Who AI tells UK founders to raise from

The short version

I wanted to know what actually happens when a founder opens ChatGPT and asks who to raise from. So I ran 40 real founder questions through it, each one twice, once with web search on and once off, and counted every citation and every fund it named. That is 80 answers.

Here is the headline. AI almost always answers with names: 92.5% of the answers recommended at least one specific fund or investor. But more than half the time, 52.5%, it did so citing no source at all. Switch web search off, so the model is working from memory, and that climbs to 77.5%.

The model isn't lying to founders. It's bluffing. And a bluff with no working still gets treated as a shortlist. I've made a version of this point before, that a source that always concludes the source is the best is not a source, it is an advert. This is the same problem in a different coat: a confident verdict with nothing behind it. The twist is that this time the machine is doing the unsourced ranking, not an agency padding a listicle it wrote itself.

I read that as an opportunity, not a scandal. A category where the recommendations are mostly unsourced is a category where the sourcing is still up for grabs.

How often AI cites nobody

The split is almost entirely about one toggle. With web search on, 72.5% of answers carried at least one source. With it off, only 22.5% did. Same questions, same model, a 3.2 times swing from whether it bothered to look at the live web.

The two modes reward different things. A grounded answer, with search on, can be earned: it reads current pages, so a clear, crawlable site can get pulled in. An ungrounded answer, with search off, is a memory test, and that test was set years ago. It rewards the funds the model already knows by reputation, and nobody else gets a look in.

What it trusts when it does cite

When AI did show a source, I sorted every one of the 284 citations by type. This is the most useful chart in the study, so it's worth sitting with.

Your own site is the single biggest source, by a distance. Nearly two in three citations point straight at a fund's own domain. And here is the irony I keep running into with investors: that is the one surface most of them treat as an afterthought.

Most VC sites are vague by design. "We back ambitious founders at the earliest stage" tells a crawler nothing about stage, cheque or sector, and the vagueness is usually deliberate. A fund's model is that the right founder arrives warm, introduced by the right person, not cold off a Google search. There is a quieter assumption underneath it, too: that a founder who has been reduced to typing "who will invest in my company" into a chatbot is probably too early to be worth the meeting. So the website stays a brochure rather than a front door, and it never spells anything out, because it was never written to be found this way.

Which leaves these funds wanting it both ways. They very much want to be visible and well regarded online. They just would rather not be caught optimising for the people doing the looking. Discovery has moved. The brochure hasn't.

The aggregators decide the long tail. One citation in five is a third-party directory or a curated list: the Sifteds, the Tech.eus, the VC-list sites. Those are the pages that decide whether a lesser-known fund gets surfaced at all. Being absent from them is being invisible.

The room you're not in

Now the one that should make every partner sit up. Reddit is under 5% of all citations, but it showed up in more than a third of every sourced answer. When a founder asks something subjective, like which investors are actually founder-friendly, the model goes looking for candid signal, and candid signal lives on forums.

Your reputation in AI search is being written on Reddit, by founders, without you in the room. And you're not even reading it.

It is easy to wave Reddit off as a nerdy backwater for people who argue about mechanical keyboards. That is a mistake, and an expensive one. This is the forum that organised the GameStop short squeeze and put a hedge fund or two on the floor. It moves markets, and it has been quietly moving reputations for far longer. The brands that win on it are the ones that turned up and joined in. No fund controls that surface. Every fund is being talked about on it anyway.

Nobody is too big for the shop window

Here is the good news for anyone who is not already a household name. AI cited 184 distinct domains across the 80 answers. This is not a cartel. A fifth of the visibility went to the top 10 sources, and the rest fragmented into a long tail of single mentions. The model is plainly willing to surface niche, regional and sector-specific funds, but only when the grounding finds them. The door is open, and an open door rewards whoever bothers to walk through it legibly.

This is the hill I will die on, and it is not only about venture capital. No brand is too big to be found. Apple does not assume everyone comes direct. It still does the basic work so that a search for a laptop with a particular spec turns up a MacBook alongside the alternatives, because the nine in ten who already know the brand are not the point. The point is the undecided person comparing options who had not thought of you yet. A VC fund is not above that. Nor is a FTSE 100. Nor is anyone.

So if you run a fund: your site is your most-cited asset, so make it say, in plain words a machine can read, what you back and what you write. That is the core of answer engine optimisation, and in this category it is still wide open. Get described accurately in the directories the model trusts, and read the forums, because your reputation there is being written without you. If you want the specific version, where your own fund actually lands when a founder asks AI who to raise from, that is what a free LLM visibility audit is for. If you are a founder: treat any AI shortlist as a starting hypothesis, not a verified list, and turn web search on before you trust a single name on it.

Nobody is too big for search. It is the biggest shop window on earth. Miss it, and you miss out.

How I ran this

I didn't run this as a stunt. I work with venture firms and read this landscape most days, which is where the curiosity comes from. I am not about to publish any client's data, so I built a clean public set instead: 40 representative UK founder questions about who to raise from, spanning stage (pre-seed and seed), sector (fintech, deeptech, climate, healthtech, consumer, dev-tools, AI), geography (London, Manchester, Edinburgh, Bristol, Oxbridge) and founder profile. Each was run twice through a ChatGPT scraper, once with web search grounding on and once off, for 80 answers in total. UK and English locale, captured 8 June 2026. For every answer I recorded the cited domains and whether any source was present. The citation-pattern findings sit on 38 sourced answers, comfortably past the 30 I treat as the bar for reporting a rate rather than a hint.

Cite this as: Dog on the Table (2026), "Who AI tells UK founders to raise from", 40 UK founder questions run web-search on and off via a ChatGPT scraper, captured 8 June 2026.