How to Find Your Founder Voice (And Why AI Search Rewards It)

Last updated: 2026-10-13

TL;DR

• The exercise is five steps anchored to a recorded conversation: record for 30 minutes, extract 3 to 5 one-sentence theses, run Rob Kaminski’s stranger-decode test, do an “AI versus you” comparison, and build a voice-provenance file of named stories, specific numbers, and decisions you reversed.
• Why bother: individual members, not company pages, made up 59 percent of LinkedIn citations on ChatGPT Search and Google AI Mode in Semrush’s March 2026 study of 89,000 cited URLs [1]. In the KDD 2024 paper that introduced generative engine optimization, adding quotations and statistics produced some of the largest visibility gains, up to 40 percent on a benchmark [2].
• Templates converge: in a Science Advances experiment with 293 writers, AI-assisted stories were rated more creative but were more alike [3], and a PNAS paper found LLM writing differs systematically from human writing in grammar and rhetoric [4].
• First-pass time cost: 60 to 90 minutes if you have never done it, plus a 20-minute weekly refresh.
• What this method is: specific-first content hygiene. No engine publishes a “founder voice” ranking factor, and I found no head-to-head study of founder-voice versus ghostwritten pages. What the method does is give engines the specific, attributable sentences they tend to extract.

In Semrush’s March 2026 study of 89,000 LinkedIn URLs cited by AI search, individual members made up 59 percent of LinkedIn citations on ChatGPT Search and Google AI Mode [1]. If you are a pre-Series B AI founder who has been heads down developing your product, that number is the problem and the opening. The engines are already quoting named people on your topic. The question is whether the named person is you.

This spoke is the method referenced in step four of The Founder-Voice Moat. The pillar argues that organic founder voice beats ghostwriters in AI search because engines retrieve what is specific, sourced, and hard to find anywhere else. This post is how you produce that material. It is not a brand workshop. It is a one-afternoon exercise you can run with a laptop, a microphone, and a transcript.

Why does AI search reward a specific voice?

Three findings I keep coming back to. First, individuals get cited: on LinkedIn, 59 percent of citations on ChatGPT Search and Google AI Mode went to members, not company pages, and about 95 percent of cited content was original rather than reshared [1]. Second, paid and advertorial content barely registers: Muck Rack’s May 2026 analysis of 1 million-plus AI-cited links put paid at 0.3 percent of citations against 27 percent for journalism [5]. Third, specifics travel: in the KDD 2024 paper, adding quotations and statistics to a page gave some of the largest visibility gains on their benchmark, up to 40 percent [2].

None of this proves an engine detects a human. Detectors have been shown to misclassify non-native English writers as machine-generated [6], so an engine that leaned on them would punish real people. What the evidence supports is narrower: a page with a named person saying something only they could say is a page an engine has a reason to pick. See What AI Engines Hear When a Founder Speaks for the signal argument. My point here is the production side.

What is the five-step method?

Step 1: Record a conversation, not a doc

Do not open a blank page. Open a recording app. Ask a smart friend, your co-founder, or an interviewer to sit with you for 30 minutes and ask you a handful of specific-to-you questions (the three below are a starter kit I use; swap in your own):

1. Who was your first paying customer, and how did you find them?
2. What is the one question you ask every prospect that most of your competitors do not?
3. What is a decision you reversed in the last 90 days, and what changed your mind?

Talk, do not pitch. Transcribe the recording the moment it is done. The transcript, not the recording, is the raw material. The reason to start here is mechanical: in a conversation you say specific things in specific words that you would never type into a doc. Jasmin Alić (Founder, Link Up) calls writing a form of thinking, and compared outsourcing it to “trying to go to the gym and having someone else lift the weights for you” [7]. The conversation is where the thinking happens; the transcript is where you can see it. If you want the longer argument for interviewing yourself first, The Quote That Carried the Episode walks through how I mine my own podcast conversations the same way.

Step 2: Extract 3 to 5 one-sentence theses

Read the transcript once, out loud, with a highlighter in hand. Mark every sentence that could stand alone as a claim: a line that would make sense quoted on its own, without the sentence before it. Aim for 3 to 5 after you cut duplicates.

AI can help here, but only up to a point. Use it to narrow the field: paste the transcript in and ask for every sentence longer than fifteen words that contains a proper noun, a specific number, or a declarative claim. That gives you a short list of candidates in a minute instead of thirty. The final pick is yours. Shankar and Husain’s O’Reilly book on evals argues that the first failure analysis of your own system has to be yours, because an AI will cluster around its own priors instead of the actual signal in the data [16]. The same applies to a founder transcript. Your brain carries context the model does not: which customer meeting that line came out of, which objection kept coming back in sales calls, which decision changed because of it. That tacit filter is what makes a sentence quotable. Let AI shortlist; you decide which three to five are the actual theses.

A thesis is not a slogan. It is a sentence a stranger could repeat in a meeting and have the room understand. Hector Hernandez (Co-founder and COO, Speakeasy) gave me one in a single breath: “the first thing that breaks when people scale their use of AI is often not cost,” and “most of the time it’s complete loss of visibility of what’s going on” [8]. That is one claim from one person, specific enough to lift whole. If your highlights look like “we help teams move faster” or “AI is transforming workflows,” you are at category jargon, not a thesis. Go back to the transcript and find the sentence underneath.

Step 3: Run the stranger-decode test

Rob Kaminski (Co-Founder, Fletch PMM) has worked with more than 600 companies on positioning. On my episode he stress-tested a pre-launch AI agent product with a starter kit of questions I now use on every thesis I write (yours can grow from here; use them as examples, not a fixed list):

“Does your audience know what that means? Do they know what they’d use it for? Do they know who the players are? And if the answer is no, it is a signal that we have not found that common ground.” [9]

Take each of your 3 to 5 theses and run them past someone outside your company who does not use your product. A friend in a different industry, a candidate you interviewed last month, an investor you have met once. Ask them in order: does the sentence tell you what it is, who it is for, and what makes it different from the alternatives? If any of the three answers is no, the sentence is still in template territory. Rewrite it with a specific noun. “AI agents for enterprise” becomes “a Chrome extension that answers sales emails for the inbound team at a 30-person series A.” The second is a sentence; the first is a category.

The reason this test matters for AI search is that a sentence a stranger can decode is also one an engine can match to a buyer’s question. The watch page is here if you want to hear the whole audit. For why positioning precedes voice at all, see the positioning pillar.

Step 4: Do the “AI versus you” comparison

Pick one of your decoded theses. Open an AI model, any model. Prompt it: “Write a 300-word LinkedIn post about [your topic] in [your category].” Save the output in a document. Do not read it yet.

Now open a second document. Write your own 300-word version. Draw from the recorded conversation. Name the customer. Give the number. Describe the morning the decision came out of. Then read both side by side.

Three things will be visible immediately. The AI version will have category nouns (“teams,” “workflows,” “platforms”). Yours will have proper nouns (“our first customer was a 12-person RevOps team in Austin”). The AI version will have round numbers (“many companies,” “most teams”). Yours will have specific ones (“3 of our last 7 pilots stalled on the same objection”). The AI version will quote no one. Yours will quote someone, by name, from a conversation that actually happened.

Jasmin’s rule for the exercise is the one I use with founders now: whenever you prompt an AI to write something, “write it yourself as well. Then mix and match” [10]. You can hear him at 38:28. The point of the comparison is not to prove AI is worse. It is to see where the median voice is so you can stay off it. Reinhart’s team at PNAS found that LLM writing differs systematically from human writing in grammar and rhetoric, and that the gap persists as models get larger [4]. The Doshi and Hauser experiment in Science Advances found AI-assisted stories were rated more creative but were more alike than human-only ones [3]. If your draft reads like the AI draft, your draft will converge with everyone else prompting the same model on the same topic. An engine has no reason to pick yours over the next ten.

Step 5: Keep extracting, from the business itself

Step 5 is the same exercise as Steps 1 to 4, applied continuously.

Steps 1 to 4 train the extraction muscle on one planned 30-minute recording so you can feel how it works: a smart friend asks you questions, you talk, you pull theses out of the transcript afterward. Once that muscle is built, you do not need to schedule a new recording to keep doing it. Every sales call, every customer meeting, every advisor conversation is already a mini-interview. You keep extracting from those as the business runs, and you log what you pull into a running file so it accumulates.

Two or three entries a week, each a sentence or two, two or three minutes to write. After a month you have 10 to 15. Over time this file becomes the single place your future posts, quotes, and buyer answers get drawn from.

I call it a voice-provenance file because every published sentence should trace back to a specific moment logged here. That trail is what makes the sentence provably yours and not something a template could produce.

The columns below are a starter kit, not THE structure. Pick the buckets that match what your week actually produces. If your material is mostly customer stories, those stay. If it is mostly objections you keep hearing, add that as a bucket. The three I use most often:

Column (starter kit)	What goes in it	Example (illustrative)
Named stories	Customer anecdotes with a name, a date, and one lesson	”The RevOps lead at Pilot Co, March 2026, who said she would pay for the product if it answered one specific question before every call.”
Specific numbers	Numbers from your own data, not from a press release	”3 of our last 7 pilots stalled on the same objection in Q3 2026.”
Decisions reversed	Something you changed your mind on and why	”Dropped the self-serve tier in June 2026 after watching four signups churn in the first week.”

Other reasonable buckets: objections you keep hearing, metaphors you reach for, questions you keep asking in meetings, specific product bets you made.

The material comes from conversations you were going to have anyway, so the per-week cost is low. Each entry is a potential post, a potential quote, and a potential answer to a buyer’s question. When a ghostwriter, an interviewer, or a model needs to produce copy, they draw from this file. Nothing in the file came from a brief or a competitor’s blog. That is what makes it a moat.

A quick note on ghostwriters, because this is where most founder-voice conversations get stuck. The usual question is “should I hire a ghostwriter or write it myself?” The more useful question is “where did the words come from?” A ghostwriter working from this voice-provenance file is working from your material, so the output traces back to you. A ghostwriter working from a content calendar and the top 10 LinkedIn posts in your category produces the median voice. The Ghostwriter Trap walks through when ghostwriting works and when it breaks.

How much time does this take?

The first full pass takes 60 to 90 minutes. 30 minutes to record, 15 to transcribe (automatic plus a cleanup pass), 15 to extract theses, 15 for the stranger-decode test over a phone call, 15 for the AI-versus-you comparison on one thesis. The voice-provenance file takes about 20 minutes a week after that, because the material comes from conversations you were going to have anyway.

That is less time than most founders spend reviewing a ghostwriter’s drafts.

What if I am a non-native English speaker, or shy on camera?

Specificity is the signal, not polish. AI-text detectors have been shown to misclassify non-native English writers as machine-generated [6], which is one reason I do not think polish is what engines reward. Record in whatever language the conversation happens in. Keep the specific nouns through translation: the names, roles, numbers, and places that make the sentence yours. A translator can smooth the connective tissue; the proper nouns and the hard numbers carry the specificity on their own.

If you are shy on camera, do audio only. The transcript is the asset. You can turn an audio conversation into a written page, chapters, a YouTube description, and a LinkedIn post without ever showing your face. See the video pillar for how the long form becomes the retrieval surface.

How do I know I have a voice, not just vocabulary?

A simple test. Hand your voice-provenance file to someone who has never met you. Ask them to write a 150-word LinkedIn post from any three entries. Read it. If the post could be about any company in your category, your entries are still too abstract. Go back to the recording and ask: what was the name, what was the number, what changed. If the post could only be about your company, you have voice.

The deeper validation is that AI engines retrieve what is quotable. In the KDD 2024 GEO paper, pages with inline quotations and statistics gained up to 40 percent on their visibility benchmark [2]. In Stockebrand’s O’Reilly book on answer engine optimization, the Vercel case shows the same pattern from the outside in: Vercel’s AI visibility is not driven by its own getting-started docs but by the Reddit recommendations, Medium posts, YouTube videos, and third-party tutorials that reinforce the same specific message about the platform [11]. Specifics travel across the web, and the engines pick them up where they land.

So, how do you find your founder voice?

You find it by recording yourself answering real questions, extracting the sentences a stranger could repeat, cutting the ones that could belong to anyone, writing your own version alongside the AI’s, and keeping a running file of the named stories, numbers, and decisions that only you have. The method does not need a brand consultant. It needs a transcript and 90 minutes.

The research here is correlational, the audits are directional, and the systems change monthly. Specifics are the one input a template cannot copy. If your sentences trace back to a conversation you really had, you are building the material the next generation of AI retrieval was designed to find. That is the Founder-Voice Moat. This spoke is the method. The pillar is where the evidence lives: The Founder-Voice Moat.

Who is TwoSetAI?

TwoSetAI is one of the top AI podcast and video studios in Silicon Valley, and the team behind this article, hosted by Angelina Yang. The show has featured 100+ AI founders and leaders on how real AI products get built, sold, and trusted. Alongside the podcast, TwoSetAI runs the Founder Visibility Program that turns each conversation into the articles, clips, and pages that make its guests easy to find. Watch full episodes on YouTube at youtube.com/@TwoSetAI, and find every guest and episode at twosetai.com.

Glossary

• Founder voice (also called: founder-led voice, organic founder voice). The specific way a founder describes their product, customers, and decisions when they are talking rather than writing to a template.
• Voice provenance (also called: source trail, voice traceability). The trail from a published sentence back to something the founder actually said, did, or decided.
• Stranger-decode test (also called: outsider legibility check, decode test). Four positioning questions from Rob Kaminski: does an outside listener know what it is, who it is for, who the players are, and what makes it different.
• AI versus you comparison (also called: median-voice calibration, AI contrast exercise). Prompting an AI for a post on your topic, writing your own in parallel, and reading them side by side to see which specifics only you have.
• Median voice (also called: template voice, AI-default voice). Writing that converges on the statistical middle because it comes from the same few models or the same brief.

About Angelina Yang

Angelina Yang is VP of AI, a fast.ai fellow, and the host of the TwoSetAI podcast, where she has interviewed 100+ AI founders and leaders about how real AI products get built, sold, and trusted. Find her on LinkedIn, talk to her AI clone on Delphi, watch the podcast on YouTube, or browse twosetai.com.

Frequently Asked Questions

How do I find my founder voice if I have been heads down developing my product? Start from a recorded conversation, not a blank page. Ask a smart friend or your co-founder to ask you a handful of specific-to-you questions (a starter kit I use: who was your first paying customer, what question do you ask every prospect, what did you reverse in the last 90 days; swap in your own). Transcribe the audio. Highlight 3 to 5 sentences that could stand alone as claims. Those are your first theses.

How long does this take the first time? 60 to 90 minutes end to end: 30 minutes to record, 15 to transcribe and clean up, 15 to extract theses, 15 for a stranger-decode call, 15 for one “AI versus you” comparison. The weekly voice-provenance update takes about 20 minutes.

Can I use AI to write my LinkedIn posts and still have a founder voice? Yes, if your source material is yours. Google says generative AI can help with research and structure, and warns against generating pages that add no value for users [12]. Jasmin Alić’s rule is to write it yourself as well, then mix and match [10]. If the post carries your story, your number, and your decision, the tool matters less than the file it drew from.

Do non-native English speakers risk sounding like AI? Detectors have misclassified non-native English writing as machine-generated [6], which is one reason I do not think polish is the signal. Record in your source language and keep the specific nouns through translation. A translator can smooth the connective tissue; the proper nouns and the hard numbers carry the specificity on their own.

What is the stranger-decode test, exactly? Four positioning questions from Rob Kaminski: does the person listening know what the product is, who it is for, who the alternatives are, and what makes yours different. If the answer to any of them is no, the sentence is still in category jargon [9].

How do I know my voice is working in AI search? Rerun your name and your category queries in ChatGPT, Perplexity, Google AI Mode, and Claude on a monthly cadence. In SparkToro’s 2026 study of 2,961 runs, ChatGPT and Google’s AI returned the same brand list in fewer than 1 in 100 runs, so track how often you appear, not where you rank [13].

Is this the same as a personal brand? No. A personal brand is an audience. The Founder-Voice Moat is the content that neither the audience nor the engines can get anywhere else. One can exist without the other, but they help each other.

What if I do not have time to record a conversation? Record the next sales call you take, with the prospect’s permission. Record the next investor meeting. Record the next customer onboarding. The method does not require a special session. Any 30-minute conversation about your product, with specifics in it, is enough to extract 3 to 5 theses from.

References

[1] Loktionova, M. (2026). “We Analyzed 89K LinkedIn URLs Cited in AI Search: Here’s What Drives Visibility.” Semrush Blog, March 10, 2026 (325,000 prompts, January to February 2026; ChatGPT Search, Google AI Mode, Perplexity; run with LinkedIn’s collaboration). https://www.semrush.com/blog/linkedin-ai-visibility-study/

[2] Aggarwal, P. et al. (2024). “GEO: Generative Engine Optimization.” KDD 2024; arXiv:2311.09735. https://arxiv.org/abs/2311.09735

[3] Doshi, A. R. and Hauser, O. P. (2024). “Generative AI enhances individual creativity but reduces the collective diversity of novel content.” Science Advances 10(28), eadn5290, July 12, 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11244532

[4] Reinhart, A. et al. (2025). “Do LLMs write like humans? Variation in grammatical and rhetorical styles.” PNAS 122(8), e2422455122, February 25, 2025; arXiv:2410.16107. https://arxiv.org/abs/2410.16107

[5] Muck Rack (2026). “How AI citations have changed in the last 6 months: New insights from ‘What is AI Reading?’” Muck Rack, May 2026 edition, page updated May 27, 2026 (1 million+ links cited by ChatGPT, Gemini, and Claude). https://muckrack.com/blog/what-is-ai-reading-new-insights-may

[6] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). “GPT detectors are biased against non-native English writers.” Patterns, 2023; arXiv:2304.02819. https://arxiv.org/abs/2304.02819

[7] TwoSetAI (2026). Jasmin Alić (Founder, Link Up) episode, recorded March 8, 2026; quote at 30:39. https://www.youtube.com/watch?v=aIKzImj6m8w

[8] TwoSetAI (2026). Hector Hernandez (Co-founder and COO, Speakeasy) episode, published October 5, 2026; quote at 22:32. https://www.youtube.com/watch?v=yhOLsoqSSEY

[9] TwoSetAI (2026). Rob Kaminski (Co-Founder, Fletch PMM) episode, published September 10, 2026; quote at 23:00. https://www.youtube.com/watch?v=SD64dL7yg7I

[10] TwoSetAI (2026). Jasmin Alić (Founder, Link Up) episode; quote at 38:28. https://www.youtube.com/watch?v=aIKzImj6m8w&t=2308s

[11] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026 (Vercel case in Chapter 1). https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[12] Google (2026). “Google Search’s guidance on using generative AI content on your website.” Google Search Central, page updated October 1, 2026. https://developers.google.com/search/docs/fundamentals/using-gen-ai-content

[13] Fishkin, R. and O’Donnell, P. (2026). “New Research: AIs are highly inconsistent when recommending brands or products.” SparkToro, January 28, 2026 (600 volunteers, 12 prompts, 2,961 runs). https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/

[14] TwoSetAI (2026). Guest search audit, checked August 10 to 25, 2026, incognito, screenshots on file. Internal data; protocol in the parent pillar’s methodology note.

[15] Edelman and LinkedIn (2024). “2024 B2B Thought Leadership Impact Report.” Reported via O’Dwyer’s PR, February 29, 2024 (nearly 3,500 management-level professionals; fieldwork December 2023). https://www.odwyerpr.com/story/public/20925/2024-02-29/thought-leadership-can-power-b2b-brands.html

[16] Shankar, S. and Husain, H. (2025). Evals for AI Engineers. O’Reilly Media (early release). On open coding and first-failure analysis: the human reviewer carries context an LLM reviewer cannot reproduce. https://www.oreilly.com/library/view/evals-for-ai/9798341660717/

Methodology note. Last verified 2026-10-13. The five-step method in this post is an operational distillation of what I have observed across 100+ founder conversations on the TwoSetAI podcast, combined with the specific exercises I have borrowed from Jasmin Alić (the “write it yourself as well, then mix and match” rule) and Rob Kaminski (the four-question stranger-decode test). It is not a controlled study. I have not run a treatment-versus-control test comparing founders who use the method against those who do not. The claims that AI engines reward specifics rest on the KDD 2024 paper, the Semrush and Muck Rack citation studies, and the Doshi and Reinhart findings on convergence; each is directional and some come from vendors who sell adjacent products, so I treat the numbers as directional, not causal. Guest quotes are verbatim from the published episodes cited, with punctuation and capitalization normalized and timestamps taken from the episode transcripts. The example rows in the voice-provenance table are illustrative, not from a real customer. “Founder-Voice Moat,” “voice provenance,” “stranger-decode test” (as a labeled exercise), “AI versus you comparison,” and “median voice” are TwoSetAI’s own terms, not industry standards. Nothing here guarantees a citation, a ranking, or a business outcome.