Who do AI assistants recommend for apprenticeships? We asked them, standard by standard

In September 2026 we asked ChatGPT and Perplexity the question employers are now asking them first: which training providers are best for a given apprenticeship? We put it 90 times, once per assistant for each of 45 standards, with web search switched on and the country set to the UK. The two assistants named 713 providers between them. We matched every name to the Department for Education's published achievement data and to the 1,125 providers listed on Find a Training Provider. This is what came back, what it agrees with, where it goes wrong, and what it was reading.

What did the assistants actually say?

Across the 45 standards, ChatGPT named an average of five providers per answer and Perplexity around ten. That is the first thing to know: an assistant does not give you the field, it gives you a shortlist, and a short one. For Business Administrator, where 126 providers have a published cohort large enough to rank, ChatGPT named two and Perplexity named ten.

The two assistants only partly agree with each other. On 29 of the 45 standards they had at least one provider in common. On the other 16 they produced entirely different lists for the same question on the same day. Ask twice and you may get two different shortlists, which is worth remembering before treating any one answer as the answer.

A small group of providers recur. BPP was named for 11 different standards, Paragon Skills for 10, and Babington, Fareport Training, and The Apprentice Academy for eight or nine each. Size and web presence carry weight with an assistant in the same way they do with a search engine.

How does that compare with the official achievement data?

This is where it gets interesting. The DfE publishes an achievement rate for every provider and standard where the cohort is big enough, and those rates are what our league tables are built from. We checked the assistants' picks against them.

Of the 713 names, 647 were providers in our directory. Only 210 of those were providers with a ranked cohort of 30 or more leavers on the standard in question. So most of what the assistants recommend is either a provider with a small or unpublished cohort, or a provider being recommended for a standard it has little published record on.

The first name in each answer is the one people act on, so we looked at that separately. In 11 of the 90 answers, the assistant's first pick was also the provider with the highest published achievement rate for that standard. In 14 it was somewhere in the top three. In 55 of the 90 answers, the first name had no rankable cohort on that standard at all, meaning fewer than 30 published leavers, or none. That is not always a mark against the provider. New and small programmes are suppressed in the official data, and some of them are excellent. But it means the assistant's confidence and the evidence behind it are two different things.

In 55 of 90 answers, the assistant's first recommendation had no rankable achievement rate for that standard.

Where the assistants and the data line up, it tends to be Perplexity. For Business Administrator its first name was Chichester College, which tops the official table at 97.1%. For Team leader both assistants led with Education Training Collective, top of that table at 94.4%. For Adult Care Worker, Perplexity led with Yorkshire Training Partnership, top of its table at 90.6%. On those standards an employer following the assistant would land on a provider the evidence supports.

Elsewhere the two diverge sharply. For Installation and maintenance electrician, Perplexity led with a provider that sits 24th of 70 in the published data, and ChatGPT with one that has no ranked cohort, while the provider with the highest published rate, Northampton College at 86.2%, was named by neither. For Multi-channel marketer, a young standard where only two providers have a rankable cohort, Perplexity named 14 providers and ChatGPT six, which tells you how much of the answer was coming from marketing copy rather than results.

What were the assistants reading?

Both assistants cite their sources when web search is on, so we can see what they read. Perplexity cited this site in 33 of its 45 answers. ChatGPT cited it in 15 of 45. Our own server logs show the same thing from the other side: ChatGPT's browsing agent fetches pages here during live conversations, not just when indexing.

That matters for two reasons. For employers, it means the shortlist an assistant gives you is partly a reading of the same tables you can check yourself, and the check takes a minute. For providers, it means your profile page is now part of the answer. Of the 336 providers the assistants named, 13 have claimed their profile here. The other 323 are being described to employers from a data-only page they have never looked at.

How should an employer use an AI shortlist?

Use it as a shortlist, not a verdict. The assistants are good at surfacing names you might not have found and at summarising what a provider says about itself. They are inconsistent with each other, they weight visibility heavily, and more often than not their first pick has no published result on the standard you asked about.

So take the three or four names it gives you and put them beside the official achievement table for your standard. If a name is in the top of the table, you have two independent signals. If it is not in the table at all, ask the provider why: a new programme, a small cohort, or a standard they have only just started delivering are all reasonable answers, and the provider-reported results some of them now publish here, labelled as such, fill the gap while the official data catches up. Then read the methodology so you know what the numbers do and do not measure, and use the provider search to check who actually delivers in your region.

If you run a training provider, what does this mean for you?

The employers asking assistants are not going to stop, and the assistants are reading your profile whether or not you have. Two things follow. First, claim the profile, which is free and takes a minute with a company email, so what gets quoted about you is yours. Second, give the assistant something to say. The reasoning we saw in the answers leaned on specialism, curriculum, and who teaches the programme, none of which appears in any government dataset. A Showcase Profile exists for exactly that: programme details per standard, case studies, and the contact details of employers who enquire, published in the structured form an assistant can read. It does not move you in any table, because nothing does, and that is what makes the tables worth being in.

We now publish what the assistants said on each standard's page, with the date and the exact question, and we will repeat the exercise monthly. It is their answer, not our ranking. Every commercial relationship this site has is listed on our commercial disclosure page.

Frequently asked questions

Which AI assistants did you ask, and how?

ChatGPT and Perplexity, through their standard interfaces with web search enabled and the location set to the United Kingdom. Each was asked the same question for each of the 45 standards on 2 September 2026: which UK training providers are best for that apprenticeship, naming specific providers. We recorded the full answer, the providers named in the order given, and the sources each assistant cited.

Why only 45 standards?

We started with the standards that have the largest ranked fields in the official data, because those are where a comparison between the assistant's answer and the evidence is most meaningful. The panel widens each month. Standards with fewer than five providers holding a rankable cohort are less informative for this exercise, though the assistants will still answer for them.

Does being named by an assistant change a provider's ranking on this site?

No. The league tables are built only from the DfE's published achievement rates, with the method set out on the methodology page. What the assistants say is shown on standard pages as their answer, dated and separate from the table, and it is not an input to anything. No provider can pay to appear in a table, a badge, or the assistant panel.

Why does a provider with a high achievement rate not get named?

Assistants weight what they can read. A provider with strong results but a thin website, an unclaimed profile, and no published programme detail gives the assistant little to work with, so it reaches for the providers with the most text about them. That is a visibility gap, not a quality judgement, and it is fixable by the provider.

How often will this be updated?

Monthly. Each standard's page shows the latest answers with the date they were collected, and this article will be refreshed with the new totals. One month's answer is weather; the trend across months is the thing worth watching.