Did your ChatGPT traffic really fall? How to audit GA4 after the October 2026 UTM bug

The answer in brief

Probably not. On the sites studied, most of the visits kept coming and GA4 stopped recognising many of them. Check your own landing pages before treating the fall as lost traffic.

• From 8 October 2026, many links in ChatGPT answers reached websites without the utm_source=chatgpt.com tag and without a referrer, and many of those visits appeared in GA4 as Direct.
• Adobe’s Cedric Huesler says OpenAI told him it was a bug.
• The loss came in two steps: a moderate dip from 8 or 9 October, then a deep trough from the evening of 14 October to early on 16 October (UTC).
• Tags came back in stages on 15 and 16 October. Glenn Gabe, who had disputed the first reports of a fix, said at 12:35 UTC on 16 October that he was finally seeing them on every link he checked.
• The fix labels new visits only. Nothing in GA4 can restore the source of visits it has already recorded as Direct, so your October ChatGPT figures keep a gap from 8 to 16 October.
• To confirm the pattern on your own property, compare ChatGPT, Direct and total sessions on the same landing pages, annotate 8 to 16 October, and report the window with a note.

What happened to ChatGPT links in October?

OpenAI began rolling out GPT-6 as the default model on paid ChatGPT plans on 7 October, and on Free and Go from 8 October. The release introduced Intelligent UI, a format that mixes text with charts, comparison tables and small working tools. Huesler found that links in these answers opened with window.open(url, "_blank", "noopener,noreferrer") and dropped the UTM tag. With neither signal, GA4 cannot tell the visit came from ChatGPT. It usually counts it as Direct, alongside typed URLs and bookmarks, unless the visitor has an earlier non-direct source within the lookback window.

The tag and the referrer are separate signals, and either one can identify a visit. During the outage most affected clicks lost both.

Not every link broke:

• Inline answer links lost the tag, Ethan Smith of Graphite found.
• Sources panel and shopping links kept it in Smith’s checks.
• Product and information panels still had tagged links, Profound found.
• ChatGPT Ads campaigns carrying your own UTM parameters should be unaffected.
• OpenAI API answers carried utm_source=openai before, during and after the outage, AEO Sherpa reports from the shop analyst’s checks, so a filter that only looks for chatgpt.com misses that traffic in normal times as well.

The loss came in two phases. Profound grouped the ChatGPT answers it collected into 30-minute intervals and recorded the share of answers with cited links that carried at least one tagged link. Times are UTC.

Compare your own daily trend with this two-step pattern, allowing for your traffic volume, link mix and reporting time zone. These percentages describe answers rather than individual links, and only the answers Profound collected. Its published series ends at 16:30 UTC on 16 October. Profound itself warns against using them to correct your traffic.

How big was the drop for other sites?

Smith reported that ChatGPT sessions in GA4 fell by between 22% and 57% on the Graphite-tracked sites with meaningful AI Assistant traffic, and estimated a fall of about 45% across them. LLM Pulse found roughly 46% fewer ChatGPT sessions on 9 to 11 October than on the same days in the previous two weeks, while total traffic on those sites barely moved.

Huesler’s analysis points to where the visits went. He took about 500 pages where ChatGPT supplied at least 5% of visits and compared 9 to 13 October with the week before:

• ChatGPT visits fell by about 31,000 a day.
• Direct visits on the same pages rose by about 26,000 a day.
• Search traffic fell 7%, in line with search elsewhere.
• Direct rose on two in three pages that lost ChatGPT traffic, against about half the week before.
• On mobile, the share of Direct visitors seeing the consent banner, his marker of a first visit, rose from 36% to 44%, moving towards the 55% he sees among ChatGPT visitors.

Server logs from one large online shop, posted anonymously on Reddit and reported by AEO Sherpa, point the same way. Against a 1 to 7 October baseline:

ChatGPT-User is the agent that loads a page when someone asks ChatGPT about it. Tagged Perplexity clicks did not drop. Fetches fell far less than tagged clicks. A fetch shows ChatGPT reading a page for someone, so it hints that interest held up, though it does not count human visits.

Three caveats apply. Huesler described his figures as early and sampled, and they predate the deeper phase from 14 October. LLM Pulse found the rise in Direct hard to spot at site level, because Direct is so much larger than ChatGPT, but visible on the pages that depended most on ChatGPT. And Graphite, LLM Pulse, Profound and Peec AI all sell SEO or AI search services or tools, Huesler works at Adobe, and the shop’s logs come from a single unnamed retailer.

When was the bug fixed?

Tagging had largely recovered by 16 October 2026. When it came back depends on whose data you read, because recovery showed up in different datasets and spot checks at different times.

On 15 October Huesler wrote on LinkedIn that OpenAI had confirmed a known bug to him the day before. He said a fix began rolling out early that morning and that by 9am Pacific time (16:00 UTC), tagged ChatGPT clicks were back to roughly 95% to 100% of the same hour a week earlier, across thousands of domains. Most coverage repeated that date.

Glenn Gabe disagreed, in a reply under Huesler’s post and on X. At 19:35 UTC that evening he wrote that he had tested paid and free accounts and was seeing no UTM parameters at all. Profound’s data supports him: at 22:00 UTC on 15 October only about 11% of cited answers carried a tag. The shop’s logs show tagged clicks starting to climb from about 14:30 UTC, but still at 41% and 35% of the previous week in the 15:00 and 16:00 hours.

The large recovery came overnight. Profound’s tagged share passed 80% by 03:00 UTC on 16 October and reached about 99% by 06:30. At 12:35 UTC Gabe posted: “I’m now finally seeing utm parameters for ALL links in ChatGPT”, covering links in the answer, the main sources and the “More” sources.

Not every link came back at once. AEO Sherpa reports that the shop analyst, asking mostly the same questions in ChatGPT from Germany, found the tag on 67 of 67 links on 3 October, none of 82 on the afternoon of 15 October and 71 of 102 on 16 October. Profound counts answers with at least one tagged link, so its figure can reach 99% while some links in those answers still lack the tag.

In our own spot checks on 19 October 2026, every ChatGPT link we clicked carried the tag. The visits arrived without a referrer, and GA4 recorded them.

We found no public statement from OpenAI as of 19 October 2026, and its ChatGPT release notes for 1 to 16 October do not mention link tracking. What is known of OpenAI’s view comes through Huesler, who says OpenAI confirmed the bug to him by 11am Pacific time on 14 October.

The referrer has not returned in the checks so far. Huesler said on 15 October that most ChatGPT clicks still arrive without one. Every link in our own test on 19 October also opened without a referrer, so for those clicks the tag is the only signal GA4 gets. Smith found that 95% of mobile ChatGPT visits to Graphite’s sites relied on the tag alone. If it disappears again, the same thing will happen, and mobile visits, which relied on the tag most in Smith’s sample, would be the most exposed.

How do I check whether my drop is real?

Look for three things that agree: ChatGPT sessions falling in two steps, Direct rising on the same landing pages, and total sessions and conversions on those pages holding steady. Then check that ChatGPT recovered from 16 October.

Set up your comparison periods

Compare periods of equal length that start on the same weekday. Both 1 to 7 October and 8 to 14 October run Thursday to Wednesday, so they share the same weekday mix. A nine-day incident window against a seven-day baseline does not.

If your volumes are small, use the four weeks from 10 September to 7 October as the baseline to see how much ChatGPT normally varies. Work in your property’s reporting time zone and convert the UTC times above before reading hourly data. In the UK during British Summer Time, 06:30 UTC on 16 October was 07:30. Give the latest days 48 hours to settle before you judge them.

1. Find every ChatGPT source value

Open Reports, then Acquisition, then Traffic acquisition, and set the primary dimension to Session source / medium. Search for chatgpt and openai, and note every value you find, such as chatgpt.com / ai-assistant, chatgpt.com / referral, chatgpt.com / (not set) or openai. Use that same list throughout the audit. Check the medium and campaign on any openai row before you count it: OpenAI’s ads documentation uses utm_source=openai in its example tracking templates, so the value can mark paid clicks as well as API answers.

Do not rely on the AI Assistant channel row alone. GA4 added that channel on 13 May 2026; Google’s channel definitions say it sets the medium to ai-assistant when the referrer matches its list of assistants. Look at Session source / medium to separate ChatGPT from the other assistants in that channel. A visit with neither a tag nor a referrer has nothing to match, so GA4 cannot identify it as ChatGPT, and the source rows are where you see the difference.

If the bug explains your drop, the daily line should step down on 8 or 9 October, fall further on 14 and 15 October, and climb back from 16 October. The recovery is the most useful cross-check. If ChatGPT sessions from 17 October look normal for those weekdays, and the landing-page check below agrees, the pattern points to lost attribution. If they stay well below baseline, recovery is not established: check for remaining tagging problems, data still processing, and how often ChatGPT cites you.

2. Read Direct beside ChatGPT

In the same report, clear the search from check 1 so Direct is visible again. Then:

1. Tick your ChatGPT rows and (direct) / (none), and click Plot rows to draw them on one chart.
2. Export the daily figures.
3. Add the two together and compare each day with the same weekday before 8 October.

A combined line that holds steady while ChatGPT falls and Direct rises points to a transfer.

Expect only a partial offset. In Huesler’s data, ChatGPT lost about 31,000 visits a day while Direct on the same pages gained about 26,000. Other traffic moves too, and GA4 adds a further reason: session reports credit a direct visit to the user’s last non-direct source within the lookback window (90 days by default), so some returning ChatGPT visitors stayed credited to ChatGPT throughout. His figures do not show how much of the gap this explains.

3. Compare the landing pages ChatGPT used to send visitors to

This is the most convincing evidence you can produce. In Explore, build a free-form exploration with these settings:

Use session segments throughout; a user segment pulls in each person’s other visits and blurs the result. GA4 leaves UTM parameters out of Landing page + query string, so use Page location when you want to see whether the tag arrived.

1. Over the baseline, list the 20 to 50 landing pages with the most ChatGPT sessions, and save the list.
2. For those pages, compare ChatGPT, Direct and all-channel sessions day by day through the incident.
3. Pick a control group of similar pages that never received much ChatGPT traffic. Their Direct line should show no matching rise.
4. Check your home page separately.

People rarely type a long article URL from memory, so a rise in Direct on deep pages that ChatGPT used to cite, with no matching rise on the control pages, points strongly to lost attribution. Untagged email or messaging links can produce the same pattern, which is why check 5 matters. The home page matters because Tomek Rudzki of Peec AI found ChatGPT put a bare website URL in 42% of responses after GPT-6, up from about 8 to 10%, and 54% of those pointed to home pages. He measured links shown, not clicks.

4. Look at who the extra Direct visitors are

Add Device category and New users to the exploration. Extra Direct traffic on former ChatGPT pages that leans towards mobile and new users fits Huesler’s finding and Smith’s mobile figure, though GA4’s New users metric is a different measure from Huesler’s consent-banner proxy. Compare its engagement rate with your baseline ChatGPT sessions. Similar behaviour supports the case, though it cannot prove that any single Direct session came from ChatGPT.

5. Rule out other causes

Before you put the whole movement down to the bug, check the rest of your traffic:

• Organic Search should look normal for the time of year.
• ChatGPT Ads: if you run them with your own UTM parameters, compare clicks in OpenAI’s Ads Manager with GA4 sessions carrying those parameters. The ratio should hold roughly steady after 7 October.
• Other assistants: tagged Perplexity, Gemini or Claude traffic should continue as before, allowing for changes of their own. LLM Pulse found Gemini referrals fell between 28 September and 2 October for an unrelated reason.
• Your own release log: look for anything that could inflate Direct in the same week, such as a consent banner change, a tag deployment, redirects or an email send.

6. Repeat the check for conversions and revenue

Run checks 1 and 2 again with key events, purchases or revenue as the metric. If ChatGPT conversions fell and Direct conversions rose by a similar amount on the same days, your session-scoped reports probably gave Direct some ChatGPT-assisted sales. GA4’s Attribution reports handle this differently: their models give Direct credit only when the whole path was Direct, so many of those sales will sit with an earlier channel instead. In Shopify, check the sales reports; its Acquisition reports cover visits only.

What your results mean

How do I annotate the window in GA4?

Annotate 8 to 16 October before anyone pastes the drop into a monthly report. If you already marked 8 to 15 October, extend it: on Profound’s data and Gabe’s tests, tagging was not back to normal until 16 October.

You need the Analyst role or above on the property. To add an annotation:

1. In Reports, open any report with a line graph.
2. Right-click a data point and choose Add annotation.
3. Enter a title of up to 60 characters, a description of up to 150, a date or date range and a colour.
4. Click Create annotation.

It appears on every report and report card with a line graph. Admin, then Data display, then Annotations lists them all for editing or export. Each property holds up to 1,000, and you can create them through the Admin API if you manage many properties.

Three annotations cover it:

Annotations do not travel to other tools on their own, although you can export them from Admin. Add the same note to any dashboard, client deck or forecast that uses these figures.

Can I recover the missing attribution?

No, not inside GA4. A visit that arrived with no tag and no referrer carried nothing that identifies ChatGPT, and the tag’s return only labels new visits. Custom channel groups can regroup traffic GA4 already identified; they cannot find a source that was never collected.

Your own server logs or a BigQuery export may hold a little more, such as the full landing URL or a referrer GA4 ignored, so they are worth a look. Do not count crawler requests as visits, though. A fetch by ChatGPT-User or OAI-SearchBot shows that ChatGPT read the page, not that a person clicked through. The shop that shared its logs tried to match untagged visits to a ChatGPT-User fetch of the same page in the previous ten minutes, and in its baseline week only about 2.5% of tagged product clicks had one. That was too few to size the hidden traffic.

Keep the figures GA4 reported, and show any estimate beside them, labelled as an estimate.

How should I report October?

Never set October’s ChatGPT total against September’s without a note. Pick one of three approaches and say which you used.

1. Exclude the window. Compare the rest of October with the same weekdays in September and leave 8 to 16 October out.
2. Combine the two lines. Report ChatGPT and Direct together for the window, labelled as such.
3. Estimate the missing sessions. Take Direct sessions on your former ChatGPT landing pages, subtract what you would expect on those weekdays from the baseline, and report the excess as an unexplained Direct increase. It is a rough guide to misfiled ChatGPT visits and can run too high or too low, because other Direct traffic moves too.

This hypothetical example shows the pattern to look for on a fixed set of landing pages over two seven-day periods:

Here the combined line barely moves, which fits a relabelling. It still does not prove that all 470 extra Direct sessions came from ChatGPT, so the estimate stays labelled as one.

Watch the rebound as well. Measured against the trough of 14 and 15 October, the recovery will look like growth. Compare late October with the same weekdays before the disruption.

A sentence you can adapt for a client report: “Identified ChatGPT sessions fell by X% between 8 and 16 October, during the ChatGPT tagging disruption. Direct traffic on the same pages rose by Y% while total sessions changed by Z%, and ChatGPT sessions returned to normal from 16 October. This points to lost attribution, though the exact amount cannot be measured. We have excluded this period from month-on-month comparisons.”

How do I test my own tracking and catch the next outage?

Test the whole route from a ChatGPT link to a GA4 session yourself. It is the same kind of manual check Gabe ran on 15 and 16 October.

1. Start clean. Close any private windows you already have open, then open a new one, so you start with no existing session on your site. Google says a session’s attribution comes from its session_start event, so a click made inside a session that has already started will not show up as a new ChatGPT session. Start a fresh private window for each link you test.
2. Use different accounts and devices. Ask ChatGPT a few questions that usually produce links to your site, on a free and a paid account and on mobile as well as desktop.
3. Vary the link position. Click links inline in the answer, in the Sources panel and in any product or information panel. Note the account, device, time and link position.
4. Check the URL. It should contain utm_source=chatgpt.com. If it does but the parameter is gone once your page loads, check your redirects and scripts, and whether GA4 collected the tag before it was removed.
5. Confirm the visit in GA4. Check that consent settings, blocking extensions and internal-traffic filters will not hide your test visit, then look for it in GA4’s Realtime report, then confirm it in Traffic acquisition once processing has caught up, which can take 24 to 48 hours.

To spot the next outage quickly, set an alert on daily sessions with a ChatGPT source, using a GA4 custom insight with email alerts (Editor role or above) or whichever reporting tool you already schedule. Flag any day that falls far below the same weekday in previous weeks. And make sure any custom channel group or dashboard filter for AI traffic includes utm_source=openai as well as chatgpt.com, kept as its own row because it marks answers built through the API rather than visits from the ChatGPT app.

Sources

• Barry Schwartz, OpenAI Drops UTM Parameters From ChatGPT Results (GPT-6), Search Engine Roundtable, 14 October 2026, updated 15 October
• Matt G. Southern, ChatGPT Traffic Drop Appears To Be A Tracking Problem, Search Engine Journal, 15 October 2026
• Matt G. Southern, Rank Tracker Data, ChatGPT Tags & AI Search Links: SEO Pulse, Search Engine Journal, 17 October 2026
• Alex Brufsky, ChatGPT’s UTM Fix Looks Uneven in Logs and Spot Checks, AEO Sherpa, 17 October 2026
• Josh Blyskal, ChatGPT’s missing UTM tags: what happened from October 8 to October 16, Profound
• Cedric Huesler, original analysis and follow-up, 15 October 2026, LinkedIn
• Glenn Gabe, reply on 15 October and post on 16 October, X
• Ethan Smith, LinkedIn post
• Tomek Rudzki, LinkedIn post
• LLM Pulse analysis
• OpenAI, ChatGPT release notes
• Google, About annotations, Analytics Help
• Google, Analytics Help on attribution models, campaign processing and data freshness
• Shopify, Acquisition reports, Shopify Help Center