What does AI say about your Firm?

In June, 2 former OpenAI researchers launched a fun tool called In the Weights.

You type a name. It asks 13 models who they are and scores the recognition up to 996.

It is a vanity search, but an excellent distribution gimmick, AND it measures something real. The weights are the billions of numbers a model learns in training. If a model can describe you with no web search, you are in “in the weights”.

There is a nuance the tool misses, though. Recognition is whether the model knows your name. Reference is whether the model thinks with you. Ask your favourite LLM how to pay off 3 credit cards and a car loan, and the answer will be the debt snowball, smallest balance first, whether or not Dave Ramsey's name appears. A person can score 200 and be the invisible skeleton of every answer in their category. A celebrity can score 900 and shape 0 answers.

In venture, this distinction is now commercial. Founders and LPs of course run fund discovery through AI - and amid this summer’s uproar around VC rankings, the line I am hearing from firms is not about these list but about AI recommendations.

I wrote in March that this would be testable and provable. So I tested it.

Here is what 2,000 answers look like.

How I tested it

I wrote 20 founder questions.

• 10 were discovery questions, the type a founder asks when building a list:

E.g. “I am building a therapeutics company in Europe and raising a seed round. Which venture capital firms should I seriously consider, and why?” or “I am a first-time technical founder raising pre-seed before I have meaningful revenue. Which venture firms are genuinely comfortable investing this early?” etc etc…

• 10 were operating questions:

E.g. “how do I know if I have product-market fit”, “should I raise again or get profitable”, “how do I find the first 10 customers” etc etc…

For these, the models were asked which firms or investors had published thinking worth drawing on.

I ran each question through 1 current model from each of 8 developer families via OpenRouter (OpenAI, Anthropic, Google, xAI, Meta, Mistral, DeepSeek and Qwen).

3 conditions:

1. Closed-book. Search and tools off. The model answers from what it already knows. 800 responses.

2. Common retriever. Every model gets the same live search layer (Exa), the same result limit and the same instruction, and is asked to cite. This holds the search engine constant so the differences are down to what each model selects and synthesises. 480 responses.

3. Specificity ladder. 5 categories, 3 levels of legitimate founder detail: the broad category, then stage plus geography plus business model, then all of that plus capital profile and a specific operating constraint. Run both with and without search. 720 responses.

2,000 answers in total, I then manually audited the 50 most-retrieved firm recommendations against current primary sources, and scored whether named frameworks kept their author and institution.

Some limitations…

• The questions I wrote influence answers (partly the point of the test).

• It measures model behaviour on 2 days in September using 1 retriever. Answers are stochastic and will drift.

• Closed-book answers are the observable trace of model knowledge. I cannot inspect a weight - "In the weights" is a metaphor for a measurable behaviour.

• The API panel is a sample of developers, not a map of what founders use. ChatGPT, Claude etc as consumer products add their own search, memory and personalisation. I ran a small separate spot check on the consumer products and kept it out of these numbers.

• A firm appearing says nothing about its quality + a firm not appearing says nothing about its expertise.

With that said, here is what the results look like. Buried in them are 3 distinct markets for a fund's reputation, each won differently.

1. Be remembered

The 400 closed-book discovery answers recommended 323 different firms. The 5 most-mentioned captured 13.1% of all recommendations.

Category owners were clear and plural:

• Breakthrough Energy Ventures in 100% of climate answers

• Forerunner in 87.5% of consumer answers

• Heavybit in 82.5% of dev-infrastructure answers

• QED in 77.5% of fintech answers

• Sofinnova in 77.5% of European-biotech answers

Clearly, the models carry a recognisable map of who backs what, and a specialist can own a category without dominating venture as a whole.

2 things stand out in the map.

1. Names that match the query win.

Breakthrough Energy and Lowercarbon in climate. Precursor on pre-seed. TinySeed, Calm Company Fund and Indie.vc on capital-efficient software, under search it was Zero Carbon and Cloudberry. The fund name contains the founder's words. The name of an idea is as important as the argument behind it, and the same applies to funds. The name is the first piece of content the weights read.

2. Discovery is won by evidence, and the canon by ideas.

Forerunner publishes a ‘normal’ amount and nothing unusual, their 87.5% is Warby, Glossier and Dollar Shave plus a decade of press about Kirsten Green. Heavybit has a name that means developer-first and a large founder-facing library. Portfolio and press get you remembered for a category… shock!

2. Become the reference frame

When I asked the operating questions, the map collapsed.

Across 400 answers the models recommended just 49 firms. A16z, Y Combinator, Sequoia, First Round and Benchmark captured 71.0% of every firm-bearing recommendation.

YC appeared in 97.5% of answers about whether a company is default alive. a16z appeared in 97.5% of answers about PMF, 97.5% about replacing an executive and 97.5% about sizing a market that does not exist yet.

The 2 halves of the study behave differently. Ask the models which firm to pitch and they spread the answer across hundreds of names, sorted by category. Ask them how to run the company and they reach for the same 5 firms almost every time. a16z was the only firm all 8 models agreed on for more than 1 question.

3 ways into the reference canon

What is interesting, is that the 5 top dogs did not all arrive by the same door.

1) a16z, First Round and YC are institutional media operations. The firm publishes, at volume, for years. Breakdowns on all 3 (here, here and here).

2) Founders Fund appears on 2 operating questions, but that is Zero to One. Likewise, USV appears on the same 2, but that is AVC. If you have a famous GP who publishes consistently, their solo-content can carry you into the results.

3) Benchmark is the third door. They appear in 0 discovery categories but 9 of 10 operating questions. Some of that is Above the Crowd. Some of it is Uber - Benchmark is in the weights for replacing a CEO because they did it, in public, in a lawsuit.
No owned media, no content machine, and a firm that exists in the canon and nowhere on the map. Reputation is quality times consistency over time, and Benchmark's judgement and character have been compounding in public for 30+ years. 1 caveat: Benchmark never reached 8 of 8 models on anything. The name has gravity but maybe the media machines have more.

So how does a firm actually get in? The mechanism is less mysterious than "the weights" sounds…

• Volume: the models have seen your thinking many times, over years.

• Distinctiveness: the idea is attached to you and nobody else.

• And, most important of all, Repetition by other people.

The models did not learn default alive from YC publishing 1 essay. They learned it from thousands of founders using the phrase in forum threads, tweets and blogs. Other people cannot repeat your idea unless it has a name to repeat. That is why a named framework beats a well-argued paragraph - the name is the handle other people use to carry your thinking into the training data for you.

3. Be retrieved

Then I gave the models the power of search. Every model got the same engine, Exa, so any difference between their answers is down to the model, not the search results.

The 2 question types reacted completely differently. For discovery questions, search essentially threw the old answer away - only 5% of the firms the models had named from memory survived, and the top recommendation changed 9/10 categories.

Simply, search changes who the models recommend. It barely changes whose thinking they use to answer.

That difference is significant - traditional search result is a spot you occupy only while your page is the best match for that exact query. Publish a clearer page and the spot is yours. A competitor publishes a better one and it is gone.

Model’s memory does not move like that. It took years of accumulated material to form, and it takes years to displace. A young firm can win search next week + they can and should… but there is a difference between a spot it is renting and an association it owns.

8 models can be 1 opinion

Every model used the same search layer, so I expected more variety and got less.

From memory, the 8 models gave 7 different first choices for a capital-efficient SaaS company.

With search, all 8 gave the same one: Calm Company Fund.

i.e. Exa repeatedly surfaced Calm’s site, and all 8 model families selected it from the same retrieval environment.

BUT...! Calm Company Fund stopped investing at the end of 2024!

To make this worse, the site went offline (404s) in February. While it was previous probably the best-written material about calm, capital-efficient investing on the internet - this feels like a miss.

I have left the chart in because that is the finding. When someone tells you "every AI recommends X", remember what it can mean - it has changed how I read consensus in an AI answer.

Therefore - citation is the start of diligence

The models produced a URL for 99.8% of their 2,268 search-enabled recommendations. Verification is where it thinned out… of the 50 most-retrieved firms, 39 held up against current public sources, and 11 did not.

The failures were all different - a firm claiming 2019 funds on a domain registered in April 2026, a holding company recommended to founders as a venture fund, a generalist served up because it had the best pre-seed page on the internet.

Search did not stop the models being wrong, in fact, the opposite. Citations show where a statement came from, not whether it is corroborated or correctly interpreted.

Specificity kills prestige

I also expected that adding stage, geography and business model to a founder prompt would surface more of the specialists the models already associated with each category.

It removed the prestige answer instead. In closed-book answers, the share going to the 5 broadest firms fell from 17.9% on the broad prompt to 1.5% once the founder added detail. With search it fell to 0%.

The handover is visible in each row. Capital-efficient software at the broad level is led by Bessemer. Add stage and place and Bessemer goes to 0 while Point Nine climbs to 23 of 24 answers. Defence at the broad level is Lux, Founders Fund and Shield. Add detail and it is the NATO Innovation Fund, Vsquared and MD One.

The specialists did not reliably arrive, though, and the third row above shows why. At the broad level, developer infrastructure belongs to Heavybit and Amplify. Add "European, seed" and both vanish, replaced by Point Nine, Crane and Seedcamp - the same firms that lead the SaaS query.

The extra detail did not make the models think harder about developer tooling. It moved the question into a different bucket - seemingly overindexing on geography.

Under search, the leading firms changed almost completely with every rewording.

So specificity moves power. Detail in the prompt takes the answer away from the famous names and gives it to whichever pages the index can parse for those exact coordinates.

For a specialist firm this is big opening… you will never out-reputation a16z on the broad question, and you do not need to, because the founder who describes their company precisely will find whoever has built the clearest page at those coordinates.

You can own the idea BUT lose the credit

The final layer is attribution.

"Do Things That Don't Scale" kept Paul Graham's name in 90.5% of mentions. "Default Alive or Default Dead", from the same author and the same firm, kept it in 67.5%. Burn Multiple and Blitzscaling lost their author roughly 1 time in 4.

The models often use an idea without naming who came up with it… in fact, the better an idea does, the more this happens - once something becomes the standard way of thinking about a problem, the models treat it as common knowledge rather than someone's work. Graham gets named less on "default alive" than on his less famous essays for exactly this reason.

The 3 markets in the models

So, there are 3 markets for venture reputation inside a model, and they compound differently.

1. Be remembered. Closed-book, category level. Won by portfolio evidence, press and a name that matches the query. Rewards time and repetition.

2. Be retrieved. The search layer. Won by current, specific, technically legible primary pages. Rewards precision and freshness. Rented, contested weekly, and cheap enough that unverified claims compete on equal terms.

3. Become the reference frame. The operating canon. Won by ideas useful enough to become part of the language. Rewards usefulness, and pays out whether or not your name survives.

Most AEO advice lives entirely in market 2. That is the one you can rent and the one that changes with every rewording. The work that compounds is in markets 1 and 3, and the search layer sits on top of it.

How much should any of this matter?

An honest caveat first: the specific answers are unstable. Different questions, a different retriever, a different Tuesday - different results…

This why your firm's individual score matters less than the structure, which held across all 2,000 answers - discovery spreads across hundreds of firms, operating advice collapses into a canon of 5, search will fluctuate, and named ideas travel while unnamed thinking dissolves into common knowledge.

So while scores will drift. The 3 markets in the models will not.

The reason it matters is simple - AI models are now distribution. Founders are running their fund discovery through these answers today and it returns a shortlist presented as a fact. For venture, like every other business, this is becoming a primary acquisition surface.

Which changes the question most firms will ask. The instinct is "how do we rank higher in AI" - the AEO instinct, and it lives entirely in market 2, the rented one.

The better question is… which answers do we have the right to belong in?

So, my take for firms is the work is not climbing a ranking, but finding and then being undeniable at your own coordinates.

Your edge, meanwhile, stays offline. The models are distribution - they put your way of seeing in front of founders you will never meet, in conversations you will never join. But the entire economic premise of a venture firm is being right before everyone agrees and the weights can only hold what everyone already agrees on. Benchmark proves the 2 assets are different things - they dominate the weights for IRL events ~10 years ago, while the model retrieves past-you, the live position on the future has no data.

So the audit is 3 questions, 1 per market:

1. If a founder describes our category to a model with search off, are we in the answer?

2. If they add stage, geography and a constraint with search on, is the page it finds ours, current, and true?

3. If they ask how to solve a problem we understand better than anyone, is the answer built from our thinking, with our name still attached?

You want your work in the answer, but you still need a reason to be right before the answer exists.

Laurie, Refinery Media

If you made it all the way through, thanks so much for reading. A thousand VCs now open this every week. If it helped you think differently about marketing, Venture, or storytelling, please send it to someone in your orbit.

The full benchmark, prompts, model identifiers, data and a self-audit checklist for firms are in the VC Answer Map repository. Please edit, build on it, run your own tests.

Read more from our top posts:

• What is Full Stack VC Marketing?

• What Should your GP be posting?

• How The Firm With The Best Returns Should Do Marketing.