Sector Index, September 2026
Beautiful websites. Nearly a third of them are invisible.
We scanned 77 architecture practice websites across Surrey the way an AI assistant reads them. Thirty one returned nothing readable at all. No practice is named.
The headline
The scores are respectable. Getting scored at all was the problem.
The forty six practices we could read scored in line with every other profession we scanned. The story here is the thirty one we could not.
Architecture sells on how a practice looks, and that has pushed the profession towards image led sites built with heavy front end frameworks. Those sites look superb in a browser. Several of them hand a plain request an empty shell, because everything on the page is assembled by JavaScript after it loads, and most AI crawlers do not run JavaScript.
The invisibility problem
A portfolio nobody can read is not a portfolio
Most were dead or misconfigured domains. Four were live, well built sites that simply serve nothing without JavaScript, or refuse any visitor that is not a browser. Those four are the ones worth worrying about, because their owners have no idea.
A site that renders only in a browser passes every check a human makes. It looks right on a phone, it looks right on a laptop, and the practice has no reason to suspect anything. The first sign of trouble is an assistant describing a competitor instead.
You can test this yourself in about ten seconds. Turn JavaScript off in your browser and load your own homepage. What you see is roughly what an AI assistant sees.
The spread
Where the 46 practices landed
Nobody cleared 80 and nobody fell below 20. Half the practices we could read sit in the 60s and 70s, which is a better shape than most sectors managed. The work for these firms is narrow and specific rather than a rebuild.
Where they differ
The checks that separate one practice from another
These are the tests some practices pass and others fail. They are what actually moves a score, because everything further down this page is failed by almost everybody.
| Failing check | Practices | Worth |
|---|---|---|
| Frame protection | 40 of 46 | 5 points |
| AI usage preferences | 40 of 46 | 10 points |
| Entity linking | 32 of 46 | 10 points |
The architects run also flagged a cluster of newer agent protocols as differentiators, which is a quirk of the scoring rather than useful advice. Almost nobody in any profession has implemented those yet, and no small practice should be adding them before the basics. They are covered further down, where they belong.
What to fix first
Three changes that would move most practices
This is the one that matters here and it is not a design compromise. The page can still animate, still be image led, still look exactly as it does now. The text and the project descriptions simply need to be in the page when it arrives, rather than assembled afterwards. Your developer will know what this means.
A typed description saying this is an architecture practice, in this town, working on these building types, at these scales. It turns a gallery of images into something an assistant can actually recommend when somebody asks for a residential architect in Farnham.
Thirty two of the forty six practices have no entity linking, which is the code that tells an assistant the practice on your about page is the same one behind every project page. Without it each page reads as an unrelated fragment.
None of these requires a new website. All three are changes to the one you have.
What comes next
Not one practice is ready for AI to act
Everything above is about reading. The next question is whether an assistant acting for a client can do anything on your website. Ask about availability. Request a fee proposal. Start a conversation about a project.
- A machine readable version of each page
- Any action an assistant could take on a client’s behalf
- A published description of what those actions are
- Discovery tags saying any of it exists
This matters less for architecture than for most professions, and it would be dishonest to pretend otherwise. Nobody appoints an architect through an AI assistant. What an assistant will realistically do is shortlist, and being on the shortlist depends on everything above this section rather than on anything in it.
Method
How this was done
The list was built from public directories and local searches covering independent architecture practices across Surrey. National multidisciplinary firms were excluded, because the marketing decisions for those sit elsewhere.
Each website was checked from the outside, exactly as an AI assistant or answer engine sees it. One anonymous request per public page, no JavaScript, no logins, nothing that is not already published. The same tests were applied to every site on the same day, 15 September 2026.
Of the 77 websites we tried, only 46 returned a readable page, and those 46 are the basis of every score above. Sixteen domains did not resolve at all, which is a high number and one we would want to verify by hand before drawing conclusions from it. The rest were a mixture of expired certificates, timeouts, connection failures, and the four live sites described above that serve nothing to a non-browser visitor.
Sites that served nothing to a plain request were retried with a browser identity, to separate a site that is genuinely down from one that refuses anything that is not a browser. That distinction matters, because a site that only answers browsers is invisible to most AI crawlers by configuration rather than by content.
No practice is named here and no individual score is published. If your firm is in this index and you want your own result, ask and we will send it, whether or not you want anything else from us.
The honest part
What a score does not tell you
A low score does not mean a bad practice, and in this profession that caveat carries more weight than usual. Architecture is won on portfolio, reputation and personal relationships. Several of the practices that scored worst here are excellent firms with full order books who have never needed search to find work.
The sample is also smaller than we would like. Thirty one of the seventy seven sites could not be read, so every average on this page describes the forty six that could, and those are likely to be the better maintained sites. The real picture across the profession is probably a little worse than the numbers here suggest.
These tests measure only what a machine can read on one day. They do not measure whether an assistant recommends you when somebody asks for an architect in Guildford, which depends on what is written about you elsewhere as much as on your own website.
We built the scanner, so we have an obvious interest in what it measures. The tests are public standards rather than our invention, and where one is for something new that almost nobody has adopted, we have said so on this page instead of quietly counting it against people.
See where your practice sits
The same scan used for this index is free on your own website. About a minute, the same tests, the same score, and a list of what to fix first.
In this index and want your own result? Email dev@shopstartup.co.uk with your domain and we will send it.