How two teams run their own GEO experiments

AI search results can vary by metric, prompt type, and platform, so a tactic that looks promising in one slice of the data may look very different in another. Bricx and OnBoard use small, measured experiments to test their assumptions and decide where to invest next.

Bricx and OnBoard started with very different ideas, but tested them the same way: turn a hunch into a hypothesis, change one deliberate thing, measure what happens, then decide what to keep, change, or drop.

How Bricx tested its way into a citation

Divij Rakhra, co-founder and CMO of Bricx, a UX design agency, uses Peec AI to track how his agency shows up in AI search answers. Running his usual prompts, he kept noticing that on “best X” prompts, AI answers were citing third-party directories and lists as sources, not agency websites themselves.

Building his own directory

So instead of trying to get a mention on someone else's list, Divij set out to test whether building the same kind of source himself would work the same way. He bought an exact-match domain, where the domain name closely matched the search phrase he was targeting, and used Claude Code to build a small directory of UX and design agencies, with Bricx placed at the top.

Over 67 days, the domain picked up just 6 referring domains, most of them scraper sites that found it on their own rather than anything Divij built, and its authority score was 0. Google indexed it fast, then parked it around page 7 and left it there. Google Search Console showed an average position of 68.6 across 18,776 impressions and 16 clicks total. Not one of its top 10 queries by impressions ever reached page one, and going by SEO measures, this domain was a failure.

But within a few weeks the exact-match domain was showing up as a cited source for "best UX agency" style prompts, at first without Bricx named alongside it, then with brand mentions appearing next to the citation.

A domain can fail every traditional SEO measure, like backlink authority and ranking, and still become one of the more reliably cited sources an AI model pulls from in the same category. 

The pattern applied to other domains in the category, too. In Peec's citation-rate tracking across the category, an established directory in the space, with over 48,000 referring domains, averaged 1.37 citations each time it was pulled in as a source. Bricx's own established site was at 1.17. Divij's zero-authority domain came in at 0.89, and a similarly built exact-match domain run by a competitor, with just 42 referring domains, hit 2.01, the highest rate of any domain Divij tracked in the comparison.

Divij has since bought more domains built the same way, and checks how they're performing in Peec as he goes. “We're doing multiple such experiments that are in progress to see how it is impacting AI search, and Peec is the starting point for everything,” he says.

Where the experiment worked, and where it didn’t

The effect was narrower than it looked: in Peec, 91% of the exact-match domain's retrievals came from ChatGPT, and even there it was retrieved in 38.6% of answers for the exact head-term prompt but under 5% for prompts closer to a real buyer’s question.

Divij's read is that once a prompt includes constraints like budget or funding stage, AI models look for evidence like case studies, portfolio, and pricing rather than a ranked list, which is where your actual site has to do the work.

Peec AI GEO expert Tomek Rudzki sees two possible explanations worth testing further. “Based on our research, ChatGPT likely uses BM25, a text-based scoring formula, as part of retrieval. Keywords in the URL, in this case the exact-match domain, can push that score higher,” he says. He also points to consensus across sources as a possible factor: when multiple sources agree on a brand or claim, an LLM may have more evidence to draw on.

But there’s still an open question. “I’d love to see more experiments on this to understand what’s the dominant factor: the exact-match domain, or whether relevant content alone would be enough.” He’d also like to see whether the same pattern appears across other LLMs and AI search engines.

What another marketer can take from it

Divij's experiment shows the value of testing a GEO idea before committing significant time or budget to it. He noticed a pattern in what was already getting cited, built something small and specific to test it, and used what happened next to decide what to do without assuming the tactic itself was proven. The early result was promising enough to justify further tests, without treating the first domain as proof that the tactic works.

They let you test an idea cheaply and use your own data to decide what's worth more time or budget, especially with a result that's still moving and could look different with a few more weeks of data.

How OnBoard tested quotable pages for better citation rates

Tyler Naples is Senior SEO and GEO strategist at OnBoard, a leading board management software serving 30,000+ boards around the globe. His team wanted to test whether making their page's claims easier to lift and quote would change how often AI platforms cited it. They tested this with a Key Takeaways section near the top of the page. The section held a handful of standalone sentences, each citable on its own, with supporting schema markup behind it.

Their working theory was that this would reduce the amount of extraction or synthesis an AI platform had to do before using the page as a source. After they added the Key Takeaways section to one page, they left a comparable page unchanged as a control. The test ran for 6 weeks with checkpoints roughly every 2 weeks, and Tyler decided up front that the test would be judged on citation rate, tracked in Peec. 

What 6 weeks of data showed 

On Tyler’s chosen metric, the result was consistent. Citation rate climbed at every checkpoint on the test page, from 1.45 to 2.33, a 61% increase over 6 weeks. Absolute citations also increased, from 289 to 387, or 34%. The control page’s citation rate rose by roughly 7% over the same period.

Retrieval told a different story. On the test page, it rose slightly at first, then finished below its starting point. On the untouched control page, it jumped from 36 to 177 before falling back to 92.

Metehan Yesilyurt, GEO researcher at Peec AI, says the results are easier to interpret when retrieval and citation are treated separately. “Getting pulled in and getting cited are two separate steps,” he says. In this test, citation rate improved even as retrieval fluctuated, suggesting the Key Takeaways format may be more relevant to what happens after a page is retrieved than to whether it gets retrieved in the first place.

What the test result changed

With citation rate improving at every checkpoint while the control page’s rate moved inconsistently, Tyler's team concluded the Key Takeaways format was worth keeping.

The test also reflects a wider shift in how Tyler's team approaches GEO. “One of my commitments to the organization was being open to testing, and committing more thought and energy to experimentation. That was really well received from our leadership,” he says.

OnBoard is now planning to roll the Key Takeaways format out more broadly as a standing part of their content process. Tyler's queue for what to test next includes schema-only changes.

What another marketer can take from it

“Don’t copy the tactic, copy the setup,” Metehan says. OnBoard changed one thing, kept a control page unchanged, and chose the metric it would use to judge the result before the test began. For a similar citation test, he recommends starting with pages that are already being retrieved but rarely cited. That gives you enough retrieval activity to see whether making the content easier to quote changes what happens next.

Keep testing the assumptions

Bricx and OnBoard tested very different ideas, but both experiments left them with a more specific understanding of where to focus next. For Bricx, what looked like a clean citation result became more nuanced once Divij looked across AI platforms and prompt types. For OnBoard, separating citation rate from retrieval helped the team identify which part of the result mattered.

That’s the value of treating GEO as something to test rather than a set of tactics to follow. A result doesn’t have to produce a universal answer to be useful. It can tell you where an idea holds up, where it doesn’t, and what question to ask next.

“Our mission is to find repeatable methods for driving broader LLM visibility. And being able to back that work up with data through Peec has been a big assist,” Tyler says.

We’d like to see more of these experiments. If you’re testing your own ideas about how AI search works and have results you’re willing to share, we’d love to hear from you.