SEO basics in 2026

SEO is the work of making your information discoverable, understandable, trustworthy and useful to search systems and to the people using them. That is the whole definition, and notice what is missing from it: Google, keywords, and the number one.

The belief I want to kill first is that SEO means ranking number one on Google. It is the most common thing a beginner arrives with and it quietly ruins every decision that follows, because it treats one position on one surface as the goal.

Organic discovery in 2026 happens across at least nine surfaces:

• Traditional Google and Bing results
• AI Overviews
• Google AI Mode
• ChatGPT search, and the other AI answer systems
• Image results
• Video results
• Maps and local packs
• Shopping results
• Forum and discussion results, which are somebody else's site carrying your name

A ranking is an intermediate metric. It sits in the middle of a chain that runs visibility, then qualified discovery, then action, then a business outcome. Every step of that chain can be measured. Only one of them pays for anything.

I stopped being able to describe my own results as rankings a while ago. Across the properties I own, the Bing AI Performance report has logged 302,037 Copilot citations in six-month windows, and Search Console has logged Google AI feature appearances since May 2026. Those are not positions. There is no position to report. The unit of measurement changed and the vocabulary has not caught up.

So the working question is never "where do I rank". It is "can the systems people use find me, understand me, trust me, and is any of that turning into customers".

Google Search works in three stages: crawling, indexing, and serving results. That is Google's own framing, and its documentation adds the sentence most SEO courses leave out: not all pages make it through each stage.

The opponent here is "submit your site to Google". There is no submission queue. Google's documentation says the vast majority of pages in its results were never manually submitted, they were found by crawlers following links.

I split those three official stages into seven below, because they fail separately and each failure looks completely different in Search Console. Discovery failing is not crawling failing. Rendering failing is not indexing failing. Treating them as one thing is why people "fix" the wrong problem for months.

Two sentences from Google's documentation are worth memorizing, because between them they close most arguments. Google does not accept payment to crawl a site more frequently or to rank it higher. And Google does not guarantee that it will crawl, index, or serve your page, even if the page follows the Search Essentials.

Read that second one again. Indexing is not a right you earn by publishing. It is a decision made about your page, every time, and "I built it properly" is not an appeal.

This is the single highest-value distinction on the page. Once someone genuinely holds those four states apart, half the questions they were going to ask answer themselves.

All of that is theory until you look at one of your own URLs, so do that now. Three minutes, free, and the rest of this lesson leans on it.

Three things on that screen are worth slowing down for, because they are the whole chapter in one screenshot.

• Discovery lists the referring pages. Four of them, one on a completely different domain. That is not a theory about how discovery works, it is the actual route, named. Chapter seven is about deliberately creating those routes.
• Crawled as Googlebot smartphone. Not the desktop crawler. If your page is different on a phone, the phone version is the one that counts.
• User-declared canonical and Google-selected canonical match. I voted, and my vote won. When those two lines disagree, that is the "Duplicate, Google chose a different canonical" case in chapter eight, and it is telling you your own signals contradicted each other.

And here is the practical version of the same knowledge. Pick your symptom, get the stage, then work on that one thing rather than everything at once.

One more thing before you move on. Later in this lesson there is a single real page of mine followed through all seven of these stages with its actual numbers at each step. It sits down there rather than here because it makes far more sense once you have read the chapters in between.

AI search changed the retrieval and presentation layers. It did not create a separate discipline with its own technical stack, and Google says so directly: its generative AI features are rooted in its core Search ranking and quality systems.

I spend most of my working week on AI search and I still think "AI SEO" is a misleading name for most of what it involves. The opponent in this chapter is the idea that there are now two disciplines and you are behind on the new one. There are two surfaces. There is one discipline.

What is genuinely new is worth naming precisely, because the precision is what makes it actionable. Google describes two techniques behind its generative features:

• Retrieval-augmented generation, also called grounding. The model relies on the core Search ranking systems to retrieve relevant, current pages from the index, then reviews the specific information in those pages to build a response, with clickable links to the pages that support it.
• Query fan-out. A set of concurrent, related queries the model generates to fetch more information than the original question would return on its own.

Fan-out is the one that changes how you plan content, so here it is with Google's own worked example rather than mine.

The consequence is not "write longer pages". It is that a single exact-match keyword is an incomplete description of the question being asked. Someone asking about the best CRM for a ten-person team is having six searches run on their behalf, and the sites that own the sub-questions get pulled into the answer even when they never targeted the headline query.

A warning I will repeat later because it is the mistake this insight causes: the answer is not a page for every sub-query. Google names that specifically as a scaled content abuse risk and as an ineffective long-term strategy. Cover the question properly, in as few pages as it honestly takes.

One more precision point. AI Overviews and AI Mode may use different models and different techniques, so the links they show will differ, and AI Overviews only appear when Google's systems judge them additive to normal Search, which means they often do not appear at all. Anyone quoting you a single "AI Overview presence" number for your whole market is quoting a sample and calling it a measurement.

For the deeper version of this mechanism, I wrote up why fan-out is a passage problem rather than a page-length problem in Query fan-out in SEO.

The query is not the intent. Two words in a search box are the compressed, lossy output of something a person wanted, and your job is to reconstruct the want rather than match the string.

The opponent here is search volume as the primary input. Volume tells you how many people typed something. It tells you nothing about what any of them were trying to do, and a page built to serve the wrong reading of a 12,000-a-month keyword earns impressions and nothing else.

The traditional classifications are worth knowing and not worth memorizing: informational, commercial investigation, transactional, navigational, local. They are a filing system, not an insight. Plenty of real queries sit in two at once and the label does not tell you what to build.

The question that does is: what is this person trying to accomplish? Then go and look at the results page, because the SERP is the engine's current published answer to exactly that question. If every result is a comparison table, Google has decided this is a comparison query, and your beautifully written definition page is going to lose to whatever is in position eight.

Intent also drifts. A query that was informational in 2023 can be commercial in 2026 because the market matured and the SERP followed. This is why intent gets re-checked on a refresh and not assumed from an old spreadsheet.

Keyword research is deciding which problems you are going to be the best answer to, and volume is the fourth-most-useful input into that decision.

The opponent is the process almost every beginner is taught: open a tool, sort by volume, filter by difficulty, export, write. That process produces a list of strings, and a list of strings is not a plan.

Four filters, in this order:

1. Relevance. Does this search belong to your business and your audience? A high-volume query you have no right to answer is a distraction with a spreadsheet attached.
2. Intent. What does this person actually need? See the chapter above.
3. Demand. How much of it exists, from every source, not just the volume column.
4. Opportunity. Can you realistically compete, and would winning matter to the business? Both halves. Winnable and worthless is the most common trap.

Demand is bigger than the volume column. It includes:

• Search volume, as an estimate, always named as an estimate
• Related queries and autocomplete
• Long-tail phrasings, where most real searches live
• Question phrasings, which tell you the shape of the confusion
• Your actual customer conversations, support tickets and sales calls
• Forums and communities where people already argue about this
• The SERP itself, and what it chooses to show
• Search Console, which is the best keyword tool I own and the only one reporting queries where a real person saw a real page of mine

Then there is the hierarchy that stops you building forty pages where four belong. Keyword, then query cluster, then topic, then user problem. These four strings look like four pages to a tool and are one page to a person:

• how to create website with AI
• AI website creator
• build site using AI
• AI WordPress site builder

Same intent, near-identical SERPs, one job to be done. One page. Google's 2026 guidance is explicit that creating separate content for every possible variation of how people might search, including fan-out queries, is both a policy risk and an ineffective long-term strategy, and that a high quantity of pages does not make a website higher quality or more relevant.

I have written the long version of this process, as step three of this roadmap, in keyword research and search demand. This chapter is the mental model. That page is the work, and it comes with the metrics pulled live so you can see volume and traffic potential disagree rather than take my word for it. The free Excel template is the working file you record the decisions in.

Read the results page before you decide what to build, because it is the only free, current, first-party statement of what the engine currently believes about this query.

The opponent is the metric that gets quoted instead: keyword difficulty. A difficulty score is a third-party model's guess, and I have watched people skip a query with a KD of 40 that two thin affiliate pages were holding, and grind for a year on a KD of 12 owned by three national brands.

Six things to record, and they take four minutes:

• Page types. Product pages, category pages, tutorials, comparisons, tools, forums, videos, news. If nine of ten results are one type, that is the format.
• Surfaces present. AI Overview, images, video, local pack, shopping, discussions. Each one is a slot you might be able to take.
• Who is winning. National brands or small specialists. A page of specialists is an invitation. A page of brands is a longer project.
• What the top results have in common. Not word count. Format, depth, evidence, freshness.
• What every one of them is missing. This is where your angle comes from.
• Whether the SERP is even about your reading of the query. If it is not, you found a different query.

Then ask the useful question: what job is the engine trying to do with this page? Everything on that results page is Google's attempt to satisfy an intent it has inferred. Understand the attempt and you know what you are competing with.

Treat volume, difficulty, CPC, DA and DR as inputs, never as answers. Two of them are third-party scores from Moz and Ahrefs, not Google metrics, and none of them has ever opened a results page and looked at it.

Your site is an information architecture, and its shape decides which of your pages ever get found. This is the most under-taught beginner topic and the one with the highest leverage per hour.

The opponent is "internal links pass juice". They move authority around your own site. They do not create it. If the whole domain has very little, you are carefully redistributing very little, and no linking pattern fixes that.

The pieces worth getting right, in order of how often they are wrong:

• Hub pages. One page per real topic that links to everything under it and is linked from your navigation. This is the fix for depth and for orphans at the same time.
• Crawl depth. Count the clicks from your home page to your most important page. If the answer is more than three, that is the project.
• Orphan pages. Pages nothing links to. Run any crawler over your site and you will find some, and they will surprise you.
• Anchor text. Describe the destination. "Click here" throws away free context for nothing.
• Breadcrumbs. They help people locate themselves and they state your hierarchy in one line. Two jobs, small cost.
• Crawlable links. An anchor tag with an href. Google lists crawlable links and making important content findable through internal links among its core best practices, and a click handler on a div is not a link.

A practical way in: I built a free AI internal linking tool because doing this by hand across a few hundred URLs is miserable, and the build takes about 30 minutes. The prompt is in the post. The suggestion quality comes from the URL list you index, not from the model.

Technical SEO for a beginner is five concepts, not a career. You need to understand crawlability, indexability, canonicalization, status codes and sitemaps well enough to recognize which one is broken. You do not need to become a server engineer.

The opponent is the 200-point audit. I have been handed several. Most of the items are hygiene worth ten minutes, presented with the same weight as the two items that were actually costing the site its traffic.

Crawlability

Can a crawler fetch the URL. The three things that stop it are robots.txt, your CDN or firewall blocking the bot, and pages behind a login. In my experience bot protection blocks more crawls than robots.txt does, and it is the one nobody checks because nobody configured it deliberately.

Indexability, and the mistake that costs the most

Whether you have told search engines they may store the page. A robots meta noindex, or an X-Robots-Tag header, is how you say no.

If you want a page out of the index, leave it crawlable and let the crawler read the noindex. Doing both at once is the most common self-inflicted wound in technical SEO, and it is worth being able to explain the mechanism rather than just remembering the rule.

Canonicalization

During indexing, Google groups near-identical URLs into a cluster and picks one to represent it. That chosen URL is the canonical, and it is the one that may be shown.

You get a vote, through rel=canonical, your internal links and your sitemap. You do not get a veto. When Search Console says "Duplicate, Google chose a different canonical than user", it is telling you your vote lost, and usually that your own signals disagreed with each other.

You can see both sides of that vote on the URL Inspection screenshot in chapter two: "User declared canonical" is my vote, "Google-selected canonical" is the verdict, and on that page they agree. Those two lines disagreeing is the entire diagnosis.

Status codes

• 200, here it is
• 301 and 308, moved permanently, use these for real moves
• 302 and 307, temporary, so do not use them for permanent moves
• 404, not here, which is a perfectly correct answer
• 410, gone deliberately, a stronger signal than 404
• 5xx, my server is unwell. Googlebot reads this as "slow down", so a flaky server costs you crawl coverage

A 404 is not a problem to be eliminated. A 404 on a URL that used to earn traffic is. Those are different jobs and treating them the same is how sites end up redirecting everything to the home page, which is worse than the 404.

Sitemaps, and what they do not do

A sitemap tells search engines which pages you think matter and when they changed. It helps discovery. It does not request indexing and it is not a ranking signal.

Google's own guidance is more relaxed than the industry's: you might not need a sitemap if your site is about 500 pages or fewer, comprehensively linked internally, and light on media and news. You probably do need one if the site is large, brand new with few external links, or heavy on video and images.

The rest, briefly, because it is genuinely brief

• HTTPS. Table stakes. Not a differentiator since about 2017.
• Mobile. Most searches are mobile. If it does not work on a phone, nothing else on this page matters.
• JavaScript. Google renders pages and runs the JavaScript it finds using a recent version of Chrome. It usually works. It breaks when the script or the API endpoint is itself blocked, so check the rendered HTML rather than trusting your browser.
• Speed and Core Web Vitals. Real, worth fixing, and not a shortcut. Page experience is one part of a larger picture, and Google says explicitly not to focus on only one or two aspects of it.
• Redirects. Point them at the closest equivalent page, and keep chains to one hop.
• Duplicate content. Not a penalty. It splits signals and wastes crawling on URLs you do not care about, which is a different and more boring problem.
• Crawl traps. Faceted navigation and calendars that generate infinite URLs. The site looks enormous and none of it is worth storing.
• URL structure. Readable, stable, lowercase, no dates you will regret. Then stop thinking about it, and never change a published URL without a redirect.

One relieving note, straight from Google's 2026 guidance: you do not need perfect semantic HTML. The web in general is not valid HTML and Google can understand it. Semantic HTML is still worth writing, mostly because screen readers and, increasingly, browser agents parse your page with it.

If you want the ordered version of this as something to work through, the technical SEO checklist is the same content as a list you can tick.

On-page SEO is making the page's subject and value obvious to a person and to a machine at the same time, which is a writing problem far more than a settings problem.

The opponent is the optimization formula: keyword in the first 100 words, keyword in an H2, density between one and two percent, exact match in the URL. None of that has a documented basis and all of it makes pages worse to read.

What Google actually asks for is one sentence in its Search Essentials, and it is worth quoting closely: use words that people would use to look for your content, and place those words in prominent locations on the page, such as the title and main heading, and other descriptive locations such as alt text and link text.

That is it. Here is what it means element by element.

• Title element. The main subject and the value, in the words a searcher would use. It is also the biggest lever you have on click-through rate, which is a different job from ranking and often the more valuable one.
• H1. One per page, and it should make the page's purpose obvious in isolation.
• Headings. A real hierarchy, because it is how both a skimmer and a retrieval system find the part of the page that answers them. Google notes that people appreciate content organized into paragraphs and sections with headings that give it a clear structure.
• Main content. Use the terminology your readers use. If your industry says "customer relationship management" and every buyer types "CRM", write CRM.
• URL. Readable and stable. Stable is the important half.
• Internal links. Meaningful anchor text, in context, pointing at the pages that genuinely help next.
• Images. Useful imagery, sensible filenames, real alt text. Alt text is for people who cannot see the image, and it happens to be one of the descriptive locations Google names.
• Video. Increasingly useful, because search is multimodal and video is a separate surface you can win. Chapter 19.
• Meta description. This is presentation, not ranking. It is your pitch in the results page, Google will often rewrite it, and stuffing keywords into it does nothing at all.

Notice there is no word count here. Google states there is no ideal page length, and that shorter or longer can both work depending on your audience and subject. Write until the argument is finished, then stop.

Generic information became a commodity the moment a model could produce a competent version of it in four seconds, which means the only content worth writing in 2026 is content a model could not have produced without you.

This is the biggest chapter on the page because it is the biggest change, and it is not a technical one. The opponent is "comprehensive content wins". Comprehensiveness was a moat when assembling everything known about a subject took a week. It costs nothing now.

Google's 2026 guidance is unusually direct about this. It says creating content people find unique, compelling and useful will likely influence your presence in generative AI search more than any of the other suggestions in the guide, and it draws a line between commodity and non-commodity content with its own examples: "7 Tips for First-Time Homebuyers" against "Why We Waived the Inspection and Saved Money: A Look Inside the Sewer Line". The first could have been written by anyone. The second could only have been written by someone who was there.

The same document tells you not to just recycle what others have already said, or what could easily be produced by a generative AI model. That sentence is in Google's documentation, and it is the most useful editorial instruction Google has published in years.

The test I actually use

Before I publish anything, one question: could ChatGPT produce essentially this page without knowing anything about me, my company or my experience?

If yes, it is commodity content and I should not be surprised when it goes nowhere. If no, I have to be able to say what the difference is in one sentence, and that sentence usually turns out to be the angle the page should have led with.

What actually counts as a difference

• Original research, and the data behind it
• First-hand experience of the thing, described specifically
• An experiment, including the ones that failed
• Screenshots and exports from your own accounts
• Benchmarks, with the method stated
• A tool, a template or a calculator that does the job rather than describing it
• A case study with real numbers and real dates
• Original photos and video of the actual thing
• A genuinely clearer explanation, or a better synthesis, which is a real contribution and the one most people underrate
• Information that did not exist before you published it

My version of this is buying things. I have bought and tested more than 500 SaaS products with my own money since 2019, and the reason I can write a useful review is not writing skill, it is the receipt. Four of them are still on my card. A model can describe a tool's feature list. It cannot tell you which four you would keep paying for after six years, and that is the only part anyone actually wants.

You do not need my history. You need yours. The support inbox, the migration that went wrong, the customer question you answer twice a week, the number in your own dashboard. Those are the assets, and almost nobody publishes them.

AI-generated content is not against the rules. Publishing many pages without adding value for users is, and it has been against the rules since long before generative AI existed.

Google's own wording is that generative AI can be particularly useful when researching a topic and for adding structure to original content, and that using such tools to generate many pages without adding value for users may violate its spam policy on scaled content abuse. Note where the line falls. Not on the tool. On the outcome.

The opponent is a two-sided myth, and both sides cost money. One side says AI content is automatically penalized, so people avoid a tool that would genuinely help them research and organize. The other side says AI content is fine, so people publish four hundred pages of nothing and lose the whole site. Neither reads what the policy says.

Where AI genuinely helps, in my own workflow:

• Research and reading things faster than I could
• Organizing a mess of notes into an argument
• Outlining, and arguing with the outline
• Editing, especially cutting
• Idea generation against a real constraint
• Data analysis, where I can check the answer
• First drafts of sections where I already know the conclusion

What it cannot do is be the value. It has no accounts, no invoices, no customers and no scars. The principle I would give a beginner: AI should increase your leverage, not replace the thing that made the page worth reading.

Two more specifics from Google's guidance. It says to focus on accuracy, quality and relevance especially when generating content automatically, including the metadata that shows up in results: title elements, meta descriptions, structured data and alt text. And it suggests that sharing how a piece of content was created can give readers useful context, which is a good idea for reasons that have nothing to do with search.

One thing to ignore completely: AI detectors. They are probabilistic classifiers that routinely flag plain human writing, Google has never described using one, and optimizing your prose to please one makes it worse. If you want to know whether a page is any good, read it.

Google also points at the sections of its Search Quality Rater guidelines covering scaled content abuse and content created with little effort, originality or added value. Worth reading, with the caveat Google itself attaches: raters' ratings do not directly influence ranking. The guidelines describe what good looks like. They are not a lever.

E-E-A-T is not a ranking factor and not a score, and Google says so in the same document that explains it: while E-E-A-T itself is not a specific ranking factor, using a mix of factors that can identify content with good E-E-A-T is useful.

The opponent is the checkbox version. Add an author bio, add a photo, add credentials to the footer, and your E-E-A-T score goes up. There is no score to go up. What you did was make the page slightly more trustworthy to a human, which is worth doing for that reason alone.

The four parts are experience, expertise, authoritativeness and trust. Google adds two things worth holding on to. Trust is the most important of the four, and the others contribute to it. And content does not have to demonstrate all four: some pages are helpful because of experience, others because of expertise, and a first-hand review of a cheap gadget needs no formal credentials at all.

There is also a weighting worth knowing. Google says its systems give even more weight to content aligning with strong E-E-A-T on topics that could significantly affect health, financial stability, safety, or the welfare of society, which it calls Your Money or Your Life. If you are writing about medication or mortgages, the bar is genuinely higher. If you are reviewing keyboards, it is not.

The three questions that replace the checklist

Google frames its own guidance as who, how and why, and it is a better framework than any tool's audit:

• Who created this? Is there a real, named, findable person or organization behind it, and can a reader tell?
• How was it produced? Was it tested, measured, researched, automated? Would you be comfortable explaining the method?
• Why does it exist? To help someone do something, or to catch search traffic? Google's phrasing is people-first content against content made primarily to manipulate rankings, and the honest answer to this one predicts more than the other two.

How that shows up in practice: real named authors with real pages, first-hand demonstration, a proper About page, transparency about how you make money, sources you actually link, original research, credentials where they are relevant, and reviews from people who used the thing. Do those because they are true. The moment they become a checklist you are producing trust theatre, which readers spot immediately.

Which raises the obvious question about this page, so here is the answer rather than the theory.

Links still matter, and the mental model that makes them work is not "get more" but "become something worth referencing".

Two opponents this time, from opposite directions. "More backlinks means higher rankings", which produces a thousand worthless placements. And "links are dead now", which is what people say when they cannot get any.

Links do three jobs for a search system:

• They are how new URLs get discovered
• They are evidence of a relationship between two things
• They contribute to an assessment of authority and popularity

Which is why quality is not a nicety. One genuine editorial reference from a relevant authoritative site is a different kind of object from a thousand directory submissions, blog comments, automated placements, private network links, or paid links, and the last category is a spam policy violation rather than a gray area.

Which brings up the part of this chapter that was missing the first time I wrote it, and it matters most if you have ever put an affiliate link on a page.

The useful widening is from links to reputation. A search or AI system assembling an answer about your category is reading everything said about you across the web, and only some of it is a hyperlink. That means:

• Digital PR, meaning giving journalists something real to report
• Research people want to cite, which is the most reliable link magnet there is
• A free tool that does something genuinely useful
• Statistics you gathered and nobody else has
• Expert commentary, on the record, with your name on it
• Industry relationships, which is the unglamorous one that works
• Real community participation, where you are a member rather than a campaign
• Brand mentions without a link, which still carry your name into the corpus
• Reviews from people who paid you money

One caution, because it is the newest bad idea in this area. Google lists seeking inauthentic mentions among the things you can ignore, noting that its generative features depend on the same quality and spam systems as the rest of Search. Buying mentions is buying links with extra steps.

"Become something worth referencing" is slower than "build 50 links this month" and it is the only version that compounds. The internal linking tool I mentioned earlier got referenced because it was a working tool somebody could use, not because I emailed anyone about it.

An entity is a distinguishable thing: a person, a company, a product, a place, an organization, a concept. Search systems do not only match strings, they try to understand things, their attributes, their relationships and their context.

The opponent is the mystical version, sold as entity stacking or entity SEO, usually involving a lot of profile creation. The real work is duller and more effective: say the same true things about yourself in every place you say anything.

Why consistency does the work: a system that reads three different descriptions of your company on three of your own pages has no confident answer to give when someone asks about you. It is not being difficult. You gave it three answers.

So keep your organization information, product information, person information, About page, external profiles, structured data and naming in agreement. When I widened my own positioning from "AI SEO expert" to "SEO and AI search expert", the actual work was not the decision, it was updating every profile so nothing disagreed. That is entity work. It is admin, and it matters.

The relationship labels are the part people skip. "Related to" is not a relationship. "Develops", "is a", "is used for", "is based in" are relationships, and they are what turns a pile of pages into something a system can reason about. This is also the layer at which AI answers decide whether you are a thing worth recommending or a page that happened to be retrieved.

Structured data is a standardized way of telling a machine what the things on your page are, and it changes how your result can be displayed rather than whether it is retrieved.

The opponent is "schema boosts rankings". It does not. Google's own framing is that structured data provides explicit clues about the meaning of a page, and that adding it can make you eligible for rich results. Eligible is the operative word. It is not a promise.

The format to use is JSON-LD. The types most sites need are a short list: Organization, Person, WebSite, BreadcrumbList, Article, Product, LocalBusiness, Recipe, Video, FAQPage. Pick the two or three your pages genuinely are and stop.

Three rules that prevent almost every problem:

• It must match the visible text. Google lists this among its practices for AI features specifically. Marking up a rating nobody left is the fastest way to lose rich result eligibility entirely.
• Never fabricate the thing to get the markup. No inventing an FAQ so you can add FAQPage. The markup describes the page, it does not conjure it.
• Validate it. The Rich Results Test tells you what you are eligible for. Search Console tells you what actually happened.

And for the AI question, since it comes up constantly: Google says structured data is not required for generative AI search and there is no special schema.org markup you need to add, while still recommending it as part of your overall SEO strategy for rich result eligibility. There is no AI schema. Anyone selling you one is selling you something else.

You are not optimizing a position, you are optimizing a presence, because a modern results page is made of a dozen surfaces and most of them are won differently.

The opponent is ten blue links, which is the page beginners still picture and which has not existed for years. On plenty of queries there is no position one worth having, and on others the winnable slot is an image, a video or a local pack entry that nobody in your market has bothered to claim.

This is also the honest introduction to image SEO, video SEO and structured data, without turning a fundamentals page into twenty tutorials. Each of those is a surface. Each surface has a specialist guide. What you need now is to know they exist and to check which ones appear for your queries, because that check takes a minute and changes what you build.

Most of what gets sold as AI SEO is SEO. Google's position is that to be eligible as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Search with a snippet, and that there are no additional technical requirements.

The opponent is a whole cottage industry. Special files, special markup, special formatting, special page structures. Google's guide has a mythbusting section that names most of it and tells you to ignore it, and I would rather send you to that page than to mine.

There is one control worth knowing about, and it is new enough that most guides do not mention it. Search Console now has a Search generative AI control under Settings, which governs whether your site can appear in AI Overviews, AI Mode and generative features in Discover. The default for all properties is to include your site. Excluding it means no impressions and no traffic from those features, it takes a few days to take effect, and it does not affect the rest of Search or model training. It is rolling out to a subset of owners, so it may not be in your account yet.

Then there are the levers that were already there:

• robots.txt for Googlebot. Google treats AI as built into Search, so this is the crawl control. There is no separate Google AI crawler to allow.
• nosnippet, data-nosnippet and max-snippet. Limit what can be shown from a page, including inside AI features. data-nosnippet is the surgical one, for a single block.
• noindex. The blunt instrument. Out of Search entirely.
• Google-Extended. Limits training and grounding in some of Google's other AI products. Google states it does not affect a site's inclusion in Google Search and is not a ranking signal there, and people repeatedly expect it to remove them from AI Overviews. It does not.

ChatGPT is a different company with a different crawler and this is where sites hurt themselves by accident. OpenAI's documentation is clear: if you want your pages available in ChatGPT's search answers, allow OAI-SearchBot. Sites opted out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links.

The distinction to carry away: crawling for AI search and crawling for model training are not the same permission. They are separate settings with separate consequences, and calling every AI crawler "GPTBot" is how a site removes itself from an answer engine while believing it protected its content.

If you want the deeper, more skeptical version of all this, including what I have measured and what I cannot, that is the AI SEO pillar.

A passage gets quoted when it is specific, checkable and self-contained, and none of the formatting rituals being sold as AI optimization change that.

The opponent is the fake formula: write exactly 40-word answer blocks, chunk every paragraph, add a summary in a specific format, and you will get cited. I have not seen anyone demonstrate that causally, and Google's guidance contradicts the mechanism directly. It says there is no requirement to break content into tiny pieces, that its systems handle multiple topics on a page and surface the relevant part, and that you do not need to write in a specific way for generative AI search because its systems understand synonyms and general meaning.

So instead of a formula, here are the properties that make a passage genuinely reference-worthy, for a machine and for a person, and which will still be true in three years:

• Explicit. The claim is stated, not implied across three paragraphs.
• Factual. It has a subject, a value and a unit.
• Organized. It is under a heading that says what it is.
• Verifiable. A reader can go and check it.
• Evidenced. There is a source, a method or a dataset behind it.
• Current. It carries a date, so staleness is visible.
• Specific. A number rather than an adjective.
• Original. It did not exist elsewhere first.
• Contextually complete. It survives being lifted out of the page, which is exactly what is going to happen to it.

The difference this makes, in two sentences about the same product:

• Useless: "Our tool is extremely fast."
• Reference-worthy: "In our test of 1,000 WordPress installations, median setup time was 41 seconds."

The second one is specific, attributable, verifiable and dated. It is also the only one a journalist, a competitor's comparison page or an AI answer can use, which is why it earns references and the first one earns nothing.

One precision note for measuring any of this. A citation is a linked source inside an answer, a mention is your name appearing with no link, and a link is a hyperlink on someone else's page. Three different things, and tools that report them as one number are reporting a number, not a measurement.

Search is not text-only, so the efficient move is to build one piece of real expertise and express it in the four or five formats different surfaces reward.

The opponent is the text-only content plan, which leaves image, video, product and place surfaces uncontested for no reason other than habit.

Google's 2026 guidance recommends supporting your textual content with high-quality relevant images and videos when it makes sense, and points out that its generative features can bring those formats in too, which means more opportunities for your site to appear beyond a page link. That is a fairly direct invitation.

Practically, instead of publishing only "How to configure X", the same afternoon of knowledge becomes:

• The written tutorial, which owns the query
• Screenshots of the actual interface, which are evidence as well as illustration
• A diagram of the thing the screenshots cannot show
• A demonstration video, which is its own surface and its own audience
• A data table, if you measured anything, which is the part people quote

I have published 617 videos, and the reason they belong in an SEO conversation is not the view count. It is that the video, the screenshots and the written page are the same expertise pointed at four different discovery surfaces, and the marginal cost of the second, third and fourth is small once you have done the work.

A caution that follows from the last chapter: do not make a video because a page exists, or a page because a video exists. Make the second format when the format genuinely serves the subject better, which for anything involving an interface is most of the time.

Both are the same fundamentals plus a second system you have to feed, and each deserves its own guide rather than a paragraph pretending to be one.

What matters here is knowing that the second system exists, because a local business doing perfect on-page SEO and ignoring its Business Profile has left the actual work undone.

Local

• Google Business Profile, complete and current, which is the system
• Reviews, in volume and recency, which is the part you cannot fake
• Location and service area, stated consistently everywhere
• Business information that agrees across every listing you appear in
• Local relevance, meaning content about the place and not just in it

Ecommerce

• Product pages that are pages, not template output with a SKU
• Product data quality, which is the thing that actually gates you
• Merchant Center, which is how the product surfaces get fed
• Availability and price, accurate, because wrong data is worse than no data
• Reviews
• Product structured data, matching what the page shows

Google's own generative-search guidance calls out both directly: where appropriate, its AI responses can include product listings, product information and information about local businesses, and it names Merchant Center and Business Profile as the way to keep that information visible. If you are either of those businesses, that is not an optional extra.

SEO performance is not rankings. It is a three-tier funnel that runs visibility, then engagement, then business outcome, and the tier everyone reports is the one that matters least.

The opponent is the ranking report. It moves, it is easy to produce, it feels like progress, and it can improve for a year while nothing happens to the business. I have seen sites go up on every tracked keyword and down on signups, because the queries were the wrong queries.

The tools you actually need, and all three are free:

• Google Search Console. Non-negotiable. It is the only place you see Google's own view of your site: what it crawled, what it indexed, what it showed, and for which queries.
• Google Analytics 4, or any analytics you trust, for what happened after the click.
• Bing Webmaster Tools. Consistently skipped, and it is free, less crowded, and its AI Performance report, in public preview, shows when your site is cited in Copilot and Bing AI answers.

Rather than describe those screens, here are four of them from one of my own properties, with the one thing to look at on each. Same account and same export window as the worked example below, so the numbers tie together.

That report is also the most useful piece of original data I can put on this page, so I counted it. GA4 files chatgpt.com under 4 different channel labels, so reading only the AI Assistant row undercounts ChatGPT by 42% on this property, and every assistant together by 41%.

For the AI surfaces specifically, Search Console now has a Generative AI performance report covering AI Overviews and AI Mode. Two details matter and both get misreported. It shows impressions, not clicks. And it is rolling out to a subset of properties, so an empty report can mean "not enabled for you yet" rather than "you are not appearing".

For ChatGPT, Perplexity and the rest, there is no per-site report. You can see referral traffic in analytics when someone clicks through, and you can test prompts yourself, but prompt testing is non-deterministic and varies by user, location and run. Treat those numbers as samples. I run them, I publish them, and I label them as estimates, because that is what they are.

Four precision points that stop most reporting arguments:

• Clicks are not sessions. Search Console clicks and GA4 sessions count different events and will never match. Stop reconciling them.
• Average position is an average of averages. It is not a rank, and it moves when your query mix changes even if nothing else did.
• Estimated traffic is not traffic. If the number came from a third-party tool, say "Ahrefs estimates" out loud.
• Server logs are the only proof of a crawl. Everything else is inference.
• One assistant, several channels. Filter GA4 by source hostname, not by the AI Assistant channel, or you will under-report by roughly what the table above shows.

SEO is a loop, not a project, and the only part that makes it work is that step ten feeds step one with evidence instead of opinion.

The opponent is the checklist mentality: a list of tasks, done once, ticked off, filed. A list has an end. Search does not.

Run this once, on one page, all the way through. That is more useful than reading every chapter above twice, because the loop is where all of it becomes yours. Then do it again with what the data told you the first time.

Most bad SEO advice is not invented, it is out of date, and knowing which category a claim falls into tells you how to argue with it.

I sorted these three ways rather than calling them all myths, because the interesting ones are not simply false. Some were never true. Some are a real mechanism inflated into a promise. Some are a true statement with a wrong conclusion bolted on, and those are the ones that survive being debunked, because the person repeating them can point at something real.

The llms.txt one deserves a sentence of its own because it is current and it is being sold hard. Google's documentation says plainly that you do not need to create new machine-readable files, AI text files, markup or Markdown to appear in Google Search including its generative capabilities, because Search itself does not use them. Creating one neither helps nor hurts your Google visibility. Whether some other system chooses to read it is a genuinely separate question, and a much smaller one than the people selling llms.txt audits would like you to think.

For what it is worth, the site in the worked example above has an llms.txt. I left it there because it costs nothing and another system might read it one day. That is the whole honest position, and it is a long way from "add llms.txt to rank in AI".

Thirteen questions, every answer sourced to a primary document. If you get one wrong, the explanation tells you the mechanism, not just the letter.

Nothing you answer leaves your browser. There is no endpoint, no email gate and no account. Your answers are saved locally so a half-finished attempt survives a closed tab, and the reset button really does clear them.