Intermediate · About 56 hours · 3 projects · Free

AI Roadmap for SEO and content

In about 56 hours you deliver an AI visibility audit of one site, a retrieval-first content brief for one page and a measurement spec for AI search visibility, each grounded in what Google, the labs and named practitioners publish.

0of 56 hours · 0%

Saved in this browser. No account, no email.

What you'll build

Go at your own pace. Stop at any milestone: each one leaves you with something to show.

  1. 1
    AI visibility audit of one site

    Measure how often AI assistants cite one site, which competitors they cite instead, and why. No paid tools.

    hour 17
  2. 2
    Retrieval-first content brief for one page

    Take the page behind your most valuable lost query and write a brief a writer can execute.

    hour 33
  3. 3
    Measurement spec for AI search visibility

    Write the spec an analytics team can build from, and prove it with a working counter on real log data.

    hour 56

Skills you'll learn

  • Query research
  • AI citations
  • Competitor analysis
  • Logging
  • Why codes
  • Chunking
  • Answer-first writing
  • Entities
  • Structured data
  • Internal links
  • Token counts
  • Metric definition
  • Repeat runs
  • Human checks
  • Log schemas
  • Data sourcing

Prerequisites

  • Working SEO knowledge: crawling, indexing, Search Console and on-page basics.
  • A spreadsheet.
  • One site you are allowed to audit.
  • No code required.

Program outline

3 milestones · 10 sections · 50 lessons · 3 projects · Every course and source is free

Learn as a feed

Milestone 1 · Hours 1 to 17 · 3 sections · 15 lessons

How AI search answers

Learn how AI search finds and cites pages, then measure one site.

Hours 1 to 60/15

By the end you can explain to a client in five sentences what a language model does and why it makes things up, and your audit has its site and a 40-question query list.

In the Encyclopedia: Large Language Models (LLMs), AI in Marketing & Search, AI-Powered Search & Ranking

  1. What a language model does, and what it does not70 min

  2. Karpathy's hour: how the models are built and used75 min

    • Watch
      Andrej Karpathy · Intro to Large Language Models (1hr) · Andrej Karpathy
    • Do
      Add three lines to the explainer, each one a thing from the talk a client should know: the model is trained on text and then tuned to follow instructions, it can call tools such as web search, and it can be wrong while sounding sure.

  3. Why models make things up, in the words of the labs70 min

    • Watch
      Why Large Language Models Hallucinate · IBM Technology
    • Read
      Why language models hallucinate · OpenAI
    • Do
      Write the audit's method note: you will check whether the site is cited, not whether the answer sounds right, because the sources explain why a fluent answer can be wrong. Then start the query list with the 10 questions already on the site's FAQ or support page.

  4. What Google says about how Search works and what it rewards70 min

    • Watch
      How does Google Search work? · Google Search Central
    • Read
      Creating Helpful, Reliable, People-First Content · Google
    • Do
      Write the audit's site profile row: what the site sells, who it is for, and its honest answers to five of Google's self-assessment questions, kept as the page words them rather than turned into ranking claims. Add 10 queries from Search Console questions (who, what, how, best, versus).

  5. The practitioners' map of AI search75 min

    • Read
      Learning AI Search: the free roadmap · Aleyda Solis
    • Read
      The AI Search Manual · iPullRank
    • Do
      Finish the query list at 40 questions, adding 20 from sales calls, support tickets and the comparison questions buyers ask. Tag each with an intent: learn, compare, buy or do. Write a client note in one paragraph: what AI search changed for this site.

Hours 7 to 110/15

By the end you can explain how a question finds a page that never uses its words, and half of your audit queries are run and logged.

In the Encyclopedia: Embeddings, Semantic Search, Vector Databases

  1. Embeddings in plain English60 min

  2. Word vectors, seen: Word2Vec65 min

  3. Google's embeddings lesson65 min

    • Course
      Embeddings (Machine Learning Crash Course) · Google
    • Do
      Run and log queries 17 to 20, then write one paragraph for the audit's method page: what an embedding is and why a page can be found for a question it never phrases, in words a client can repeat.

  4. Semantic search and question answering with embeddings65 min

  5. Vector databases, and try an embedding model yourself65 min

Hours 12 to 170/16

By the end you can draw the retrieve, rerank and cite pipeline on a whiteboard, and your audit is complete with a why code for every lost query and five actions.

In the Encyclopedia: Retrieval-Augmented Generation (RAG), Chunking, Hybrid Search & Reranking

  1. RAG in one hour75 min

  2. Chunks: the unit a model actually reads70 min

  3. Hybrid search and reranking: why exact words still count65 min

  4. Contextual retrieval and agentic RAG: the context around a passage70 min

    • Read
      Contextual retrieval in AI systems · Anthropic
    • Watch
      What is Agentic RAG? · IBM Technology
    • Do
      Re-run the 10 most valuable lost queries once more in fresh chats and add a stable column: same citation, different citation, or no citation. Note on the method page that one run is a sample, not a verdict.

  5. Late interaction retrieval, and the audit summary65 min

Milestone 1 project · about 6 hours

AI visibility audit of one site

Measure how often AI assistants cite one site, which competitors they cite instead, and why. No paid tools.

  • Query research
  • AI citations
  • Competitor analysis
  • Logging
  • Why codes

Courses that teach this

Build it

  1. Pick one site you are allowed to audit and build a 40-question list, tagged by intent.

  2. Take the questions from its FAQ, Search Console, support tickets and buyer comparisons.

  3. Run every query signed out, in a fresh chat, on two or more assistants that show sources.

  4. Log whether the site is cited, which URL, which competitor, and the passage used.

  5. Give every lost query a why code, and re-run the ten most valuable to test the repeat.

  6. Write the summary: cited share, top competitors cited and five ordered actions.

You finish with

A workbook with the query log, a method page and a summary page: cited share by surface, top competitors cited, why code counts and five ordered actions.

Check your work

  • Every query is logged with surface, citation and passage.

  • Each lost query has a why code.

  • The ten most valuable queries were re-run.

  • It ends in five ordered actions.

See how others did it

Milestone 2 · Hours 18 to 33 · 3 sections · 15 lessons

Content for retrieval

Learn tokens, chunks and entities, then write content built for retrieval.

Hours 18 to 230/17

By the end you can explain tokens and token density to a writer, and your brief has its page, purpose, lead, scope and chunk plan.

In the Encyclopedia: Tokenization, Token Density and Information Density, Topic Density and Topical Coverage

  1. Tokens: what a model counts75 min

  2. Token density: say more per token60 min

  3. Topic density: one page, one subject70 min

  4. Chunk plan: headings as retrieval units70 min

  5. The question list the page must answer60 min

Hours 24 to 280/16

By the end you can explain to a client what an entity is, and your brief names the page entities and facts, their relationships, the structured data that states them and the links that name them.

In the Encyclopedia: Named Entity Recognition (NER), Information Extraction, Knowledge Graphs

  1. NLP and named entities: the nouns machines track65 min

  2. Information extraction: from text to facts65 min

  3. Knowledge graphs: facts with relationships70 min

  4. Structured data: the facts a parser can read65 min

    • Read
      Structured outputs · Anthropic
    • Read
      Structured model outputs · OpenAI
    • Do
      Write brief section 8, the structured data plan: the page type and each property to state (name, description, author, date, price, address, FAQ pairs), with the visible sentence each value comes from, so the markup never says what the page does not.

  5. GraphRAG: why relationships get retrieved65 min

Hours 29 to 330/15

By the end you can explain how an assistant writes an answer from your page, and the brief is finished: Q&A blocks, a lead that survives summarization, plain facts, a usage note and a writer checklist.

In the Encyclopedia: Question Answering, Text Summarization, AI, Copyright and Training Data

  1. Question answering: the direct answer60 min

  2. Summarization: what survives compression60 min

  3. When the assistant gets your brand wrong60 min

    • Read
      What Are AI Hallucinations? | IBM · IBM
    • Read
      Anthropic: Reducing Hallucinations · platform.claude.com
    • Do
      Ask two assistants ten factual questions about the brand (founded, locations, prices, people, hours). Log each wrong answer against the page that should have corrected it. This is brief section 12, the facts the page must state in plain sentences.

  4. Copyright, training data and what you allow70 min

  5. Context windows and long pages; assemble the brief60 min

Milestone 2 project · about 6 hours

Retrieval-first content brief for one page

Take the page behind your most valuable lost query and write a brief a writer can execute.

  • Chunking
  • Answer-first writing
  • Entities
  • Structured data
  • Internal links
  • Token counts

Courses that teach this

Build it

  1. Take the page behind your most valuable lost query.

  2. Write a lead that answers in 60 words, and the scope of the page.

  3. Plan the chunks, with question headings and self-contained answers.

  4. List the entities and a fact table, with canonical names.

  5. Plan structured data whose values match the visible text, and internal links that name the entity.

  6. Test the Q&A blocks in a question answering demo, then write the writer's checklist.

You finish with

A 13-section brief plus a one-page writer checklist, with before and after token counts for the lead.

Check your work

  • The lead answers the question in 60 words.

  • Each chunk answers one question on its own.

  • Structured data matches the visible text.

  • A writer can execute it without asking you.

See how others did it

Milestone 3 · Hours 34 to 56 · 4 sections · 20 lessons

Measure and automate

Learn why one run is not a number, then build the measurement.

Hours 34 to 390/14

By the end your audit prompt is written down, versioned and returns the same shape of answer every run, and the spec has its metric definitions, run conditions and log schema.

In the Encyclopedia: Prompt Engineering, Context Engineering, Structured Outputs

  1. Prompting 101 from Anthropic70 min

  2. The prompt engineering short course70 min

    • Course
      DeepLearning.AI · ChatGPT Prompt Engineering (Free) · DeepLearning.AI · free, certificate extra
    • Do
      Write the audit's standard prompt, version 1: the exact wording sent per query with no brand hint, plus the follow-up that asks for sources when none are shown. This is spec paragraph 2, the instrument.

  3. Context engineering: everything else in the window65 min

  4. Structured outputs: answers you can count65 min

  5. Practice prompts and set the baseline70 min

    • Read
      Learn Prompting: free guide · learnprompting.org
    • Read
      Practical Prompt Engineering · Cameron R. Wolfe
    • Do
      Run the standard prompt on 10 audit queries using the log schema. Compare with your audit log and record what changed. This is spec paragraph 5, the baseline and its date.

Hours 40 to 450/15

By the end you can explain why one run is not a number, and the spec says how many queries, how many runs, how a person checks the log and what counts as a change.

In the Encyclopedia: LLM Evaluation & Benchmarks, Offline and Online Evaluation, Evaluation Metrics & Cross-Validation

  1. What a benchmark is65 min

  2. Evals in practice70 min

  3. Precision, recall and the confusion matrix for cited65 min

  4. Offline numbers versus live traffic70 min

  5. Judges: letting a model grade an answer60 min

    • Read
      Demystifying evals for AI agents · Anthropic
    • Try
      LMArena leaderboard · LMArena
    • Do
      Write the judge prompt that labels a logged answer stance toward the brand (positive, neutral, negative, absent). Test it on 10 logged answers against your own labels and record the agreement. This is spec paragraph 10.

Hours 46 to 500/15

By the end you can explain what an agent is and is not, and the spec names every data source, how it is collected, by whom or what, how often, at what cost and under which safety rules.

In the Encyclopedia: AI Agents, Agentic Workflows & Patterns, Computer Use & Browser Agents

  1. What an agent is, and the tools it can call75 min

  2. Workflows before agents65 min

  3. Browser agents and computer use65 min

    • Read
      Computer use tool · Anthropic
    • Read
      Anthropic: Developing computer use · Anthropic
    • Do
      Write spec paragraph 13, the automated run: what a browser agent would do per query, what it must never do (sign in to client accounts, accept terms, click ads, enter personal data), and the checks before its log is trusted.

  4. Prompt injection: the risk when agents read the web65 min

  5. Andrew Ng's agentic patterns, applied70 min

    • Course
      Agentic AI, Andrew Ng (free, certificate extra) · DeepLearning.AI · free, certificate extra
    • Do
      Write spec paragraph 15, the other data sources: Search Console, analytics referrals from assistants, server logs for crawler user agents, each with owner, access, cadence and the free tier you will use.

Hours 51 to 560/14

By the end the spec is complete with a datasheet for its data set, a working or scoped collector, and a client-ready explainer of what the numbers can and cannot say.

In the Encyclopedia: Agentic Coding, Vibe Coding, Datasheets, Data Cards and Model Cards

  1. Vibe coding: what it is and when it is enough65 min

  2. Agentic coding: have a model write the counter70 min

    • Course
      Anthropic Academy · Claude Code, MCP, Agent SDK · anthropic.skilljar.com
    • Do
      With any free coding assistant, generate a script or spreadsheet formulas that turn the log into the three metrics with their ranges. Run it on your audit data. Keep only what you can read and explain.

  3. Datasheets: documenting the data set70 min

  4. Model cards: what the labs disclose, and what you will65 min

    • Read
      Model cards · Google DeepMind
    • Read
      Model Card Guidebook · Hugging Face
    • Do
      Write spec paragraph 18, the report layout: one page, three metrics with ranges, top cited competitors, the why code mix, changes since last period, and a box titled what this cannot tell you.

  5. Deliver75 min

Milestone 3 project · about 8 hours

Measurement spec for AI search visibility

Write the spec an analytics team can build from, and prove it with a working counter on real log data.

  • Metric definition
  • Repeat runs
  • Human checks
  • Log schemas
  • Data sourcing

Courses that teach this

  • AI Search Operating System

    Semrush Academy, Leigh McKenzie

    Audit four layers of AI visibility, define KPIs, build a 90-day plan.

  • AEO Fundamentals

    HubSpot Academy, Jorie Munroe · 2 hr 20 min

    Plan content for search and AI, optimize pages for citations, measure visibility.

  • SEO in the Age of AI Learning Path

    HubSpot Academy

    Get cited on AI Overviews, ChatGPT, Perplexity, Gemini; measure an AEO plan.

Build it

  1. Define your three metrics in plain words, and what each one can and cannot show.

  2. Write the standard prompt and run conditions, with a log schema.

  3. Set the query set and the repeat runs, and report results as ranges, not single numbers.

  4. Add a human check and record its agreement rate.

  5. List each data source with its owner, cadence and cost, and the safety rules for any agent.

  6. Build a working counter, as a script or spreadsheet, and run it on real log data.

You finish with

An 18-paragraph spec, a working counter (script or spreadsheet) run on real log data, and a one-page client explainer.

Check your work

  • Each metric has a definition and a limit.

  • Results are ranges from repeat runs, not single numbers.

  • A person checked a sample, with the agreement rate.

  • The counter runs on real log data.

See how others did it

Check yourself at the end

  • You can explain to a client, in plain words and without invented ranking claims, how an assistant turns a question into retrieved passages and a cited answer, and where their site falls out of that chain.

  • You can run a repeatable, free AI visibility audit: a stratified query set, a fixed prompt, logged runs across assistants, a why code for every lost query and five ordered actions.

  • You can write a content brief a writer can execute that is built for retrieval: self-contained chunks, a lead that survives summarization, canonical entities and a fact table, structured data that matches the page, and internal links that name the entity.

  • You can specify AI visibility measurement your analytics team can build: metric definitions, run conditions, sample size, repeat runs reported as ranges, a human check, a judge prompt with its agreement rate, and a datasheet for the data.

  • You can decide where an agent belongs in SEO work, write its safety rules, and brief a developer or a coding assistant to build the collector without pretending the result is a ranking signal.

  • You can read a lab document, a Google Search Central page or a practitioner manual and tell a client which claims are documented, which are measured and which are opinion.