Gemini dropped its UTM tags and GA4 may be missing orders. A holiday measurement check for CPG brands

Holiday reporting season is about to start, and two stories from the past week say the same uncomfortable thing: the numbers in your analytics may be shrinking for reasons that have nothing to do with shoppers.

The first is small and specific. Search Engine Roundtable reported on October 30 that Google's Gemini stopped attaching UTM tags to the links it cites. The second is broad. An audit published in Search Engine Journal on October 28 found GA4 recording a falling share of real orders at several online stores, in one case by enough to show a sales drop that never happened.

For a packaged goods brand selling direct, both matter. AI assistants are a growing source of high-intent visits, and holiday budgets for next year will be set on this quarter's reports. Here is what changed and how to check your own numbers.

What changed with Gemini

In early September, Google began adding utm_source=gemini to links cited in Gemini answers. That made Gemini visits easy to find in any analytics tool. According to data from SEO consultant Natzir that Search Engine Roundtable reported, the tags stopped appearing around September 28.

What happens to those visits now depends on where the shopper is.

• Desktop and web Gemini still pass a gemini.google.com referrer, and Natzir's data showed those visits holding steady. GA4 still attributes them to Gemini as a referral.
• The Gemini iOS app opens links in its own in-app browser and passes no referrer. Without a UTM tag, GA4 has nothing to go on, so those visits are counted as Direct.

The only remaining trace for iOS app visits, per Natzir, is a GeminiiOS string in the user agent, which shows up in server or CDN logs but not in standard GA4 reports.

Search Engine Roundtable's report includes no explanation from Google. The practical upshot is that if your GA4 showed a Gemini source in September and it shrank in October, some of that traffic probably did not disappear. It moved to Direct.

Why a small change matters at holiday time

Gemini is not the largest source of AI visits for most brands. But AI referral traffic is valuable out of proportion to its size. Adobe reported on October 29 that AI-referred visits to US retail sites converted 90% better than other traffic in September, which we covered in our look at the Adobe numbers. Losing attribution for a high-converting slice makes AI look less useful than it is, and makes Direct look better than it is.

We walked through the general method for finding AI visits in GA4 in our earlier GA4 guide, including a channel group for assistant referrers. That guide still works for referrer-based traffic. What it cannot catch is a visit that arrives with no referrer and no tag.

The bigger problem: GA4 missing orders

The Search Engine Journal audit, by Christian Goodrich of the agency SOZO, looked at eight UK online stores and compared what GA4 recorded with what the store platforms actually took. The gaps were larger than most teams would guess.

Retailer ARetailer B

Show the data

Period	Retailer A	Retailer B
Q4 2024	81%	94%
Q4 2025	41%	36%

Source: Search Engine Journal, October 28, 2026. Over the Christmas period, Retailer A's capture fell to 32%.

For Retailer A, GA4 captured 81% of web orders in Q4 2024, 41% in Q4 2025, and 32% over Christmas. Between Christmas Day 2025 and January 4, 2026, GA4 showed revenue down 68% on the year before, when the store's own records showed a 13% fall. Across the quarter, actual web orders rose from 5,874 to 7,686 while GA4 showed them falling by a third.

Retailer B fell from 94% capture in Q4 2024 to 36% in Q4 2025, and GA4 showed revenue down 69% while net sales rose 18.5%. The audit tied much of that to the cookie banner. Both stores moved to a new consent setup in 2025, and each consent plan covered a set number of page views a month; once the allowance ran out, the banner stopped showing until the next billing date. At Retailer B, GA4 captured up to 68% of orders in weeks the banner stayed up and 10% to 22% in weeks it switched off.

Retailer C had a different failure. GA4 recorded zero purchases from October 28 to December 4, 2025, while the store took 2,215 orders. On December 9, GA4 recorded 1,261 orders in a single day, 1,000 of them from inside the gap. Across the quarter, 657 of the 3,365 order numbers GA4 recorded had more than one purchase event, adding 735 purchases. In the week of November 3, Google sent 4,244 clicks according to Search Console, while GA4 recorded 57 Google organic sessions.

Table: What went wrong at each audited store

Store	What the audit found	How it showed up in GA4
Retailer A	Capture fell from 81% to 41% of web orders, 32% over Christmas	Christmas revenue down 68% when store records showed down 13%
Retailer B	Capture fell from 94% to 36%; cookie banner switching itself off	Revenue down 69% while net sales rose 18.5%
Retailer C	No purchases recorded for five weeks, then a late backfill	657 of 3,365 order numbers counted more than once
Retailer D	Bot spikes of 23,187 and 25,622 sessions on Sept 27 and 28, 2026	Normal days ran 300 to 900 sessions
Others	Steadier gaps	Roughly 72% to 84% of orders captured

Source: Search Engine Journal, October 28, 2026. UK stores; numbers as reported in the audit.

These are UK stores, where consent rules expect refusing tracking to be as easy as accepting it. In the US, Goodrich notes, opt-out signals mainly affect advertising tags, while browser privacy features and ad blockers chip away at analytics, so US gaps may be smaller. But the failure modes he found, a consent tool that quietly switches off, a tag that stops firing, duplicate events and bot traffic, can happen at any store.

Check one: your weekly capture rate

The audit's main recommendation is simple enough to set up this week. Each week, divide the number of distinct purchase transaction IDs in GA4 by the number of orders your store platform took. That ratio is your capture rate.

1. Export weekly web orders from your store platform for Q4 2024 and Q4 2025, leaving out marketplace, subscription and manual orders.
2. Pull GA4 purchases for the same weeks as distinct transaction IDs on purchase events. The audit warns against the transactions metric, which counts refunds as well as purchases.
3. Divide GA4 purchases by store orders and chart the ratio. A steady rate, even a low one, is usable for trends. A step change, a near-zero week or a week above 100% is a tracking problem, and makes year-over-year comparisons wrong.
4. Note the weeks when anything changed: consent banner updates, theme releases, checkout app installs, tag manager publishes. Most drops line up with one of these.

The audit also suggests using Search Console as an early warning: divide GA4's Google organic sessions by Search Console clicks each week, since Search Console counts clicks whatever your tags are doing. And it recommends sending the purchase from the store's own system when the order is created, through the Measurement Protocol or a server-side container fed by the store, keyed on the order number, while keeping the browser tag for behavior.

Check two: where your Gemini visits went

For the AI side, three checks will tell you how much Gemini traffic you are losing to Direct.

1. In GA4, compare sessions with source gemini or referrer gemini.google.com for September and October. A drop after late September that is not matched by a drop in Search Console impressions points to the tagging change.
2. In the same weeks, look at Direct sessions landing on deep product pages rather than the home page. Shoppers rarely type a product URL by hand, so a rise there is a clue.
3. If you can read server or CDN logs, search the user agent field for GeminiiOS. That is the clearest sign of app visits GA4 cannot attribute. Our tip on checking whether AI crawlers visit covers where to find those logs on common hosts.

Put both checks in the holiday report

The point is not to build a perfect number. It is to stop a broken one from driving decisions. A holiday report that says "AI referrals were flat and online sales fell" may really mean "GA4 lost a third of orders to consent and Gemini visits moved to Direct."

We would add two lines to every weekly report from now through January: the capture rate, and the share of sessions arriving with no source. If either moves sharply, look at the tags before you look at the market.

Referral data has never been the whole picture for AI search anyway. Many shoppers read an assistant's answer and then buy at a retailer, which no brand analytics will see. Teams that want the answer side of the picture can track it directly; XENA Intelligence, for example, records what assistants say about products when asked shopper questions, product by product, in SearchPanel. Our holiday AI shopping playbook has the rest of the season's checklist.