Who it’s for: solo operators and small sites that want to show up in ChatGPT, Gemini and Claude answers without an SEO budget.
Tools these people use: FAQPage schema, GA4 referral reports and monthly manual checks (Anastasiia Kyslenko); markdown twins, rel=“alternate” links, llms.txt and IP-verified server logs (Umar Gora); Next.js, a Supabase view, Article/Breadcrumb/FAQ schema and scheduled re-verification (Mark Fulton, written up by his ops agent); an AI agent with the Google Search Console API and a public experiment log (Sunny Patel).
The steps
-
Answer the headline in the first 40-80 words. Anastasiia’s first sentence answers the question in the title. “AI pulls exactly the opening sentences.” She ranks this as what worked fastest. Source: Anastasiia (teardown, original)
-
Structure every page. She ends every article with a 5-6 question FAQ marked up with FAQPage schema, uses tables and numbered lists, and puts a named, dated source link right next to every external number. Source: Anastasiia (teardown, original)
-
Use one description of yourself everywhere. Her byline, About page and service pages carry the same core line about her experience and credentials. Source: Anastasiia (teardown, original)
-
Publish a markdown twin of every post. Umar publishes each post again at the same path with .md added, links it from the page head with rel=“alternate” and lists every twin in llms.txt. Over 63 days, verified AI and search agents read the twins 75 times and requested /llms.txt zero times. Source: Umar (teardown, original)
-
Build pages from data you already verify. Mark’s favors.dev turned an internal table of submission directories into 46 templated pages, served only from verified, active rows, each with its own content block and FAQ and breadcrumb schema. A scheduled job re-checks rows and drops dead ones. Source: Mark (teardown, original)
-
Give each search intent one page, and a next step. Mark’s linking rules: one surface owns each intent, templated pages link up to the hub and sideways to 2-3 related rows, and every page has a next step. Source: Mark (teardown, original)
-
Let an agent make one change at a time and log a prediction. Sunny’s agent reads Search Console through the API, makes one change, logs a prediction and scores itself 7-14 days later, with the log public. Source: Sunny (teardown, original, proof log)
-
Measure in three places. Anastasiia checks GA4 for referrals from chatgpt.com and gemini.google.com, and once a month asks ChatGPT and Gemini 10-15 niche questions to see if she appears. Umar verifies crawler IPs before counting requests and tabulates per agent. Sources: Anastasiia (teardown, original); Umar (teardown, original)
Postby’s suggestion: ship llms.txt because it’s cheap, but don’t count on it. Put the effort into steps 1, 2 and 4 and check your own logs, as Umar did.
Human approval points
- Sunny: iteration 4 was a sprint he directed himself. The post doesn’t describe approving each agent change.
- The other sources describe manual work rather than an approval gate.
Common failure points (as reported)
- An editorial post aimed at the same intent as the hub ranked around 76 while the hub sat near 27 (Mark).
- Impressions didn’t turn into clicks until a next step was added to each template (Mark).
- Sunny’s log flags iteration 4 as three changes at once, so attribution is page-level only.
Caveats carried over
- Anastasiia: “a handful of sessions a month”, no controlled test, and the post promotes her services. Outside figures she cites are unchecked. X: Unverified.
- Umar: one site’s logs.
- Mark: the writeup is bylined to his ops agent (“Mark’s Claude Ops”); the link to Mark comes from byline metadata.
- Sunny: outcomes Unverified · pending. He writes that he “will not claim wins I have not measured.” X: Unverified.
Examples
Get one teardown a week.
Same format, a new operator every issue. Double opt-in.