Cork café websites checked: what 24 sites show an AI assistant

Most of the 24 sites were in order on https (all 24 after a hand check on 10 October), opening hours (17) and the Google search basics (16). The gaps were the menu (13 of 24 failed the menu check) and the phone number and address as data (20 of 24 did not have both: 18 had neither, 2 had the address but not the phone). On 9 October 2026 I ran six checks on the websites of 24 cafés and roasters in Cork city and county, starting from each home page, and no robots.txt rule closed the home page to any of the 10 named AI crawlers. Of the 13 that failed the menu check, six had no menu found from the home page, six had a menu page with fewer than five price-like figures the check recognises, and one had PDF menus only (found on a hand re-read). In our view, fix the menu first (13 of 24 failed), then the phone and address as data (20 of 24 failed), then opening hours (7 of 24 failed).

What we checked, and which shops

This is not a count of Cork cafés. The checks read each website as it is served, without running the page's code (no browser). On 9 October 2026 I found 30 cafés and roasters in Cork city and county that each have a website of their own, through a Google Places text search. They are in Cork city and its suburbs and in towns around the county. So this page says nothing about how many Cork cafés have a website. It is about what the shops that do have a website show.

All 30 were audited. Two returned an access-denied response to our checker, so I left them out. After a hand re-read I left out 4 more. The reasons are of a general kind: two businesses with no public premises to visit (one sells wholesale and online, one is a coffee trailer for hire), where checks written for a café's opening hours and menu do not fit; a site whose home address returned "not found", so the checker read another page in its place and its results would not describe the home page; and a site where the hand re-read and the checker disagreed on whether it had a menu at all (dishes named as text, no prices) and on its one dead link. Where a hand re-read showed what the check should have returned, I corrected the result and kept the site; for that last site it did not settle the result either way, so I left it out rather than pick one. That leaves 24. Every count of sites below is out of 24 unless it says otherwise, and every one was read on 9 October 2026, and some results were corrected by hand re-read on 9 and 10 October 2026.

Check (as named on our free check)	Passed	Failed
A menu an assistant can reach	11	13
Opening hours on your site	17	7
Phone and address an assistant can read	4	20
An llms.txt, or AI crawlers let in	24	0
Ready to show up in Google search	16	8
Secure, with no dead links	23	1

The search check tests for a page title, a meta description and the mobile viewport setting on the home page. The phone-and-address check counts only schema.org data (JSON-LD or microdata). A phone number and address shown as plain text on the page do not count.

What the 24 sites got right

On 9 October 2026, the AI check passed a site if no robots.txt rule closed the home page to any of 10 named AI crawlers, or if it had an llms.txt. A robots.txt that answers with a server error or cannot be reached also counts as keep out; none of the 24 did. All 24 passed. No robots.txt rule closed the home page to any of the 10 named crawlers, and 5 of the 24 had an llms.txt. The check only catches rules that close the home page (/, /* or /$), so it would miss a rule that shuts out only part of a site. It reads robots.txt and looks for an llms.txt, and the two sites that turned our checker away are not among the 24, so all 24 passing describes the sites that answered, not every Cork site we looked at.

On 9 October 2026, 23 of the 24 were audited on https. The 24th was audited at the address we had for it, a plain http address; a hand check on 10 October 2026 found it also answers on https, so all 24 count as https here. The secure, no dead links check passed 23: that means https plus no dead link among the links tried. The checker tried the links on the home page that stay on the same site, up to 30 per home page, and counted one as dead only when it returned "not found" or "gone" or would not connect. Links that returned other errors, for example "forbidden", "too many requests" or a server error, were not counted as dead, so a pass means no link was found dead, not that every link was verified. The 1 that failed has a home-page link that leads to a "not found" page. For 5 sites no links were tried, so they passed that part trivially. Sixteen had a title, a meta description and the mobile viewport setting, the three things the search check tests. Seventeen showed opening hours on the site, as text or as data. The hours check looks for a day word close to a time on the home page, the menu page or up to three further pages. Its passes were mostly not re-read by hand (one was a hand correction) and can include false positives.

What most were missing

The common gap is not a robots.txt that closes the site to AI. It is what there is to read. The menu check passes when a page reachable from the home page (the home page itself, or the first page whose link text or address contains the word menu) shows at least five price-like figures as text, or carries menu schema data. The price count only recognises figures written like €5, 5.10 or 5,10, so a price written "EUR5.10" is missed. On 9 October 2026, 13 of the 24 sites failed the check. For six of them, no menu was found that way. Six had a menu page with fewer than five price-like figures the check recognises, and one had PDF menus only (found on a hand re-read). A pass has a limit too. It means the checker found at least five price-like figures as text, or menu schema data, on the home page or the first page whose link text or address contains the word menu. Of the 11 passes, 9 are the checker's count and 2 are hand-read passes. A checker pass is not always a café menu: the figures can come from shop prices for coffee beans or from other figures that look like prices, including times written like 8.30. Hand re-reads went mostly to the failures, not the passes. So the checker's passes can include false positives (figures that are not menu prices), and its fails can include false negatives (menus it could not see, for example on pages that are almost empty without running code). In our view, for AI assistants the fix is a menu page linked from the home page, with dishes and prices as text.

On 9 October 2026, 20 of the 24 did not have both the phone number and the address as schema.org data (JSON-LD or microdata): 18 had neither as data and 2 had the address but not the phone. The contact check reads the home page and the menu page and, only while it is still looking for opening hours, up to three further home-page links whose text or address mention hours, opening, visit, contact, find, location or about. It fails a site if the phone or the address is not there as schema.org data, even if it is shown as text. This check is strict, because it asks for data and not text. The phone number and address may well be on the page as visible text, and a person reading the site would find them. I am not saying they are missing from those sites. In our view, without that data an assistant has nothing labelled to read from.

Our 10-point AI readiness score is ten checks chosen by our own tool. It gave a median of 2 out of 10 for the 24 sites on 9 October 2026. Fourteen scored 2, eight scored 3 and two scored 4. None scored above 4. These scores were not re-read by hand.

How fast the home pages loaded

This is a lab test, one run per site from our machine on 9 October 2026, using Lighthouse mobile emulation with the throttled network Lighthouse sets. Real phones on real connections vary, so treat it as a rough picture and not what any customer saw. These results were not re-read by hand.

In that test on 9 October 2026, the largest content on the home page took a median of 9.3 seconds to appear. The range was 2.8 to 43.6 seconds. Twenty-two of the 24 took over 4 seconds. I am giving our numbers and not a pass mark.

How does your café's site compare?

Our free check uses the same rules on any café's website, but it reads fewer pages and links (it stops after a set number of requests and tries 10 links), so a result on /check can differ from this audit. In our view, from the counts on 9 October 2026, the order to fix things is: the menu first (13 of the 24 failed), then the phone number and address as data (20 failed), then opening hours (7 failed). Google's side of this is in how Google's AI Overviews treat a café. Our Dublin count of 77 Google listings, read on 20 September 2026, used a different method: it read Google listings, not websites.

Method and limits

The sample was 30 cafés and roasters in Cork city and county with websites of their own, found by Google Places text search on 9 October 2026. All 30 were audited. Two returned an access-denied response to our checker, and I did not check whether they load for people. Four more were left out after a hand re-read, for the general reasons given above. That leaves 24. The tool is site-audit.mjs, run between 09:08 and 09:21 UTC that day, using the same rules as the free check. The free check reads fewer pages and links (it stops after a set number of requests and tries 10 links), so a result on /check can differ from this audit; this audit tried up to 30 links per home page. The checker fetches pages with its own user agent and stops reading a page after 1.5 MB or 8 seconds. The AI check covers 10 named AI crawlers (GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User, Google-Extended and Applebot-Extended) and only catches robots.txt rules that close the home page. The search check tests only for a page title, a meta description and the mobile viewport setting on the home page. Our 10-point AI readiness score comes from agent-check.mjs. No shop is named here and I have not broken the figures down by town, because the sample is too small to do that fairly.

The automatic check can be wrong. Some results were re-read by hand on 9 and 10 October 2026 and either corrected or confirmed, and the counts on this page use the results after that. The hand re-reads changed six results, each listed once, on four of the 24 sites: five moved to pass and one moved to fail. Site A had its opening hours corrected to pass and its menu corrected to fail; the menu fail was a site whose only menus are PDFs. Site B had its menu corrected to pass and its secure check corrected to pass: it was audited at the address we had for it, a plain http address, but it also answers on https. Site C had its menu corrected to pass. Site D had its secure check corrected to pass, because the only "dead link" was a misread by our checker: the page writes its link to its menu page in an escaped form the checker did not decode, and the real menu page answers normally. The two menu passes (sites B and C) were sites whose menu page did not load for the checker (the menu page answered with a server error), and a hand re-read found the menu with prices as text. The secure-check corrections were made on 10 October 2026. So 2 of the 11 menu passes and 1 of the 13 menu fails are the hand reading, not the automatic one. On some other sites the main fault was re-read and confirmed. Not every result was re-read by hand, and the re-reads went mostly to the failures, not the passes. Of the 11 menu passes, 9 are the checker's count and 2 are hand-read passes. A checker pass means only that it found at least five price-like figures as text (or menu schema data) on the home page or the first page whose link text or address contains the word menu; they can come from shop prices for coffee beans or from other figures that look like prices, including times written like 8.30, so the passes can include false positives (figures that are not menu prices) and the fails can include false negatives (menus it could not see, for example on pages that are almost empty without running code). The price count only recognises certain EUR formats, as described above. The AI readiness score and Lighthouse figures were not re-read.

One more limit. On 9 October 2026, 4 of the 24 sites had pages that were almost empty to the checker without running code. On those sites a checker that does not run code can miss a menu, opening hours and the phone and address. Those 4 account for 4 of the 6 sites with no menu found, 4 of the 7 opening-hours failures and 4 of the 20 phone-and-address failures. For one of the four, the pages returned an error when re-read by hand.

The aggregate as a CSV, if you want to use it. Cite it as: Cupán Studio, Cork café website audit, 24 sites, 9 October 2026. Please carry the sampling caveat with it.

Cupán Studio builds websites for cafés in Ireland, audited these sites while looking for cafés to contact, and has since contacted some of them about what their audit found, which is why we have the data. That is a conflict of interest, and you should read the findings with it in mind. This is a summary of our own audit.