We analyzed 10 million AI visits to B2B websites. The optimized pages got 26x more.

For 13 months we've been logging every AI agent visit to our customers' websites at the edge: which agent, which page, when. The log crossed 10 million visits this month. We ran one question against all of it: what decides which pages AI agents actually read?

The answer, across 27 B2B sites with enough traffic to measure: pages optimized for AI agents get 26 times more AI visits than non-optimized pages on the same domain. And the attention isn't just unequal, it's winner-take-most. On the median site, optimized pages are a minority of URLs but capture roughly 9 in 10 of all AI visits. On one cybersecurity customer's site, 43% of URLs took 98% of the AI traffic.

We know how that sounds coming from the company that sells the optimization. We didn't trust the number either, so we attacked it. Most of this post is about the test it survived.

First, the obvious objection

A 26x gap between optimized and non-optimized pages proves nothing by itself. Companies optimize their important pages first, and important pages attract more visits anyway. Maybe optimization causes AI traffic, or maybe we just optimize winners. Any vendor stat that ignores this deserves your contempt.

Our data has one property that let us settle it: every optimized page carries a timestamp. So we could stop comparing good pages against bad pages and instead compare the same pages against themselves.

The test: same pages, before and after

The design: take every batch of pages optimized on a given date, measure its AI visits in the 30 days before and the 30 days after, and compare that change against untouched pages on the same site over the same calendar windows. Same pages, so popularity is controlled. Same site and dates, so seasonality and the overall rise of AI traffic are controlled.

We ran it on 208 optimization batches across 22 customer sites. The result:

Optimized pages tripled their AI visits within 30 days (3.1x). Untouched pages on the same sites grew 1.7x over the same windows. Net of each site's own baseline, that's an 82% lift attributable to the optimization itself, in the first month. 72% of batches beat their site's baseline. The extreme end: one batch of 12 pages went from 70 AI visits in the month before to 7,900 in the month after, while the rest of that site's pages declined.

The batches that showed nothing were just as informative. One customer's biggest batch was flat, and the reason is instructive: those pages were already getting six AI visits per page per day before optimization. They optimized their already-famous pages. The effect concentrates where pages start near zero, which is where most of every B2B site sits.

So the 26x is the observed gap, and the before/after test says a real causal lift sits inside it: visible within 30 days, compounding after, largest on the pages agents currently ignore.

What "optimized" actually means

No mystery box: an optimized page is an agent-legible version of the page. Direct answers instead of marketing build-up. The specifics agents get asked about (pricing, integrations, limits, compliance) stated instead of gestured at. Structure a machine can parse without executing your JavaScript. Same facts, arranged for a reader that doesn't scroll, doesn't infer, and quotes exactly what it finds.

Humans and search crawlers keep getting your original pages, byte for byte. Only identified AI agents get the legible version. The 10 million visits in our log are those agents showing up to read.

Who's doing the visiting

From the last 90 days alone, 3.9 million AI agent visits across 93 sites and 141,000 URLs. ChatGPT dominates, and not with a background crawler: ChatGPT-User, the fetcher OpenAI sends when a live person's question needs your page right now, accounts for 2.7 million of the 3.9 million. Claude and Perplexity agents make up most of the rest. A large share of AI visits to your site happen because a buyer just asked about your category, and an agent went to read you mid-answer.

By the time that buyer lands on your site, if they land at all, the comparison is over. Which pages the agent could read decided it.

Visibility is not conversion

One finding cuts against us, and we'd rather you hear it from us: in one customer's 90-day window, non-optimized pages converted their human visitors at more than twice the rate of optimized ones.

Your website now has two audiences: buyers, and the agents advising them. A page can work beautifully for one and be nearly invisible to the other. Optimization for the agent gets you selected; the human experience gets you chosen. Across our dataset, AI-referred visitors convert at 4.4x the rate of organic visitors, but that's mostly a property of who they are: pre-educated, mid-evaluation, arriving with the comparison already done. If your conversion path is weak, AI visibility fills the top of a leaky funnel. Anyone selling you visibility without saying this is selling you half a system.

Why you've never seen this traffic

None of the 10 million visits in our log fired an analytics tag. An agent fetching your page isn't a browser session; GA4 has no row for it. And when the human behind the question clicks through from the AI's answer, the referrer usually dies on the way, so the session that does get a row lands in Direct. Part of this traffic is invisible to your analytics, and the rest is filed under the wrong name.

You can see the second half in your own data today: pull your top 100 Direct landing pages and flag the ones nobody would type or bookmark. A no-referrer session landing on /careers/careers-post/4913293101 followed a link that lost its referrer. That count is your mislabeled bucket. What GA4 can't tell you is how much of it is AI, because the confirmation lives on the other side of the tag: in which pages agents are actually fetching.

What to do this week

1. Ask an AI the question your buyer asks. Put your category's mid-evaluation question to ChatGPT and watch which pages it reads and cites, if any of them are yours. That's your real front door now.
2. Run the plausibility check on your Direct traffic. Top 100 Direct landing pages, flag the untypeable ones. That's the size of your mislabeled bucket, from data you already own.
3. Look at your actual agent traffic. Our agent traffic analytics are free, forever: which agents visit, which pages they read, what they're told. Every number in this post came from that log. Whether the 26x pattern is happening on your site is checkable in an afternoon.

AI agents are already deciding which parts of your website get seen. On the sites we measure, nine tenths of their attention goes to the pages built to be read by them. The rest fight for scraps, invisibly.

Data: Salespeak edge-level AI agent visit logs, 10 million+ visits across customer websites, August 2025 to September 2026. Cross-site comparison covers 27 customer sites with optimization live; the before/after analysis covers 208 optimization batches on 22 sites, each measured 30 days pre and post against same-site controls. Salespeak-owned properties were excluded from all portfolio statistics.