Visiting a company website, reading it, and forming a judgement used to be work done by people. The same work is now done by AI. This report examines what a website can and cannot be expected to deliver after that shift, using the published measurements.
10 September 2026
Being surfaced and described by AI search does happen. It rarely produces the revenue or the enquiries a company expected. Three reasons account for this: the number of people who actually arrive is an order of magnitude smaller than assumed, half of those who arrive leave immediately, and most of the remaining effect is never measured because it surfaces through a different channel.
None of this makes a website unnecessary. The reader has changed. The role — being the place where a company is judged — has not.
The number of steps is unchanged. What changed is who performs the middle two. Sites are still being read. They are no longer being read by people.
In Contentsquare's 2026 benchmark, drawn from billions of sessions, visits arriving from generative AI platforms accounted for 0.2% of all visits. Year on year that figure grew 632%. The base it grew from is the number above.
On a site receiving 100,000 visits a month, AI referrals amount to roughly 200. The growth rate exceeds six-fold a year. The absolute volume starts here.
In the same benchmark, 53.6% of AI-referred visits bounced — behind only paid and organic social, and advertising. More than half of those who arrived left without going further than the landing page.
The channel is also concentrated in a single provider. Previsible tracked 6.77 million LLM-referred sessions across 166 properties between November 2024 and May 2026, and found 92.4% of them came from ChatGPT. A policy change at one company moves most of the channel.
Volume is small; quality has inverted. Adobe Analytics, measuring across more than a trillion visits to US retail sites, found AI-referred traffic converting 42% better than non-AI traffic in March 2026, rising to 54% by May. Twelve months earlier, the same measurement had AI-referred visitors converting at roughly half the rate of everyone else.
| Measure, against non-AI traffic | Difference | Source |
|---|---|---|
| Conversion rate (May 2026) | +54% | Adobe Analytics |
| Revenue per visit | +37% | Adobe Analytics |
| Time on site | +48% | Adobe Analytics |
| Pages per visit | +13% | Adobe Analytics |
| Share of all visits | 0.2% | Contentsquare |
The two findings do not contradict each other. People arrive after the AI has already done the comparison and the narrowing, so only those who have finished the research stage make the trip. The quality is high because the volume has been filtered.
A 54% lift on 0.2% of visits leaves very little behind in total revenue.
Much of the effect is also never recorded as AI at all. Similarweb found that brands recommended by ChatGPT in its answers received 2.5 times more visits within seven days than comparable brands that were not recommended. Those visits arrive as branded search and are credited to search, not to the recommendation. The study establishes correlation; whether the recommendation caused the visit, or whether already-known brands are recommended more often, cannot be separated from the data.
The standard account in this field is that results take three to six months or longer. It rests on three dependencies: waiting for a model's training data to update, waiting for search indexing, and accumulating mentions on third-party sites.
The ground under that account is moving. Analysis of the relationship between domain authority and AI citation puts the correlation at r=0.18, down from r=0.43 before 2024 — squared, an explanation of roughly 3% of the variation in whether a brand is cited at all. Ahrefs, examining 75,000 brands, found backlinks correlating at 0.218 with AI visibility, well below the 0.664 of unlinked brand mentions.
Other work points the opposite way. SE Ranking, analysing 2.3 million pages, identifies domain authority as the single strongest predictor of AI citation. The findings are split.
These studies are not contradicting each other. They measured at different times. The assumptions are moving fast enough that a correlation coefficient halved in two years.
The source libraries barely overlap. A comparison of citations across Google's AI Overviews and several LLMs, covering 8,090 keywords in 25 verticals, found that only 7.2% of the 22,410 cited domains appeared in both systems. Success in one system does not carry to another.
The platforms themselves are moving. On Similarweb's figures, ChatGPT's share of generative AI web traffic fell from roughly 76% in June 2025 to about 53% by May 2026, while Claude rose from around 2% to close to 9% in the same period. Designing for one assistant now misses about half the audience.
It is worth returning to the original premise. Companies did not build websites solely to acquire customers.
In surveys classifying the primary purpose of small-business websites in the United States, roughly 24% exist to present products and services and roughly 17% to sell online, while about 14% were built primarily for search visibility. Ownership itself is high: NFIB's 2025 survey found 75% of firms with one to nine employees and 97% of firms with fifty or more had a site. In Wix's 2026 survey of small-business owners, more than 75% described their website as essential or important to growth.
The reasons for not having one are not about failed acquisition either. Around 27% consider a website irrelevant to their industry, about 35% consider their business too small to need one, and roughly 26% cite cost.
For most companies the website was never a machine for producing customers. It was the place where the business could be verified when somebody went looking.
That role has not disappeared. It has moved. People used to visit and verify for themselves. An AI now visits, reads, summarises, and hands the result to the person.
Judged as an acquisition channel, AI search visibility disappoints. Arrivals are 0.2% of visits, half of them bounce, the effect surfaces as branded search, and the rankings move week to week. Counting appearances will not produce a number that justifies the work.
What can be judged is the content of the account. When an AI describes the company, is the description accurate, current, and free of confusion with competitors? Unlike appearance counts, whose measurement is still unsettled, this can simply be read and assessed as right or wrong.
And the cost of being described incorrectly is now higher than the cost of not being described at all — because people are deciding from the account rather than from the site.