The AI web is one company

HostingBrain analyst brief · August 2026 · data snapshot 2026-08-16 · aggregate-level, no named providers

Describe a business, get a website. That pitch has been the loudest thing in web software for about a year — and across Europe's business websites, about six in seven of the sites it produced come from one company.

The category is real, with real customers. It is also, on the European web we can measure, very small. AI builders leave markers on the pages they generate, the same way hosted builders and content systems do, so their footprint is countable from outside. Here is the count.

8,928business websites running an AI website builder, across thirteen European markets

43.0xas many carry a conventional hosted builder

85.5%of the AI-built sites are a single product — 92.6% on the wider answered base

The gap is two orders of magnitude

Read the widest way instead, over every active domain whose pages we read rather than only those showing a homepage or shop, and the shares change but the shape does not: 0.42% AI builders against 7.281% hosted builders, 17.3 times.

Two things are true at once: this category is growing quickly, and it is currently a rounding error on the installed European web.

It is not an artefact of when we looked

The obvious objection is that a new category looks small because measurement lags it. So take only the sites whose pages we re-read since 24 July — 1,680,015 of the 5,450,090, 30.8% of the book, every one of them observed recently. AI builders are 3,690 of that cohort: 0.22%. That is the higher of the two readings, and both of them sit near two tenths of one percent.

The fresh cohort moves the comparison further than it moves the level: conventional builders read 5.211% there rather than 7.042%, so the gap narrows to 23.7 times. That cohort is a re-read queue rather than a random sample of the web, which is why the level is the reading to take from it and the ratio is not. Looking only at what we read most recently makes this category bigger. It does not make it big.

The category is mostly one product

Eight AI builders are detected on this web at all. One of them, Lovable, accounts for 7,636 of the 8,928 detections — 85.5%: about six in seven. On the wider reading, every page we read rather than only company sites and shops, the same product holds 92.6%. The next, Base44, has 838 at the narrower scope, or 9.4%. The six after that share the remainder.

That figure needs its observation window stated, because the eight markers did not enter our reading at the same time. Seven of the eight were added on 24 July; the eighth — the one in front — has been read for far longer. So the seven are counted only on pages we have read since, while the leader is counted on the whole book. Restrict every product to the same window, the freshly re-read cohort above, and the leader holds 65.0% instead of 85.5%.

Both readings say the same thing in the end. Whichever window you give it, "AI-built websites" in Europe today is one product's customer base, and a long tail that has not happened yet.

What AI builders are used for

Mostly the front door. Take every page we read, not only those showing a homepage or a shop, and 82,859 carry an AI builder. 71,994 of them — 86.9% — are simple brochure pages: a company's face on the web, not the machinery behind it. Only 4.6% are shopfronts.

Conventional builders put up plenty of brochure pages too, 68.9% of the 1,435,466 pages carrying one. But they sell 2.5 times as often: 11.6% shops against 4.6%. So far the AI category has taken the simpler half of the job — the landing pages and one-pagers a business puts up to exist online, rather than the site it runs the business on.

Where the markets differ

The spread across markets is wider than the headline suggests. Sweden is the highest, at 0.352% — 643 domains of 182,495. Germany, the largest book we measure at 1,359,099 domains, reads 0.122%. Italy is the lowest at 0.095%.

The United Kingdom is the market worth pausing on. AI builders read 0.254% there, well above the thirteen-market average — and conventional hosted builders read 18.024%, 155,744 domains and the highest share of any market here. The market most receptive to a builder is the market where builders already won, which is an argument about who the customer is rather than about which tool is new.

Who keeps the customer when AI builds the site?

A website is not one product. It is four, sold by different people: something builds the pages, something serves them, someone holds the domain and the DNS behind it, and someone runs the mailbox. The AI builder took the first one. Here is what happened to the other three.

Take the 76,736 domains on the widest reading that carry the leading AI builder's marker, and follow each layer. 80.5% — 61,751 of them — are served from a content-delivery edge rather than from a host's own machines, which is the AI builder's own infrastructure doing what it promises. That is the layer that moved.

The domain did not move with it. 58,460 of the same sites — 76.2% — keep their nameservers with a hosting company or a registrar: an operator in this industry, billing this customer, every year. Only 9.3% sit on a standalone DNS or edge platform. And the striking part is the direction: across the thirteen markets as a whole, 60.6% of active domains are held that way, so the AI-built sites are 15.6 points more tied to the traditional industry than the web they are being built alongside, not less. 1,288 distinct operators hold the control plane of this one product's customer base.

The AI builder took the most visible layer of the stack. It did not take the billing relationship, and on this evidence it did not try.

The mailbox never moved at all. 14.4% of these sites — 11,022 — carry no mail record whatsoever, which is its own finding: a seventh of them are a published page and nothing else yet, the shape of a prototype rather than a business. Of the 65,714 that do run mail, 75.8% run it with a hosting company or registrar and 23.9% with a productivity suite. Read against the whole set instead, that is 65.0% with the industry and 20.5% on suites. No AI builder sells mail, and none of these customers went looking for one that did.

There is a second door, and it runs the same way. A customer can bring a domain they already own — then their registrar simply keeps them. Or they can buy one inside the product, in which case the AI builder hands that domain business to a hosting company under a commercial partnership. Either route ends at an incumbent's invoice. That is also why one large European operator's block in this data cannot be read as customer loyalty, and we do not read it as any: see the method card.

The traffic runs the other way too, in smaller numbers. 7,357 of these sites — 9.6% — are served not from an edge and not from a cloud platform but from a named hosting company's own network: sites generated by the AI tool and then deployed on ordinary hosting. That is an AI-built workload arriving as a hosting customer.

What this changes

For a hosting operator: the threat is not the installed base yet. A category at 0.164% is not taking your renewals this year. The number to watch is the new-registration cohort, not the book.
For an investor: the category is one company's revenue. A thesis about "AI site generation" priced off category growth is, on this evidence, a thesis about one product's execution and one product's churn.
Ask for the denominator before the growth rate. Percentage growth on a base this small is easy, and every number here is stated with the base it divides.
The AI builder is an acquisition channel, not only a competitor. Three quarters of these customers already pay someone in this industry for the domain, and two thirds pay someone for mail. A business that arrives through an AI builder arrives with a control-plane relationship attached — bought from an incumbent, or brought to one. The question for an operator is not how to stop the category, but what the second product is once the site itself has stopped being the thing they sell.

What we cannot see. Detection depends on a marker in the page, so a site generated by an AI tool and then exported, self-hosted or stripped of its markers counts as absent here. Every AI figure in this brief is therefore a floor, and the true share is higher by an amount we cannot bound. The population is the other limit: this is the web of active domains that answered us with a page, not a traffic-weighted sample of the sites people actually visit. A small tool with a large audience would read small here.

How the population is defined

The book is active domains in thirteen European markets whose pages we read. The headline reading keeps only pages classified as a homepage or a shop, which is the closest thing we have to "a real site" rather than a login screen, a parked page or an error. The widest reading keeps every page we read. Each share divides by the base for its own reading, and the two are never mixed inside one sentence.

Method & caveats. Categories and products are detected from markers in the page we read — presence, not billing, so a product with no marker is undercounted and one whose marker survives an abandoned site is overcounted. Page shape comes from the same read: shop wording makes a page a shop, a form together with a call to action makes it a homepage, and login screens, parked pages and error pages are taken out first. Everything left is the brochure shape, which is wide enough to hold a company site whose only contact route is an email address in the footer. The thirteen markets are the national domains we publish market pages for. Our per-market answers publish a cell only at 50 domains or more, and a market below it returns no cell at all rather than a zero. On an earlier reading that made the market-by-market sum come up short of the aggregate. On this one it does not: the smallest of the thirteen sits exactly at that floor, with 50 domains, so it is published and adding the thirteen cells up returns the same 8,928 this page states. The freshly-read cohort is every domain in the book whose most recent page read came after 24 July, which is the date the seven newer markers were added. Aggregate figures only: this brief names products and publicly announced partnerships, never our per-provider measurements — that detail is the paid product. Snapshot 2026-08-16.

Four layers, four instruments. The layer section reads each layer with a different signal, and they are not interchangeable. What BUILT the page is a marker in the page itself. Who HOLDS the domain is the nameserver the domain delegates to, folded to the company that owns the operator — not the address the site resolves to, and not a reverse lookup. Where it is SERVED is the network that owns the IP address the name resolves to. Who runs MAIL is the mail record. A site can be built by one company, served by a second, delegated to a third and mailed by a fourth, which is the whole point of the section.

What the layer section cannot tell you. Three limits, in the order they bite. The edge hides the origin. Where a site is served from a content-delivery network, the machine actually holding the files is not observable from outside — the finding is that these sites run on the AI builder's own delivery rails, not a claim about what sits behind them. One operator's block is not customer loyalty, and is excluded from every reading that would treat it as such. The leading AI builder sells domains inside its product through a commercial partnership with IONOS — their own published documentation, not a measurement of ours — so that operator's share here mixes its ordinary customers with purchases the AI builder routed to it. The test is in the data rather than asserted: an operator's organic book concentrates in its home market, and this one does not. Germany is 68.6% of that operator's whole book across these markets but only 43.2% of its AI-built block — a ratio of 0.63, against 0.84 for the comparable operator with no such partnership. Every per-operator count is a floor of a different depth. We have read the pages of 86.4% of the active domains in these markets — a deeper reading than the one an earlier version of this brief was written on — but not evenly. Among the ten operators holding most of these domains, the best-read is at 95.3% of its own book and the least-read at 57.7%, so the raw ranking still flatters the well-covered operator, and the shallowest-read operator's count is the one most likely to be understated. What has changed is where the partner operator sits in that range: its own book is now read at 90.0%, near the top rather than near the bottom, so its block is no longer among the shallowest floors here and cannot be explained away as one. Each count is published beside the share of that operator's own book we have read, and we publish no coverage-adjusted estimate. For the same reason this brief publishes no per-operator retention rate: the share of an operator's AI-building customers still served on its own network is a measurement we can take, and on this release we cannot take it to the same depth for every operator, so it is withheld rather than ranked.

Reproduce this — or ask it yourself

Every figure here is queryable through the HostingBrain connector. In Claude or any MCP-compatible assistant, this is the whole brief in one prompt:

Prompt · paste into an MCP client with HostingBrain connected

“Using HostingBrain, compare AI website builders with conventional hosted builders and WordPress across the European markets you cover — the counts, the denominators and the per-product breakdown inside the AI-builder category. Then show me the same thing for Germany, Sweden and the United Kingdom.”

Resolves to technology_adoption — the category totals are free, the per-product breakdown is Pro. definitions(term='technology_adoption') explains the two page-scope readings, on the free tier.

Reproduce this analysis: the technology_adoption tool returns every count, denominator and per-product row above, market by market — with the free tier covering the category totals, market structure and the definitions.

Sizing a category before you price it? HostingBrain answers questions like this — with the date, denominator and caveats attached — inside Claude and any MCP-compatible assistant.

Get free access