How to Fix ChatGPT 404 Referrals in Your Analytics: Broken Links, Redirects, Branded URLs
How to find ChatGPT 404 referrals in GA4, which invented URLs to redirect, which to build and which to leave, with the measured 404 rates behind each decision.
Quick answer
To fix ChatGPT 404 referrals under this method, Canlah AI sorts every invented URL into one of three fixes rather than applying a blanket redirect: a 301 when a close live page exists or external links point at the URL, a real page when one URL draws three or more sessions and a clean 404 with a site-search box for the rest. The sorting matters because Ahrefs checked 16 million cited URLs and found that 1.01% of clicked ChatGPT links and 2.38% of cited ChatGPT links return a 404, against Google baselines of 0.15% and 0.84%.
Find the URLs first: filter Google Analytics 4 on session source containing chatgpt.com, cross it with the page title of your 404 template and reconcile the export against your server logs. Canlah AI runs this as AI-referral attribution inside its SEO + GEO retainer and publishes the manual version here so any team can work through it in one afternoon.
Treat the result as directional. One month of referral logs cannot prove that an invented URL will recur, that a redirect recovered the visit or that the pattern survives the next model update. The figures below were reviewed on September 23, 2026; freeze your URL list, record the date, engine and market, then repeat the count before paying for a rebuild.
Disclosure: Canlah AI publishes this guide and sells the AI-referral measurement described below. Details for Google Analytics 4, Google Search Console, Ahrefs, Semrush, SE Ranking, Screaming Frog and Cloudflare come from their public pages and published studies, reviewed on September 23, 2026, and were not tested through paid accounts. The Ahrefs, Semrush and SE Ranking figures are their measurements, not Canlah AI measurements.
What is a ChatGPT 404 referral?
A ChatGPT 404 referral is a visit that Google Analytics 4 attributes to chatgpt.com and that lands on a page returning HTTP 404, which Ahrefs measured at 1.01% of clicked ChatGPT links in a study published September 2, 2025. Canlah AI codes each one on six properties, because each property points at a different fix.
- Referrer: chatgpt.com, or Perplexity, Claude, Copilot or Gemini, which have their own rates.
- Status: 404, 410, a redirect chain or a 200 response on an error page that analytics cannot tell apart from real content.
- Origin: a path that was live and then deleted, or a path that was never published.
- Repetition: how many sessions hit the same invented URL, and across how many distinct days.
- Links: whether external pages already link to the dead path, so a redirect recovers link equity as well as a visitor.
- Pattern: whether the invented URL follows your own URL grammar, such as /pricing, /blog/
or /features/ .
Buyers underrate the last property. An invented URL that matches your branded URL pattern is a statement about what a reader expected to find on your domain, which makes it a content brief rather than an error.
How much ChatGPT referral traffic lands on a 404?
ChatGPT has the highest 404 rate of the six assistants Ahrefs measured, at 1.01% of clicked URLs and 2.38% of cited URLs, against Google baselines of 0.15% and 0.84%. Ahrefs built the clicked column from anonymised Web Analytics sessions and the cited column from its Brand Radar citation database. Across all six assistants the clicked-URL rate averaged 0.43%, which Ahrefs reported as 2.87 times the Google referral baseline.
404 rates by referrer, from Ahrefs’ study of 16 million cited URLs, published September 2, 2025.
| Referrer | Clicked URLs returning 404 | Cited URLs returning 404 | What to do with the number |
|---|---|---|---|
| ChatGPT | 1.01% of 8,332,436 URLs | 2.38% of 1,524,277 crawled URLs | Highest of the set; build the audit on this referrer first |
| Claude | 0.58% of 95,293 URLs | Not in the published sample | Add as a second column once volume justifies it |
| Copilot | 0.34% of 431,319 URLs | 0.54% of 1,120,780 crawled URLs | Lowest cited-URL rate of the set; audit only if Copilot sends real volume |
| Perplexity | 0.31% of 1,133,084 URLs | 0.87% of 2,450,016 crawled URLs | Within 0.03 points of the Google SERP baseline |
| Gemini | 0.21% of 351,242 URLs | 0.86% of 641,603 crawled URLs | Within 0.02 points of the Google SERP baseline |
| Mistral | 0.12% of 6,760 URLs | Not in the published sample | Sample too small to act on |
| Google Search (baseline) | 0.15% of 629,000,000 URLs | 0.84% of top-20 SERP URLs | Read every row above against this one |
Source: Ahrefs study published September 2, 2025 and reviewed September 23, 2026; a page counted as a likely 404 when its title contained “404” or “not found”. “Not in the published sample” means Ahrefs did not report that figure for that referrer in that study. It is not a claim that the rate is zero.
Single-site numbers run far above the population average, which is why your own log matters more than the benchmark. Semrush reports that Dan Hinckley of Go Fish Digital analysed more than 18,000 ChatGPT landing pages and found about 3.35% returned a 404, and that Alex Galinos recorded 57% of one client’s ChatGPT referrals landing on 404s. SE Ranking logged 70 invented URLs over three months on one property, of which 20 had backlinks and a handful drew more than 20 sessions each.
How to find ChatGPT referral traffic and its 404s in GA4
Finding ChatGPT 404 referrals in Google Analytics 4 takes six steps, and the first four put every invented URL into one exportable table in about ten minutes. Steps 5 and 6 reconcile it against server logs and backlink data.
Step 1: Match on the source, not the source and medium together
ChatGPT appends utm_source=chatgpt.com to outbound links but sets no utm_medium and no utm_campaign, so the same traffic splits across chatgpt.com / referral, chatgpt.com / organic and chatgpt.com / (not set) depending on channel rules. Filter on Session source containing chatgpt.com rather than an exact source and medium string, or the audit misses part of the volume.
Step 2: Build the report
Open Reports, then Life cycle, then Engagement, then Pages and screens. Set the primary dimension to Page title and screen class, add Session source / medium as a secondary dimension and search the table for chatgpt.com. Your 404 template appears by its title, commonly “Page Not Found”, beside the referrer that sent it.
Step 3: Save it and add the other assistants
Save the view as a report named ChatGPT referrals and drag it into the Engagement folder in the Library so it does not need rebuilding. To cover the other assistants, use an Exploration with the Session source regular expression that Ahrefs publishes from Thierry Ngutegure at SALT.agency, and update it whenever an assistant changes its referrer string.
Step 4: Export the full landing-page paths
Switch the dimension to Landing page and screen class so you capture the exact path, not only the page title. Export sessions, first date and last date for every path that resolves to your 404 template.
Step 5: Cross-check against server logs
Analytics only sees requests where the 404 template loaded and the tag fired, so a hard 404 served at the edge, a rate-limited request or a blocked script leaves no GA4 row. Pull 404 responses with a chatgpt.com referrer from Cloudflare logs, your CDN or your origin access log for the same window and reconcile the two lists. Canlah AI treats the log file as the denominator and GA4 as the sample, and reports the gap between them rather than picking one number.
Step 6: Check which dead URLs have links
Run the exported paths through Ahrefs Site Explorer or Semrush Backlink Analytics to see which have referring domains, and crawl the live site with Screaming Frog SEO Spider to find the nearest existing page for each. A dead path with links is a redirect; a dead path without links goes to the decision table below.
Redirect, build or leave it: the decision table
Every invented URL in the Step 4 export gets one of three decisions: redirect it when a close live page exists or links point at it, build a page when it draws three or more sessions and leave a clean 404 for the rest. SE Ranking’s published playbook supplies the session thresholds and Semrush’s decision framework adds the backlink condition; the table below merges them into one pass over the export.
Decision rules for invented URLs in the GA4 and log export, compiled from SE Ranking and Semrush guidance reviewed September 23, 2026.
| Signal in the export | What to record | Where to check it | Decision it supports |
|---|---|---|---|
| A close live page exists | Dead path, matched live path | Screaming Frog SEO Spider crawl | 301 to the matched page |
| Referring domains point at the dead path | Linking domains, anchor text | Ahrefs Site Explorer or Semrush Backlink Analytics | 301 to the nearest live page, then monitor for repeat links |
| Three or more sessions, or repeat referrers | Session count, first and last date | GA4 Pages and screens | Open a content brief for a real page; 301 to the nearest live page until it ships |
| One or two sessions, no links | Path, date | GA4 | Leave the 404 and re-check in the next monthly export |
| The path matches a service you actually sell | Path, pattern, frequency | GA4 plus your own service list | Publish a real page at that exact path |
| Many dead paths pointed at the homepage | Redirect count | Search Console Page indexing report | Undo it; bulk redirects to an irrelevant page can be treated as soft 404s |
Source: SE Ranking and Semrush playbooks reviewed September 23, 2026. The thresholds are their rules of thumb, not Canlah AI measurements.
Google’s John Mueller calls 404 or 410 the technically correct response for a URL that does not exist, and in March 2025 he predicted a slight uptick in hallucinated-link clicks for six to twelve months. That is a reason to log the pattern, not to re-architect the site around it.
How to read five common ChatGPT 404 patterns
Five patterns cover most ChatGPT 404 referrals, and each maps to one fix: invented pricing paths get a real page, invented blog slugs get a grouped brief, deleted paths get a 301, tracked paths get traced to their base path and foreign paths stay 404.
Invented pricing or product path: The reader asked what you charge and the model guessed /pricing, /plans or /pricing/enterprise. If you sell the thing and have no page at that path, publish one, because a published price page is the page type answer engines cite most often on commercial questions.
Invented blog slug: The path follows your blog’s URL grammar and names a topic you write about but have not covered, which makes it a keyword with a click already attached. Group these by theme before writing, because one good page usually absorbs several near-identical invented slugs.
Path that used to exist: The URL was live, was deleted or moved without a redirect and survives in model memory and in third-party crawl databases. This is ordinary technical debt, and a 301 to the replacement page resolves it permanently.
Path with a tracking parameter attached: The request carries utm_source=chatgpt.com on a path that does not exist, confirming the referrer even when GA4 filed the session under another channel. Treat the parameter as evidence of origin, and fix the base path under whichever of the other four patterns it matches.
Path that is not yours at all: The model mixed your domain with a competitor’s path or a directory listing, producing something your site never had a grammar for. Canlah AI leaves these as 404s and records the frequency, because a rising count is a signal about how the model describes your category rather than a defect in your site.
What your 404 page should do for a ChatGPT visitor
A 404 page for ChatGPT visitors should do three things while still returning HTTP status 404: explain what the site offers, run the requested URL as a site-search query and link three to five live pages matched to the requested path. The first two come from John Mueller’s published guidance, and the third turns a guessed URL into a route to a real page. Keep the status at 404, because a 200 response on an error page creates a soft 404 that search engines index as thin content.
Log every 404 request with its full path, referrer and timestamp at the same time, so the next monthly export takes minutes rather than hours. On Cloudflare the same data comes from HTTP request logs filtered on status 404; on an origin server it is the access log filtered on the referrer field.
What a ChatGPT 404 audit cannot prove
A ChatGPT 404 audit cannot prove five things, and Canlah AI states all five on the first page of any report that contains one.
- It cannot prove an invented URL came from a hallucination rather than a deleted page, unless you hold a crawl history for the path.
- It cannot prove a redirect recovered the visitor, because the model may never surface that link again.
- It cannot count invented URLs cited but never clicked, which Ahrefs found at 2.38% of cited ChatGPT URLs against 1.01% of clicked ones.
- It cannot establish causation between a new page and a change in ChatGPT citations, only association across a recorded window.
- It cannot see referrals stripped of their referrer by privacy settings, in-app browsers or proxying.
When to move from a manual check to ongoing monitoring
A team should move from a manual ChatGPT 404 check to ongoing monitoring when three conditions hold together: ChatGPT referrals are large enough that losing 1.01% of them is money, invented URLs repeat across consecutive monthly exports and more than one team publishes pages that could answer them. Below that, a saved GA4 report and a monthly log export are enough.
Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Its GEO service is priced per AI platform: Starter from $1,500 a month (ChatGPT, 4 buyer questions), Growth from $3,000 a month (adds Google AI Overviews, 8 buyer questions) and Full customised (adds Google AI Mode and Gemini, 16 buyer questions). The questions are re-measured monthly, and each month closes with before-and-after reporting and AI-referral attribution; a free AI visibility report, delivered within 48 hours, comes before you choose a plan. The limitation is scope: Canlah AI measures which answers cite you and where those clicks land, and it does not ship the redirects and pages inside your CMS for you.
Whichever route you take, verify the list yourself before acting on it. Pull one week of raw 404 log lines, check five paths by hand and confirm the referrer is real, because a redirect map built on an unverified export is worse than no redirect map at all.
Frequently asked questions
Why is ChatGPT sending traffic to pages that do not exist on my site?
ChatGPT combines web search with internal knowledge, so some cited links are pages that were deleted or moved without a redirect and others are guesses that fit your site’s URL pattern without ever having existed. OpenAI’s own research attributes the second kind to training and evaluation procedures that reward a confident guess over an admission of uncertainty. Both kinds land on the same 404 page and need different fixes.
How do I fix ChatGPT 404 referrals in Google Analytics 4?
Fix ChatGPT 404 referrals in two passes: find them in Google Analytics 4, then decide per URL. Filter the Pages and screens report on session source containing chatgpt.com, because ChatGPT sets utm_source without utm_medium. Then 301 each invented URL that has a close live page or external links, publish a page when one URL draws three or more sessions and leave a clean 404 for the rest.
Should I redirect every hallucinated URL to my homepage?
No. Bulk redirects to an irrelevant destination can be treated as soft 404s, which wastes crawl budget and can be read as a site-quality signal, and Google’s guidance is that 404 or 410 is the correct response for a URL that never existed. Redirect a hallucinated URL only where a close live page exists or where external links point at the dead path.
Do ChatGPT 404 referrals hurt my Google rankings?
A 404 on a URL that never existed is normal and is not itself a Google ranking problem. The measurable costs of ChatGPT 404 referrals are the lost session, any link equity pointing at a dead path and the soft 404 risk created by redirecting in bulk to a page that does not match the request. Redirecting only where links or a close live page exist removes the last two costs.
How much ChatGPT referral traffic normally ends on a 404?
Ahrefs measured 1.01% of clicked ChatGPT URLs returning a 404 across 8,332,436 unique URLs, against 0.15% for Google referrals. Individual sites vary far more: Semrush reports one analysis of more than 18,000 landing pages at about 3.35% and one client account at 57%.
Can Canlah AI stop ChatGPT from inventing URLs for my domain?
No. Nobody controls how an independent model constructs a link. What Canlah AI can do is measure which invented URLs recur, which carry links and which match a page worth publishing, then report that beside the citation data for the same buyer questions.
Method and sources
Tool behaviour, study figures and Google guidance were checked on public pages on September 23, 2026 and were not tested through paid accounts. The 404 rates are Ahrefs measurements from 2025 describing a population of websites, not any single domain; treat them as a benchmark for your own export, not as a prediction of your rate.
- Ahrefs: How often do AI assistants hallucinate links, September 2, 2025. Clicked and cited 404 rates, the 0.43% six-assistant average, Google baselines and sample sizes in the first table.
- Semrush: How to find and fix ChatGPT 404 referrals, October 24, 2025. GA4 report path, the redirect-or-leave framework and the Go Fish Digital and Galinos single-site figures.
- SE Ranking: Hallucinated URLs, September 24, 2025. The 70-URL three-month log, the backlink count and the session thresholds in the decision table.
- Search Engine Journal: Google’s Mueller predicts uptick of hallucinated links, March 17, 2025. Mueller’s 404-page guidance and his six-to-twelve-month prediction.
- Google Search Central: Farewell to soft 404s. Why an error page must return 404 rather than 200.
- Google Search Console. Page indexing report, used to confirm soft 404 classifications.
- Canlah AI pricing and Canlah AI free AI visibility report. AI platforms and buyer questions by tier, monthly cadence, the free AI visibility report and AI-referral attribution scope.
See where your brand stands in AI answers. Run the free AI visibility audit →
Related articles
FREE WHITEPAPER
Marketing in the Agent Era
All 13 chapters public — no email wall. Includes an original dataset on the agent-readiness of 50 cross-border DTC brands.
Read it free →