Why Did My ChatGPT Referral Traffic Drop in October 2026?

October's ChatGPT referral numbers are a floor. Sistrix's Oct. 30, 2026, analysis found only 1% to 24% of ChatGPT's source links tagged each day from Oct. 9 to 15, and links in its new answer format pass no referrer, so those visits show up as Direct. If your report shows ChatGPT traffic falling that week, flag the dates before anyone explains the drop.

Last verified: November 1, 2026.

ChatGPT's US local answers changed the same week. In the Cheers Market Baseline, our own research panel, most carried a Google Maps link for 15 days and almost all carried a phone number, and neither path shows up as ChatGPT in your analytics. Those answer paths help explain why website referrals alone understate the ways a customer can reach you. Combine the answer research with your branch's analytics, profile activity and call records.

If you haven't built an AI referral segment yet, start with our tracking guide.

Five checks before you send the October report

Run checks 2 and 3 location by location, because a company-wide total can bury one branch's Direct bump under another branch's slow month.

• Check 1: Flag Oct. 8 to 17 on every ChatGPT chart, then confirm the tag is back for you. Profound saw tagged answers start to fall on Oct. 8 and Sistrix found tags back on 99% to 100% of links from Oct. 17, so the window covers both. To confirm the fix, ask ChatGPT about your business by name, click its link to your site, and check that a utm_source tag set to chatgpt.com survives your redirects and shows in GA4 as session source chatgpt.com. Build the AI line on session source, too, since GA4's AI Assistant channel can miss a tagged ChatGPT click with no referrer.
• Check 2: Look in Direct for the missing visits. Direct landings on service, city and location pages that rise from Oct. 9 and fall back around Oct. 17 are a rough sign of untagged ChatGPT visits, though an email send, a postcard drop or a site change can make the same bump. For visits from the Gemini iPhone app, ask your developer or host to count server-log requests with "GeminiiOS" in the user agent, the one trace Natzir found (more on Gemini below).
• Check 3: Read Business Profile Performance by location for Oct. 9 to 29. Those are the weeks our panel's ChatGPT answers often linked Google Maps, so compare views, calls, website clicks and directions requests with Sept. 11 to Oct. 1 (the same 21 days) and the same weeks last year; our guide to Business Profile metrics explains which ones to compare across branches. October demand moves with the season (furnace season for HVAC), and Google's help page doesn't say whether a visit from another app's Maps link counts, so report any rise as a possible ChatGPT effect, not a measured one.
• Check 4: Count AI calls in two separate tallies. One is for customers who say an AI assistant sent them, so add "an AI assistant (ChatGPT, Gemini, other)" to the "how did you hear about us" options your call team logs. The other is for AI agents that announce themselves. Google says Gemini opens every call by disclosing that it's an AI assistant from Google calling on a recorded line, so search transcribed recordings for "AI assistant" and "recorded line" instead of an exact script.
• Check 5: Write the unknowns into the report as "not measured," never as zero. Add a note like this on the chart: "Oct. 8 to 17: for part of this window most ChatGPT links arrived with no tracking tag or referrer (Sistrix, Profound), so ChatGPT visits are undercounted and some sit in Direct." And don't scale October up with a correction factor; Profound says its own percentages "do not provide a direct multiplier for correcting your traffic or revenue reports."

What broke in October, and when

None of this is evidence that fewer people asked an AI for a plumber in October, only that the usual ways of counting them broke on dates you can name. Keep the flag on every Q4, full-year and year-over-year report that includes October, or the same week turns up in January's deck as a story about falling demand.

ChatGPT's untagged week

Up to Oct. 8, Sistrix found the tag on every source link in its sample; after the low week it was back to 42% on Oct. 16 and near-full from Oct. 17. Profound counted answers rather than links (see Method), but both trackers date the start to Oct. 8 to 9 and the end to Oct. 17.

Sistrix traces the missing referrer to the new format's noopener noreferrer links, and its advice is to flag Oct. 9 to 16 "before anyone reads a trend into them."

On pages ChatGPT already sent a lot of traffic, Cedric Huesler, who works at Adobe, counted about 31,000 fewer ChatGPT visits a day from Oct. 9 to 13 than the week before and about 26,000 more Direct visits (via Search Engine Roundtable), so these figures describe those selected pages, not every site's traffic.

Huesler also posted on Oct. 15 that OpenAI had told him on Oct. 14 the missing tag was a known bug, with a fix rolling out from about 2 a.m. PT on Oct. 15 (Relevant Audience reported it the next day). It's secondhand, and OpenAI hasn't said so publicly. He added that most ChatGPT clicks still had no referrer, and nothing published since says whether it came back.

Gemini: September was the unusual month

Natzir's chart, reported by Search Engine Roundtable on Oct. 30, shows Google adding utm_source=gemini to cited links in early September and stopping around Sept. 28. The Gemini iPhone app opens links in its own window without a referrer, so outside those weeks its visits land in Direct, while desktop Gemini still sends one and its gemini.google.com referrals stayed flat. A Gemini line that rose in September and fell back around Sept. 28 is showing you the unusual weeks, not a loss, so it needs no October flag. It's one analyst's chart; SER reports no Google statement.

For 15 days, most ChatGPT local answers linked Google Maps

From Oct. 9 to 23, 80.7% of US ChatGPT answers in the Cheers Market Baseline carried a Google Maps link, against 0.2% from Oct. 2 to 8. Most were directions links, which open a Google Maps route (77.9% of answers), while 32.9% linked a place page, the business's own Maps listing; an answer could carry both. They came on top of website links, not instead of them (see Method).

The links arrived with the per-business layout Sistrix saw ChatGPT switch to overnight on Oct. 8 to 9, so read them as part of a layout, not a choice of source. The panel saw a Maps link in 0.0% of answers on Oct. 8 and 80.4% on Oct. 9.

After Oct. 23 the links changed in two steps: from Oct. 24 to 29, directions links showed up in 0.1% of answers and place-page links in 37.5%, and on Oct. 30 and 31 any Maps link showed up in 2.0%. They may have changed again since, so before a November deck says anything about Maps, ask ChatGPT one of your own trade-and-city questions and see what it links.

Follow the customer journey through the listing as well as the website. A person can open a ChatGPT Maps link, call from the listing or continue to your site. A website session that starts on Google Maps will not automatically carry the original ChatGPT source. Check what your profile's website link carries, since any UTM tags on it decide which channel that visit lands in.

Callers who dial from ChatGPT never reach your analytics

In the Cheers Market Baseline, 12.6% of US ChatGPT local answers carried at least one phone number from Oct. 2 to 8 and 92.5% from Oct. 9 to 31. Across all 100 metros, answers after Oct. 9 printed about 6.4 times as many phone numbers each. A customer who dials from the answer creates no website session, and the call rings whichever number ChatGPT printed.

That number isn't always one your tracking owns. Of the US numbers ChatGPT printed in October for businesses whose sites run a call-tracking script, 41.0% weren't on the business's own site in the pages we checked, against 12.1% for sites without one. That doesn't make them wrong, but your call tracking may not see those calls, which is why check 4 asks the caller. If you run call tracking, the call-tracking standard for AI search matters more after October than before.

Combine reports for a fuller view of AI discovery

The tag gap closed by Oct. 17; the rest were still open on Nov. 1.

• Tagged or referred AI clicks: GA4 session source and the AI Assistant channel count them, but not untagged clicks with no referrer (ChatGPT Oct. 9 to 16, the Gemini iPhone app).
• AI Overviews and AI Mode: Search Console's Generative AI report shows impressions, and the help page describes no clicks.
• Copilot and Bing's AI summaries: Bing Webmaster Tools AI Performance, a public preview announced Feb. 10, 2026, reports citations; Microsoft's announcement doesn't mention clicks.
• Google Business Profile: Performance doesn't say where the visitor came from, and Search Engine Roundtable's write-up of reports Google tested on Oct. 23 mentions no AI metric.
• Calls to a number ChatGPT printed: "how did you hear about us" and call tracking catch some, but not calls to a printed number your tracking doesn't own.
• AI agents that call: transcribed recordings catch the ones that say what they are, never the ones that don't.

Calls from AI agents are the newest gap. TechCrunch reported Google's test of Gemini Call for Me on Sept. 24, 2026, and each call comes from the customer's own number, so caller ID shows the customer. It works differently from Google's own price-check calls to businesses, which our earlier piece on Google's AI calls covers.

NextPhone, which sells AI receptionists, counted about 4 disclosed AI callers per 10,000 calls in September 2026. That isn't a home-service rate (the Method has NextPhone's line mix and trend), and NextPhone calls every figure a floor because AI callers that don't say what they are go uncounted. Start a branch-level count using the call and intake checks above.

Method

The Cheers Market Baseline is our own research panel: every week it asks ChatGPT, Google AI Mode and Gemini the same 32 home-service questions in 100 metro areas, 60 in the US and 40 in Canada. Perplexity isn't in this edition. This article uses US ChatGPT answers from Oct. 2 to 31, 2026: 8,217 answers, about 240 to 300 a day, split into 1,910 for Oct. 2 to 8, 4,087 for Oct. 9 to 23, 1,680 for Oct. 24 to 29 and 540 for Oct. 30 to 31. The phone-number share for Oct. 9 to 31 is 5,835 of 6,307 answers.

Two figures use all 100 metros instead. One is phone numbers per answer, counting repeats: 0.71 for Oct. 2 to 8 against 4.53 for Oct. 9 to 31. The other is the share of answers citing a shown business's own website, 49.0% against 79.8%, which is why we say the Maps links came on top of website links. They rest on 3,185 answers from Oct. 2 to 8 and 10,507 from Oct. 9 to 31.

Each day of the panel runs a different set of our questions, so single days and short stretches are what the panel saw on those days, never trend points. The Maps chart starts on Oct. 3 so its periods break on Oct. 9, Oct. 24 and Oct. 30, the days the links changed; each covers three days and 777 to 899 answers, except the last, Oct. 30 to 31, which covers two days and 540 answers.

A Google Maps link is any cited link to Google Maps. We split directions links, which open a Maps route with the destination filled in, from place-page links, which open a business's Maps listing, and one answer can hold both; from Oct. 24 to 29, 38.3% of answers had a Maps link of either kind. A phone number is a North American number pattern in the answer text.

The call-tracking figures come from our Nov. 1 check of 900 phone numbers ChatGPT printed from Oct. 2 to 31 (US and Canada) against each business's own website. This article quotes the US cut: 32 of 78 numbers on sites running a call-tracking script, and 34 of 280 on sites without one. A mismatch means the number wasn't on the business's own website in the pages we checked, never that it's wrong.

The Baseline records answers, not clicks. It sees what answers say and link to, but it doesn't record whether a link carried a tracking tag and never sees a reader's click, so the tag-gap figures here come from Sistrix and Profound. Neither tracker's sample is local or US-only, and neither states a sample size; Sistrix samples logged-out queries in six countries, the US among them.

Profound checked every 30 minutes and counted, of the answers with cited links, the share with at least one tagged link. That share was about 97% at 18:30 UTC on Oct. 8, about 75% across Oct. 9 and about 67% until 19:00 UTC on Oct. 14, then fell to 5% to 12% for the next 27 hours or so and was back to about 99% by 06:30 UTC on Oct. 16.

Huesler's count covers about 500 pages where ChatGPT drove at least 5% of visits. NextPhone's lines are mainly small-business lines (law firms, home services and local services) plus personal phones. On phones it tracked all year, its rate went from 1.2 (October to December 2025) to 1.8 (July to September 2026); business lines joined from November 2025, and an August 2026 transcription upgrade may have lifted the August and September figures.

Sources

• What changed in ChatGPT in October: GPT-6, UTM tracking and sources, Sistrix, Oct. 30, 2026. Daily tag shares, the missing referrer and the dates to flag.
• ChatGPT's missing UTM tags: what happened from October 8 to October 16, Profound, Oct. 14, 2026. The Oct. 8 start, the 30-minute series and the no-multiplier warning.
• Default channel group, Google Analytics Help, undated. The AI Assistant and Direct channel rules.
• Understand your Business Profile performance & insights, Google Business Profile Help, undated. What views, calls, website clicks and directions count.
• Ask Gemini to handle your everyday phone calls, Gemini Help, undated. Call for Me's disclosure and the caller's own number.
• OpenAI Drops UTM Parameters From ChatGPT Results (GPT-6), Search Engine Roundtable, Oct. 14, 2026. Huesler's shift from ChatGPT visits into Direct.
• ChatGPT UTM tag loss was a bug, Adobe's Huesler says; fix rolling out, Relevant Audience, Oct. 16, 2026. Huesler's secondhand bug account and fix timing.
• Google Gemini Stopped Attaching UTM Parameters To Cited URLs, Search Engine Roundtable, Oct. 30, 2026. Natzir's Gemini tag dates and the iPhone app's missing referrer.
• Generative AI performance report (Search), Search Console Help, undated. AI Overviews and AI Mode impressions.
• Introducing AI Performance in Bing Webmaster Tools Public Preview, Bing Webmaster Blog, Feb. 10, 2026. Copilot and Bing AI summary citations.
• Google Business Profiles Tests New Advanced Performance Reports, Search Engine Roundtable, Oct. 23, 2026. No AI metric in the tested reports.
• Google tests letting Gemini make phone calls, initially for US Pixel owners, TechCrunch, Sept. 24, 2026. Google's Call for Me test.
• The Comprehensive 2026 AI Receptionist Call Report, NextPhone, Oct. 27, 2026. Disclosed AI callers per 10,000 calls.

Sources read November 1, 2026.

Dylan Allen-Arnegård is the CEO & Co-Founder of Cheers, the done-for-you local visibility partner for established home-service companies and PE-backed portfolios.