Your pageviews are flat, and yet the door people arrive through has quietly changed.
Fewer inbound conversations start with “I googled you.” Something about the mix of how people arrive has shifted, and nobody can point at a single cause.
I keep hearing some version of this lately. “Organic feels off.” “By how much?” “…hard to say, exactly.”
But there might be a very literal explanation, and it’s running on your server right now: AI companies are reading your site at a volume nothing in your analytics is built to show you.
Cloudflare sits in front of enough of the web to just count this. So let’s look at what they counted.
What “Crawl-To-Refer” Actually Measures
Cloudflare published a number they call the crawl-to-refer ratio (their data).
It’s about as simple as a metric gets. Take how many pages of your site a given AI company crawled, divide by how many visitors that company’s product actually sent you. A bigger number means “reads a lot, returns almost nobody.”
And since Cloudflare’s network carries the real traffic, these aren’t models or projections. They’re counts, not estimates — which I think is rarer than it should be around here.
So What Happens When You Actually Count It?
Here’s where the four big platforms sat as of July 2025:
- Anthropic: 38,065 to 1 (38,000 pages read, one visitor back)
- OpenAI: 1,091 to 1
- Perplexity: 195 to 1
- Google: 5.4 to 1
Those aren’t four different numbers so much as four different orders of magnitude. Google — the old-fashioned search engine in the group — reads about five pages for every person it hands over. Anthropic reads a small library.
So “an AI read you” and “a person showed up” have become two separate events. Your content can get consumed at enormous volume and produce almost nothing that shows up as traffic in the sense anyone’s dashboard means it.
So Is This Getting Better, Or Just Less Absurd?
Better, actually. Just not better enough to change what you do about it.
In January 2025 the Anthropic ratio was 286,930 to 1. By July it was 38,065 to 1 — roughly an 86.7% improvement in half a year. The reported reason: shipping web search, and citing sources directly inside answers.
And still 38,000 to 1 after all that. Reading the whole library and walking out without borrowing a thing.
And Perplexity went the other way, from 54 to 1 up to 195 to 1 — more crawling per visitor, not less. So the platforms aren’t converging on some shared policy here. How much any of them hands back is a choice, and they’re making different ones.
The 20% Search Drop Is The Same Story From The Other End
But maybe the crawl side is a weird artifact of how bots identify themselves. Fine — let’s check from the receiving end.
A team from London Business School and UCLA looked at 2,014 households and roughly 1.2 million page views (the research).
Households that started using ChatGPT cut their traditional search volume by more than 20% within twenty weeks. The steepest drops were in question-shaped searches — the “how do I actually do this” queries. Smaller sites took the worst of it.
One category barely moved, though: searches for a specific brand by name.
Which I think is the most useful sentence in either study. What the AI takes is the vague browsing. What it doesn’t take is somebody who already knows your name and wants you specifically.
Conclusion: Add A Column For How The AI Reads You
So: AI reads your content at a scale your access logs can barely parse, sends back nearly nobody, and for the households that switched, search volume dropped more than 20% underneath all of it. Watch only the visit count, and this whole shift arrives as “hm, traffic’s been a bit soft lately.”
I think the answer is a column, not a panic. Keep counting visits — they still pay the bills. But add the thing that happens before the visit: how the models read you, how they describe you, and which sources they cite when your category comes up. The reason to bother is sitting in that branded-search finding: name recognition survived the shift, generic discovery didn’t.
Cheapest possible start: ask ChatGPT and Claude a handful of questions about your category, write down how you get introduced and which sites get cited, then do it again next month. Park that next to the referral report and give it the same weight.
Maybe I’m wrong about how long the ratios stay this lopsided — Anthropic moved 86.7% in six months, so these numbers obviously aren’t fixed. But the “before the visit” column is worth having either way.
Sources
- Cloudflare (2025), “The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals”, blog.cloudflare.com
- Lambrecht, A., Padilla, N. (London Business School) & Lam, H.T., Hollenbeck, B. (UCLA) (2026), “The Impact of LLM Adoption on Online User Behavior”, london.edu
