The Ghostwriter Trap: When Outsourcing Your Voice Kills AI Citation

Last updated: 2026-10-13

TL;DR

• The trap: ghostwritten founder content fails in AI search when none of its sentences trace back to the founder. Not because engines detect ghostwriting, but because the output converges on a median voice ten other companies also publish [1][2].
• When it works: a ghostwriter working from your recorded interviews, decision logs, or sales calls can be indistinguishable from you. Jasmin Alic told me a VA “can now own your voice and engage as you” if you have taught them your voice [6]. Adam Biddlecombe uses AI for all his LinkedIn posts and said HubSpot came to him through his personal brand [7].
• When it fails: a ghostwriter working from a brief, a keyword list, or a competitor’s top ten. In Muck Rack’s May 2026 analysis of 1 million+ AI-cited links, paid and advertorial content made up 0.3 percent of citations against 27 percent for journalism [4]. Weak content costs twice: the engine passes it over, and the Edelman and LinkedIn 2024 study found 70 percent of C-suite leaders said thought leadership had at least occasionally made them question a supplier [8].
• The test I use: for every claim-level sentence, can you point to the transcript, interview, or decision log it came from? If no, you have decoupled claims, not voice.
• Scope: No study I found compares ghostwritten and founder-written pages head to head in AI citations. The failure modes below are my pattern-spotting across 100+ founder interviews, grounded in adjacent evidence.

In Muck Rack’s May 2026 analysis of over 1 million AI-cited links across ChatGPT, Gemini, and Claude, paid and advertorial content made up 0.3 percent of citations, while journalism made up 27 percent [4].

That gap is the clue. The question is not whether AI engines punish ghostwriting as a category. They do not, and I doubt they can tell. The question is why a specific kind of ghostwritten content keeps missing. In this post I lay out when ghostwriting works, the three failure modes I see when it does not, and the voice-provenance test and 5-item contract I use before I let anyone else write a word under my name. This is the deeper look at the problem introduced in The Founder-Voice Moat.

What is the ghostwriter trap?

The ghostwriter trap is the quiet failure where a founder pays for content that reads cleanly, posts on schedule, and still produces no AI citations and no inbound. The output looks professional. Nothing points to the founder actually having said any of it.

I use one anchor term for this post. The voice-provenance test asks a single question of every claim-level sentence in a published piece: can you point to the recording, transcript, interview, or decision log where this sentence came from? If yes for every claim, the piece has voice provenance. If no, the piece has decoupled claims, specific-looking words that trace to no one in particular. Voice provenance is what separates a ghostwriter working from your interviews from one working from a content calendar, and it is what separates the ghostwritten pages engines retrieve from the ones they skip.

Two limits up front. Google’s guidance says generating pages that add no value may violate its spam policy, but it does not single out authorship, and its generative-AI page encourages manual fact-checking rather than banning AI-assisted work [10]. And the voice-provenance test is my own framing, not an industry standard.

When does ghostwriting actually work?

Often, if the raw material is yours. The strongest counter-evidence to a naive “no ghostwriters” rule came from two of my own guests.

Jasmin Alic (Founder, Link Up) coaches founders on LinkedIn. He told me that if you cannot spare 30 minutes a day, hire a virtual assistant, because “they can now own your voice and they can engage as you” [6]. He was fine with outsourcing engagement and even batching the writing. His hard line was elsewhere: do not let AI do your thinking. Jasmin was describing a VA who had absorbed enough of a founder’s conversations and preferences to produce sentences that trace back to the founder’s own positions. That is voice provenance, delivered by a human writer who is not the founder.

Adam Biddlecombe (Founder, Mindstream) ran a newsletter that signed off every issue as written by humans. He then told me he uses AI to write all of his own LinkedIn content [7]. Adam also told me that HubSpot said it loved Mindstream’s editorial when it came knocking, and that the person who introduced him to HubSpot came through his personal brand [7]. His AI prompts carry his takes on the news he actually reads that day. The claims trace back to him.

Both cases pass the voice-provenance test. The pen is in someone else’s hand. The source is still the founder. Semrush’s March 2026 study of 89,000 LinkedIn URLs cited by AI search found that individual members made up 59 percent of LinkedIn citations on ChatGPT Search and Google AI Mode, and that about 95 percent of cited LinkedIn content was original, not reshared [13]. Engines retrieve individual-named content at scale, no matter who typed the words, as long as the content has source behind it.

The research backs this up with limits. Graphite’s October 2025 study found 86 percent of Google results and 82 percent of ChatGPT and Perplexity citations were human-written, but it did not evaluate heavily edited AI-assisted content and said that approach may work [3]. Google’s helpful-content guidance targets content without original research or added value, not AI assistance as such [11]. And AI-text detectors, the tool a strict rule would depend on, have been shown to misclassify non-native English writing as machine-generated [5], which is one reason I think no engine seriously tries to detect ghostwriting at the sentence level.

Who holds the pen matters less than where the words came from.

Three ways the trap closes

Here is where I see it fail, in rough order of how often I meet each one. Each failure mode has adjacent evidence; none has a head-to-head study.

1. Median voice: templates converge on the same page

A ghostwriter who takes a brief, a keyword list, and the top ten competitor posts as inputs produces the median post for that query. So does an AI model prompted on the same inputs. The result is a page a dozen companies could publish, so an engine has no reason to quote yours. In a 2024 Science Advances experiment with 293 writers, AI-assisted stories scored higher with evaluators but were more alike than human-only stories, and the gain in individual creativity came with a measurable drop in collective diversity [1]. A 2025 PNAS study by Reinhart and colleagues found LLM writing differs systematically from human writing in grammar and rhetoric and that the gap persists as models get larger [2]. Those two papers describe the mechanism, not the citation outcome, but the direction is consistent with Muck Rack’s finding that paid and advertorial content earns a 0.3 percent share of AI citations [4]. When a page says what ten others already say, there is nothing for an engine to lift.

2. Decoupled claims: specific words that came from nowhere

The second failure mode is more subtle, and it is the one that confused me when I first saw it. A ghostwriter adds names, numbers, and anecdotes to a draft to make it sound specific. None of them came from the founder. Sometimes the writer invented a plausible example to fill a hole. Sometimes an AI model hallucinated a stat that sounds right. Sometimes the writer paraphrased a competitor’s case study and swapped the logo. On the page the claim reads like firsthand knowledge. Press on it in a sales call, and no one can source it. That is a decoupled claim. It is also the fastest way to burn trust with the reader who arrives after the engine, because in the Edelman and LinkedIn 2024 study 70 percent of C-suite leaders said thought leadership had at least occasionally made them question whether to keep working with an existing supplier [8]. Specific-looking content without a source is worse than generic content, because the moment a buyer asks “where did you see that,” the answer collapses.

3. Stage-managed authenticity: the “humans wrote this” badge without the specifics

The third failure mode is the hardest to spot because it tries to look like the fix. A page carries a “written by a human” banner, or a founder photo, or an “unlike other newsletters, we are real people” preface. Then the body is still a brief-driven template. The badge is signal theater: a credential layered on top of content that would still fail the voice-provenance test if you ran it. AI engines do not reward “human” as a label. In one benchmark, neural retrievers actually ranked LLM-written text above human-written text [12]. Google’s guidance asks who created the content, how, and why, and treats first-hand experience as a quality signal [11], but the signal lives inside the sentences, not inside a disclosure line.

The pattern across all three: the pen is irrelevant. The source trail is the thing. The post that passes is the one whose claims point somewhere a buyer, an auditor, or a model could follow.

What is the voice-provenance test?

A one-question test you can run on any draft before it ships.

For every claim-level sentence in this piece, can I point to the recording, transcript, interview, or decision log where it came from?

If every sentence passes, you have voice. If a batch of sentences cannot be sourced, you have median voice or decoupled claims and you fix those sentences before you publish. The test does not care who typed the words. It cares whether the words trace to the founder.

The test is cheap to run. The hard work is keeping the source material in the first place: recording your calls, saving your Slack notes, keeping decision logs, publishing transcripts of the podcast or video interviews you do. That is why the Founder-Voice Moat compounds. The raw material is a stock you add to every week whether or not you ever write. Rodrigo Stockebrand’s O’Reilly book on answer engine optimization makes the related point from the engine side: “in AI search, wide beats deep. To win a top mention on a given query, you usually need to show up across many cited sources, not just on your own pages” [14]. Voice provenance is what makes a founder’s material usable across all those sources without going median.

What a sentence looks like	What it traces to	Voice-provenance verdict
”Companies are adopting AI to unlock operational efficiency.”	A keyword tool and three competitor posts	Fails (median voice)
“In Q3 we cut our support deflection goal by 40 percent after a customer told us the AI apologies were making complaints worse.”	A customer email saved October 7 and the Q3 strategy doc	Passes
”Our research shows most founders waste 12 hours a week on social media.”	No source cited; the number was invented to fit the hook	Fails (decoupled claim)
“Jasmin Alic told me, ‘writing is thinking,’ and compared outsourcing it to someone else lifting your weights at the gym.”	The published episode transcript at 38:28	Passes
”As real humans, we care about our readers.”	A disclosure line, not a sentence that carries a claim	Not applicable (not a claim)

The 5-item ghostwriter contract

If you are going to use a ghostwriter, a VA, or an AI drafter, give them this rubric before the first draft. I use a version of it with every writer who touches our pages.

1. Source every claim-level sentence. Deliver the recording, transcript, or decision log behind every number, story, or opinion. If a claim has no source, cut it or ask the founder for 10 minutes to produce one.
2. Record the founder on the topic before the first draft. Thirty minutes of the founder talking out loud is the cheapest and highest-leverage input you have. In the KDD 2024 GEO paper, adding quotations and statistics was among the techniques that produced the largest visibility gains on a benchmark, up to 40 percent [9]. The transcript is where quotations come from.
3. Prompt from the founder’s words, not from a template. If an AI model is in the loop, its input is the founder transcript, the founder’s past posts, and the founder’s own notes. Not a style guide and three competitor URLs. Jasmin’s exercise is a useful discipline: whenever you prompt an AI to write something, “write it yourself as well. Then mix and match” [6].
4. Keep the artifact. Transcripts, recordings, decision logs, and the first-pass draft stay in a shared folder tagged to the piece. Six months later, when a buyer asks “where did you see that,” you can answer in under a minute. This is also where positioning work stays sharp over time: the folder is the record of how your story has changed.
5. One-sentence voice check before publish. The founder reads the draft out loud and asks, “Could I say this in a meeting without feeling like a stranger?” If no, rewrite. If yes, ship. This is the single step that catches stage-managed authenticity, because posture does not survive being spoken.

Pin the contract to the brief. If a writer or a tool cannot work this way, that is useful information too. A ghostwriter who refuses to show sources is a template vendor with a founder photo on top. A ghostwriter who insists on recordings and keeps the artifact is a leveraged version of you. The second is the one worth paying for.

So, when does outsourcing your voice kill AI citation?

When nothing on the page traces back to you. The pen is a distraction. The source trail is the whole thing. Median voice converges by construction [1][2], paid and advertorial content shows up at a tiny fraction of citations [4], and the reader who arrives after the engine still asks where you saw that [8]. The founders who keep showing up in AI answers have content with voice provenance, whether they typed it, dictated it, or handed it to a writer who worked from their recordings. That is also what makes a strong voice compound into AI presence over time, which the citation gap post and what AI engines hear when a founder speaks treat from the engine side.

If you are still not sure what your voice sounds like on the page, start with How to Find Your Founder Voice and work backward from a recorded conversation. The pen can go to anyone. The source stays with you.

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

• Ghostwriter trap (also called: ghostwriting failure, outsourced-voice failure). The quiet failure where a founder pays for content that reads cleanly, posts on schedule, and still produces no AI citations because none of its sentences trace back to the founder.
• Voice provenance (also called: source trail, voice traceability). The trail from a published sentence back to a recording, transcript, interview, or decision log the founder actually produced.
• Voice-provenance test (also called: provenance check, source-trail check). A one-question check run on every claim-level sentence before publish: can you point to the recording, transcript, interview, or decision log it came from?
• Median voice (also called: template voice, AI-default voice). Writing that converges on the statistical middle because it comes from the same brief or the same few models.
• Decoupled claims (also called: unsourced specifics, orphan facts). Sentences that read as specific but trace to no source, often added by a writer or model to fill a hole.
• Stage-managed authenticity (also called: authenticity theater, human-badge content). Content that layers “written by a human” signals on top of a template, without the sentences themselves carrying founder-sourced material.

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

Is it OK to use a ghostwriter for founder LinkedIn content? Yes, if the ghostwriter works from your recorded conversations and can source every claim. Jasmin Alic told me a VA “can now own your voice and they can engage as you” when they have absorbed enough of the founder [6]. The practice that fails is a ghostwriter working from a brief and competitor posts, which produces a median voice [1][2].

Can AI engines detect ghostwritten or AI-generated content? I found no vendor documentation saying they try at the sentence level. AI-text detectors, the obvious tool, have been shown to misclassify non-native English writing as AI-generated [5], so an engine that leaned on them would punish real people. What engines can observe is the page itself: how specific it is and whether other sources corroborate it [4].

Does AI-assisted writing hurt my AI search visibility? Not on its own. Graphite’s October 2025 study found human-written articles made up 86 percent of Google results and 82 percent of ChatGPT and Perplexity citations, but it explicitly did not test heavily edited AI-assisted content and noted that approach may work [3]. Google’s guidance warns against generating pages that add no value, not AI assistance as such [10]. Adam Biddlecombe uses AI for all his LinkedIn posts and told me HubSpot came to him through his personal brand [7].

What is a voice-provenance test and how do I run it? It is one question you ask of every claim-level sentence in a draft: can you point to the recording, transcript, interview, or decision log it came from? If every claim passes, the piece has voice provenance. If a batch of claims cannot be sourced, you have median voice or decoupled claims, and those sentences get rewritten before publish.

How can I tell if my ghostwriter is working from my words or from a template? Ask for the sources behind every number and story in your last three posts. A ghostwriter working from your interviews can produce the recordings or transcripts in under a day. A template-driven writer will reach for “research” or “industry data” without pointing to a specific input you gave them. The 5-item ghostwriter contract above is what I hand new writers.

What should go in a ghostwriter brief for AI search? A recording of the founder on the topic, a list of the founder’s past posts, decision logs or customer notes, and the single thesis the piece has to carry. Not a keyword list or a top-ten competitor sweep as the primary input. In the KDD 2024 GEO paper, adding quotations and statistics produced some of the largest visibility gains, up to 40 percent on a benchmark [9]. Those come from a transcript, not a style guide.

Is a “written by humans” disclosure enough to signal authenticity to AI engines? No. In one benchmark, neural retrievers ranked LLM-written text above human-written text [12]. Google’s helpful-content guidance treats first-hand experience as a quality signal and asks who created the content, how, and why [11], and that signal lives inside the sentences, not inside a disclosure line.

I am a non-native English speaker. Will my writing get flagged as AI? Detectors have misclassified non-native English writing as machine-generated [5], which is one reason I do not think polish is the real signal. Voice is specificity: your story, your number, your phrase. Those travel across any accent.

References

[1] 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

[2] 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

[3] Graphite (2025). “How Does AI-Generated Content Perform in Search and Answer Engines?” Graphite, October 14, 2025 (31,493 keywords; AI detection by Surfer; English articles and listicles only). https://graphite.io/five-percent/ai-content-in-search-and-llms

[4] 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

[5] 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

[6] TwoSetAI (2026). Jasmin Alic (Founder, Link Up) episode, recorded March 8, 2026; quotes at 30:39, 38:28, and 39:36. https://www.youtube.com/watch?v=aIKzImj6m8w

[7] TwoSetAI (2026). Adam Biddlecombe (Founder, Mindstream) episode; quotes at 15:43, 39:23, and 39:48. https://www.youtube.com/watch?v=cV8CYk2YvsA

[8] O’Dwyer’s (2024). “Thought Leadership Can Power B2B Brands.” O’Dwyer’s PR, February 29, 2024, reporting the sixth annual Edelman and LinkedIn B2B Thought Leadership study (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

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

[10] 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

[11] Google (2026). “Creating helpful, reliable, people-first content.” Google Search Central, page updated October 5, 2026. https://developers.google.com/search/docs/fundamentals/creating-helpful-content

[12] Dai, S. et al. (2023). “LLMs May Dominate Information Access: Neural Retrievers are Biased Towards LLM-Generated Texts.” arXiv:2310.20501. https://arxiv.org/abs/2310.20501

[13] 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/

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

Methodology note. Last verified 2026-10-13. The three failure modes and the voice-provenance test are my framings from hosting 100+ founder conversations, not coded data. No study I found compares ghostwritten and founder-written pages head to head in AI citations, so the thesis rests on adjacent evidence (convergence findings in [1][2], paid-content citation share in [4], AI-detector bias in [5]) and is labeled as pattern-spotting where it goes beyond that. Guest quotes are verbatim from the published episodes cited, with punctuation and capitalization normalized and timestamps taken from the episode transcripts. Several third-party studies (Semrush, Muck Rack, Graphite) come from companies that sell search, PR, or AI-visibility products, so I treat their numbers as directional. The Edelman and LinkedIn figures are cited through O’Dwyer’s report of the study [8]. Google’s pages were read on October 12, 2026 and may change. “Ghostwriter trap,” “voice-provenance test,” “decoupled claims,” “stage-managed authenticity,” and “median voice” are TwoSetAI’s own terms, not industry standards. Nothing here guarantees a citation, a ranking, or a business outcome.