AI crawlers and bots

Checking whether your ChatGPT visits moved to direct traffic in October 2026

By Emily Parker · Oct 23, 2026 · 8 min read

If your ChatGPT visits fell off a cliff on Oct 9, 2026 while Direct grew, your site may not have lost anyone. Search Engine Roundtable reported on Oct 14 that OpenAI seemed to have removed or replaced the UTM tags on ChatGPT links after the GPT-6 release about a week earlier. Cedric Huesler of Adobe wrote that the links opened with no tag and no referrer, "so your analytics files the visit under direct." Peeko works the same way. It counts ChatGPT visits from the referrer on ordinary page views, so a visit that arrives with no referrer lands under Direct in Peeko too. On Oct 15 the report was updated: Huesler said ChatGPT was tagging its links again and called the change a bug.

Your own numbers for Oct 9 to 14 still show what happened on your site. This post walks through how to read them. It also works for the next time a referrer goes missing, whatever the cause.

Two paths, two kinds of number

Peeko sees ChatGPT in two separate ways, and the difference is the whole method.

Visits are ordinary page views. A person clicks a link in ChatGPT, their browser loads your page, the script tag records the view, and the referrer says where they came from. No referrer means no ChatGPT label.

Answer fetches come from a server side path, because crawlers do not run JavaScript. The @usepeeko/bot-track middleware matches the user agent against a curated list of 55 agents from 21 companies. OpenAI describes the agent behind these fetches this way: "When users ask ChatGPT or a CustomGPT a question, it may visit a web page with a ChatGPT-User agent." A fetch says a page was read while an assistant was replying to somebody. The question itself never reaches your server. No address and no visitor hash exist on this path, and it does not depend on a referrer.

That independence makes fetches a useful second signal. It also means the two numbers measure different things. A fetch is not a visit: an assistant can read your page and the person never clicks. Never add the two. Cloudflare's 2025 review shows how far apart crawling and referring can sit.

500,0003,700700AnthropicOpenAIPerplexity

Source: Cloudflare, Cloudflare Radar 2025 Year in Review, 2025-12-15. Retrieved Oct 20, 2026.

The chart has limits. Cloudflare's ratios cover all crawling by each company, training included, not ChatGPT-User alone, and they are peaks from 2025, not a current figure. The point is only that crawl counts and referred visits are different quantities.

What was reported, and what is not known

The reporting so far comes from a few sources, and each has limits.

  • Adobe's figures. Huesler looked at about 500 pages where ChatGPT drove 5% or more of visits, and compared Oct 9 to 13 with the prior week. ChatGPT visits fell by about 31,000 a day, Direct rose by about 26,000 a day, and search fell 7%. Direct rose on 2 in 3 of the pages that lost ChatGPT traffic. His own caveat, quoted by mean.ceo: "It's early: 4 days of data, sampled."
  • The gap. A fall of 31,000 and a rise of 26,000 leave about 5,000 a day that did not show up in Direct on those pages. None of the three reports we read explains the difference.
  • The cause. Adobe says GPT-6 opens links with no tag and no referrer. The Prompt Insider reports that Profound points instead at how the new Intelligent UI renders links. Profound's figures: the share of ChatGPT answers with at least one tagged source link was about 98% on Oct 8, 75% on Oct 9 and 66% by Oct 14. Neither explanation is confirmed, and OpenAI had not commented in the reports we read.
  • A contrary signal. Sébastien Billard reported a sharp drop in both ChatGPT human traffic and ChatGPT-User bot traffic on his site since Oct 9, while OAI-SearchBot traffic rose. So a fall in fetches on your site is not proof of lost demand, and steady fetches are not guaranteed either.
  • The fix. The Oct 15 update says tags are back. We have no figure for how fully.

The same Prompt Insider piece says a different analytics tool cannot recover source data that was never sent with the visit. Nothing you do now will relabel those visits. You can only read around them.

Check your own numbers

You need three daily series for the same window. Use Oct 2 to Oct 20, 2026. Oct 9 was a Friday, so Oct 9 to 14 (Friday to Wednesday) lines up weekday for weekday with Oct 2 to 7. Weekday patterns can swing Direct on their own, and this keeps them out of the comparison. Page views and crawler hits are kept three years by default, so the baseline is there.

1. ChatGPT visits per day. In Peeko, set the date range and read the daily count for the ChatGPT referrer. If your coding agent is connected to the MCP server, ask it for the same series. Connecting Claude Code or Cursor takes a few minutes, and npx @usepeeko/cli speaks the same endpoint, so a command and a tool call return the same numbers. For how the referrer match works, see counting visits from AI assistants by referrer.

2. Direct visits per day. Same range, same tool. Then go one level down: which pages gained Direct visits? Adobe's strongest hint was page level, since Direct rose on pages that had been earning ChatGPT traffic. A rise on your front page is probably something else. A rise on the deep article that ChatGPT used to send people to is the pattern to look for. If your tool separates new from returning visitors, check that too. Peeko tells a returning visitor from a new one with a second hash that does not mix in the date. Adobe used a consent banner rate as a proxy for new visitors, and a jump in new visitors inside Direct is the same kind of hint.

3. ChatGPT-User fetches per day. In Peeko this is the AI crawler tab, filed under AI answers. If you are not using the middleware yet, your access log has the same information. This works on an nginx or Apache log in the combined format, and counts fetches per day:

grep 'ChatGPT-User' access.log \
  | awk -F'"' '{ match($1, /\[[^:]+/); print substr($1, RSTART+1, RLENGTH-1) }' \
  | uniq -c

And this counts requests that carried a ChatGPT referrer or a utm_source=chatgpt.com tag, per day:

awk -F'"' '$4 ~ /chatgpt\.com/ || $2 ~ /utm_source=chatgpt\.com/ { match($1, /\[[^:]+/); print substr($1, RSTART+1, RLENGTH-1) }' access.log \
  | uniq -c

Both assume one log file in time order. Rotated logs need zcat and a loop. They count requests, not people.

Reading the three lines

Put the three series in one table, one row per day, and compare Oct 9 to 14 against the week before. Then read the pattern.

ChatGPT visitsDirectChatGPT-User fetchesMost likely reading
DownUp by a similar amountSteadyVisits moved to Direct. Demand held.
DownFlatDownFewer answers are reading your pages. A real drop, or the same UI change reducing fetches as Billard saw.
DownFlatSteadyNot explained by this change. Check that the script is still on the page.
FlatFlatFlatYour site was not affected.

Two notes on the table. In the third row, a broken tracker would also pull Direct and every other referrer down, so check total page views. In the second row, do not conclude too fast: Billard's site shows fetches can fall together with referred visits, so look at whether OAI-SearchBot hits, which are filed under Search indexing, moved too.

Small sites should be careful with the first row. If you get 20 ChatGPT visits on a normal day, a swing of five is noise, and Direct gets visits from newsletters, chat apps and bookmarks that nobody tags. The pattern is only convincing with a clear size match, on the right pages, across several days. Adobe's own sample was 500 pages and it still cautioned that four days of sampled data is early.

A shopkeeper counts pebbles into two jars while visitors with small tags enter one door and visitors with blank coats enter a plain door beside it.

After the fix

If the tags came back on your site, the ChatGPT line should recover from Oct 15 and Direct should ease. If ChatGPT stays low and Direct stays high, the move hypothesis gets weaker, not stronger, and the table's second row becomes more likely. Profound's figure of 66% tagged by Oct 14 was still a drop from 98%, so a partial recovery is the expected shape if the sources are right, not a full one.

Do not edit history. Do not hand-label Direct visits as ChatGPT to make the line look whole. Write down the window instead: ChatGPT visits undercounted, Oct 9 to 14, 2026. When you report the number, report the range and the fetch count beside it, and keep them as two numbers.

The loss also reaches past the page view. A visit without a source cannot be followed to a signup or a sale under its real origin. If that matters to you, following an AI referral to a signup without cookies covers what survives. For why a visibility score from a vendor and your own server side counts can disagree, see AI visibility scores versus server side counts.

The method outlasts this week. Whenever a referrer disappears, whether from an app update, a browser change or a link redesign, the same three lines tell you whether people stopped coming or only stopped signing their visits.