The State of Knowledge Base Traffic (2026)

AI agents now make one in five knowledge base pageviews. Here's which AI crawlers read help centers, why they read them, and how many people AI assistants send back.

Research reportPublished November 2, 2026

Key findings

  1. AI agents made 20.5% of knowledge base pageviews in our October 2026 sample, about one in five. That's a floor, because some AI tools can't be counted.

  2. AI agents read 70% of the help centers that had any traffic during our October 2026 sample.

  3. Help centers get a more useful kind of AI traffic than the web at large: 20% of AI fetches are an assistant answering a person's question, against 2% across the web in Cloudflare's data. 46% is still model training.

  4. OpenAI is the biggest reader at 40% of AI agent traffic, ahead of Perplexity (22%) and Anthropic (19%).

  5. AI doesn't take weekends off. Human traffic drops 43% on Saturday and Sunday, AI traffic drops 19%, so agents make 27% of weekend pageviews.

  6. AI assistants doubled their referral share in 2026, from 0.75 visits per 100 search engine visits in the first half to about 1.5 in August and September.

  7. Of the AI assistant referrals we can trace by referring site, ChatGPT sends 73%, then Gemini (12%), Perplexity (6%), Claude (4%), and Copilot (4%).

  8. Half of knowledge base pageviews are on a phone, and two in three visits are a single page, so the first article someone lands on has to do the whole job.

One in five knowledge base pageviews is now an AI agent.

Behind that traffic sit three different jobs. One group of crawlers grabs text to feed the next model’s training, another indexes your pages for AI search, and sometimes an assistant opens your article because a customer of yours just asked it something. This report counts each of those three, and we’re drawing on HelpDocs’ own analytics across every help center we host.

We pulled these AI traffic statistics from 2026 and verified the biggest AI agents (OpenAI, Anthropic, and Perplexity) server-side, since some agents never execute JavaScript and would slip past browser-based tracking. We’re looking at how much of your knowledge base traffic AI agents produce, which AI crawlers read help centers and what draws them in, how many people AI assistants send your way, and who the remaining human readers turn out to be. The methodology walks through how we tallied each figure, including the places we simply couldn’t measure.

Last year’s 2025 edition looked at where human visitors come from. This year the most interesting reader isn’t human.

How much knowledge base traffic comes from AI agents

In a sample we took in October 2026, AI agents made 20.5% of all help center pageviews. People made the other 79.5%. AI agents read 70% of the help centers that had any traffic during the sample.

Who reads knowledge bases: people vs AI agents

20.5%

of pageviews were AI agents (all days)

  • AI agents 20.5%
  • People 79.5%

Each square is 1% of pageviews.

Share of help center pageviews in our October 2026 sample. All days covers the whole sample, including part days at either end. Weekdays and weekends use one full week only, so 20.5% isn't an average of the other two.
Show the data
ReaderShare of pageviews
People79.5%
AI agents20.5%
Help centers read by at least one AI agent70%

Note

Treat 20.5% as a lower bound, not a precise reading. Coding assistants and comparable tools fetch pages from the visitor’s own machine, so those requests go unverified and we only catch a portion of them. The spam filter we used also left the human tally 1 to 5% too high. Each factor nudges the real AI share upward.

Zoom out to the rest of the web, and the picture changes. In 2025, Cloudflare’s network-wide data showed AI bots taking 4.2% of all HTML requests, with Googlebot contributing another 4.5%. The two numbers come from different methods (Cloudflare’s figure sweeps in every other bot too), so don’t treat them as an exact ratio. Even so, that difference is large enough to conclude that help docs draw far more AI traffic than a typical page does, and that’s because assistants hunt for clear answers to specific questions.

AI doesn’t take weekends off

Most people read help articles during work hours. On weekends, human pageviews drop 43%, but AI agent pageviews dip only 19%. That pushes AI’s share of weekend traffic up to 27%, compared with about 21% on weekdays.

AI agents' share of help center pageviews, by day
  • Weekdays
  • Saturday and Sunday
  • Whole-sample share 20.5%
0%10%20%30%MonTueWedThuFri27.4%Sat26.5%Sun
One sample week in October 2026, chosen deliberately. Hover or focus a day for its value.
Show the data
DayAI agents’ share of pageviews
Monday20.3%
Tuesday19.7%
Wednesday20.5%
Thursday21.4%
Friday21.3%
Saturday27.4%
Sunday26.5%

The fact that the load stays steady over the weekend is a strong hint about what the traffic actually is. Crawlers follow schedules, not office hours, so the bulk of AI agent traffic comes from machines gathering pages, not from people asking questions. The next section gets into how much.

Which AI crawlers read knowledge bases, and why

All the AI agents we’ve identified fill one of three roles:

  • Model training. Crawlers like GPTBot and ClaudeBot collect text that may end up in a future model.
  • AI search indexing. Agents like OAI-SearchBot and PerplexityBot index pages so an AI search engine can cite them later.
  • Fetching for a person. Agents like ChatGPT-User and Claude-User open a page right now because someone asked a question.

Training crawlers make almost half of all AI traffic to help centers, and one AI fetch in five is a person’s question being answered in real time. That second number is what sets knowledge bases apart. Across the web, Cloudflare found 80% of AI crawling was for training, 18% for search, and just 2% for user actions. Help centers get ten times the share of fetches made on behalf of a real person.

AI agent traffic to knowledge bases
  • Model training46.4%
  • AI search indexing33.6%
  • Fetching for a person20.0%
Share of AI agent pageviews in our October 2026 sample. The web-wide view is Cloudflare's 12-month split of AI crawling by purpose, published August 2025. Switch views to compare.
Show the data
PurposeKnowledge bases (HelpDocs)Across the web (Cloudflare)
Model training46.4%80%
AI search indexing33.6%18%
Fetching for a person20.0%2%
CompanyShare of AI agent pageviews
OpenAI39.6%
Perplexity22.0%
Anthropic18.7%
ByteDance12.2%
Meta6.7%
Amazon0.7%
DuckDuckGo0.1%

OpenAI reads more than anyone else, by a wide margin, since it puts agents to work on all three jobs. Perplexity comes in second, and nearly all of that traffic comes from one indexer, PerplexityBot.

Every AI agent, ranked

PerplexityBot, ChatGPT-User, and ClaudeBot together account for 55% of all AI agent traffic. Change the measure and you’ll see how far each bot spreads its reading: Bytespider, ChatGPT-User, and ClaudeBot all reached around 60% of help centers in the sample.

AI agents reading knowledge bases
  • PerplexityBotPerplexity22.0%
  • ChatGPT-UserOpenAI17.3%
  • ClaudeBotAnthropic16.0%
  • BytespiderByteDance12.2%
  • OAI-SearchBotOpenAI11.6%
  • GPTBotOpenAI10.7%
  • Meta-ExternalAgentMeta6.7%
  • Claude-UserAnthropic2.7%
  • AmazonbotAmazon0.7%
  • DuckAssistBotDuckDuckGo0.1%

Bars show of all AI agent pageviews. Hover or focus a row for the other measure.

October 2026 sample. Help centers reached is the share of help centers with any traffic that the agent read at least once.
Show the data
AI agentCompanyPurposeShare of AI trafficHelp centers reached
PerplexityBotPerplexityAI search indexing22.0%52.9%
ChatGPT-UserOpenAIFetching for a person17.3%60.0%
ClaudeBotAnthropicModel training16.0%60.0%
BytespiderByteDanceModel training12.2%61.1%
OAI-SearchBotOpenAIAI search indexing11.6%56.7%
GPTBotOpenAIModel training10.7%21.8%
Meta-ExternalAgentMetaModel training6.7%27.8%
Claude-UserAnthropicFetching for a person2.7%47.7%
AmazonbotAmazonModel training0.7%11.7%
DuckAssistBotDuckDuckGoFetching for a person0.1%12.6%

These crawlers behave differently when they browse. ClaudeBot opened about 2.2 times as many unique articles as ChatGPT-User. Training crawlers work their way through an entire site, while ChatGPT-User only visits the pages people are actually asking about.

Important

We identify ByteDance, Meta, Amazon, and DuckDuckGo agents purely from their user agent string, and only when they ran our tracking script. So you’re seeing an undercount of their real traffic, and a few of those hits could be impostors wearing the same name. OpenAI, Anthropic, and Perplexity got a stricter check, since we matched their requests against each company’s publicly listed IP addresses.

ChatGPT does most of the asking

If you look at fetches triggered by a person’s question, ChatGPT’s almost always the assistant behind them. ChatGPT-User handled 86% of those fetches. On weekdays, it fetched once for every 18 human pageviews.

86%of fetches made for a person came from ChatGPT

1:18ChatGPT fetches to human pageviews on weekdays

You’ll want to hold onto that figure when you sit down to write help articles. A real slice of your readership now gets your docs through an assistant, and the assistant reads the article, supplies the answer itself, and might never send the visitor to your site.

AI referral traffic: how many people AI assistants send

AI referral traffic means someone clicking a link that lives inside an AI assistant’s response. Over the first nine months of 2026, AI assistants delivered about one visit for every 111 that search engines sent. Treat that as a floor. We identify an AI referral by its referring site, but most links from AI assistants arrive with only a tracking tag and no referring site, so they’re counted as direct visits instead. The real figure could be several times higher.

Still, the trend points up. AI assistants generated 0.75 visits per 100 search visits across the first half of the year. August and September pushed that to about 1.5, double the earlier rate.

1:111AI assistant visits to search engine visits, January to September, counted by referring site

2×AI’s referral share in August and September versus the first half

Visits from AI assistants per 100 visits from search engines, 2026
00.511.52Jan to Jun average 0.75JanFebMarAprMayJunJulAugSepSep 1.50
People arriving at help centers from AI assistant answers, per 100 arriving from search engines. October is left out (see Methodology).
Show the data
Month (2026)AI assistant visits per 100 search visits
January0.75
February0.80
March0.76
April0.58
May0.70
June0.92
July1.01
August1.60
September1.50

Read the trend with some care. AI referrals held roughly steady month to month while visits from search engines fell, so most of the rise in the ratio comes from search shrinking. Part of that drop is likely AI answers inside Google itself: Pew Research Center found people clicked a search result in 8% of visits when Google showed an AI summary, against 15% when it didn’t. The help centers in our data also change as customers come and go, though, so treat the decline as directional.

Traffic from ChatGPT and the other AI assistants

ChatGPT sends almost three in four AI referrals. Gemini is a distant second, and Perplexity, Claude, and Copilot share most of the rest. That’s almost exactly the split SE Ranking found across 101,574 websites in early 2026, with one exception: Claude sends help centers 4.4% of AI referrals, against 2.6% for websites in general. Claude’s users seem to lean toward the technical questions help docs answer.

Which AI assistants send people to knowledge bases
  • ChatGPT73.3%
  • Gemini11.5%
  • Perplexity6.4%
  • Claude4.4%
  • Copilot4.1%
  • Others DeepSeek, Grok, Mistral, Poe, and more0.2%
Share of AI assistant referrals. Knowledge bases: January to October 2026. All websites: SE Ranking, January to April 2026. Switch views to compare.
Show the data
AI assistantKnowledge bases (HelpDocs)All websites (SE Ranking)
ChatGPT (chatgpt.com)73.3%74.8%
Gemini (gemini.google.com)11.5%11.6%
Perplexity (perplexity.ai)6.4%7.2%
Claude (claude.ai)4.4%2.6%
Copilot (copilot.microsoft.com)4.1%3.5%
DeepSeek, Grok, Mistral, Poe, AI Studio, and You.com0.2%0.3%

Now look at the readers. Perplexity is the second biggest AI to read knowledge bases but drives just 6% of AI referrals. Its agent tends to index pages, and its answers seldom send anyone onward.

How people use knowledge bases

People still outweigh the bots, since they’re the ones who open a support ticket when an article falls short. This section keeps the comparison fair: it covers 1 August to 12 October 2026, a period with no tracking changes, and counts only people.

49.5%of pageviews were on a phone

67%of visits were a single page

43%of article views went to a typical help center’s 10 most-read articles

Half of help center traffic is on a phone

Phones and computers now split pageviews almost exactly in half: 49.5% on a phone, 49.9% on a computer and 0.5% on a tablet. iOS is the second biggest platform at 31%, behind only Windows. Another 7% of pageviews come from in-app browsers, which is a link opened inside another app such as an email or chat client.

What people read knowledge bases on
  • Computer49.9%
  • Phone49.5%
  • Tablet0.5%
Share of human pageviews, 1 August to 12 October 2026.
Show the data
DeviceShare of pageviews
Computer49.9%
Phone49.5%
Tablet0.5%
Operating systemShare of pageviews
Windows36.6%
iOS31.4%
Android18.6%
macOS11.4%
Linux0.9%
ChromeOS0.9%
BrowserShare of pageviews
Chrome47.3%
Safari24.0%
Edge10.9%
In-app browsers6.7%
Android browser6.3%
Firefox2.9%
Google app1.2%

Mobile made up 37.6% of traffic in last year’s report. The two editions weren’t measured the same way, so treat the rise as a broad direction instead of a precise shift. The trend still points the same place, though: your help center now gets opened on a small screen about as often as on a big one.

Two in three visits are a single page

Help centers don’t get browsed. 67% of visits look at exactly one page, and just 5% reach six or more. That puts the average at 1.9 pages per visit. You show up via search, a product link, or an AI answer, get your answer or don’t, and you’re gone.

Where visits drop off
  • Still reading
  • Left
1 page100.0%66.8%left after 1 page2+ pages33.2%15.2%left after 23+ pages18.0%13.1%left after 3 to 56+ pages4.9%3.7%left after 6 to 1011+ pages1.2%
Share of human visits that reached each number of pages, and the share that ended at each step. 1 August to 12 October 2026.
Show the data
Pages in the visitShare of visitsVisits that read at least this many
166.8%100%
215.2%33.2%
3 to 513.1%18.0%
6 to 103.7%4.9%
11 or more1.2%1.2%

These are averages across every help center we host. Your own drop-off will look different, and the visits worth studying are the ones that ended badly. Reader Insights shows how each visit to your own help center ended, the last article people read before they gave up, and the searches they reworded or abandoned.

Most reading happens on your own domain

Two thirds of pageviews happen on a customer’s own domain, such as help.yourcompany.com. Another 31% are on a helpdocs.io address, and 2.4% happen inside an in-app help widget without the reader leaving the product. And most of what people open is an article: 64% of pageviews, against 13% for category pages.

Where and what people read
  • Your own domain66.9%
  • A helpdocs.io address30.8%
  • An in-app help widget2.4%
Share of human pageviews, 1 August to 12 October 2026. Switch between where the page was read and what kind of page it was.
Show the data
WhereShare of pageviews
Your own domain66.9%
A helpdocs.io address30.8%
An in-app help widget2.4%
Page typeShare of pageviews
Articles64.1%
Home, search results, and other pages22.5%
Category pages13.4%

Monday is the busiest day

Reading follows the working week. Monday gets 17% of the week’s pageviews, then each day is a little quieter until the weekend, when Saturday and Sunday get about 10% each. Across the day, reading peaks between 14:00 and 16:00 UTC, when Europe’s afternoon overlaps the Americas’ morning.

Share of the week's human pageviews, by day
  • Weekdays
  • Saturday and Sunday
  • Even split 14.3%
0%10%20%MonTueWedThuFri10.3%Sat9.8%Sun
Days in UTC, 1 August to 12 October 2026. Hover or focus a day for its value.
When people read: share of the day's pageviews, by hour (UTC)
0%2%4%6%8%Busiest: 14:00 to 16:0000:0003:0006:0009:0012:0015:0018:0021:00
Hours in UTC, so the curve blends time zones. 1 August to 12 October 2026. Hover for any hour.
Show the data
Day (UTC)Share of the week
Monday17.0%
Tuesday16.4%
Wednesday16.3%
Thursday15.8%
Friday14.4%
Saturday10.3%
Sunday9.8%
Hour (UTC)Share of the dayHour (UTC)Share of the day
00:003.19%12:004.99%
01:003.07%13:005.92%
02:002.73%14:006.35%
03:002.38%15:006.35%
04:002.30%16:005.90%
05:002.34%17:005.68%
06:002.68%18:005.65%
07:003.20%19:005.42%
08:003.66%20:005.04%
09:003.74%21:004.39%
10:003.67%22:003.82%
11:004.08%23:003.44%

The weekly shape is the mirror image of the AI agents earlier in this report. People read on weekdays, and crawlers keep going through the weekend.

A few articles do most of the work

In a typical help center, the 10 most-read articles get 43% of all article views, and 32% of articles account for 80% of views. Most knowledge bases have a short list of articles that matter far more than the rest.

This pattern shows up across help centers, too. The top tenth of them account for 78% of all the reading.

32%of a typical help center’s articles make up 80% of its views

78%of all reading happens on the busiest tenth of help centers

“Typical” here means the median help center, not an average weighted by traffic.

Search, feedback, and contact

Roughly one search happens for every five article views, and those searches run short: the median is two words, with only 13% reaching five words or more. 23% of searches end in a click on a result, while 8% turn up nothing (here’s how to fix no-result searches). On-site search is still mostly keyword search.

Per 1,000 article views:

199searches

11article feedback votes

5.6contact form submissions

Here’s a closer look at those searches:

23%end with a click on a result

8%return no results

13%are five words or longer

A search only registers a click once the reader opens a result, so if you grabbed the answer straight from the list, that’s a no-click in the books. Contact forms are optional, which means help centers that include one see more than 5.6 per 1,000.

Was this article helpful?
  • Helpful48.2%
  • Neutral13.9%
  • Not helpful37.9%
Article votes left by readers, January to October 2026.
Show the data
VoteShare of article votes
Helpful48.2%
Neutral13.9%
Not helpful37.9%

Nearly two in five votes come back saying the article missed the mark. The people who take the time to cast one lean toward the already annoyed, so read the results as your repair list, not a satisfaction score. Those same shaky articles now get picked up and recited by AI agents, too.

Key takeaways

  1. Write for the reader who isn’t on your site

    An assistant reads your article, answers from it, and often sends nobody back. Put the answer in the first two sentences, use the words your customers use, and keep one task per article. Our guide to how AI uses your knowledge base goes deeper.

  2. Design for one page on a phone

    Half of pageviews are on a phone and two in three visits never reach a second page. Keep the answer above the fold, paragraphs short, and screenshots readable at phone width. More in our guide to help center design.

  3. Look after your top 10 articles

    They get 43% of article views in a typical help center. Review them first, keep them current, and check each one answers the question its title asks.

  4. Decide what to allow by purpose, not by bot

    Blocking training crawlers keeps your docs out of future models. Blocking AI search indexers and agents that fetch for a person stops assistants from citing your help center when customers ask about you. An llms.txt file is fine to publish if it costs you nothing, but few agents fetch one yet, so clear, crawlable articles matter more.

  5. Measure on the server, not just in the browser

    If your analytics only counts visits that run a script, it’s missing most of your AI traffic. HelpDocs records verified AI fetches on the server and shows them in the AI references tab of your knowledge base analytics. Our guide to the metrics that matter covers the rest.

  6. Fix the articles people vote down

    With 38% of votes saying “not helpful”, your feedback report is the fastest content to-do list you have, and fixing those articles helps the AI agents that quote them too. Connecting docs to your own AI agent? See our MCP server and AI support agents guide.

Tip

Want your help center ready for AI readers without the busywork? HelpDocs can publish an llms.txt and a clean markdown copy of every article, free on every plan, and shows you which articles AI agents read.

Methodology

The figures in this report come straight from HelpDocs’ production analytics, gathered during October 2026. Each one is a rollup across every help center hosted on HelpDocs. No number singles out any customer, and we keep the total count of help centers in the dataset private.

Time windows. The AI agent numbers come from a deliberately chosen slice of October 2026, starting on the day we began verifying AI agents on our servers, so every major agent gets counted the same way throughout. Day-of-week patterns for AI agents draw on one full week within that sample. AI referral and article feedback figures span January through October 2026, with the referral trend ending in September. The section on how people use knowledge bases covers 1 August to 12 October 2026, a window that sits between two tracking changes, so all of its figures are like-for-like.

How shares are counted. Nearly all of our percentages are calculated from one combined pool of pageviews, which means the busiest help centers dominate: the top tenth of them alone accounts for 78% of human pageviews. That’s why you’ll see “of pageviews” instead of “of help centers”. When we describe a typical help center, we switch to medians across help centers: pageviews a month, top 10 articles, and the share of articles making up 80% of views.

Identifying AI agents. Every request’s user agent gets compared with a list of 19 AI bots, and each bot carries a label for why it’s fetching: on a person’s behalf, for AI search indexing, or for model training. During the sample, our servers logged fetches from OpenAI, Anthropic, and Perplexity agents whenever the request came from an IP address on that company’s published list. Every other AI agent counts only if it runs our tracking script.

People. We count a person whenever a pageview comes in without an AI agent tag and gets past our regular spam filter. Generic search engine crawlers, Googlebot for example, are thrown out the moment they’re logged, so they never show up in either bucket.

AI referrals. An AI referral is a person arriving from a link in an AI assistant’s answer, identified by the referring site (for example chatgpt.com or gemini.google.com). This has always been recorded, so it covers the whole year. Links that carry only a tracking tag and no referring site count as direct visits, so these figures undercount AI referrals.

Help centers reached. Of the help centers that saw some traffic during the sample, the portion a specific AI agent opened at least once.

Industry benchmarks. Our wider-web comparisons draw on published research, and each source measured things its own way. That means Cloudflare’s 2025 Year in Review covers AI bots as a share of HTML requests across Cloudflare’s network in 2025, while Cloudflare’s analysis of AI crawling by purpose spans 12 months to mid-2025. SE Ranking’s AI traffic study tracks AI referrals across 101,574 websites from January to April 2026, and Pew Research Center looked at Google clicks with and without an AI summary among 900 US adults in March 2025. Because those time windows and definitions don’t match ours, treat each comparison as a direction, not an exact gap.

What this data can’t tell you

  • The AI share is a floor. Coding assistants and tools like them pull from the user’s own machine, which we can’t verify, so they get only partially counted. The spam filter we used also pushes the human count 1 to 5% high.
  • Some agents are undercounted. ByteDance, Meta, Amazon, and DuckDuckGo agents get counted by user agent only, and only when they ran our script. Before server-side verification, we saw just the 2 to 4% of AI agents that happened to run that script, which is why the AI numbers come from a recent sample instead of the full year.
  • AI referrals are a floor. We count an AI referral by its referring site. Most links from AI assistants arrive with only a tracking tag and no referring site, so they’re counted as direct visits. The real volume could be several times one in 111, and the split between assistants could shift too.
  • No year-over-year AI trend yet. We need a second year of server-side data for that, so look for it in the 2027 edition.
  • Tracking changed in mid-July 2026. Spam detection and session tracking got rebuilt, so totals across that point don’t line up cleanly. Comparing shares inside a single month is fine, and the reader behavior section sidesteps the change altogether.
  • Times are in UTC. The day and hour numbers mix time zones, so the 14:00 to 16:00 peak marks Europe’s afternoon overlapping the Americas’ morning, not a single region’s lunchtime.
  • The fall in search referrals is fleet-wide. The lineup of help centers shifts as customers join and leave, so read it as directional rather than a decline per customer.
  • October referrals are left out of the trend. In October, ChatGPT links stopped carrying a referring site and, for about a week, their tracking tag too, so counted referrals fell sharply. Counting the tag, AI visits were back above September’s level later in the month. Our referral figures go by referring site, so the trend ends in September.
  • What we left out on purpose. Reading language (half of all pageviews carry no language), signed-in readers (they include teams looking at their own help center), Ask AI answer counts, how often agents pick markdown copies of articles (too few help centers to publish), and any trend in zero-result searches (a recording fix in August leaves only the current 8% rate as trustworthy).
  • Comparing with the 2025 edition. Last year’s report blended our traffic data with industry benchmarks without drawing a line between them. This year, our numbers come only from our analytics, and any outside benchmark gets named and linked where it appears, so read changes between editions as directional.

Frequently asked questions

How much website traffic comes from AI bots?
Across HelpDocs knowledge bases, AI agents made 20.5% of pageviews in a sample we took in October 2026, about one in five. On Saturdays and Sundays that rose to 27%, because people read far less at weekends and AI agents don't. Treat 20.5% as a floor: tools that fetch from a person's own computer, such as coding assistants, can't be verified and are only partly counted. For context, Cloudflare measured AI bots at 4.2% of all HTML requests across its network in 2025, so help docs look like a much heavier target than the average web page, though the two figures are counted differently.
What's the difference between GPTBot, OAI-SearchBot, and ChatGPT-User?
They are three OpenAI agents with three jobs. GPTBot crawls pages to train models (10.7% of AI traffic in our data). OAI-SearchBot indexes pages so ChatGPT search can cite them (11.6%). ChatGPT-User fetches a page in the moment because someone asked ChatGPT a question (17.3%). Anthropic (ClaudeBot and Claude-User) and Perplexity (PerplexityBot) follow the same split.
Should you block AI crawlers from your knowledge base?
Decide per purpose rather than all or nothing. Blocking training crawlers keeps your docs out of future models but doesn't change today's answers. Blocking AI search indexers and agents that fetch for a person means assistants can't read or cite your help center when your customers ask about your product. In our data ChatGPT-User alone made one fetch for every 18 human pageviews on weekdays, so for most support teams being readable is worth more than being hidden.
How much traffic does ChatGPT send to websites?
For knowledge bases, less than search, but more than referral counts show. From January to September 2026, AI assistants sent at least one visit for every 111 from search engines, and ChatGPT sent 73% of those we could trace. Most AI links arrive without a referring site, so the real volume is likely higher. The share doubled during the year, from 0.75 AI visits per 100 search visits in the first half to about 1.5 in August and September.
How much knowledge base traffic is on mobile?
About half. From August to mid-October 2026, 49.5% of human pageviews on HelpDocs knowledge bases were on a phone, 49.9% on a computer and 0.5% on a tablet. Two in three visits were a single page, so a phone reader usually sees one article and leaves. Put the answer near the top and keep paragraphs short.
Why doesn't AI traffic show up in my analytics?
Most analytics tools count a visit when a tracking script runs in the browser, and most AI agents never run JavaScript. Before we started recording verified AI fetches on the server, our own script-based tracking saw only 2 to 4% of them. If your analytics can't see server requests, assume it is missing most of your AI traffic.

Cite this report

Found this useful? A link back to the report helps others find it too.

Turn what you learn into trusted answers.

Create, organize, translate, and connect your docs without the admin sprawl of a support suite.

Start free 14-day trial

No credit card required.

Migrate from