Founders are being asked to fund AI visibility before the traffic math is clear.
That is the problem.
Some vendors point to growth inside the AI slice, then imply the slice itself matters at scale. The denominator is the missing piece.
Cloudflare Radar gives a cleaner view. For May 2026, Google accounted for 87.63% of observed search referrals, while ChatGPT, Gemini, Claude, and Perplexity combined accounted for 0.29% of referrals to Cloudflare-protected origins, according to a widely cited Cloudflare Radar analysis published by Technology Checker, which reproduces the underlying figures and method from Radar, source.
That is the number most buyers need before they sign an AEO contract.
The right denominator
One common claim is that ChatGPT dominates AI referrals. In the AI slice, that can be true. In the full referral market, it means something else.
The distinction matters. ChatGPT may take roughly three quarters of AI referrals, but AI referrals themselves are still a very small share of total search referrals. Seventy-five percent of a small number is still a small number.
For a seed-stage team, this is not a philosophical debate. It is a budget question.
If a channel sends 0.29% of observed referrals, then the first duty is not to optimize it. The first duty is to ask whether it deserves a line item at all.
What Radar measures
Cloudflare Radar measures referrals landing on Cloudflare-protected origins. It is not a census of all web traffic, and it is not a count of search queries.
That limitation is not a flaw, it is the point. The sample is large, public, and repeatable. Anyone can inspect the method and rerun the query.
There are two caveats you should not skip.
First, the sample skews toward Western e-commerce, media, and SaaS traffic on Cloudflare-protected sites. That means it is useful, but not universal.
Second, a February 2026 TikTok reclassification makes the Google and TikTok year-over-year lines non-comparable in some summaries. If you ignore that, you are not doing analysis.
Credibility here depends on saying what the data cannot prove.
Per-model share
In the May 2026 Cloudflare Radar data, the AI referral mix was small and concentrated.
| Model | Share of referrals |
|---|---|
| ChatGPT | 0.23% |
| Gemini | 0.028% |
| Claude.ai | 0.016% |
| Perplexity.ai | 0.013% |
Combined, the major AI chatbots reached 0.29%.
Only ChatGPT clears Radar’s quarterly reporting threshold. The rest sit in the rounding-error zone.
That matters because many AI search narratives are built from the inside of the AI slice, not from the outside looking in.
74.78% of a 0.29% slice is not a market transformation, it is a distribution detail.
The growth story
The honest version is not that AI referrals are irrelevant. It is that they are growing from a low base.
ChatGPT rose from 0.19% to 0.28% year over year in the Radar dataset, a 47% gain. That is the fastest-growing line in the table.
And it is still small.
Both statements can be true. Good analysis holds them together instead of choosing the one that supports a thesis.
The crawl cost
Referral share is only half the equation. The other half is extraction cost.
Cloudflare’s crawl-to-refer ratios, also surfaced in the Radar data, show a wide gap between bot activity and downstream referral value.
| Bot | Crawl-to-refer ratio |
|---|---|
| Anthropic | 11,122:1 |
| OpenAI | 857:1 |
| Perplexity | 190:1 |
| Microsoft | 34:1 |
| 5:1 | |
| DuckDuckGo | 1.5:1 |
Quarter over quarter, those ratios improved sharply. Anthropic moved from 91,702:1 in Q2 2025 to 5,143:1 in Q2 2026, based on the same Radar reporting discussed in the Technology Checker analysis, source.
That is a real trend. It is not yet a reason to overfund the channel.
For a small site, the implication is simple. You may be paying in content effort, technical work, and brand clarity to feed systems that still return very little traffic.
If you do not have a content engine that already wins Google, you are unlikely to win the AI layer by adding a separate tool.
Where it matters
The broad market share picture can hide useful exceptions.
Cloudflare Radar shows ChatGPT referral share is materially higher in finance, where it reaches 0.60%. That is more than double its all-web share. Computer and electronics reaches 0.29%, and shopping reaches 0.14%.
So the right answer is not, “ignore AI search.” The right answer is, “check whether your vertical is one of the few where the channel already punches above its weight.”
If you sell into finance, your math is different. If you sell productivity software to a narrow technical audience, it probably is not.
- Finance, strongest current exception
- Computer and electronics, modest lift
- Shopping, some lift, still small
Run it yourself
The best part of the Cloudflare method is that it is public.
You do not need to trust a vendor summary. You can inspect the data directly through Cloudflare Radar’s public API endpoints, including:
- Radar documentation
bots/crawlers/summary/refererbots/crawlers/summary/crawl_refer_ratioai/bots/summary/crawl_purpose
A simple query pattern looks like this, with the exact path depending on the endpoint:
GET https://api.cloudflare.com/client/v4/radar/...
The point is not the syntax. The point is that the evidence is reusable.
That is what makes this useful to founders. If a channel claim cannot survive a free rerun, it probably should not survive a budget review.
What to fund
For an early-stage team, the decision is usually not between SEO and AEO. It is between building durable content systems and buying a tool that promises shortcut visibility.
Fund the work that earns citation in both Google and answer engines, because those systems still reward clear, specific, source-backed content.
Do not buy a separate AEO program just because the label is new.
Do not hire for “answer engine optimization” if you have eight organic keywords and no repeatable content process.
Do not treat AI visibility as a standalone channel before you have a page worth citing.
In practical terms, the near-term budget call is this:
- invest in one or two genuinely citable pages
- build a clean source-and-method habit
- track Google referrals and branded demand first
- test AI visibility as a byproduct, not a separate religion
That is especially true for low-domain-authority companies. If the site is not already trusted enough to rank, it is unlikely to be trusted enough to be quoted.
What would change
Good strategy names its falsification condition.
We would change our view if one of two things became true.
- combined AI referral share crossed 2% of observed referrals
- OpenAI’s crawl-to-refer ratio fell below 100:1 and stayed there
Those thresholds would not mean AI search has replaced Google. They would mean it had become large enough to deserve a distinct budget conversation.
Takeaway
The cleanest current number is this, in May 2026, Cloudflare Radar shows Google at 87.63% of observed search referrals and major AI chatbots combined at 0.29%.
For most early-stage SaaS teams, that does not justify a separate AEO spend. It justifies better content, clearer sourcing, and a watchful test plan.
If you are going to fund anything this quarter, fund the page that deserves to be cited, then measure whether AI systems pick it up.
That is how small teams should approach this market, with evidence first, labels second.