The Katman method: research, plan, build, verify

Three gates, four steps

An AI coding tool does exactly what you ask. If it doesn’t know what to ask for, “fix the SEO” gets you one meta tag. The method gives it a sequence and a pass-or-fail check at every step. The rule from our book: never tell your coding tool to “improve rankings”. Give it a measurable job (“a unique title on every page”) and an acceptance criterion (“no two pages share a title”).

The three gates

Two layers sit outside the page: mention (what others say about you) and measure (whether any of it shows up in answers). The background is on What is GEO?.

The four steps

Step 1: Research

The most common vibe-coder mistake is naming pages after the product. Customers don’t search for your product; they describe their situation.

The research file

• 10–12 situation sentences, without your brand, as a customer would type them into ChatGPT. Take them from sales calls, support emails and reviews.
• What assistants say today. Ask each question logged out with web search on, and write down who is named and which pages are cited. That is your competitor list and your source list.
• What people type into Google: autocomplete, “People also ask”, Bing Webmaster Tools keyword research, Search Console queries. Name URLs with those words: katman.pro’s Turkish consulting page became /tr/geo-danismanligi because Turkish buyers write it that way.
• Competitors only from their own pages, with a check date; proof only if it can be verified, with source and date.

In Katman, katman_research writes this file to .katman/research.md and flags every [FILL] it can’t answer from your code.

Your one sentence and the not-X note

Write one sentence in four parts: what the customer brings, what you do, what they get, which fear is unnecessary. An invented example: “Upload your invoice. We move the line items into your accounting software in 30 seconds and flag the ones that don’t match. Your document is never stored.” Our sample site’s real sentence had the same structure, and ChatGPT repeated all four parts in two sessions on 28 September 2026 (own measurement).

Then ask ChatGPT “What is [your brand]?”. On 22 September 2026, Perplexity took the sample site for another company’s product with the same short name. After a two-word distinctive name and a “we’re not that product” line, it described the site correctly on 24 September (own measurement).

Step 2: Plan

The page plan

Few pages, but the right ones.

Login, signup, checkout, thank-you and site-search pages get noindex,follow and stay out of the sitemap. On a new domain, publish 5–10 distinct pages a week, and stop any template whose “Crawled – currently not indexed” rate passes about 30%.

Anatomy of a situation page

1. Title: situation first, brand last, about 60 characters. The meta description is the sentence you want the assistant to say.
2. H1: the customer’s own sentence; the URL uses the same words.
3. First 1–3 sentences: the answer, in four parts. This is the block assistants quote.
4. One immediate action: a file drop box, a trial or a price.
5. Proof: one verifiable number with its date.
6. How it works: 3–5 steps in plain HTML text.
7. Who it’s not for: 2–3 lines.
8. Alternatives, named honestly.
9. FAQ: 5–8 questions in the customer’s words, each answered in its first sentence.
10. Related pages: text links; every page gets at least one inbound link.
11. A real update date (moved only when the words change), an author and valid JSON-LD.

The brief you hand to your coding tool before writing:

Situation:  [the customer's sentence]
URL:        /[words-from-the-sentence]     Title: [situation] — [brand]
First line: [brings] → [we do] → [they get]. [fear unnecessary].
Action:     [file box / trial / price]
Proof:      [number, source, date]
Not for:    [2–3 lines]
Alternatives: [competitor: who it suits]
FAQ:        [5–8 questions in the customer's words]
Related:    [3–5 URLs]

katman_page_brief fills this in from your research file.

Don’t ask the visitor again what they want

A visitor from ChatGPT arrives with the situation they already described. On our sample site (own measurement, 21–27 September 2026):

• The first quiz question lost about 40% of visitors, so visitors from ChatGPT now skip the quiz.
• A file drop box went on the two pages ChatGPT linked to most; 52 of the first 68 generations used the visitor’s own file.
• Four people who arrived with a file bought a plan that didn’t accept files, and two cancelled within an hour. Checkout now opens on the plan that matches their intent.

The rule: don’t make someone from ChatGPT explain themselves twice. The page’s job is to start solving their situation in the first minute.

Step 3: Build with the prompt book

Give the prompts to your AI coding tool in order and don’t move on until the acceptance criterion passes. P1 and P14 change nothing; they only check.

katman_prompt returns each prompt filled with your own data, and Katman’s audit and lint tools check the criterion (all tools).

Step 4: Verify

Don’t take “done” from your coding tool on trust. Run the checks yourself: the one-minute tests are on My vibe-coded site is invisible to ChatGPT.

Tell the search engines

After 14 days, “Discovered – currently not indexed” usually means missing internal links: that was the cause for 35 pages on the sample site, fixed with footer links. “Crawled – currently not indexed” means the pages aren’t distinct enough.

Measure the right way

• Question set: 12 unbranded questions (3 situation, 2 tool, 2 “best X”, 2 alternative, 2 language, 1 problem) and 2 diagnostics: “What is [brand]?” and “What is [brand] ([domain])?”.
• Human filter: a session counts as human only if it scrolled, saw more than 2 pages, finished a step, signed up or paid. In the sample site’s first week, 528 of 1,546 sessions passed (own measurement).
• Channel rule: ChatGPT means a chatgpt.com referrer or utm_source=chatgpt.com in the link; 79% of the sample site’s ChatGPT visitors had no referrer.
• Bots on the server: crawlers don’t run your analytics script, so count them from server logs.
• Official reports: Search Console’s generative AI performance report (Google) and Bing’s AI Performance report (Bing).
• Re-ask every 14 days. One run is a small sample; see visibility runs.

Off-site: mentions instead of bought links

Publish YouTube explainers on the pages they explain, get listed in real directories, and ask the assistant “which articles did you use?” so you can send each list author one short, factual note. On Reddit: no fake accounts, disclose every mention, and only help for the first 14 days. Don’t buy links or reviews: Google’s spam policies cover link schemes and attempts to manipulate AI answers, and the UK’s DMCC Act 2024 bans fake reviews. Our own weak spot: katman.pro had 3 external links from 2 sites in Search Console on 29 September 2026.

What worked and what didn’t

Own measurement on our sample site, 21–29 September 2026. The strength labels are ours.

Mistakes we made

The book lists 21; these cost us the most:

• Every date was launch day in the sitemap, Article schema and RSS. Dates now move only when the words change.
• llms.txt contradicted a page, with two prices for one feature. Pages and llms.txt now come from one facts file.
• A claim bigger than the proof: “we pick the best model for every category and language”, based on 7 samples rated by 1 person. We narrowed it and stated the test size.
• One fix broke another: Cache-Control: no-transform switched off HTML compression, and pages went out about 4× larger.
• Counting bots as people: a PDF link checker created 704 fake sessions in one day.
• Scaled AI blogging: another of our sites published 203 AI-generated posts on a new domain between 19 July and 10 September 2026. By 10 September, 103 of its 106 non-indexed pages were “Crawled – currently not indexed” (own Search Console data).

The full numbers are in the case study, and all 21 mistakes are in the book.

Sources. The Katman book (29 September 2026); Google Search Central (AI optimization guide, spam policies, JavaScript SEO basics); Bing Webmaster Guidelines; IndexNow. Sample-site figures come from the site’s own analytics (21–28 September 2026, human sessions only); the site is Rnify Limited’s own product, not a client. Written by the Katman team, Rnify Limited.