# For about a week, ChatGPT stopped tagging its citation links, so GA4 filed those clicks as Direct

> From GPT-6 on October 7, 2026 until a fix reported October 15, many ChatGPT citation links lost their utm_source tag, so GA4 filed the clicks as Direct.

Canonical: https://brandonlazovic.dev/articles/chatgpt-clicks-arrive-as-direct/  
Author: Brandon Lazovic  
Published: 2026-10-15

## The short version

- From GPT-6's arrival on October 7 and 8, 2026 until a fix reported on October 15, a growing share of ChatGPT's citation links carried no utm_source=chatgpt.com tag and told the browser to send no referrer, so the visit arrived with no source and GA4 recorded it as Direct.
- Profound counted answers with at least one tagged source link falling from about 98% on October 8 to 66% by October 14, while answers in the new Intelligent UI layout rose from 2% to about 35%.
- A click I made on October 15 showed both halves, an untagged link marked noreferrer in ChatGPT's own markup and an empty referrer on the landing page, and GA4's AI Assistant channel cannot catch such a visit because it matches on the referrer.
- The untagged window ran about a week, so mark October 7 to 15 in every series, estimate the displaced share from landing pages ChatGPT already sent traffic to, and keep tagging the links you supply yourself.

From GPT-6's arrival in ChatGPT on October 7 and 8, 2026, a growing share of the links in ChatGPT's answers carried no source tag (`utm_source=chatgpt.com`) and told the browser to send no referrer, the signal that names the page a click came from. [2] [3] [4] A link I clicked on October 15 had no tag, and the retailer's page got an empty referrer. [1] Google Analytics 4 (GA4) had nothing left to classify, so it recorded the visit as Direct. Later on October 15, Cedric Huesler of Adobe reported the tags were back and called the drop a bug. [3] If your chatgpt.com sessions dipped that week, read the dip as lost attribution before you read it as lost visibility.

> ChatGPT answers carrying at least one tagged source link fell from about 98% on October 8 to 66% by October 14, 2026, by Profound's count, and the untagged clicks land in Direct.

## What changed in ChatGPT's citation links in October 2026?

ChatGPT's citation links changed in two ways after GPT-6 and a new answer layout called Intelligent UI rolled out on October 7 and 8, 2026. [2] Many links lost the `utm_source=chatgpt.com` tag, a UTM (Urchin Tracking Module) parameter that analytics tools read as the visit's source. Many also told the browser to send no referrer, the header naming the page a click came from. [3]

Cedric Huesler of Adobe described both changes on October 14, and the link I inspected a day later showed both. [3] [1] Paid plans got the update on October 7 and the Free and Go plans on October 8. [2]

The two changes are separate, and each one used to be enough on its own. A tagged link told GA4 the source even when the referrer was missing, which is common on mobile apps. A plain link with a referrer still showed up as chatgpt.com in referral reports. With both gone, nothing in the request points back to ChatGPT.

The referrer half is a standard browser instruction. The Mozilla Developer Network (MDN) Web Docs reference says `noreferrer` "instructs the browser, when navigating to the target resource, to omit the Referer header and otherwise leak no referrer information." [6] The attribute sits on the link itself, so it travels with every click on that link.

The change was partial and spreading. Profound, which sells AI-visibility monitoring, counted answers with at least one tagged source link falling from about 98% on October 8 to 75% on October 9 and 66% by October 14. [4] Over the same days, answers rendered in Intelligent UI went from 2% to about 35% of the answers it collected. [4] OpenAI's release note describes that layout as answers that "combine text, visuals, and interactive elements," and it says nothing about links or tracking. [2]

## How much ChatGPT traffic moved into Direct?

The ChatGPT traffic that moved into Direct is large on the pages that measured it, and part of the drop has not reappeared anywhere. Cedric Huesler of Adobe looked at about 500 pages where ChatGPT drove 5% or more of visits, comparing October 9 to 13 with the week before. [3] ChatGPT visits fell about 31,000 a day, and Direct rose about 26,000 a day on the same pages. [3]

Those are his figures, from sampled data for October 9 to 13, and I found no independent check of them. The gap matters as much as the overlap. By my arithmetic on his two figures, about 5,000 visits a day fell out of the ChatGPT line and did not show up as Direct, so some of the decline may be real.

A second dataset points the same way. LLM Pulse, which syncs GA4 referral data for the sites it monitors, found daily ChatGPT sessions down about 46% from October 9 to 11 against the same days in the two prior weeks, with Perplexity and Claude referrals steady. [5] Its own summary says "we cannot tell whether the visits are lost or just arriving untagged as Direct." [5]

## How did I confirm that ChatGPT clicks arrive with no referrer?

I confirmed it on October 15, 2026 with citation clicks in desktop Chrome while logged in to ChatGPT, checking the link in ChatGPT's page, the landing page in the browser's developer tools, and an analytics request from the same session. None of the three carried a ChatGPT source. [1]

1. In ChatGPT's page, the link markup had no tag and carried `noreferrer`.
2. On the landing page, the browser's developer tools showed an empty referrer (`document.referrer`) and nothing after the `?` in the address (`location.search`).
3. In a GA4 request from the same session, the address carried no campaign tag and the visitor was on a first visit.

Copied from ChatGPT's page with the retailer's domain replaced, the link read like this.

```html
<a data-d-component="link" href="https://www.[retailer].com/collections/mens-shirts" target="_blank" rel="noopener noreferrer" class="x1heor9g x1i9j9tv ...">
```

The analytics request I captured came later in the same session, an engagement event on a second collection page of the same retailer. The address field in the request GA4 sends held that collection URL with no tag. It also showed a first session for a visitor ID created that day, with analytics consent granted, so neither a cookie banner nor an older traffic source explains the result.

The console reading, not this request, carries the referrer claim, because GA4 may omit the referrer field on an engagement event either way.

The test has limits. It is one session, on desktop, while logged in. I have not tested the ChatGPT phone app or a logged-out session, and I cannot say whether that answer used Intelligent UI. Profound's count says some answers still carry tagged links, so my click shows the mechanism, and its share comes from Profound.

## Why does GA4 record these ChatGPT visits as Direct?

GA4 records these ChatGPT visits as Direct because Direct is the label it gives a session that arrives with no campaign tag, no ad click identifier and no referrer. Google's channel definition makes Direct the bucket where the source is `(direct)` and the medium is `(none)` or `(not set)`, which is what a request with nothing in it produces. [7]

Even GA4's AI Assistant channel cannot rescue the visit. Google says the channel covers "sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok," and assigns it when the medium is `ai-assistant` or when "the referrer matches a list of AI Assistants." [7] A blocked referrer leaves that rule nothing to match. A custom channel group built on chatgpt.com as a referrer has the same blind spot, because it reads the same empty field.

Returning visitors blur the picture further. Google says "Sessions initiated by direct entrance are attributed to the UTM values for that user," and the session lookback follows the attribution setting, 90 days by default. [8] [9] If a returning visitor clicks a ChatGPT link, GA4 can credit the click to whatever tagged source brought them last, such as an email or a paid click. Only a first visit, like mine, lands cleanly in Direct.

The label is fixed when the data is collected. GA4 reads the source from fields the browser sends, and the BigQuery export stores what was collected, including the campaign source in `collected_traffic_source.manual_source`. [10] No later join or report setting can restore a referrer the browser never sent. Any recovery from here is an estimate.

## Do ChatGPT shopping and ad links still carry tracking parameters?

ChatGPT's shopping links and its sponsored ad links still carried UTM parameters when I inspected them on October 15, 2026, in the same session where the citation links carried none. [1] The commercial surfaces stayed measurable. The citations a brand earns by being the source an answer relies on are the ones losing their tag.

For ads, the reason is documented. OpenAI's ads documentation tells advertisers to add a `query_string_template` at the campaign, ad-group or ad level, and its own examples begin `utm_source=openai`. [11] The advertiser writes that tag into its own landing URL, so it survives a `noreferrer` link. I have not found an OpenAI document on tracking for organic shopping links, so I cannot say who adds the parameters I saw there.

The asymmetry has a cost for anyone reading a dashboard. Paid and shopping clicks keep a clean source, while earned citations slide into Direct. A team comparing the two after October 8 is likely to see ads holding steady and organic AI traffic falling, and part of that gap is a tagging artifact. Budget moved on that comparison would follow the measurement and miss the demand.

A tag you write into a URL you control does not depend on the platform's goodwill. A tag the platform adds for you can disappear in a release note that never mentions it.

## Is ChatGPT surfacing fewer citations than it did earlier in 2026?

ChatGPT has surfaced fewer citations at several points in 2026, though the trend has swung both ways instead of falling in a straight line. Two tracking datasets measured fewer sources per answer after the model changes of March and May, and a third saw citations drop in March and April, then rebound in May. The October loss of UTM tags adds to those drops, so fewer links and fewer tagged links compound.

Each dataset measured something different, in a different window.

| Who measured | Window | What moved |
|---|---|---|
| Resoneo, using Meteoria data (400 prompts a day, 27,000 comparable responses) | 14 weeks around the early-March switch to GPT-5.3 Instant | Unique domains per response fell from 19 to 15 [12] |
| SISTRIX (about 800,000 German-language responses compared) | 4 days on each side of the May 22 to 23 change tied to GPT-5.5 | Cited sources per response fell from 30.9 to 28.4 [13] |
| seoClarity (five markets) | February to May 2026, with drops dated March 8 and April 19 | US share of answers with no citation rose from 28% to 48% in March, then citations rebounded in May [14] |

Each of these vendors sells the tracking product behind its numbers, and the three measure different things in different markets. seoClarity's own conclusion is that "what first looked like a sustained decline now looks like volatility." [14] My read: cited sites per answer fell after both named model changes, rebounded once, and each release can move the count without notice.

The tag was a thin instrument even before October. Profound's panel study, which joined real AI conversations to the same users' later browsing, found that about 1% of visits after a ChatGPT mention carried a trackable AI referral parameter from January through April, rising to 2.47% in June, a partial month Profound calls directional. [15]

I covered that study in [Why most AI-driven site visits skip the referral tag](/articles/ai-referral-tag-undercount/). It counted visits carrying a parameter, so it leaves out the referrer, which was the other signal GA4 used. October removed both at once on a growing share of links, which may explain why the drops above are larger than a 2.47% tag rate would suggest (suspected, not measured). Where those visits still arrive, the change moved their label and not the demand behind them.

## Is the missing UTM tag a bug or a permanent change?

The missing UTM tag was a bug, by the account of Cedric Huesler of Adobe, who updated his post on October 15, 2026. "Relieved. It was a bug. ChatGPT is tagging its links with utm_source=chatgpt(.)com again," he wrote, as Search Engine Roundtable reported. [3] OpenAI has not commented, and its release notes carried no entry on link tags when I read them on October 16. So the bug label is Huesler's reading, and the return of the tag is the confirmed part.

Before that update the two public explanations differed. Huesler had said on October 14 that sources told him it might be a bug, while the monitoring vendor Profound attributed it to how Intelligent UI renders links. [4]

The markup fits Profound's account. The link I copied carries `data-d-component="link"` and machine-generated class names, which look like a rebuilt link component, and the untagged share rose in step with Intelligent UI. A rebuilt component that dropped a tag by accident is plausible. One that added `noreferrer` by accident is harder to picture. My read, from markup only: the referrer block was written on purpose, and whether anyone weighed the attribution cost is unknown.

The tag's return does not close the referrer question. Huesler's update names the tag only, and I have not re-run my click test since, so whether links still carry `noreferrer` is unconfirmed. Either way, your series carries a gap from October 7 to about October 15, and measurement built to work without the tag lost nothing when the tag came back.

## How should you measure ChatGPT referrals without the UTM tag?

To measure ChatGPT referrals across the untagged week, treat chatgpt.com sessions from October 7 to 15 as a floor and stop reading them as a count. Mark that window as a break in every series, estimate the displaced share from landing pages ChatGPT sent traffic to before the break, tag the links you supply, and track citations separately from clicks. The first three steps use data a site already owns, and the last three add a tag, a form question and a citation measure.

1. Annotate the break. Add an annotation for October 7 to 15, 2026 in GA4 and in every dashboard that shows AI referrals. A week-over-week chart that spans the break compares two different instruments.
2. Build a matched-page baseline. List the landing pages that received ChatGPT-tagged or chatgpt.com-referred sessions in the four weeks before October 7. Those are the pages where hidden ChatGPT traffic will surface.
3. Measure the swap on those pages. For each page, compare the change in ChatGPT sessions with the change in sessions that arrived with no source at all. Where ChatGPT sessions fall and no-source sessions rise by a similar amount, the rise in no-source sessions is your estimate of displaced ChatGPT traffic. Report it in its own row, labeled as an estimate.
4. Tag what you control. Where you supply a URL that ChatGPT shows, such as an ad's query string template or a product feed's link field, write your own campaign tag into it. Then click through once to confirm it survives.
5. Ask at conversion. Add "AI assistant" as an answer to a "How did you hear about us?" field on forms or at checkout. It is self-reported and imperfect, and it does not depend on any link.
6. Track citations on their own. Keep a prompt-level citation measure, how often ChatGPT cites your pages for the questions you care about, from sampled prompts or a monitoring tool. If citations hold while ChatGPT sessions fall, the loss is in the label.

Each method still sees a different slice of ChatGPT's traffic after the break.

| Method | Still saw during October 8 to 15 | Misses |
|---|---|---|
| chatgpt.com referral or `utm_source` sessions | Answers that still tag, shopping and ad clicks, and any tag you add | Untagged citation clicks |
| GA4 AI Assistant channel | Visits whose referrer still names an assistant | Every `noreferrer` click, by its own rule |
| Matched-page estimate | The displaced share on pages ChatGPT already sent traffic to | Pages ChatGPT starts citing after the break |
| Your own tag in a supplied URL | Every click on that link, whatever the referrer | Links ChatGPT chooses on its own |
| Self-reported "how did you hear" | Converting visitors who remember | Visitors who never convert |
| Prompt-level citation tracking | Whether ChatGPT still cites you | Whether anyone clicked |

For a team with the GA4 export in BigQuery, step 3 is one query. The sketch below counts session-opening page views per landing page, once for views carrying a ChatGPT tag or referrer and once for views carrying nothing. It counts a tag and a referrer together because Google says `manual_source` can also hold values parsed from referral parameters. Replace the table name with your own, and adjust the dates. I have not run it on this site, whose GA4 export is not connected.

```sql
-- Session-opening page views per landing page: ChatGPT-sourced versus no source at all.
WITH first_hits AS (
  SELECT
    PARSE_DATE('%Y%m%d', event_date) AS day,
    (SELECT value.string_value FROM UNNEST(event_params) WHERE key = 'page_location') AS page_location,
    (SELECT value.string_value FROM UNNEST(event_params) WHERE key = 'page_referrer') AS page_referrer,
    collected_traffic_source.manual_source AS utm_source,
    collected_traffic_source.gclid AS gclid
  FROM `your-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN '20260909' AND '20261014'
    AND event_name = 'page_view'
    AND (SELECT value.int_value FROM UNNEST(event_params) WHERE key = 'entrances') = 1
)
SELECT
  REGEXP_EXTRACT(page_location, r'^[^?#]+') AS landing_page,
  day >= DATE '2026-10-08' AS after_break,
  COUNTIF(utm_source = 'chatgpt.com' OR page_referrer LIKE '%chatgpt.com%') AS chatgpt_sessions,
  COUNTIF(utm_source IS NULL AND gclid IS NULL AND page_referrer IS NULL) AS no_source
FROM first_hits
GROUP BY landing_page, after_break
ORDER BY landing_page, after_break;
```

The `no_source` column is wider than GA4's Direct, since it ignores the returning-visitor rule. That makes it the better input here, because it counts every session that arrived empty, whatever label GA4 later gave it. Divide each side by the number of days in its window before you compare them.

## What should you tell stakeholders about ChatGPT traffic since October 8?

Tell stakeholders that ChatGPT traffic from October 8 to about October 15, 2026 is undercounted, and that the size of any real decline in that week is still unknown. In that window, the chatgpt.com number in analytics covered answers that still tagged, shopping and ad clicks, and whatever you tagged yourself, while untagged citation clicks landed in Direct or under an older source. Report the tagged count, the matched-page estimate and the citation measure as three separate lines for that week.

Put this sentence in the monthly report. "Many ChatGPT citation clicks were recorded as Direct from October 8 to about October 15, 2026, so AI referral figures for that week are a floor." It keeps a falling line from being read as a lost channel, and it keeps the estimate from being mistaken for a measurement.

## Sources

1. First-hand: ChatGPT citation clicks in desktop Chrome, logged in, checked in the link markup, the landing page's browser console and one GA4 request from the same session; shopping and ad links inspected in the same session (site owner, 2026-10-15)
2. OpenAI Help Center: ChatGPT Release Notes, entry for October 7, 2026 (GPT-6 and Intelligent UI in ChatGPT) — https://help.openai.com/en/articles/6825453-chatgpt-release-notes
3. Search Engine Roundtable: OpenAI Drops UTM Parameters From ChatGPT Results (GPT-6), quoting Cedric Huesler of Adobe (Barry Schwartz, October 14, 2026, updated October 15) — https://www.seroundtable.com/utm-parameters-missing-chatgpt-42299.html
4. The Prompt Insider: where ChatGPT traffic went after the UTM tags disappeared, with Profound's and Cedric Huesler's figures (October 14, 2026) — https://thepromptinsider.com/news/chatgpt-missing-utm-tags-referral-traffic/
5. LLM Pulse: ChatGPT Referral Traffic Fell by Almost Half on October 9, 2026 (Esteve Castells, October 13, 2026) — https://llmpulse.ai/blog/chatgpt-referral-traffic-fell/
6. MDN Web Docs: rel=noreferrer — https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Attributes/rel/noreferrer
7. Google Analytics Help: Default channel group — https://support.google.com/analytics/answer/9756891?hl=en
8. Google Analytics Help: how sessions are attributed when a session starts from direct traffic — https://support.google.com/analytics/answer/11080067?hl=en
9. Google Analytics Help: About Analytics sessions (session attribution and its 90-day default lookback) — https://support.google.com/analytics/answer/9191807?hl=en
10. Google Analytics Help: BigQuery Export schema — https://support.google.com/analytics/answer/7029846?hl=en
11. OpenAI Ads documentation: campaign management, landing page query string templates — https://developers.openai.com/ads/campaign-management
12. Search Engine Journal: ChatGPT Search Is Citing Fewer Sites, Data Shows (Matt G. Southern, on Resoneo's analysis of Meteoria data) — https://www.searchenginejournal.com/chatgpt-search-is-citing-fewer-sites-data-shows/571219/
13. Search Engine Journal: GPT-5.5 Update Changes How ChatGPT Cites Sources (SISTRIX data, German-language responses) — https://www.searchenginejournal.com/chatgpt-citations-changed-after-gpt-5-5-sistrix-data-shows/577694/
14. seoClarity: ChatGPT citation decline analysis, February to May 2026 — https://www.seoclarity.net/chatgpt-citation-decline-analysis
15. Profound: The AI mention effect (Nikolas Laskaris, July 1, 2026) — https://www.tryprofound.com/blog/the-ai-mention-effect