In 2026 every page you publish is read by three very different readers: a human, a search crawler, and an AI answer engine that will paraphrase you to someone who never visits your site. This is a working guide to being found — and quoted — by all three.
learn.projekinaja.com, an education platform with 8 courses and 340 lessons, as the worked example.You can still use this whole page. Three tips: (1) read the plain-words dictionary first — every acronym here is defined in one sentence; (2) Parts 1 and 3 require no code at all and are where most of the gain is; (3) Part 2 is the technical one — you can hand it to a developer verbatim, since every item there names the exact file and the exact check.
For twenty years, search optimisation had a single objective function: appear as high as possible in a list of ten blue links, because position determined clicks and clicks determined everything else. That objective still exists. It is simply no longer the only one, and on a growing share of queries it is no longer the most valuable one.
Here is what happens to a page you publish today. A human may land on it and decide within seconds whether it answers the question they actually had. A search crawler fetches it, renders it, extracts its structure, and files it in an index that decides whether it is eligible to appear at all. And an AI answer engine — Google's AI Overviews and AI Mode, ChatGPT's search, Perplexity, Claude, Copilot, Gemini — retrieves it as one of several sources, reads a few paragraphs of it, and writes a new answer that may quote you, may paraphrase you, may cite you with a small link, or may use your work without any visible credit at all.
Three shifts matter, and it is worth separating them because they call for different responses.
1. The answer moved above the links. Google began rolling out AI Overviews broadly in 2024 and followed with a conversational AI Mode; ChatGPT, Perplexity, and Copilot built search directly into chat. On many informational queries, the top of the page is now a synthesised paragraph with a handful of citation chips. Your competitor for attention is no longer the result above you. It is a summary that answers the question before the results begin.
2. Click-through rates on informational queries fell. This is the honest, uncomfortable part. Independent measurements through 2025 consistently pointed the same direction: when an AI summary is present, fewer people click anything. The Pew Research Center's July 2025 analysis of real browsing behaviour reported that users clicked a traditional search result on roughly 8% of visits where an AI summary appeared, versus about 15% where it did not — roughly half. Treat the exact figures as one study of one sample rather than a law of nature, but the direction has been reproduced widely enough to plan around. Google's public position is that the clicks that remain are higher quality; that may well be true, and it is also not independently verifiable from outside.
3. The impression became the prize. If someone asks Perplexity "what is the best way to learn agentic AI" and the answer says "ProjekinAja Learn's agentic AI course structures this in five stages…", you got something real even with no click: you were named, in front of a person actively researching, by a source they currently trust more than an ad. This is the new top position, and it does not appear in your traffic graph.
AI Overviews are generated largely from the same search index that produces the blue links. Pages that rank well are dramatically more likely to be retrieved and cited. There is no separate "AI SEO" back door that lets a page nobody can crawl, nobody links to, and nobody ranks show up inside an AI answer. GEO is mostly classic SEO plus extractability plus off-site presence. Anyone selling you a wholly new discipline is selling you the old discipline in a new hat.
Be crawlable, be relevant to the intent, be genuinely better than what already ranks, be linked to. Everything downstream depends on this.
Write passages a machine can lift correctly — self-contained, factual, structured, unambiguous about who said what. Then exist in the sources these engines retrieve from.
That is the entire strategy of this tutorial. Part 1 covers earning the rank, Part 2 the technical floor that makes it possible, Part 3 being quotable, and Part 4 applies all of it to a real site.
Read these once and the rest of the page is plain instructions. No prior knowledge assumed.
<script> tag) that say this is a Course, it costs this, it takes this long, this person wrote it. Humans never see it; machines rely on it.Start here, and do not skip to Part 3 because it sounds newer. Every credible analysis of AI citations lands on the same unglamorous conclusion: pages that get cited by answer engines are overwhelmingly pages that already rank. The foundations are not the boring prerequisite to the AI part. They are the AI part, viewed from a different angle.
Every query carries an intent: the searcher wants to know something, go somewhere specific, do something, or buy something. Serving the wrong format to an intent loses before you start — a sales page cannot win an explanatory query, and a 4,000-word essay cannot win "reset my password".
The diagnostic takes two minutes and costs nothing: search your target query and look at what already ranks. If the top ten are all step-by-step tutorials with numbered headings, the engine has already told you what its users accept as an answer to that query. Publishing a philosophical think-piece into that result set is not being differentiated, it is being off-format.
| Query shape | Intent | Format that wins |
|---|---|---|
| "what is generative engine optimization" | Informational | Definition first, then explanation, then example |
| "how to add course schema" | Instructional | Numbered steps with real code and a verification step |
| "projekinaja learn" | Navigational | Your homepage, ranking for your own brand |
| "best agentic ai course" | Commercial investigation | Comparison with criteria, honest trade-offs, a table |
Answer engines fan out. A single question like "is an AI course worth it in 2026" may trigger a dozen sub-searches — cost, duration, job outcomes, alternatives, prerequisites. A page that answers one question well can be cited; a page that answers the surrounding cluster of questions well can be cited repeatedly, across many different user prompts. Structure your page as a set of clearly-labelled answers to real sub-questions, not one undifferentiated wall.
This is the highest-leverage idea in Part 1. If your page restates what the current top ten already say, there is no reason for a ranking system — or a summarising model — to prefer it. Rewriting a competitor's article more nicely produces a page that is redundant by construction. Something has to be new in it.
Sources of genuine information gain, roughly in order of difficulty:
Take your draft and the current top five results for its target query. Write down, in one sentence, what your page contains that none of those five contain. Passing: you can name a specific thing — a dataset, a first-hand result, a caveat, a worked artifact — in under fifteen seconds. Failing sign: your honest answer is "mine is clearer" or "mine is more up to date". Clarity is table stakes, not gain; recency ages out in a month. If you fail, do not publish yet — add the missing thing, because it is far cheaper than trying to promote a redundant page later.
E-E-A-T is not a score you can inspect and it is not a ranking factor you can set. It is the framework human quality raters use, and the systems are built to approximate what those raters reward. Which makes it, in practice, a very good checklist — with the important caveat that it works by being true, not by being declared.
The concrete signals worth implementing:
Invented author personas, fake credentials, and AI-generated "expert" headshots are a category of risk with no upside. They fail the exact test they are meant to pass, and Google's spam policies explicitly cover misrepresenting who is behind content. If a real person cannot put their name on it, that is useful information about whether it should be published.
Internal links do three jobs simultaneously, which is why they are the best return per hour of any technical task on this page. They route discovery (a page linked from nowhere may never be crawled), they distribute authority (links pass ranking signals internally, not just from other sites), and they declare topical relationships (they tell engines that these twenty pages form one coherent body of work on one subject).
The pattern that works, and it is deliberately simple:
Anchor text is the free part people waste. The words inside the link are a strong description of the destination. Twenty lesson pages all linking to a course index with the anchor "here" throw away twenty descriptive signals.
None of this wins you anything. All of it prevents self-inflicted losses, which is a category most sites underestimate.
| Item | What good looks like | What it costs you when wrong |
|---|---|---|
| Canonical tag | Every page self-canonicals to one absolute HTTPS URL; duplicates point at the original | Your own variants compete; the engine picks a URL you didn't choose |
| One URL per page | Pick with-or-without trailing slash, www or not, and 301 the rest | Signals split across four addresses for the same content |
| XML sitemap | Only indexable, canonical, 200-status URLs; referenced from robots.txt | A sitemap full of redirects and 404s teaches the crawler to distrust it |
| Redirects | 301 for permanent moves, direct to the final destination | Redirect chains lose signal and waste crawl time |
| Pagination / filters | Faceted URLs excluded or canonicalised deliberately | Thousands of near-duplicate URLs diluting a small site |
Open Google Search Console → Pages (the indexing report). Passing: the number of indexed pages is close to the number of pages you actually want indexed, and the "Not indexed" reasons are ones you intended (noindex on utility pages, canonicalised duplicates). Failing sign: a large "Crawled — currently not indexed" or "Discovered — currently not indexed" bucket. Those two mean the engine saw your pages and declined to keep them — which is nearly always a content-quality or duplication verdict, not a technical bug, and no amount of schema will fix it.
Technical SEO is a floor, not a ceiling. A perfect technical implementation of a mediocre page ranks nowhere. But a technical fault can make an excellent page ineligible, and ineligibility is absolute — no amount of quality survives a page that cannot be fetched, rendered, or indexed. This part is the floor, item by item, each with the exact check.
Start with the simplest possible test, because it catches the most expensive class of failure. Disable JavaScript in your browser and load a key page. What remains is close to what a text-based crawler and most AI fetchers see first.
The failure that matters here is client-side-only content. If your lesson text is fetched by JavaScript after load and injected into an empty <div id="root">, Googlebot will usually still get it — it renders JavaScript, on a delayed second pass. Many AI crawlers, as of early 2026, are far less patient: several fetch the raw HTML and do not execute JavaScript at all. A page that renders only in a browser is a page that answer engines see as blank.
# The 30-second version of this test, from a terminal: curl -sL "https://learn.projekinaja.com/tutorials/seo-in-the-ai-era.html" | wc -c # then eyeball the actual text a bot receives: curl -sL "https://learn.projekinaja.com/tutorials/seo-in-the-ai-era.html" \ | sed 's/<[^>]*>//g' | head -60
If your headline, your first paragraph, and your key answers are not in that output, fix that before touching anything else on this page. Server-side rendering or static HTML is the fix; there is no clever workaround.
Every page on the ProjekinAja Learn platform is pre-rendered static HTML with the content in the markup. That was originally a simplicity decision. In the AI-crawler era it is also an optimisation: the content is available to every fetcher, including the ones that will never run a line of your JavaScript.
Google measures three aspects of real-world experience, collected from actual Chrome users rather than a lab test. As of 2026 the trio is:
| Metric | Measures | "Good" threshold | Usual culprit |
|---|---|---|---|
| LCP — Largest Contentful Paint | When the main content becomes visible | ≤ 2.5 seconds | Huge unoptimised hero image; slow server response |
| INP — Interaction to Next Paint | How fast the page responds to taps and clicks, across the whole visit | ≤ 200 milliseconds | Heavy JavaScript blocking the main thread |
| CLS — Cumulative Layout Shift | How much content jumps around while loading | ≤ 0.1 | Images without width/height; late-injected banners |
INP replaced FID in March 2024. This matters more than a metric swap sounds, because the two measure genuinely different things. FID only measured the delay before the browser began processing your first interaction — a forgiving metric that most sites passed. INP measures the full latency from interaction to visual update, across all interactions in the visit, and reports near the worst of them. Sites that comfortably passed FID routinely fail INP. If you last checked your vitals before 2024, you have not checked them.
Core Web Vitals are a real ranking signal and a genuinely real user-experience factor — but they are a tiebreaker, not a lever. Moving LCP from 3.1s to 2.4s will not lift a page that lacks information gain. Fix red vitals because slow pages lose readers; do not expect a green score to substitute for a reason to exist. Measure with field data (Search Console's Core Web Vitals report or the Chrome UX Report), not only Lighthouse — lab scores on your fast laptop are not what your users experience.
Google has indexed the mobile version of pages as the primary version for years; by 2024 the transition to mobile-only indexing was effectively complete. The practical rule is blunt: if content is hidden or absent on mobile, it may as well not exist. Desktop-only content is invisible content.
There is a bonus most teams miss. Semantic, accessible HTML — real <h1>–<h3> hierarchy, real <table> elements for tabular data, real lists, descriptive alt text, meaningful link text — is the same thing a language model needs to parse your page correctly. Accessibility work and machine-extractability work are largely the same work. A page built from styled <div>s is hostile to a screen reader and to an answer engine for the identical reason: the structure exists only visually.
Structured data is JSON-LD in a script tag that states, unambiguously and in a machine's own vocabulary, what your page is. It does not make you rank higher directly. It does two other things that matter: it makes you eligible for rich results, and it removes ambiguity for any system trying to extract facts — which now includes answer engines.
The types worth implementing for an education platform, in priority order:
| Type | Use it on | Realistic benefit |
|---|---|---|
Course + CourseInstance | Every course landing page | Eligible for Google's course rich results; states duration, mode, provider, price unambiguously |
LearningResource | Individual lessons | No Google rich result, but a precise schema.org description of educational level, learning type, and audience |
Article / TechArticle | Tutorials and posts | Author, dates, publisher stated explicitly rather than guessed from the page |
FAQPage | Pages with genuine Q&A | Rich results heavily restricted since 2023 — but the markup still labels your Q&A blocks cleanly for extraction |
HowTo | Step-by-step procedures | Google deprecated the rich result in 2023; the markup remains a valid structural description |
BreadcrumbList | Every page | Breadcrumb display in results; states your site hierarchy explicitly |
Organization + Person | Homepage, author pages | Entity consistency — see Part 3, this is more valuable than it looks |
First: Google has been steadily retiring rich result types, not adding them. FAQ rich results were restricted to well-known health and government sites in 2023, and HowTo rich results were deprecated the same year. Implement FAQPage and HowTo because they describe your content accurately for any consumer, not because you expect stars in the SERP. Anyone promising you FAQ rich results in 2026 has not checked since 2022.
Second: schema must describe what is actually on the page. Marking up reviews you don't display, an FAQ that isn't visible, or a price you don't charge is a spam violation that can earn a manual action. Schema spam is one of the few things on this page that can actively hurt you rather than merely waste your time.
Here is a complete, valid Course block — the shape you want on a course landing page:
<script type="application/ld+json"> { "@context": "https://schema.org", "@type": "Course", "name": "Agentic AI Mastery", "description": "A five-stage course on how AI agents work and how to build them.", "url": "https://learn.projekinaja.com/agentic-ai-mastery/site/index.html", "provider": { "@type": "Organization", "name": "ProjekinAja Learn", "url": "https://learn.projekinaja.com/" }, "inLanguage": "en", "isAccessibleForFree": true, "educationalLevel": "Beginner to intermediate", "teaches": ["AI agents", "tool use", "multi-agent systems"], "hasCourseInstance": { "@type": "CourseInstance", "courseMode": "online", "courseWorkload": "PT12H" } } </script>
Two details people get wrong. courseWorkload uses ISO 8601 duration format, so twelve hours is PT12H, not "12 hours". And isAccessibleForFree: true should only appear if the course really is free — it is a factual claim that a machine will repeat on your behalf.
Run the page through Google's Rich Results Test and validator.schema.org. They check different things: Google tells you whether you qualify for a rich result, schema.org tells you whether your markup is valid at all. Passing: zero errors, and the detected type matches what the page actually is. Failing sign: "no items detected" — which almost always means invalid JSON (a trailing comma is the usual culprit) rather than a missing feature. Warnings about recommended-but-absent properties are fine to leave; errors are not.
This is the section with a genuine strategic decision in it, and it is the one most sites have never consciously made. Their robots.txt still says User-agent: * / Allow: / from the day the site launched, which means they are opted in to everything by default — possibly correctly, but not deliberately.
The decision is easier once you separate one question that is usually collapsed into one:
Collect text to train future models. Your content may influence a model's general knowledge, with no link back, ever. Blocking costs you nothing in traffic.
Fetch pages so the product can cite you right now, with a visible link. Blocking these removes you from AI answers entirely. This is usually a mistake.
Conflating those two is how sites accidentally delete themselves from ChatGPT's search results while trying to opt out of model training. Here is the roster as documented by the operators as of early 2026 — verify current tokens against each vendor's own documentation before you deploy, because this list changes several times a year:
| User-agent | Operator | What it's for | Reasonable default |
|---|---|---|---|
GPTBot | OpenAI | Crawls content for model training | Your call — pure training, no citation back |
OAI-SearchBot | OpenAI | Indexes pages so ChatGPT search can surface and link them | Allow — this is the one that cites you |
ChatGPT-User | OpenAI | Live fetch when a user's prompt makes ChatGPT open your page | Allow — a real person is asking for your page |
ClaudeBot | Anthropic | General crawling, including for training | Your call |
Claude-User / Claude-SearchBot | Anthropic | User-initiated fetches and search indexing | Allow — these produce citations |
PerplexityBot | Perplexity | Indexes pages for Perplexity's answers and citations | Allow — Perplexity cites prominently |
Perplexity-User | Perplexity | Live fetch triggered by a user's question | Allow |
Google-Extended | A robots token, not a crawler: controls whether your content trains and grounds Gemini | Read the caveat below before touching this | |
Applebot-Extended | Apple | Opt-out token for Apple's generative model training (Applebot itself serves Siri/Spotlight) | Your call — does not affect Apple search features |
CCBot | Common Crawl | Builds the open web corpus that many models train on | Your call — note it also feeds legitimate research |
Bytespider | ByteDance | Crawling, widely reported as aggressive | Block if server load is a concern |
Meta-ExternalAgent | Meta | Crawls for AI training | Your call |
You cannot opt out of Google's AI Overviews while staying in Google Search. AI Overviews are generated from the regular search index, so the only way to be excluded is to be excluded from search itself — via nosnippet, max-snippet, or blocking Googlebot, all of which cost you ordinary organic traffic. Google-Extended is a narrower control: it governs whether your content is used to train and ground Gemini, and Google states it does not affect your Search ranking. Blocking it is a defensible training-data choice. It is not a way to escape AI Overviews, and anyone who told you it was has cost their clients real traffic.
A robots.txt that reflects the "let them cite me, decide separately about training" posture:
# robots.txt — learn.projekinaja.com # Posture: everything that can cite us is welcome. # Training-only crawlers are a separate, deliberate decision. User-agent: * Allow: / # --- search & retrieval crawlers: these link back to us --- User-agent: OAI-SearchBot Allow: / User-agent: ChatGPT-User Allow: / User-agent: PerplexityBot Allow: / User-agent: Claude-SearchBot Allow: / # --- training crawlers: edit these two lines to match YOUR decision --- # We allow them: our goal is reach, and our courses are free. # To opt out instead, change each Allow to Disallow. User-agent: GPTBot Allow: / User-agent: ClaudeBot Allow: / User-agent: CCBot Allow: / Sitemap: https://learn.projekinaja.com/sitemap.xml
Ask one question: is your content your product, or your marketing? If people pay for the words themselves — a paywalled research service, a subscription publication — training crawlers extract value and return none, and blocking is rational. If the content exists to make people aware of something else you sell, being inside a model's general knowledge is closer to distribution than to theft. ProjekinAja Learn's courses are free and exist to build reputation, so the roster above allows everything. A paid-course platform could reasonably reach the opposite conclusion with the same reasoning.
It has no enforcement mechanism. Reputable operators honour it; scrapers ignore it and can spoof user-agents freely. If you genuinely need to stop a crawler, that is a job for your CDN or WAF — Cloudflare, for instance, ships managed rules for AI bots and began blocking them by default for new domains in 2025. Also note that Disallow stops fetching, which is not the same as removing content already ingested into a model trained last year. Nothing in robots.txt is retroactive.
llms.txt is a proposal, introduced in late 2024, for a markdown file at your site root that gives language models a clean, curated map of your most useful content — a table of contents written for a machine, free of navigation chrome and marketing copy.
# llms.txt — https://learn.projekinaja.com/llms.txt # ProjekinAja Learn > A free education platform with 8 courses and 340 lessons on AI, robotics, > software engineering, finance, and clear thinking. All content is free, > browser-based, and requires no account. ## Courses - [Agentic AI Mastery](https://learn.projekinaja.com/agentic-ai-mastery/site/index.html): how AI agents work and how to build them, in 5 stages - [Financial Literacy Mastery](https://learn.projekinaja.com/financial-literacy-mastery/site/index.html): personal finance from budgeting to investing - [Clear Thinking](https://learn.projekinaja.com/clear-thinking/site/index.html): reasoning, bias, and decision-making ## Tutorials - [SEO in the AI Era](https://learn.projekinaja.com/tutorials/seo-in-the-ai-era.html): search visibility for humans, crawlers, and answer engines ## About - Publisher: ProjekinAja (PT ProjekinAja), Indonesia - Contact: learn@projekinaja.com
As of early 2026, no major AI engine has publicly confirmed it uses llms.txt, and Google search representatives have said plainly that their systems do not. It is a community convention with real adoption among documentation sites and effectively zero confirmed consumption by the large engines. The reason to ship it anyway is that it takes twenty minutes, cannot hurt, and forces a genuinely useful exercise: writing down what your best content actually is. Ship it as a cheap option on a possible future, not as a growth tactic. If someone tells you llms.txt is how you get into ChatGPT, they are guessing.
Generative Engine Optimization is a young practice with more confident advice than evidence behind it. What follows is split deliberately: mechanisms that follow from how these systems demonstrably work, and tactics that are plausible but unproven. Treat the first group as engineering and the second as experiments.
Strip away the marketing and the pipeline is roughly the same across products: interpret the question, fan out into several searches, retrieve a set of candidate pages, chunk them into passages, select the passages that best support an answer, synthesise, and attach citations to the passages actually used.
The selection step is where GEO lives, and it has a specific bias worth internalising: the model is looking for a passage it can lift with minimal risk of being wrong. That preference favours text with particular properties — and those properties are things you control entirely.
A quotable passage makes sense on its own, torn out of the page, with no surrounding context. A passage that begins "As we saw above, this is why it matters…" is unusable — the model would have to reconstruct three paragraphs of context to quote it safely, so it takes someone else's cleaner sentence instead.
The shape that gets lifted is consistently the same three moves:
"As mentioned earlier, this is something a lot of people get wrong, and it's worth thinking carefully about before you decide either way, since it really depends on your situation."
"INP replaced FID as a Core Web Vital in March 2024. Unlike FID, which measured only the delay before the first interaction was processed, INP measures full interaction-to-paint latency across the whole visit — so sites that passed FID often fail INP."
Read those two again. The second is not better writing in some abstract sense; it is better because a machine can quote it and be correct. Every self-contained paragraph on your page is an independent chance to be cited.
Journalists call it the inverted pyramid and it was always good practice for human readers, who decide in seconds whether to stay. It is now also machine-readable structure. Under a heading that poses a question, the first sentence should answer it completely.
## What is generative engine optimization? Generative engine optimization (GEO) is the practice of making web content more likely to be retrieved, quoted, and cited by AI answer engines such as Google AI Overviews, ChatGPT search, and Perplexity. // ← everything above is quotable on its own. // Now elaborate for the human who kept reading. It overlaps heavily with classic SEO, because most answer engines retrieve from conventional search indexes. It differs in emphasis: GEO optimises for being *extractable and citable*, not only for ranking position...
Notice what this costs you: nothing. Human readers get their answer sooner, which improves engagement; machines get a clean self-contained definition. There is no tension between the two audiences here, which is why this is the first tactic to adopt.
Retrieval systems split pages into chunks. Formats that carry their own context through that split survive it; prose that depends on distant context does not.
| Format | Why it survives extraction | Use it for |
|---|---|---|
| Q&A blocks | The question is in the text, so the passage is pre-matched to a query | Real questions users ask, in their words |
| Tables | Each row is a labelled fact with its own header context | Comparisons, specs, thresholds, options |
| Definition sentences | "X is Y" is unambiguous and directly answerable | Every term you introduce |
| Numbered steps | Ordered, self-labelled, individually meaningful | Procedures |
| Descriptive headings | Acts as the passage's title after chunking | Every section — write "How to add Course schema", not "Getting started" |
| Explicit numbers with units and dates | Checkable, quotable, hard to garble | Any claim involving a quantity |
Compare "studies show CTR is falling" with "Pew Research Center's July 2025 analysis found users clicked a result on about 8% of visits with an AI summary versus 15% without." The second is quotable because the attribution travels with the fact. A model can repeat it without inheriting the risk of an unsourced claim — and your name stays attached to the passage that carried it.
Search and AI systems model the world as entities: distinct things with attributes and relationships. If your brand appears as "ProjekinAja", "Projekin Aja", "PT ProjekinAja", and "Projekinaja Learn" across your own properties, you are asking every system to guess whether those are one organisation or four. Some will guess wrong, and your accumulated signals split.
The fix is unglamorous bookkeeping, and it is worth an afternoon:
Organization schema, your social profiles, your directory listings, and your email signature.Organization schema on your homepage with sameAs links to every profile you control — LinkedIn, GitHub, YouTube, Instagram. sameAs is how you assert "these accounts are all the same entity as me", which is exactly the disambiguation problem you're solving.Person schema and sameAs to their professional profiles.<script type="application/ld+json"> { "@context": "https://schema.org", "@type": "Organization", "name": "ProjekinAja Learn", "url": "https://learn.projekinaja.com/", "description": "A free education platform with 8 courses and 340 lessons on AI, robotics, software engineering, finance, and clear thinking.", "parentOrganization": { "@type": "Organization", "name": "PT ProjekinAja" }, "sameAs": [ "https://www.linkedin.com/company/projekinaja", "https://github.com/projekinaja-hub" ] } </script>
Only list profiles that genuinely exist and that you control. A sameAs pointing at a dead or wrong account is worse than an omission, because you have asserted a false identity relationship.
This is the tactic that most resembles old-fashioned public relations, and it is probably the highest-ceiling item in Part 3. Independent analyses of AI citations through 2025 converged on a striking pattern: a small set of sources is cited far out of proportion to their share of the web — Wikipedia, Reddit, YouTube, major reference sites, and established niche authorities in each vertical. Community and user-generated discussion appears especially heavily in recommendation-shaped questions ("best X for Y").
The uncomfortable implication: if the answer to "best free AI course" is being assembled from a Reddit thread, three review sites, and a YouTube comparison, then your beautifully optimised landing page is not in the running. You are not competing for that citation. You are absent from the retrieval set entirely.
What actually moves this, in order of honesty rather than convenience:
Astroturfing community platforms with promotional accounts, paying for fake reviews, or spinning up a network of thin sites to cite yourself. Beyond being dishonest, these are exactly the patterns platforms are best at detecting, and the failure mode is a ban that removes you from the source you were trying to appear in — a strictly worse position than where you started. Google's spam policies specifically cover site reputation abuse and scaled content abuse; both were enforcement priorities through 2024–2025.
Answer engines favour recent sources for questions where recency matters — pricing, model versions, "as of" claims, anything with a year in the query. They do not reward recency for questions where it is irrelevant; a 2019 explanation of compound interest is not worse for being from 2019.
So: date your pages honestly, update the ones where the world moved, and state the date inside the text as well as in metadata ("as of early 2026, no major engine has confirmed…"). That in-text stamp travels with the passage when it is extracted, which both helps a model judge whether to use it and protects you from being quoted as an authority on a fact that expired. Do not rotate dates on unchanged pages — it is a transparent trick, and it destroys the one signal you were trying to send.
Here is the honest starting position: you cannot fully measure this yet. Anyone presenting a complete AI visibility dashboard is presenting estimates. What follows is what is genuinely measurable, what is only sampleable, and what is invisible — because knowing which is which prevents a whole category of bad decisions.
| Signal | Where to see it | Trust level |
|---|---|---|
| Referral traffic from AI products | GA4 or your analytics → referrers: chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com | Real and exact. Small numbers, but these are people who chose to click through from an answer |
| Clicks from Google AI Overviews | Search Console — but not separable | Not isolatable. AI Overview clicks are folded into ordinary Search performance with no dedicated filter |
| CTR decline at stable position | Search Console → Performance, compare period over period | Strong inference. Same average position, falling CTR on informational queries, is the signature of an AI answer absorbing the clicks |
| Brand mentions inside AI answers | Manual prompt testing, or a third-party AI-visibility tracker | Sampling only. Answers are non-deterministic and personalised; one test is an anecdote |
| AI crawler hits | Server access logs, filtered by user-agent | Real. Tells you who is actually fetching you and how often — the most underused data source here |
| Branded search volume | Search Console → queries containing your brand | Good proxy. If AI answers are naming you, some of those people search your brand afterwards |
Since no reliable universal tool exists, build the cheapest honest version: a fixed panel of prompts, run on a fixed schedule, recorded in a spreadsheet.
These systems are non-deterministic. The same prompt can produce different sources on two consecutive runs, and results vary by account, location, and model version. A single bad answer is noise. A pattern across twenty prompts and three months is signal. Track the trend line, and resist the urge to rewrite a page because one chat session didn't mention you.
Passing: you have a dated spreadsheet with 15+ prompts, mention and citation columns filled for at least one engine, and a note of the top three competing sources per prompt — plus a filtered analytics view for AI referrers. Failing sign: "we asked ChatGPT once and it mentioned us, so we're doing well." That is a coin flip you happened to win. Without the baseline you will not be able to tell whether anything in this tutorial worked, which makes the next three months of effort unfalsifiable.
Abstract advice is easy to agree with and impossible to act on. So we apply everything above to a specific property: learn.projekinaja.com — a free education platform with 8 courses and roughly 340 lessons, published as static HTML, covering AI, robotics, software engineering, finance, and thinking skills.
Its situation is typical of content-rich sites: a lot of genuinely useful pages, a small brand, no link-building budget, and a topic area where AI answer engines are heavily used by exactly the people it wants to reach. Someone learning about AI agents is, almost by definition, already asking an AI where to learn about AI agents.
For a platform like this, the money query is not "agentic AI course" as a keyword. It is the recommendation prompt: "what's a good free course to learn how AI agents work?" That question is now answered far more often by a chat interface than by a results page, and the winner is whoever the engine can retrieve, understand, and safely recommend. Everything below serves that one goal.
Ten items. Each names the exact file, the exact check, and what failing looks like — so this can be handed to someone else and finished without you.
Every course index page carries a Course block with provider, description, language, workload, and free-access status. This is the single highest-value schema on the site: it turns "a page about a course" into "a course, with these properties" for every machine that reads it.
Passing: Google's Rich Results Test detects a valid Course on all 8, with no errors. Failing: "no items detected" (nearly always malformed JSON), or a courseWorkload written as "12 hours" instead of the ISO 8601 PT12H.
Each of the ~340 lesson pages describes itself as a learning resource belonging to its parent course. There is no rich result for this and there is no promise of one — the value is a precise, machine-readable statement of what the page teaches and to whom.
<script type="application/ld+json"> { "@context": "https://schema.org", "@type": "LearningResource", "name": "Lesson 2.3 — How agents use tools", "description": "How an AI agent calls external tools, what a tool schema is, and why tool errors are the most common failure mode.", "url": "https://learn.projekinaja.com/agentic-ai-mastery/site/stage-2-how-agents-work/lesson-2-3.html", "learningResourceType": "Lesson", "educationalLevel": "Beginner", "teaches": "Tool use in AI agents", "inLanguage": "en", "isPartOf": { "@type": "Course", "name": "Agentic AI Mastery", "url": "https://learn.projekinaja.com/agentic-ai-mastery/site/index.html" }, "isAccessibleForFree": true } </script>
Passing: generated automatically by the build script from lesson front-matter, so no lesson can ship without it. Failing: hand-written on a few pages — 340 pages will never be maintained by hand, and half-coverage is how you get a schema that lies about your site.
Every lesson gets a unique title and description. The title pattern is Lesson topic — Course name · ProjekinAja Learn: specific first, brand last, because the beginning of a title carries the most weight in both search results and extraction.
Passing: crawl the site and confirm zero duplicate titles and zero duplicate descriptions across all ~350 pages. Failing: 340 lessons all titled "ProjekinAja Learn" or "Lesson" — which is both a ranking problem and a sign the build template is doing nothing.
The description tag is not a ranking factor and search engines frequently rewrite it. It is still worth writing, for two reasons: it is often the snippet a human reads before deciding to click, and it is a clean one-sentence summary of the page that any extraction system can use. Write it as a real summary, not a keyword list.
The structure is already natural for a course platform, which makes this cheap: the site index links to all 8 course indexes; each course index links to all of its lessons; each lesson links back to its course index and to its previous and next siblings. The gap to close is contextual links — a lesson on tool use should link to the lesson on agent memory when it mentions it, in body text, with descriptive anchor text.
Passing: every lesson is reachable from the homepage within three clicks, and no page's only inbound reference is the sitemap. Failing: an orphan lesson that was published, linked from nowhere, and quietly never indexed — the most common silent failure on large content sites.
Each course index answers the five questions people actually ask before starting: Is it free? How long does it take? Do I need to code? Is there a certificate? What should I do first? Visible on the page for humans, marked up with FAQPage for machines.
Passing: the questions are phrased the way a person would type them, the answers are complete in one or two sentences, and the marked-up text matches the visible text exactly. Failing: invented questions nobody asks ("Why is ProjekinAja the leading platform?"), or — the disqualifying version — FAQ schema describing content that isn't on the page.
Ship the file from Part 2b at /llms.txt, listing the 8 courses and the tutorials with a one-line description each.
Passing: it loads as plain text, every link resolves with a 200, and it is regenerated by the build so it cannot drift. Failing: writing it once by hand, adding a ninth course six months later, and shipping a file that now misrepresents the site. Expected benefit: unproven — see the caveat in Part 2b. Twenty minutes, low confidence, cheap option.
Replace the default User-agent: * / Allow: / with the explicit roster from Part 2b, including the comment explaining why each decision was made. For this platform the reasoning is clear: the courses are free and exist to build reputation, so every crawler that can cite or recommend the platform is welcome, and training crawlers are allowed too.
Passing: a person reading the file can tell what the policy is and why. Failing: a copied block from a blog post that blocks Googlebot or OAI-SearchBot by accident. Before deploying any robots.txt change, read it line by line and ask "which of these removes me from something I want to be in?"
With ~350 URLs, the sitemap must contain only canonical, indexable, 200-status URLs, and every page must self-canonical to its absolute HTTPS address.
Passing: Search Console reports the sitemap as read with a discovered count matching your published count, and the indexing report shows no "Alternate page with proper canonical tag" surprises. Failing: a sitemap listing both /index.html and the directory form of the same page, teaching the crawler that your site has twice as many pages as it does.
This is the only manual item, and it is the one that most directly targets citation. Take the 20 lessons with the most search impressions and rewrite the first paragraph of each so it fully answers the question in the title, standalone, in two or three sentences. Do not touch the other 320 yet — prove the pattern first.
Passing: for each of the 20, you can delete the entire rest of the page and the opening paragraph still correctly answers the title's question. Failing: intros that begin "In this lesson, we'll explore…" — a sentence that contains no answer, tells the reader nothing, and gives an extraction system nothing to lift.
Open any lesson you've written. Copy the first paragraph into a blank document. Read it as if you'd never seen the page. Expected result: it answers the question in the heading, on its own, without needing the page around it. If it instead announces what the page is going to do, you found a paragraph that will never be quoted — rewrite it as claim, then evidence, then boundary.
Record, on a dated sheet: current indexed page count, current impressions and CTR from Search Console for the top 50 queries, current AI referral sessions, and the manual prompt panel from Part 3b. Do this first. A baseline taken after the changes is not a baseline.
Passing: a dated snapshot you can diff against in 90 days. Failing: shipping ten changes at once with no before-state, then arguing about which one caused a movement you cannot attribute.
Technical fixes (schema, canonicals, sitemap) are read within days to weeks. Content improvements take a month or more to show in rankings. Off-site presence — the Part 3 Step 5 work that has the highest ceiling — takes months and does not move in a straight line. Anyone promising results in two weeks is describing a coincidence they intend to take credit for. Set the review date 90 days out and do not panic-tune in between.
Four walls almost everyone hits, and what each one actually means.
| Symptom | Most likely cause | What to do |
|---|---|---|
| "Crawled — currently not indexed" on many pages | A quality/duplication verdict, not a bug | Consolidate thin pages, add information gain, improve internal links to them. Resubmitting does nothing |
| Rich Results Test says "no items detected" | Invalid JSON — usually a trailing comma or a smart quote | Paste the JSON-LD into any JSON validator first; fix the syntax, then re-test |
| Rankings stable, traffic falling | An AI answer is absorbing the clicks above you | Confirm with a CTR-at-stable-position check, then shift the goal from clicks to citations for those queries |
| Never mentioned in AI answers despite ranking | Retrieved but not extractable, or absent from the off-site sources being retrieved | Run the deletion test on your intros; then check which sources are cited and get present there |
| Mistake | Why it hurts | Fix |
|---|---|---|
| Blocking Google-Extended to "opt out of AI Overviews" | It doesn't do that — you lose Gemini grounding and gain nothing | Understand the token first; AI Overviews cannot be exited without exiting Search |
| Blocking OAI-SearchBot / PerplexityBot with the training crawlers | Deletes you from the AI results that would have linked to you | Separate the two decisions; allow everything that cites |
| Publishing bulk AI-generated pages | Google's scaled content abuse policy targets mass-produced low-value content regardless of author | Fewer pages, real information gain, a named human accountable for each |
| Schema for content that isn't on the page | Structured data spam is a manual-action offence | Markup must mirror visible content, exactly |
| Chasing FAQ rich results | Restricted since 2023 — the stars aren't coming back | Use FAQ markup for extraction clarity, not for SERP decoration |
| Client-side-rendered content | Many AI crawlers don't execute JavaScript; your page reads as empty | Server-side render or ship static HTML with the text in the markup |
| Optimising vitals instead of content | A fast page with nothing new still has no reason to rank | Fix red vitals, then spend the remaining time on information gain |
| "In this lesson we'll explore…" intros | Announces instead of answering; nothing to extract | Answer in sentence one; elaborate after |
| Inconsistent brand naming across properties | Splits your entity; engines can't tell it's one organisation | One canonical name everywhere, plus Organization schema with sameAs |
| Judging AI visibility from one chat session | These systems are non-deterministic and personalised | A fixed prompt panel, run monthly, logged out |
| Date-rotating unchanged pages | A transparent trick that burns your freshness signal | Update the content, then the date — in that order |
| Treating llms.txt as a growth tactic | No major engine has confirmed using it | Ship it as a cheap option; expect nothing from it yet |
WEEK 1 □ Baseline: indexed pages, top-50 impressions & CTR, AI referrals
□ Build the 15–25 prompt panel; run it once, logged out; record sources
□ JS-off test on your three most important pages
WEEK 2 □ Fix crawl blockers found in week 1 (rendering, canonicals, sitemap)
□ Write a deliberate robots.txt; decide training vs citing separately
□ Ship Organization + Person schema; unify brand naming everywhere
WEEK 3 □ Course / Article / LearningResource schema, generated by the build
□ Rewrite the top 20 page intros for the deletion test
□ Add real FAQ blocks where genuine questions exist
WEEK 4 □ Internal links: kill orphans, fix "click here" anchors, add context links
□ Ship llms.txt (cheap option, low expectation)
□ Pick ONE off-site presence to build, and start showing up there weekly
DAY 90 □ Re-run the prompt panel and the analytics baseline. Compare. Then decide.
Where to go next. The highest-ceiling work is the least technical: being genuinely present and useful in the places answer engines retrieve from, under a real name, over months. The second is publishing something worth referencing — original data, a real benchmark, a definitive guide to a narrow thing. Both are slow. Both compound. Neither can be bought in a two-week sprint, which is precisely why they still work.
Nobody can guarantee you a ranking or a citation, and any tool, agency, or tutorial that promises "first position" is selling certainty that does not exist. Search and AI systems change monthly; specific crawler names, schema support, and engine behaviours in this tutorial were accurate as of August 2026 and some of them will be wrong within a year — verify against primary documentation before you deploy. What ages well is the underlying logic: be findable, be genuinely useful, be the clearest available answer, and be honest about what you don't know. That has survived every algorithm update so far, and it is the only strategy that doesn't need to be replaced when the next one lands.
Primary documentation only — these are the pages to check when something in this tutorial has aged.