AI visibility for Welsh websites: can an assistant read your Welsh at all?
Ask an assistant for a builder in Bangor, in Welsh, and see which names come back. If the answer arrives in English, about businesses two counties away, that is not the assistant being awkward. It is what happens when nothing on your site tells a machine that the Welsh version exists.
Search gave you ten links. An assistant gives a handful of names
A search box returns a list and leaves the choosing to the customer. An assistant is asked a question and answers it with a few names, which means something has to read your site and decide whether you belong in that answer. Ranking eleventh was a position. Not appearing at all is an absence.
That reading happens without a person involved, and it happens in whatever language the question was asked in. For a Welsh business the second half matters more than the first.
A bilingual site is two documents, and a machine has to be told so
Most Welsh sites hold both languages. Far fewer say so in a way a machine can act on. The declaration is an `hreflang` annotation on each page naming its counterpart, and it has to work both ways: the English page points at the Welsh, the Welsh points back. A one way annotation is treated as unreliable and dropped, which is a quiet failure that looks like nothing at all.
Without it a Welsh page reads as a near duplicate of the English one rather than as the Welsh version of it. Two addresses compete instead of reinforcing each other, and a Welsh speaker searching in Welsh has nothing to find.
This is not a fault in every case. Not every business needs a Welsh site. But the ones that do are rarely told they are missing it, which is why the assessment reports it as a finding rather than an error.
How to check whether AI can read your Welsh site
Ask before you measure. Open an assistant and type what a customer would type, in Welsh, naming your town and your trade. Do it three or four times with different wording. If competitors appear and you do not, you have your answer without any tooling.
Then look at what the machine had to work with. Does the page declare its language on the `html` element. Does it name its counterpart in the other language, both ways. Does it carry a description written for it, or one lifted from the first paragraph. Are the services named in prose a person could read, or only in a picture. Is there a single main heading that says what the page is about.
None of that needs a technical audit. It needs knowing which questions to ask, which is the part most guidance leaves out.
The three files that decide what an assistant may read
A `robots.txt` file tells crawlers what they may fetch, and the AI crawlers are named separately from the search ones. A site can be perfectly open to Google and closed to everything else without anyone intending it.
Structured data, the JSON-LD block in a page head, states in a fixed vocabulary what the search result already shows a person: the name, the location, the opening hours, the services. It removes the guessing.
An `llms.txt` file is newer and not a standard yet: a curated markdown summary at the root of a site, giving an assistant a short and reliable account of what is there. The proposal is public and adoption is growing rather than universal.
The assessment generates all three from what is already on your page rather than describing them. A field with no source is left out rather than guessed at.
What the free assessment reports
It reads one address and answers three questions in turn: whether the page was found, whether what it says can be understood, and whether there is enough to make a recommendation. Underneath are six categories, and every finding says what was checked, why it matters, and what the fix is worth in points.
The score is weighted by what kind of business you are. A studio, a shop, a platform and a professional service are not judged on the same things, so a profile does not raise or lower the number so much as move the weight to what that business is actually judged on.
The rubric carries a version, and scores are comparable only within one. That is deliberate. A score that quietly changes meaning between runs is worse than no score.
Frequently asked questions
- Does having a Welsh version of my site help AI recommend me?
- Only if the site says the Welsh version exists. The two versions have to name each other in their markup, both ways. Where that declaration is missing, a Welsh speaker searching in Welsh, and any system serving them, has nothing to find.
- Can I check this without a developer?
- Yes. Ask an assistant what your customers would ask, in Welsh, and see which names come back. For the technical half, run a free assessment against your address. Nothing to install, and no access to your site needed.
- What is llms.txt, and do I need one?
- A short markdown file at the root of a site giving an assistant a curated account of what is there. It is a proposal rather than a standard, and adoption is growing rather than universal. It is worth having because it is cheap, not because it is required.
- Is this the same as SEO?
- It overlaps and it is not the same. Search engine optimisation rewards relevance and authority against a query. An assistant has to understand what a business is well enough to name it in an answer, so clarity, structure and stated facts count for more than keywords.
- What is GEO, and do I need it as well as SEO?
- Generative engine optimisation, GEO, is the name people have started using for being findable when an AI system answers rather than lists. It is not a separate discipline you buy on top of SEO: the same clean structure, stated facts and readable prose serve both. The difference is what counts as success. Search engine optimisation aims at a position in a list. Generative engine optimisation aims at being one of the few names an assistant says out loud.
- My site is Welsh only. Is that a problem?
- Not in itself. The assessment checks that a page declares what language it is in and that any counterpart is named, not that both languages exist. A Welsh only site stating its language clearly is better placed than a bilingual one declaring neither.
- How long does the assessment take?
- Seconds. It reads the address you give it, scores it, and shows the findings with the evidence for each one. You see the score before you give an email address.
Check what an assistant can read on your site
A free assessment of one address. The score first, the findings with their evidence, and three files generated from your own page.