Every Brand That Won AI Search Tested Their Way There
Vercel, Tally and Gymshark tested their way into a new channel. AI search rewards iteration speed, and the limit is how fast you ship work a person will sign.
machina sculpsit · homo probavit · MMXXVIIn 1901 Wilbur and Orville Wright built a six-foot wooden box with a fan on one end and ran about two hundred wing shapes through it, one at a time, writing down what each one lifted. Samuel Langley, who had a War Department grant and a houseboat catapult on the Potomac, worked from theory and launched twice. Both machines went in the river.
Langley had the grant. The Wrights had the notebook, and two years later they flew.
Every brand that has won a new distribution channel did the same thing. They treated it as a laboratory while everyone else treated it as a broadcast.
This matters now because the buyer stopped searching and started asking. The next decision gets made inside an answer that returns three names, and the job is being one of the three. A page of ten blue links gave everyone a seat somewhere. An answer does not.
AI search is a new channel, it is non-deterministic, and nobody has the theory yet. So the question for a marketing team is not what the right GEO strategy is. It is how many tests you can run before your competitors work it out.
Three Brands Treated a New Channel as a Laboratory
A $500 test that proved a concept in a week beat a $50,000 campaign that took three months to plan.
Dollar Shave Club spent $4,500 on a launch video in 2012 and acquired 12,000 customers in 48 hours (Optimonk, 2022). Four years later Unilever bought the company for a billion dollars in cash. The video was not a campaign. It was a test that worked.
Gymshark ran the same year on the opposite shape: dozens of small influencer collaborations, measured, losers killed, winners funded. It reached a billion-dollar valuation without ever running a traditional advertising campaign (TrendTrack, 2026). The budget was never the advantage; the loop was. Warby Parker did it in 2010 with small-budget Facebook tests for its home try-on program, and had run hundreds of creative variations before legacy eyewear brands bought their first ad.
None of the three had a bigger budget than the incumbents they took share from. They ran more tests per dollar.
Micro-Tests Beat Brand Campaigns on Facebook
The first wave of direct-to-consumer brands between 2010 and 2015 bypassed retail and traditional advertising entirely, and the winners were not the best funded (TrendTrack, 2026). On Facebook specifically, early adopters who tested new features found outsized wins, and the tactic returning the most six months out was routinely not the one working today (Matchnode, 2020).
A $500 test that proved a concept in a week was worth more than a $50,000 campaign that took three months to plan and three more to measure. Advertisers who treated the platform as an experiment adapted every time it changed. Advertisers who planned quarterly from last quarter’s numbers fell behind, on a schedule.
| Channel | What won | What lost |
|---|---|---|
| Facebook ads, 2012 to 2015 | Small tests on seven-day cycles, losers killed | Quarterly brand campaigns measured after 90 days |
| Google Ads, 2005 to 2010 | Rapid keyword testing, daily bid changes | Annual keyword strategies, monthly reviews |
| AI search, 2025 to 2026 | Publish, measure citations, iterate weekly | Months of dashboards, six articles, wait a quarter |
The pattern holds because it is about the age of the channel. The medium has little to do with it. New channels reward speed. Mature channels reward optimization. AI search is new.
Vercel Iterated From Under 1% to 10% of Signups
Vercel grew ChatGPT referrals from under 1% of new signups in October 2024 to 4.8% in March 2025 and 10% by April 2025 (Chirag Garg analysis of Guillermo Rauch data, 2025). Six months, no rebrand, no acquisition. Its CEO described the work as owning a concept clearly and consistently, with the structure a model can follow.
Tally, a form builder, ran the other end of the same idea. It grew from $2M to $3M in annual recurring revenue in four months after ChatGPT and Perplexity became its largest acquisition channels, which its founder put down to years of showing up and answering questions like a person (Vercel, 2025).
Neither company called this a monitoring strategy. Both described publishing, watching what the models picked up, and adjusting.
When we ran 1,000 queries through Perplexity’s Sonar API, Vercel held the #1 recommendation against Netlify and Cloudflare in 8 of 10 runs (Res, 1,000-query Perplexity study, 2026). That was a reading taken at that time, and citation positions move, which is the point of the next section rather than an asterisk on this one.
The System Randomizes Everything Below Position One
Ask an AI engine the same question ten times and you get a different list most of the time. Across our 1,000-query study a single query surfaced 8.2 unique brands on average, and only 3.1 of them appeared in all ten runs (Res, 1,000-query Perplexity study, 2026).
The instability is not evenly spread. On 75 of 100 queries the same brand held #1 in at least seven of ten runs, while 25 queries had no stable leader at all (Res, 1,000-query Perplexity study, 2026). The top slot is close to earned. The seats behind it are close to a raffle, which is why a program aimed at second place is aimed at nothing in particular.
Practically, that means one check is not evidence. Run a query once, see your page, write it down as a win, and you have recorded a coin landing heads. Ten runs is the smallest reliable sample, and it is cheap enough that taking fewer is a choice.
The wider picture is the same. Between 40% and 60% of domains cited in AI answers change month over month, rising to 70% to 90% across six months (Profound, 2026), and only 11% of cited domains overlap between ChatGPT and Perplexity across 680 million cited URLs (Averi, 2026). Adding statistics to a page lifted AI visibility 41% in the Princeton GEO benchmark, the largest single-tactic gain measured (Princeton KDD, 2024).
You cannot plan your way through that. Nobody can tell you in advance which of your pages will be the one an engine reaches for on a Tuesday, and any vendor who says otherwise is selling a forecast for a coin.
Test the Prompts That Pay, Not the Ones With Volume
Every paid search practitioner already knows this distinction. Top five cars owned by Jay Leno might draw 50,000 searches and sell nothing. Best used SUV under $30,000 near me draws a fraction and books test drives. AI search sorts the same way, so a program needs five to ten prompts that sit on the buyer’s decision. Ten thousand of them across ten engines is noise with a subscription.
| Prompt type | Example | Revenue signal |
|---|---|---|
| Category evaluation | Best tools in a category for a use case | High, the buyer is comparing |
| Direct comparison | One brand against a named rival | High, the buyer is deciding |
| Problem to solution | How to solve the problem you fix | Medium, the buyer is scoping |
| General education | What a broad industry term means | Low, the reader is learning |
| Tangential | The history of a related topic | None |
Watch the core five to ten daily, fan out to thirty for conquest, check those monthly, and let the rest go. A dashboard will happily report a rising number while the pipeline stays empty, because being mentioned is not the same as being bought. Paying to monitor every prompt variation on every engine to publish six articles a quarter is a Bloomberg terminal bought to trade three stocks.
Throughput Is the Constraint, and a Person Is Inside It
A hundred and sixty machine-made pages is not a hundred and sixty chances. It is one bad idea, run a hundred and sixty times.
If iteration decides the outcome, the binding constraint is how many articles you can ship in a quarter. That is a throughput question, and the obvious answer is the wrong one.
The obvious answer is to make the articles cheap. Generate at volume, publish everything, let the engines sort it out. It fails, and not for taste reasons. The models and Google were built to see through volume, so flooding a site with machine-made pages produces a library nothing cites. A hundred and sixty of those is not a hundred and sixty chances to learn.
It is one bad idea, run a hundred and sixty times.
What the engines reward is the one input a model cannot manufacture: somebody who knows the subject, says what is true, and attaches the evidence. Our own count of 852 cited pages found the shape of a page separating the cited from the invisible, and that shape is not hard to produce quickly. It is hard to produce quickly and have a person stand behind every claim in it.
That second condition is the one most vendors quietly drop, because it is the expensive half of the sentence.
So the real constraint was never cost per article.
It is how fast you can ship work somebody will sign.
That is a solvable problem, and the solution is not a faster generator. It is an operator with good tools pointed at the right strategy, which is why Res is a service. Conductor reported a 448% increase in AI citations and a 185% rise in AI mentions over several months running an enterprise program (Conductor, 2025). Programs move when somebody runs them.
Choose a Cadence You Can Actually Sustain
The right cadence is the fastest one you can hold without dropping the human check. Weekly publication with a weekly citation scan gives a program roughly twelve readings a quarter, which is enough to tell a working structural change from a lucky week.
Our own smallest possible test makes the point. Fifteen days after launch, with two articles and no backlinks, Perplexity ranked one of them #1 on “domain authority in AI citations” and this page #7 on “brands winning AI search” beside Search Engine Land, Reddit, Adweek, LinkedIn, and Forbes, against 408 Google impressions and zero clicks in the same window (our day-15 launch proof, 2026).
Two articles, two citations. The curve Vercel rode over six months, run at the smallest scale that still counts as a test.
The reason to hold a cadence rather than a campaign is that a single reading tells you almost nothing. One week a page is cited, the next it is not, and neither week is the truth. Twelve readings a quarter separate the changes that moved something from the ones that happened to land in a good week. That is the difference between running an experiment and telling a story about one.
Where This Sits Against the Other Approaches
Every team spending here settles into one of four postures, and they differ less in price than in how many readings a quarter they generate. The axes that matter are what each one publishes, how fast it learns, and what it cannot see.
| Approach | What it produces | Readings per quarter | The blind spot |
|---|---|---|---|
| Visibility dashboard | A chart of your absence | Many readings, all about other sites | Watching is not testing |
| Traditional SEO | Rankings and links | One or two | Optimizes for clicks that stopped coming |
| Generated volume | Pages, quickly | Many, all the same test | Engines see through volume |
| Human-signed publishing, how Res works | Articles somebody stands behind | Weekly | Needs an operator who knows the subject |
How Res AI Gets 43% of a Launch Cohort to #1
There is no auction here and no budget to outspend. A contractor with 100 crews and a Fortune 500 group get the same hearing from the same engine, and the better, more honest answer takes the slot. You cannot buy the position.
The only variable left is the work.
Two things reliably fail. Flooding a site with generated pages fails, because the engines were built to see through volume. Watching a dashboard fails too, because it reports impressions while the pipeline stays empty.
A chart of your absence has never once ended it.
If you are looking for a vendor who can get you cited, Res runs content programs for companies of every size and category: enterprise software, a Fortune 500 automotive group, a regional contractor with more than 100 crews. Different markets, different buyers, the same program.
Across three of those programs we have scanned 419 published articles inside Google AI Overviews, and 162 of them earned a citation at an average position of 2.4. One client published 76 articles in six weeks from a standing start; 43% became the #1 answer in AI search and 28% took the #1 result on Google, above Microsoft, Gartner, IBM, Anthropic, and OpenAI on the questions that pay.
We do not sell a dashboard. Content goes in, citations come out, and a person signs every page.
Questions
Why does testing beat planning here?
Because the system is non-deterministic. Between 40% and 60% of cited domains change month over month (Profound, 2026), and one query returned 8.2 brands across ten runs with only 3.1 appearing every time. A plan assumes a stable target, and there is not one.
How many prompts should we actually track?
Five to ten that sit on the buyer’s decision, with a fan-out of about thirty checked monthly. The Wrights did not test every shape of wood in Ohio. They tested wings.
Does publishing more articles always mean more citations?
Only if a person stands behind them. Volume alone loses, because the engines were built to see through it. More articles helps when each one is something you would put your name on.
How long before a new program shows a reading?
Weeks. Our own domain earned two Perplexity page-one citations fifteen days after launch with two articles published, and Vercel’s climb from under 1% to 10% of signups took six months of iteration.
Is a citation on one engine worth anything on another?
Not much on current evidence. Only 11% of cited domains overlap between ChatGPT and Perplexity across 680 million cited URLs (Averi, 2026), so an engine has to be won on its own terms.
What single change lifts citations most?
Adding real statistics to the page produced a 41% improvement in the Princeton GEO benchmark, the largest single-tactic gain measured (Princeton KDD, 2024). Attribution and a year matter as much as the number.
What does one test actually look like?
Publish an article built to the cited shape, then run its target query ten times a week later and count how often your page comes back. Ten runs, because one is a coin flip: the same query surfaced 8.2 brands across ten runs in our study, and only 3.1 of them appeared every time.
What should we stop doing?
Paying to monitor thousands of prompts while publishing six articles a quarter. That budget buys readings you cannot act on, and the readings are about other people’s pages.
Res AI is how a team gets more readings per quarter without dropping the human check. Ten articles free, published into the CMS you already run.