We noticed something strange
Ask ChatGPT, Claude, or Perplexity what Zypsy is, and the answer comes back right. A design and investment firm for founders. The client work, the model, the funds and VCs we co-invest alongside. The machines had read the site and understood it.
Then ask the question a founder actually types. "Best design agency for a venture-backed startup." "Who should build my brand and product before a raise." We were rarely in the answer. The same names came back, over and over, and ours usually wasn't one of them.
Understood, and not recommended. That gap is the whole subject of this piece.
We measured it instead of guessing
We instrumented two things. A daily tracker ran founder-facing category questions across four engines, ChatGPT, Perplexity, Claude, and Google's AI answers. It logged several numbers. The two that mattered most were how often Zypsy was mentioned, and how often an answer cited one of our own pages as a source. Separately, our engineering team built an auditor that ran a broad neutral sweep of discovery questions with no brand names in them, so we were not just grading our own homework.
June was the baseline. On neutral discovery questions, our unbranded visibility was about 2 percent. Asked about Zypsy by name, recognition was 100 percent. The engines could describe us in detail. They just did not surface us on their own.
We changed the website, then measured again
Then we did the work you are supposed to do. We wrote new pages that led with a direct, quotable answer in plain language. Blog posts built around the questions founders ask, an updated capabilities page, a client directory. We published structured data, added FAQs, and reconciled facts that had drifted between pages. We added an llms.txt and opened the AI crawlers that were quietly blocked. A free readiness checker moved our score from 54 to 80 over the project, a useful directional check that the technical work was landing.
None of this was new proof. Zypsy already had the client work, the case studies, the outcomes. We were making evidence that already existed easier for a machine to read.
Citations jumped
The rate at which an AI answer pointed to a Zypsy page as a source went from about 1 percent in June to about 10 percent in August, across the same set of founder questions in both months. Close to a tenfold jump. The count of citations to our own pages roughly tripled.
It was strongest on ChatGPT, which went from about 6 percent to about 25. Perplexity went from nothing to about 9. Claude barely moved, but its crawler was blocked for most of the run, so we would not read much into that.
This is the half of the problem a website controls, and it moved fast. The gain was already showing up in July, before we touched the crawler settings at the end of the run, which points more strongly to the new pages and the structured content doing the work, not the technical switches. Give an engine a page that answers the question in plain terms, and you give it much more to quote.
Discovery didn't follow
Take the seven questions we tracked from the first run to the last. Citations rose on every question where the rate changed at all. Mentions rose too, but almost all of it landed on two questions, the firms that run a venture fund and the firms that co-invest alongside VCs. On the rest, mentions were flat or lower. Those two are the narrow corner we already own, and almost no one asks them.
Against our competitors, our share of voice did not grow. It sat at 13.9 percent in June and 13.5 in August. Our gains stayed inside that niche, so our share of the broader conversation held rather than climbed. On a broad neutral sweep of discovery questions, we stayed near 2 percent.
We became much easier to quote without becoming much easier to discover.
Where AI actually gets its recommendations
So we used the same in-house auditor to look at what the engines cite when they answer a discovery question. Not for us. For the whole category. The source mix gave us our clearest clue.
About 82 percent of the sources the engines cited were companies writing about themselves. Their own websites, and the "best agency" roundups that agencies publish about their own field, usually ranking themselves near the top. Around 13 percent were directories and review sites. About 3 percent was community, mostly Reddit. Independent editorial, a writer or an outlet covering the space, was a low single-digit share.
The lesson is not that we needed a prestige press hit. It is that getting recommended appears to depend on being named across more of these third-party pages, and so far we are on only some of them.
We could have added to that 82 percent ourselves, a Zypsy list with Zypsy near the top. We chose not to. It is a cheap way to get cited, and a hollow one.
The honest version is slower, and we have started it. We verified our Clutch profile and gathered the first client reviews. It has not moved discovery yet. Reviews show up when someone asks about us by name, not on the open questions, and that is the nature of the layer. It is the slower layer, and we cannot yet say how long it takes to move.
That is the distinction the whole experiment comes down to, and we had missed it at the start. Almost all our proof, the outcomes and the case studies, lived on zypsy.com. That is why citations climbed: we made those pages legible and the engines quoted them. But an engine deciding who to recommend draws on the wider web, and there our footprint was thin. The proof existed. Too much of it existed only on pages we control. Making a site legible to AI and getting recommended by AI are two different problems, and a website solves only the first.
It is easy to misread this as a verdict on quality. It is not. The names that surface are not necessarily better, they are the ones with more corroborating evidence across the open web. That is distribution, not judgment.
It is harder than it looks, in a few specific ways
The first is language. One of the seven questions did this to us. We asked the engines about firms that "build and engineer the product," and across our tests the phrasing pulled every one of them toward hardware and industrial design rather than software. Your industry shorthand carries the web's meaning, not the one you intend. Say the plain thing, then read what comes back.
The second is demand. It is easy to win a query nobody types and mistake it for visibility. Track the questions real buyers ask in volume, not the ones that happen to describe you.
And there is no single AI answer to optimize for. We asked the same question of four engines and got four different top picks. Not four different worlds. The same handful of firms led everywhere, just reshuffled, with a different one on top of each engine. The shortlist a founder sees depends on which one they open, so measure more than one.
The playbook
Here is what we would tell a founder, ordered from what you fully control to what you do not.
Make sure the machine can read you. Open the AI crawlers, keep your answer in plain HTML, and mark up your pages with schema so the engines read your facts as structured data instead of guessing at them. Our readiness score moved from 54 to 80 over the project. It is necessary, not an advantage.
Answer the real question, in your buyer's words. If none of your pages answers the question a founder actually asks, you make it much harder for your own site to become part of the answer. Over the same period, our citation rate on those questions moved from about 1 percent to about 10.
Measure citation and discovery separately. They are different results and they move differently. Our owned citation count roughly tripled while discovery stayed broadly flat. Do not let a jump in one flatter you about the other.
Earn corroboration beyond your own domain. Reviews on the platforms engines cite, directory listings, community, and the occasional real piece of press. This is the slower layer, and our audit suggests it is the one we now need to strengthen.
Where that leaves us
We set out to see if we could optimize our way into AI answers. We can, up to a point. A website makes you understood and easy to quote, which is now table stakes. It does not make you recommended on its own. That last mile does not happen on your site alone. It happens across the wider web, and much of it has to be earned.
The same changes also make the site clearer for human readers. Clearer answers, straighter facts, less buried in marketing build-up. Agents turned out to be honest feedback on work we thought was already finished.
That is the spirit we borrowed from the developer-tools world. Make something agents want. We would add one line. Being wanted by an agent starts with being understood, and ends somewhere you do not control. We are still closing that gap.
