How Much AI Traffic Is People? 15 Small Business Sites

On a typical one of the 15 small business websites we host, about one in six AI bot requests came from an assistant fetching a page, the kind of fetch ChatGPT makes when someone asks it a question. The rest were crawlers collecting pages for model training or for a search index. Pooled across all 15 sites the share was a quarter, mostly because of one busy site. Most of those fetches were for the homepage.

The figures come from 30 days of server logs, 9 October to 8 November 2026, and about two weeks of visitor counts, on 15 Canadian client sites we build and host: 10 local service businesses, mostly home building and renovation, plus trades, specialist installers and hospitality, and 5 small B2B companies. We measured them with cite-met, the hosting and measurement platform we run, and we sell AI search work, so read this as a vendor's data with the method laid out at the end. The AI traffic studies that rank today measure networks of large sites. Most of these sites are small local businesses with modest traffic.

What should a small business take from this?

• Your homepage is the page an assistant opens. On these sites, 70.8% of assistant fetches went to the homepage. That makes it the safest place to say, in plain text, what you do, where, for whom, a rough price range and how to reach you.
• Most assistant reads don't produce a visit. In the last week of the window, ChatGPT fetched pages about 15 times for each visit it referred. That figure rests on a small number of visits, so treat it as rough.
• Blocking is a per-bot choice. Blocking a training crawler doesn't stop the assistants that fetch pages for people, and some of those don't follow robots.txt anyway.
• Your own AI traffic chart mostly shows who crawled you that month. The busiest bot changed from site to site, and on a typical site one bot made 42% of the AI requests.

What counts as "AI traffic" in your dashboard?

Three different kinds of bot, and most dashboards add them together. Each AI company runs separate user agents for separate jobs:

The first column is a model being trained. The second is a search engine deciding whether your page can appear in answers later. Only the third is meant to mean that someone has just asked the assistant something and it went to read your page. Meta says Meta-ExternalAgent is used both for training and for "improving products by indexing content directly", so we count all of it as training.

Two names you may see in a robots.txt file should never appear as crawlers in your logs. Google-Extended and Applebot-Extended are permission tokens that Google and Apple read to decide what they may use for AI, while the crawling is done by their existing crawlers. Our platform labelled a handful of requests with those names in 30 days, under 0.1% of the total, and those were mislabelled or borrowed names. Our counts leave out Googlebot and Bingbot, and the fetchers Google and Microsoft use for Gemini and Copilot, so "AI bots" here means the companies in the table and a handful of smaller ones.

How much of the AI traffic is people?

On the median site, 17.4% of AI bot requests were assistant fetches. The middle half of sites ran from 3.8% to 25.9%, and the full range was 0% to 47.2%. Pooled across every request, the share was 24.6%.

Medians are for each column on its own, so they don't add to 100%. "Other or unclear" is mostly Velen's crawler, plus DeepSeek, Diffbot and a few stray labels.

An "assistant fetch" here is a request that named an assistant's user agent and asked for a page. We didn't check IP addresses, so we can't prove a person was behind each one. One B2B site shows why that matters: its ChatGPT-User fetches were within 2% of each other in three straight weeks, mostly of its homepage, which is steadier than real visits usually are. That looks like a tool asking ChatGPT the same question on a schedule, or a tool borrowing ChatGPT's name. That one site made 40% of all assistant fetches in the sample. So the share depends on what you assume:

Small sites look different from big networks. Cloudflare's analysis of its network, published in August 2025 from July 2025 data, found nearly 80% of AI crawling was for training, and user-triggered fetches plus undeclared crawling were under 5%. Its figure is weighted by request volume, so large sites dominate it, and it is 15 months older than ours. Our own studio site, which has more than a hundred blog posts, is closer to Cloudflare's picture for a narrower reason: Meta's crawler made 81% of its AI requests over the same 30 days, and assistant fetches were 2.6%. One likely reason for the gap is that a local contractor's site has little for a training crawler to collect.

Which AI companies crawl small business sites most?

OpenAI, Anthropic and Perplexity reached all 15 sites. OpenAI made the most requests, 40.7%.

With Claude-User included, Anthropic would have made about 1.5 times as many requests.

Which bot was busiest changed from site to site. ClaudeBot topped 6 of the 15, ChatGPT-User 3 and Meta's crawler 2, with GPTBot, OAI-SearchBot, PerplexityBot and Common Crawl's CCBot on one each. On a typical site, the single busiest AI bot made 42% of its AI requests, so one site's chart mostly reflects which company happened to crawl it heavily, not a trend.

Does ChatGPT send anyone back?

A few people. Across the 15 sites, AI assistants referred about 1% of visits, and 93% of those came from ChatGPT. The median site got 0.3%, and 7 of the 15 got none in the two weeks the visitor counts cover.

In the last 7 days of the window, ChatGPT fetched pages on 10 of the 15 sites and made 15.2 fetches for each visit it referred. We can't tell which visit followed which fetch; we divided one count by the other. Those visits came from only 8 sites, which is a thin base. Counting noise alone puts the ratio somewhere between about 10 and 23, the per-site figures ran from 1 to 60 with a median of about 13, and because some visits from ChatGPT's app arrive with no referrer, the real ratio is probably lower.

It doesn't hold for every site. Lotso Travel, an eSIM publisher of ours with far more traffic than these sites, got visits from ChatGPT in the same 30 days with not one ChatGPT-User fetch logged. Its answers about Lotso were probably built from a search index instead: OAI-SearchBot crawled it heavily. Lotso also logged no requests from Anthropic or Perplexity, so we treat it as a contrast, not evidence.

Our ratio is not Cloudflare's. Cloudflare reported 887 HTML page requests from OpenAI's crawlers for every referral in the first week of August 2025. We divided only ChatGPT-User fetches by ChatGPT's referred visits. The two answer different questions.

For an owner, it means that most of the time ChatGPT read one of these sites live, no visit followed that we could see. Part of what it told the person came from what it read, so what your pages say matters even when the visit count is small.

What do AI assistants read when they fetch a small business site?

The homepage, mostly. Of the assistant fetches we could see page by page, on 13 sites, 70.8% went to the homepage. Page types come from the 50 most-fetched pages on each site.

Guides were fetched only on the five sites that publish a lot of them, and cost pages only on the four that have one. Location pages were almost never opened.

This doesn't prove what an assistant needs, only what it opened. But on most of these sites the homepage is the page it opens, so that is where to put what it would need to answer about you, in plain text that loads without JavaScript. In tests published in December 2024, Vercel found that the AI crawlers it checked from OpenAI, Anthropic, Meta, ByteDance and Perplexity didn't run JavaScript; Google's and Apple's, which render pages, were the exceptions. Our post on how ChatGPT recommends local businesses covers what else feeds those answers.

Should you block AI crawlers?

Decide per bot, because blocking one job leaves the others running. Blocking GPTBot keeps your pages out of OpenAI's training and leaves you in ChatGPT search; OpenAI says blocking OAI-SearchBot takes you out of ChatGPT's search answers, though your pages can still appear as navigational links. The person fetchers follow different rules: OpenAI says robots.txt rules may not apply to ChatGPT-User, Perplexity says Perplexity-User generally ignores them, and Meta says its fetcher may bypass them. Anthropic says Claude-User follows robots.txt. And robots.txt is a request, not a lock: it works only on crawlers that honour it.

In our view, for most small businesses the training crawlers are the only real choice to make. Keeping the search indexers and person fetchers open is what lets an assistant read you when someone asks.

How we counted, and what we couldn't see

The sites. 15 client websites we build and host on cite-met, our own platform: 10 local service businesses and 5 small B2B companies, all in Canada. Visitor counts cover 30 days on one site and 8 to 14 days on the rest. We name none of them. Of the 26 projects our platform tracks, we left out our own sites, because we change them constantly and visit them with our own tools, and mention three only as contrasts.

Requests. A request is any response a bot received from a site in the 30 days, 9 October to 8 November 2026: pages, files, redirects and errors, not HTML pages only. We counted requests whose user agent belongs to a known AI bot. We didn't check each site's robots.txt or bot settings, so a site that blocks a crawler would show a lower share for it.

How we sorted bots. We sorted each bot by the job its operator documents. Where the operator documents none (ByteDance's Bytespider) or the output is used by others (Common Crawl's CCBot), we counted it as training. Meta's dual-purpose crawler counts as training. Assistant fetches are ChatGPT-User, Meta-ExternalFetcher, DuckAssistBot, Mistral's fetcher and Perplexity-User.

People. A visitor script on each site counts visits by people and records the referring site. A visit is AI-referred when it came from ChatGPT, Gemini, Claude, Perplexity, Copilot or a similar assistant. On 14 of the 15 sites the script started 8 to 14 days before we took the numbers. For the fetches-per-visit ratio we used fetches in the week starting 2 November and visits in the 7 days before the pull, which overlap for six days.

• What we removed.
• All Claude-User requests. We use Claude ourselves to research client sites, and our sessions look the same as anyone else's. In the last two weeks of the window Claude-User ran at about 25 times its weekly rate in ordinary weeks of late August and September (apart from two batch runs), which we believe is mostly our own work.
• Requests under assistant user agents that asked for passwords and configuration files rather than pages, on 2 of the 15 sites, under 1% of the total. Those are vulnerability scanners using a borrowed name. About a third of them were Claude-User requests, already removed; we removed the rest.

• What we couldn't see.
• We identify bots by the name they send, without checking IP addresses, so some requests that claim to be an assistant are not.
• In our window, the only Perplexity-User requests we could see page by page were scanner probes, so Perplexity's real person fetches are missing from these figures, not zero.
• People who click through from an assistant's app often arrive with no referrer and count as direct visits. Our AI visit figures are a floor and the fetches-per-visit ratio is a ceiling.
• Fifteen small sites are a small sample. These numbers describe these sites, not the web.

Every figure comes from one saved pull of cite-met's reporting tools on 8 November 2026, between 01:28 and 01:33 UTC, with the rules above applied by one script. We plan to repeat the count once the visitor script has a full month on every site.

If you're weighing what it costs to act on this, our AI search cost guide lays out published prices, and our free audit has a person read your homepage with these numbers in mind.