Quick Answer
Cloudflare Radar, Cloudflare's free public view of Internet traffic trends, is the cheapest way to size AI search traffic before you fund it. In the 28 days to 19 September 2026 it put Google at 90.15 percent of tracked search and AI referrals, and ChatGPT, Claude.ai, Gemini and Perplexity combined at 0.73 percent. That slice nearly tripled in a year and is still under one percent.
Founders are being asked to fund AI visibility before anyone shows them the denominator.
A vendor points at growth inside the AI slice and lets you assume the slice itself is large. Both halves of that move survive because the figure that would settle it rarely reaches the slide.
It is not a secret figure. Cloudflare publishes it, refreshes it daily, and lets anyone query the same public endpoints we did.
The first version of this post took its numbers from a third-party blog instead of the source. Going back to the source broke several of them.
What follows is the current reading, what the data cannot prove, and the two thresholds that would change the budget call.
Cloudflare Radar Puts AI Referrals Under One Percent
The four major AI chatbots together sent 0.73 percent of the referrals Cloudflare Radar tracks in the 28 days to 19 September 2026. Google sent 90.15 percent of the same set, and Bing, TikTok, Yandex and DuckDuckGo take most of what is left.
So the first question is not how to win the AI channel. It is whether a channel that size has earned a line item yet.
Growth inside that slice is real, and it is fast. The size of the slice is the part the deck skips.
1. Claude.ai Now Leads the AI Referral Slice
Claude.ai is the largest single AI referrer in the set Cloudflare Radar tracks, at 0.352 percent against ChatGPT's 0.340 percent. Gemini sits at 0.029 percent and Perplexity at 0.0067 percent.
A year earlier the same query returned chatgpt.com at 0.234 percent and claude.ai at 0.0034 percent. Claude.ai grew roughly a hundredfold over those twelve months while ChatGPT grew by about half.
The common mistake is to quote a per-model ranking from a 2025 report. That ranking has reversed, and a recommendation resting on it is now aimed at the wrong model.
You can tell you have the current picture when the table you are looking at names its window. Any figure presented without one is a figure somebody stopped checking.
2. The Slice Nearly Tripled and Is Still Tiny
Combined AI referral share went from 0.263 percent a year ago to 0.728 percent now, on matched 28-day windows in Cloudflare Radar. That is a 2.8 times increase in twelve months.
In percentage points it is a move of less than half a point. Both readings are true at once, and a budget decision needs both of them.
Growth rates are what get quoted, because growth rates are flattering at this size. Ask for the share alongside the rate, every time, and watch what happens to the room.
3. Google Still Takes Nine Referrals in Ten
Google's share of the set Cloudflare Radar tracks rose from 81.23 percent to 90.15 percent over the same year. That is not Google taking ground from AI.
TikTok fell from 10.23 percent to 1.79 percent inside the same fixed set, and these shares are relative, so everyone else rises when one neighbour drops. Reading Google's rise as evidence about AI is the error the arithmetic invites.
The check is cheap. Run the referral query for your own window and read the whole table, not the two rows that support the argument you walked in with.
What Cloudflare Radar Actually Counts and What It Does Not
Cloudflare Radar counts HTTP requests crossing Cloudflare's own network, which is large, public and repeatable, and is not the whole web. Referrals are identified by the Referer header, so a platform that does not send one is invisible to the count.
That matters more than it sounds. Cloudflare says traffic referred by Claude's native app carries no Referer header, and believes the same holds for other native apps.
So the referral shares are a floor for AI and the crawl ratios are a ceiling. In Cloudflare's own words, the calculations "may overstate the respective ratios, but it is unclear by how much".
The shares are also relative to the platforms Cloudflare Radar tracks, not to all Internet traffic. Cloudflare states this plainly, and it is the line most often dropped when the dataset is recapped elsewhere.
One more exclusion is worth knowing. Referral traffic from Google's own network, AS15169, is left out, because prefetching driven by speculation rules is not a person reading anything.
On the crawler side, user agents are identified by name, so Googlebot, GPTBot and their search-specific siblings are counted separately rather than lumped together. That separation is what makes purpose-level reporting possible at all.
Three Cloudflare Radar endpoints carry every figure in this post, and all three are documented and free to query.
/radar/bots/crawlers/summary/REFERERfor referral share by platform/radar/bots/crawlers/summary/CRAWL_REFER_RATIOfor the crawl-to-refer table/radar/ai/bots/summary/CRAWL_PURPOSEfor what the crawling was declared for
Each takes a date range and an optional vertical filter. If a channel claim cannot survive a free rerun of one of those queries, it should not survive a budget review either.
The Crawl-to-Refer Ratios Collapsed in One Year
In Cloudflare Radar, Anthropic's crawl-to-refer ratio fell from 35,371:1 to 629:1 in twelve months, an improvement of fifty-six times. Perplexity went the other way, from 164:1 up to 1,231:1, and is now the operator taking the most and returning the least.
Cloudflare defines the ratio as HTML page requests from a platform's crawlers divided by HTML page requests whose Referer header names that platform, normalized to a single referral. It is extraction measured against return.
"Legacy search index crawlers would scan your content a couple of times, or less, for each visitor sent."
That is David Belson and Sam Rhea of Cloudflare, writing in July 2025. Google's ratio in the current window is 4.7:1, which is close to what they describe, and no AI operator is anywhere near it.
| Operator | Aug-Sept 2025 | Aug-Sept 2026 |
|---|---|---|
| Perplexity | 164:1 | 1,231:1 |
| Anthropic | 35,371:1 | 629:1 |
| OpenAI | 813:1 | 444:1 |
| Microsoft | 38.5:1 | 38.5:1 |
| 4.6:1 | 4.7:1 | |
| DuckDuckGo | 0.5:1 | 2.5:1 |
The figure still circulating in vendor decks is Anthropic at eleven thousand to one or worse. It has not been true for months, and the operator now taking the most per referral is Perplexity.
None of this makes the channel large. It does mean the extraction argument against AI crawlers is weaker than it was, and an argument that has moved by a factor of fifty-six is a poor foundation for a two-year plan.
Where the AI Referral Share Breaks Its Average
The overall average is the wrong number for some sites, and Cloudflare Radar lets you check which by filtering the same referral query to a vertical. One vertical breaks it badly.
In Computer and Electronics, claude.ai accounts for 4.41 percent of referrals, twelve times its share across the tracked set. Anthropic's crawl-to-refer ratio inside that vertical is 19.6:1, against 629:1 overall.
If you sell developer tools or infrastructure, that is your vertical. The aggregate understates your exposure by an order of magnitude, and this is the one case where a founder should look before concluding the channel is too small to matter.
- Computer and Electronics: claude.ai at 4.41 percent of referrals, the one genuine exception in the current window
- Finance: ChatGPT at 0.32 percent, marginally below its 0.34 percent across the tracked set
- Shopping and general merchandise: ChatGPT at 0.27 percent, against Google at 90.93 percent in the same vertical
Finance is worth naming because it used to be the exception every deck quoted. In the current window it sits marginally below average, which is what happens to an exception nobody re-checks.
A B2B demand generation agency still building a plan on the finance figure is building on a number that expired. Re-run the vertical filter before you accept any claim about your own category.
What Separate AEO Spend Buys Compared With Content Work
A separate answer engine program buys you a vocabulary, a dashboard and a retainer. The work that earns a citation is the same work that earns a ranking, and the largest AI surface says so in writing.
Google Search Central is direct about it: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." Cloudflare Radar supplies the size of the prize, and Google supplies the method, and neither of them describes a second discipline.
| What you are buying | A separate AEO program | Content work that earns citation |
|---|---|---|
| Unit of work | A visibility score per prompt | A page specific enough to quote |
| What changes on the site | Usually nothing | The page itself |
| Value when traffic is near zero | Reports anyway | Compounds anyway |
| Google's stated position | No special optimizations needed | The same practices as ever |
Better Marketing runs its AI SEO service inside the content work, not beside it. Content marketing for tech companies has not changed shape, and neither has the job a B2B demand generation agency is hired to do: publish something specific enough that quoting it is easier than paraphrasing it.
The counter-case is real, and it is about the tactic rather than the channel. If you already rank for nothing, a citation program has nothing to work with, because a site a crawler does not trust enough to rank is not a site a model trusts enough to quote.
Two things are worth reading before you sign anything. Start with what an AI visibility tool can and cannot tell you, then how AEO and SEO differ for a startup.
What Would Change Our View on AI Search
Two thresholds would change our view on AI search, and Cloudflare Radar says neither has been crossed. Good strategy names the evidence that would overturn it, and these two are unchanged from the first version of this post.
- Combined AI referral share crosses 2 percent of the tracked set. It stands at 0.73 percent, up from 0.26 percent a year ago.
- OpenAI's crawl-to-refer ratio falls below 100:1 and stays there. It stands at 444:1, down from 813:1 a year ago.
- One indicator has already moved a long way: crawling declared for search was 1.56 percent of AI crawler requests a year ago and is 9.69 percent now.
The third line is the one to watch. Crawling declared for search is crawling that can produce a referral, and it has multiplied by six while the share declared for training moved only a few points.
Crossing either threshold would not mean AI search had replaced Google. It would mean the channel had grown large enough to argue about on its own, which today it has not.
Fund the Page That Deserves a Citation
The current reading from Cloudflare Radar does not justify a separate AEO line for most seed-stage teams. It justifies better pages, clearer sourcing and a standing test.
Publish one or two pages specific enough to quote. Keep the habit of showing your method. Track Google referrals and branded demand first, and treat AI citation as the byproduct you measure rather than the channel you fund.
The order we use at Better Marketing puts the AI SEO service after the page and never before it, because content marketing for tech companies still starts with something worth citing.