Handout for the webinar on November 10, 2026 · 19:00 Kyiv time
20 SEO and GEO rules for an AI-written blog
I run a bilingual blog written with AI. These rules come from our publishing workflow and help us prepare useful content for search and AI answers. They do not guarantee rankings or citations.
No signup, no email
Author: Yaroslav Maxymovych · Published
Our blog in numbers
961 articles, each available in two languages.
People visiting the blog (Google Analytics): February 2026 — 35; October 2026 — 4,174.
Visits from Google in October 2026 — 1,525.
Top article — End of Day Report Template: 1,693 visits from search.
Data from our database and Google Analytics as of November 9, 2026
Article structure
-
Answer one specific reader question per article. Choose a topic from your plan and check what you have already published on related subjects. Why: a distinct reader need helps you create something useful without duplicating existing pages. How to check: express the question in one sentence and compare your planned article with existing answers.
-
Use keywords naturally. Include the main search phrase in the title, introduction, an early subheading and the SEO title or description; related terms should clarify the topic. Why: readers and search engines should immediately understand what the page covers. How to check: read it aloud; keep the SEO title within 60 characters and the description within 160, without promising anything the article does not deliver.
-
Keep the introduction short and get to the point. If the author has relevant, verified experience, open with one short paragraph, followed by a TL;DR of up to three points within the first 100 words. Why: readers can quickly see the answer and its practical value. How to check: the opening and key points explain what the reader will learn; the author's experience is real.
-
Use questions as subheadings and answer them immediately. The first one or two sentences under a question-style H2 should give a direct answer. Why: the passage is easier to read on its own and use in an assistant's response. How to check: keep only the heading and opening sentences; together they should form a complete question and answer.
-
Turn explanations into tools. Use paragraphs of two or three sentences, steps for processes and tables for comparisons; add a template, checklist or example. Why: practical material can be put to work and shared with others. How to check: readers can take something from the article into their own work, beyond a general idea.
-
Finish with a useful FAQ. Add four to six real questions with short answers of two to four sentences each. Why: an FAQ addresses questions left open by the main text. How to check: every answer makes sense on its own, answers its question and does more than repeat a marketing slogan.
GEO: preparing content for citations in ChatGPT and Perplexity
-
Support facts with sources. Put a date and a working link to the original source next to external figures; explain how you measured your own data. Why: a verified fact is more valuable than a convincing claim without evidence. How to check: open the source and find the exact figure; remove unsupported claims or clearly label them as assumptions.
-
Show who is speaking and what their experience is based on. Name the author; for a quotation, include the speaker, their role, the date and the source. Why: clear attribution helps readers assess a claim's reliability. How to check: every quotation and first-person story can be traced to a real person or approved source material.
-
Define terms directly. Give key concepts a separate one- or two-sentence definition, such as “A term is…”. Why: a concise definition can be understood and quoted without a long introduction. How to check: someone encountering the term for the first time can understand it from that block alone.
-
Make passages self-contained and keep promotion separate. Aim for 60–120 words in the main explanation under an H2; put calls to action in separate blocks between sections. Why: passages that do not depend on “as mentioned above” or contain promotional interruptions are easier to cite. How to check: copy a passage into a separate document; it should retain its meaning without the rest of the article.
-
Measure results with a fixed set of questions. Repeat the same questions in ChatGPT and Perplexity each month and separately track visits from people following their answers. Why: readable structure alone does not prove that an assistant links to your content. How to check: save the date, exact question and link in the response; count bot requests separately, since a crawl alone does not prove a citation.
Technical foundations
-
Serve the main content in the HTML response. The heading, answer and links should be readable without JavaScript; check page access and permissions for the crawlers you need. Why: some bots cannot see text that only appears after scripts run or a user signs in. How to check: inspect the original HTML without executing JavaScript and find the article content; check robots.txt and the HTTP status.
-
Include only ready, public pages in the sitemap. Set an accurate lastmod and submit the sitemap in Search Console. Why: a sitemap should lead search engines to material people can already read. How to check: each URL is in the sitemap, loads successfully, is not a draft and has a date that reflects actual content changes.
-
Match Schema.org markup to visible content. Use Article with the author and dates, BreadcrumbList for breadcrumbs and FAQPage when a corresponding FAQ is visible. Why: structured data explains the page's type and content to machines. How to check: validate the JSON-LD and compare its author, dates, questions and answers with what readers see; markup does not guarantee a special search appearance.
-
Give each language its own URL. Use language-specific URLs, a self-referencing canonical and reciprocal hreflang only between completed translations. Why: search engines need unambiguous addresses for each language version. How to check: both pages load, reference each other through hreflang and use the correct language; a query parameter such as
?lang=ukdoes not replace a translated page. -
Check speed and readability on a phone. Reduce unnecessary image and script weight and make sure readers can access the text immediately. Why: slow pages frustrate readers and make crawling harder. How to check: open the page on a phone and measure it with browser tools or PageSpeed; record the actual result instead of simply calling it “fast”.
Mistakes we have made
-
Do not publish your entire backlog in one day. I once published 50 articles in a day, and traffic fell for a week and a half. Publication now follows a schedule. Why: this is my experience, not a proven universal Google threshold. How to check: draft generation is separate from publication, every draft has a future slot and daily output is capped.
-
Avoid duplicates, keyword stuffing and “magic” files. Check existing topics and URLs before generating content, avoid cramming in search phrases and do not treat llms.txt as a substitute for content and technical work. Why: a narrow topic list and weak review have already produced near-duplicates for us. How to check: each new article adds a distinct answer or practical benefit, its URL is not a duplicate with a numeric suffix and success is measured through reading and visits.
-
Do not present an illustrative scenario as a client result. Label scenarios and estimates as examples and show the assumptions. Why: invented clients, precise percentages and falsely attributed research create misplaced trust. How to check: real results have evidence; simulations carry a visible “illustrative example” label rather than the name of an imaginary company.
-
Do not publish a model response just because it reports success. Check format, completeness, links and the readiness of every advertised language. Why: we have received translations containing only “...” while their pages were already entering the sitemap. How to check: automated checks reject broken formats, empty or truncated text and unusable translations. Unfinished language versions must stay out of the sitemap and hreflang; after automated checks, read the material and verify the facts.
Claude skill
Copy the text below into your Claude project instructions. Or save it as .claude/skills/seo-geo-article/SKILL.md in your project to use it as a skill.
---
name: seo-geo-article
description: Prepare or review a blog article using practical SEO and GEO rules for sources, answer structure, language versions and publication checks.
---
# Claude instructions: articles for search and AI answers
Paste this text into your project instructions. To use it as a Claude skill, save it as `.claude/skills/seo-geo-article/SKILL.md`. Adding these instructions does not authorize publication or website changes.
Work with the specific project's topic, audience, facts and existing implementation. First find its editorial rules, article template, checks, language routes, sitemap and structured-data components. Reuse them instead of building a parallel system. If you cannot access a source, mark the claim as unverified or remove it. Do not promise rankings or citations.
Prepare the article and a short verification table: rule → result → evidence, or the reason a check is unavailable. The same 20 rules from the handout follow.
## Structure
1. **One reader need.** Choose one specific question from the topic plan and check existing content. Goal: a distinct, useful answer. Check: state the reader's need in one sentence and explain how the article adds to existing material.
2. **Natural keywords.** Use the main search phrase in the title, introduction, an early H2 and the SEO title or description; add related terms only where relevant. Goal: an unambiguous topic. Check: read aloud; SEO title up to 60 characters, description up to 160; no unsupported promises.
3. **Short introduction and TL;DR.** If the author has real experience, start with a short paragraph, then give up to three key points within the first 100 words. Goal: explain the answer and value immediately. Check: the points stand on their own and the experience is real; without a verified story, use a direct introduction without pretending to speak from experience.
4. **Question-style H2s with immediate answers.** Answer in the first one or two sentences under the question. Goal: a clear passage for readers and assistants. Check: the heading and opening sentences form a complete answer.
5. **Practical structure.** Use two- or three-sentence paragraphs, steps for processes and tables for comparisons; add a usable template, example or checklist. Goal: help readers do the work. Check: the resource can be used without further explanation.
6. **FAQ.** Give four to six real follow-up questions with answers of two to four sentences. Goal: resolve remaining questions. Check: answers stand on their own and are not marketing slogans.
## GEO
7. **Evidence for facts.** External figures need a working URL, date and source that supports the claim; your own figures need a measurement explanation. Goal: verifiability. Check: open the source and find the exact claim, or remove it or label it an assumption; a link alone is not a fact-check.
8. **Real authorship and attribution.** Name the author; attribute quotations with the person's name, role, date and source. Goal: let readers assess experience. Check: quotations and first-person stories are verifiable; do not speak on someone's behalf without their material.
9. **Direct definitions.** Put key terms in separate one- or two-sentence “A term is…” blocks. Goal: explain without relying on context. Check: a reader encountering the term for the first time understands the definition.
10. **Self-contained passages and separate CTAs.** Aim for 60–120 words in the main explanation under an H2; place calls to action in separate blocks between substantive sections. Goal: passages that can be cited. Check: a passage retains its meaning when copied on its own and does not open with “as mentioned above”.
11. **Visibility checks.** Propose a repeatable monthly set of questions in ChatGPT/Perplexity and separate tracking of human visits from answers. Goal: measure results. Check: save the date, exact question and response link; bot requests, human visits and confirmed citations are different signals. Do not invent assistant-check results or call paid services without existing authorization.
## Technical checks
12. **Accessible HTML.** Serve the main text, heading and links without executing JavaScript; allow access for relevant crawlers. Goal: readable, crawlable content. Check: original HTML, HTTP status, robots.txt and no mandatory login for public text.
13. **Ready URLs in the sitemap.** Include only published, usable pages with an accurate lastmod. Goal: lead search engines to working content. Check: sitemap URLs load and dates are accurate; check Search Console submission only if access is available, otherwise mark it “not verified”.
14. **Accurate Schema.org markup.** Reuse existing Article components with author/dates, BreadcrumbList and FAQPage when an FAQ is visible. Goal: explain the page type to machines. Check: valid JSON-LD matching visible content; no hidden invented answers or promises of rich results.
15. **Separate language URLs.** Completed translations have their own URLs, canonicals and reciprocal hreflang. Goal: clear language mapping. Check: each URL works in its language and both reference each other; `?lang=uk` does not replace a separate translated page.
16. **Speed and phones.** Check image weight, scripts and immediate text access. Goal: easy reading and crawling. Check: actually open on a phone and measure with browser tools or PageSpeed; never invent a score when access is unavailable.
## Pitfalls and publication
17. **Separate draft inventory from publication.** Schedule releases in slots with a project-specific limit. Goal: avoid accidentally releasing the entire backlog. Check: slots are spread out and rerunning the job does not publish the whole queue. Yaroslav's experience: “I once published 50 articles in a day, and traffic fell for a week and a half. Publication now follows a schedule.” This is the author's experience, not a universal Google threshold or proof of causation.
18. **No duplicates, keyword stuffing or reliance on llms.txt.** Check topics and URLs before generation, avoid multiplying versions of the same article and do not replace GEO work with an auxiliary file. Goal: independent value in each article. Check: new substance, no duplicates with artificial suffixes and evaluation based on actual reading and visits.
19. **Label illustrative scenarios.** For scenarios and estimates, show assumptions and a visible “illustrative example” label. Goal: avoid substituting a plausible story for evidence. Check: real results have supporting evidence; simulations do not present invented clients or precise metrics as facts.
20. **Check before release.** Verify format, content, links and advertised language readiness, beyond a successful model response. Goal: avoid publishing truncated or empty pages. Check: valid format, complete meaningful text and usable translations with meaningful headings. Use the project's own checks. Unusable language versions must stay out of the sitemap/hreflang; check both languages for a bilingual release. Read the material and verify facts after automated checks.
## Deliverables
Return the article, SEO title and description, verified sources and a table of 20 checks. For technical points you cannot access, write “not verified” and state the required action. Do not install integrations, publish or change production based on these instructions alone.Frequently asked questions
Will Google penalize AI-written content?
Google evaluates how useful and reliable content is, rather than how it was written. Publishing large volumes without review is a risk. After we published 50 articles in one day, traffic dipped for a week and a half; that is our experience, not proof of causation or a universal Google threshold.
How is GEO different from SEO?
SEO helps readers find a page through search. GEO helps an AI assistant understand and cite a specific passage. Both rely on useful content, clear structure and verified facts.
How many articles should I publish per day?
Spread reviewed articles across a steady schedule instead of publishing your entire backlog at once. We publish a few articles a day. That is our working pace, not a guaranteed safe limit for every website.
Do I need a developer?
You do not need a developer to apply these rules to an article. An automated publishing pipeline needs a one-time setup, or you can order a fully managed blog. Get a fully managed blog
How can I tell whether ChatGPT cites my website?
Repeat a fixed set of questions every month and save responses that link to your content. A bot visiting your website does not prove that it cited you. Track visits from people following AI answers separately.

Implements AI agents in companies and teaches founders and their teams to work with them — through courses and corporate programs.
This article was prepared with AI assistance, based on Yaroslav Maxymovych's methodology and materials. Spotted an inaccuracy — let us know via the form below.