How to Get Into ChatGPT Recommendations: A 90-Day GEO Experiment Plan Betting on External MediaArticle · digest for November 19
Blueprint of a 90-day GEO experiment plan for promoting a brand in ChatGPT recommendations with an external media schedule.
Article4 sections≈ 7 min

How to Get Into ChatGPT Recommendations: A 90-Day GEO Experiment Plan Betting on External Media

Picture this: you've spent five years in crypto marketing, open a fresh ChatGPT account with no chat history, and ask it to recommend an author in your field. You get zero mentions in the top five, and when you search your own name, the model digs up a college side hustle selling colored smoke bombs and labels you a wedding pyrotechnician. That is exactly how Vadim Budkov launched his 90-day experiment on November 18, 2026, on vc.ru.

Vadim has worked in content since 2019: starting as a copywriter for a sex shop, writing video scripts for Gazprom, working in info-business, and spending five years promoting crypto projects. When he asked a clean ChatGPT to find an author covering P2P payments, AML, and KYT, the bot listed five other names. Past achievements won't save you if a client looks for a contractor through an AI assistant.

Why Your Website Loses to Third-Party Publications

Traditional SEO works in a straight line: grow your own blog, cover keywords, and push your domain to the top of Google. In generative search—where optimization is called GEO (Generative Engine Optimization)—the rules are different. Research by University of Toronto scholars, analyzed on Britopian and Influencer Marketing Hub, revealed a sharp skew in AI sources. On commercial software queries in the US, GPT with web search pulled 72.7% of its citations from earned media: independent outlets, review sites, and roundups.

In Google results for the same category, external sites made up 45.4%, with brand websites taking the rest. Depending on the model, topic, and country, earned media accounts for 57% to over 74% of AI answers—reaching 74.2% in Canadian queries versus 43.7% in Google. Bots cite corporate blogs less often because they lean on third-party consensus. Even unlinked mentions work in GEO: simply having your brand name on someone else's site without a hyperlink already shapes recommendations.

Where does the model pull an answer from?

  • Chat context and memory: OpenAI confirms that chat history even influences how search queries are phrased before hitting the web.
  • Live web search: when the model decides to go online, it splits the question into multiple queries and pulls fresh pages.
  • Training data: knowledge baked into model weights, which new articles won't reach if the bot never opens search.

That creates the main risk factor in GEO: randomness. A marketer can't force a user's ChatGPT to turn on web search, and fresh articles only help when the model actually queries the web.

Day 0: What 30 Prompts Revealed and the 90-Day Plan

From his work account, ChatGPT knew Vadim well. On clean profiles without history, the picture fell apart: ChatGPT recognized the expert, Grok doubted whether it was the same person, Claude pulled a different biography, and Perplexity blurred the profile. Meanwhile, branded searches by surname bring no new clients because the market doesn't know the name yet.

For an objective baseline, Budkov froze 30 non-branded prompts in Russian and English across five groups: broad topics, hiring, narrow expertise, specific tasks, and author discovery. Tests run in fresh chats without memory, separately with web search on and off. The Day 0 snapshot on November 18 exposed four failures:

  • Grok, asked in Russian for five crypto marketers, returned only English-speaking experts.
  • Perplexity, without an explicit language instruction in the prompt, cut out the entire Russian-speaking web.
  • Claude, asked to name five Russian-speaking content strategists, listed four agencies instead of solo specialists.
  • ChatGPT, searching for an author on P2P, bank compliance checks, and AML/KYT, left Budkov out of the top 5 with zero mentions.

Where his name did appear, models uniformly complained about the lack of public case studies with numbers. The author notes honestly that Day 0 only tests hypotheses, while results will be measured at days 30, 60, and 90 across six tracks:

  1. Personal hub website with case studies, open indexing, and Schema.org markup (Person, ProfilePage, Article) to link with external texts.
  2. Articles on vc.ru and other platforms to test indexing speed and domain weight differences like VC vs. RBC.
  3. Public case studies with numbers, closing the main gap bots flagged.
  4. Media commentary via Pressfeed on crypto and payments—the hardest track, targeting the earned media share.
  5. Unifying LinkedIn and other profiles so algorithms stop confusing the marketer with a wedding pyrotechnician.
  6. Original research that other writers cite, creating secondary mentions.

How to Rebuild Digital PR for AI Answers

An international breakdown by Kalin Anastasov on Influencer Marketing Hub adds a practical PR workflow to Budkov's plan. Instead of picking outlets by traffic reach, you build pitching around actual links cited in AI answers using a three-step loop:

  1. Source map. Run commercial prompts and list the domains and articles in footnotes from ChatGPT, Perplexity, and Google AI Overviews, spotting gaps where competitors appear and you don't.
  2. Editorial pitch list. Use that map to build a list of publications and journalists for quotes, reviews, and case studies.
  3. Re-measurement. Log the publication date of each article and re-run the exact same prompts on the same engines in the same region.

This approach has two cold limits. Frequent bot citations don't guarantee value on their own: audience fit and editorial odds still decide everything. Even landing in a top outlet gives no 100% guarantee of a recommendation, so without before-and-after tracking, AI PR turns into guesswork.

How to Measure AI Visibility and Where to Start

Classic SEO rank trackers check a fixed position for a keyword, whereas AI engines have no stable top-ten list. AI visibility trackers measure your share of mentions against competitors, cited domains, URLs, and sentiment. Seven Western tools from the review map neatly onto the three PR stages.

For source mapping, teams use SE Visible by SE Ranking starting at $89/month (Core at $189, Agency at $399) and Ahrefs Brand Radar with custom prompts starting at $50/month on a 475-million-query base. SE Visible tracks domain and page citations across ChatGPT, Google AI Overviews, AI Mode, Gemini, and Copilot, while Ahrefs helps size up the market map.

For editorial outreach, teams use Peec AI ($95/month for brands, $245 for agencies), which separates crawled pages from cited ones to flag outreach targets. Larger teams plug in Cision AI Visibility across nine AI platforms tied to the CisionOne journalist database, and Muck Rack AI Visibility (formerly Generative Pulse) to find top authors cited in ChatGPT, Claude, and Gemini.

To check post-release impact, Otterly.AI and Meltwater GenAI Lens fit well. Otterly.AI offers the lowest entry point: Lite is $29/month, Standard is $189, with a 7-day trial, seven AI engines, and export via API, MCP, and Looker Studio. Meltwater ties AI citations to media monitoring, separating news articles from press releases, Reddit, and Wikipedia.

All seven tools share one drawback for small businesses and non-English markets: dollar pricing and an English-first focus. With zero budget, the working minimum mirrors Budkov's manual protocol.

Minimal Protocol for Clean GEO Measurement

  • Lock in 30 target prompts matching real hiring or buying scenarios and keep them unchanged for 90 days.
  • Run checks in clean chats with memory and personalization turned off.
  • Separate runs with web search enabled and disabled so you don't mix live index results with base model weights.
  • Log publication dates of case studies and media mentions right next to your tracking sheet.
  • Repeat runs on days 0, 30, 60, and 90, comparing answers and cited domains.
Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 3