible position states what can be measured, identifies what cannot, and offers a transparent estimate of the difference.

Frequently asked questions

Did Gemini stop adding UTM parameters?
According to server-log data from Natzir Turrado and a Search Engine Roundtable report dated October 30, 2026, yes. The parameter appeared on cited links for roughly three weeks in September and was then removed. Google has not commented.

Can Gemini traffic still be tracked?
In part. Desktop and web visits retain a gemini.google.com referrer. Visits from the mobile app may be recorded as Direct.

Why does AI traffic appear as Direct?
Some AI applications open links in their own browser without passing a referrer. With neither a referrer nor a UTM parameter, analytics platforms have no basis for attribution and default to Direct.

Did ChatGPT also lose its tracking parameter?
Reportedly, for about a week in October, between roughly October 7 and 16. The interruption was later described as a bug.

Should organisations stop using UTM parameters?
No. They remain essential for links an organisation controls. They should simply not be assumed on links generated by third-party platforms.

Conclusion

The removal of Gemini’s UTM parameter is a minor technical event, but it illustrates a broader limitation. AI assistants pass visitors to websites inconsistently, and most analytics tools were designed for an earlier model of web traffic. Parameters will continue to appear and disappear, applications will continue to share limited information, and new assistants will continue to emerge.

Organisations that adapt best will be those that rely on several imperfect signals rather than a single tag, and that report their limitations candidly. A clearly qualified estimate is more useful than a confident figure that is incorrect.

Adapted from a longer article originally published at kilowott.com.

This is a common story right now, and it often traces back to an old assumption: that targeting does the heavy lifting and creative follows. Many advertisers now report the reverse. They upload a range of genuinely different ads and let Meta decide who sees which one.

The system usually credited with this shift is Andromeda. This post explains what it is, what Meta has said publicly, where industry advice goes beyond that, and what to do before Black Friday.

What is Meta Andromeda?

Andromeda is Meta’s ads retrieval system. Before any bidding or ranking happens, Meta has to narrow tens of millions of eligible ads down to a shortlist of a few thousand worth scoring for a specific person. Andromeda performs that first cut.

Meta described it on its engineering blog on December 2, 2024, and has since said it has rolled out across its apps. It isn’t a campaign type or a setting in Ads Manager. It’s part of how the auction works.

The practical implication is simple: retrieval happens before the auction. An ad that isn’t shortlisted never gets the chance to win a bid. That’s why the quality and variety of your creative matters more than it once did.

What Meta has published

According to Meta’s announcement, Andromeda delivered:

• 10,000x more retrieval model capacity
• A 6% improvement in recall
• An 8% improvement in ads quality on selected segments

The quality gain applies to selected segments, not to every advertiser. These numbers describe system performance, not the results any single account will see.

Meta also published a newer engineering post on October 28, 2026. Summaries of it say the current system can process more than a billion ads per second on a single GPU server, a 100x lift over the 2024 system. Reporting also says a first generative model was deployed into ads retrieval in July 2026. The details will keep changing, but the direction is clear: more capacity to consider more ads, matched more precisely.

Creative diversification: what Meta recommends

Meta’s guidance since early 2025 has been consistent. As advertisers upload more creatives, Andromeda powers the models that choose the right one for each person. Meta has described the focus as moving from niche targeting to creative diversification.

Diversification means genuinely different ads, not small edits to one ad:

• Format: video, static images, partnership content
• Theme: humour, nostalgia, savings, social proof, problem-led storytelling

A new caption on the same footage is iteration, not diversification. The idea is that distinct ads reach distinct kinds of people, while near-duplicates compete for the same audience.

Meta’s guidance vs. industry advice

It’s worth separating what Meta has said from what practitioners report. Meta hasn’t published a rulebook. Its posts recommend diversification, but they don’t say how many ads to run, how to structure ad sets or how to read results.

Much of the specific advice online goes further: fewer ad sets, more ads per ad set, campaign-level budgets, and creatives stacked across awareness stages. Many agencies report good results from these approaches, and the logic is plausible, since pooling data may help the system learn. But those results come from individual accounts, not from Meta, and they may not transfer to every brand.

Treat them as hypotheses to test, not rules. The same applies to exact creative counts or similarity thresholds. If a number isn’t in Meta’s own material, it’s an observation.

A practical approach

Whatever structure you choose, these steps make sense on the available evidence:

1. Fix your tracking first. Meta can only learn from the data it receives. Check that purchase events are sent reliably, including server-side, and that customer details are matched correctly.
2. Build a creative map. List the angles you could take (problem, solution, proof, offer) and the formats you could use. Cover several of each, rather than ten versions of one idea.
3. Test structure changes one at a time. If you want to try consolidating ad sets, change one campaign and compare it against a control.
4. Plan a steady supply of new creative. Creative fatigues, so a pipeline works better than a one-off batch.
5. Match the landing page to the ad. A story-led ad needs a page that continues the story.
6. Read results at the ad level. See which ads carry the delivery, and check platform numbers against your own analytics.
7. Keep product feeds clean. For catalogue and Advantage+ shopping campaigns, accurate product data feeds the system’s decisions directly.

Why timing matters

With Black Friday on November 27, there’s a practical deadline. Holiday periods are when creative fatigues fastest and competition is highest, so having distinct ads ready before the peak is more useful than scrambling during it.

Quick readiness checklist

• Purchase events sent reliably, with customer details matched
• A creative map covering several angles and formats
• Genuinely different ads, not edits of the same footage
• A plan for fresh creative every few weeks
• Landing pages matched to each ad’s message
• Structure changes tested against a control
• Results checked against independent analytics
• Product feed accurate and complete
• Holiday creative approved before the peak

FAQ

What is the Meta Andromeda update?
Andromeda is Meta’s ads retrieval system. It narrows tens of millions of eligible ads to a few thousand candidates before ranking. Meta announced it on December 2, 2024.

Does Andromeda replace targeting?
Not entirely, but it changes where the work happens. Meta’s guidance says the focus has shifted from niche targeting to creative diversification.

What is creative diversification?
Running ads that are genuinely different in format, theme and message, not minor edits of one ad.

How many ads should I run?
Meta hasn’t published a number. Test what works for your account and focus on making each ad genuinely different.

Should I consolidate my ad sets?
Many agencies report good results from fewer ad sets, but that’s practitioner advice, not a Meta rule. Test it on one campaign first.

Final thoughts

Andromeda rewards having something distinct to say. The advertisers who benefit won’t necessarily be the ones with the cleverest account structure. They’ll be the ones who treat creative as the main lever, keep data reliable and test carefully.

Meta’s systems will keep changing, and advice that works this month may look different by spring. Different ads, clean data and honest measurement hold up through each update.

This article is adapted from a longer piece originally published at kilowott.com.