A chatbot nearly killed a $50 million penthouse deal before a single human ever picked up the phone. The buyer asked ChatGPT whether the price was too high, and got told yes. The seller asked the same question of the same tool and was told to hold out for more. Both parties nearly walked away from the table simultaneously, steered by one AI that handed each side a different answer. That is the shape of the new market, and Marina Mogilko, the entrepreneur and podcast host behind Silicon Valley Girl, decided to stop watching it happen to others and start reverse-engineering it for herself.
The stakes, she found, are larger than most creators or founders have registered. According to her research, 35% of people in the US now begin shopping inside an AI assistant rather than a search engine. Visitors arriving from ChatGPT convert at 15.9%, compared to 1.76% for visitors coming from organic Google. The math is not subtle. If the AI does not name you, the sale is already over.
When 50 interviews and a famous name still left you invisible
Mogilko had spent years building Silicon Valley Girl into a podcast with genuinely notable guests: the CEO of LinkedIn, the founder of LinkedIn appearing twice, the CEO of GitHub, and the founder of Perplexity, among others. Then she ran a simple test, asking ChatGPT, Claude, and Perplexity to recommend AI podcasts. Her show did not appear once. ‘We were never ever on the list,’ she said, ‘and knowing how people convert when ChatGPT recommends them something, we decided to run a full GEO.’ GEO, short for generative engine optimization, is the practice of structuring your online content so that AI crawlers can find, read, and cite it.
The first diagnostic was brutally clarifying. Her team pasted the site URL into Claude and asked a single question: why is this site invisible to AI search? The answer came back as a specific list of failures. HTML parameters that AI crawlers require were missing. Podcast transcripts were hidden behind JavaScript, so when a bot arrived, it saw an empty shell. To a human visitor, the site looked fine. To an AI reading the source code, it was a blank page.
The five-step rebuild, in the order it actually happened
Step one was that diagnostic prompt. Before anything else, Mogilko recommends pausing and pasting your own URL into a chatbot with a direct question about AI visibility and what is blocking it. The output becomes the specific road map for your site.
Step two was building a crawlable home from scratch. Her team used Claude’s coding tool to write the actual code in plain English instructions, hosted it on Vercel, and stored all episode data on Railway so that each new episode generates its own indexed page automatically. Every episode now has a full transcript sitting in plain HTML, no click required to open it. That single change adds between 8,000 and 15,000 words per episode that AI can read and attribute back to the show by name.
Step three was fixing her identity inside the structured databases that models actually trust. Her Wikidata entry, which is the machine-readable record that feeds Gemini and sits behind Google’s knowledge panel, still listed her occupation as vlogger and YouTuber from entries made a decade ago. Gemini read that field and categorized Silicon Valley Girl as a vlog, and vlogs do not appear in podcast recommendation answers. She updated the occupation to podcast host, entrepreneur, and angel investor, added the podcast as its own Wikidata entry, and translated the record into 11 languages. Then she rewrote the show descriptions on Apple Podcasts and Spotify to match Wikidata word for word on the phrases that matter. ‘When all three say the same sentence, it repeats your sentence back,’ she noted. When sources disagree, the model names someone else.
Step four was the plumbing: submitting a sitemap to Google Search Console. That one hour of administrative work led Google to discover 91 pages it had never indexed. The result was roughly 700 additional clicks per month that simply had not existed before.
Step five was tracking the movement with a dedicated AI visibility tool she identifies as Peec AI, which monitors presence across Google, Google AI Overviews, ChatGPT, and Perplexity. The tool surfaces the exact keyword strings the models use internally while generating answers, and her team began writing those phrases directly into episode titles. The first confirmed win came when the Reid Hoffman episode was retitled to match what models search for, and it started winning that answer category.
When the rebuild started, the show appeared in about 7 of 48 tracked questions. As of the most recent measurement, Silicon Valley Girl ranked fifth in AI visibility across the 11 podcast brands the team monitors, ahead of Gary V and Mel Robbins in the tracked set. Overall AI search visibility doubled across the period of the work. The show now owns specific answer slots, appearing as the named result 34% of the time for queries about a head of Google research interview and 31% of the time for Perplexity CEO podcast queries.
Mogilko is clear about where the ceiling is. Broad category questions like ‘best AI podcast in 2026’ remain out of reach because those answers draw from major editorial listicles, and the show has not yet been cited enough in earned media to compete. A Forbes listicle mention in April moved the needle on broad queries more than most technical fixes. Eighty-two percent of what AI cites, she notes, is earned media: other people talking about you in articles, threads, and roundups, not your own website talking about itself.
The Forbes mention that mattered more than the website rebuild
One line in a Forbes listicle published in April shifted the show’s performance on the broad competitive queries that all the technical infrastructure could not crack on its own.
The gap Mogilko is pointing at is still wide open: 92% of marketers say they plan to optimize for AI search, and only 40% are actually doing it. The first-mover advantage is still available, and for now, a blank page in the source code is the most expensive thing a business can have.
