The citation economy is the new shape of search visibility. The unit is no longer the ranked page receiving a click, but the cited passage feeding a generated answer. AI Overviews, ChatGPT, Perplexity, Claude, and Gemini retrieve a small set of pages from the index, lift specific passages from them, and synthesise an answer the user reads without ever clicking through. Winning is being the canonical source those passages come from, on the entities your buyers care about. Everything else in this guide is the operating manual for that.

On May 15, 2026, Google published the document the SEO industry has been arguing about for two years. It is called Optimizing your website for generative AI features on Google Search. It is roughly 1,800 words long. It contains almost no surprises, which is itself the most important thing about it.

The takeaway most people will pull from the document is "SEO still works, AI optimization is mostly the same thing." That is true, and also incomplete. The mechanics of how a page becomes visible have changed in ways the official guide acknowledges but does not fully explore. The pages that win in this new layer are not just well-ranked. They are well-structured for retrieval, written with enough specificity to be cited, and built on enough technical foundation to be machine-readable without sacrificing human readability.

This is the long-form companion to that guide. It is the version we wish Google had written. It exists for one reason: most of the operators we talk to are still measuring the wrong things, optimizing for the wrong layer, and treating AI search like a marketing problem when it is actually a content architecture problem.

We are going to walk through what generative AI search actually is, how the retrieval layer works, why content quality is now a structural property rather than an aesthetic one, what to ignore from the "GEO and AEO hacks" industry, how to think about agents that will browse your site on behalf of users, and how to build a measurement system that does not collapse the moment AI Overviews start appearing on your money queries.

If you read this end to end, you will know what to do tomorrow. Not in theory. In production.

The shift nobody can keep ignoring

For two decades, search visibility meant ranking. Ten blue links, one user, one click. You optimized a page, earned links, climbed positions, and traffic followed. The model was simple enough that an entire industry calcified around it.

That model is breaking. Not slowly. Right now.

The numbers tell the story. Sixty percent of Google searches now end without a click. AI Overviews appear on roughly half of all informational queries. When they appear, click-through rates on the underlying links drop by nearly half. ChatGPT, Perplexity, Claude, Gemini, and a dozen smaller systems are doing the same thing inside their own apps: pulling content from indexed pages, synthesizing answers, and returning a summary with citations.

We covered the surface of this in Your Traffic Is Down and Your Rankings Haven't Moved. What we want to do here is go deeper. Because the question "why is my traffic dropping" leads, eventually, to a much more useful question: "what is the unit of visibility now, if it is not the ranked page?"

The answer is the citation.

A citation is what happens when an AI system reads your page, decides that a specific passage of it answers a user's question well enough, and either quotes it directly or links to it as a supporting source. The user does not have to click. They might. They probably will not. The visibility was earned anyway, because your brand was the source of the answer they got.

This is what we mean when we say the discovery layer of the internet is being restructured. The old goal was ranking. The new goal is being chosen as the source of the answer. These overlap, but they are not the same job.

What Google actually said

Google's official guide makes four arguments. We will summarize them, then add what they leave out.

First, SEO best practices still apply. Generative AI features on Google Search are built on top of Google's existing ranking and retrieval systems. They use the same index. They rely on the same quality signals. A page that does not rank well in conventional search will not be retrieved into an AI Overview either, because the retrieval step uses the same underlying machinery.

Second, the dominant signal is content quality, defined specifically as content that is unique, non-commodity, helpful, and people-first. Google's exact framing is worth reading. They contrast a commodity title like "7 Tips for First-Time Homebuyers" against a non-commodity title like "Why We Waived the Inspection and Saved Money: A Look Inside the Sewer Line." The first one any agency could write. The second one only the person who lived it can write. Generative AI systems, Google argues, are getting better at distinguishing the two.

Third, the technical foundation matters. A page must be crawlable, indexable, eligible for snippets, semantically structured enough to be parsed, and performant enough to deliver a good page experience. None of this is new. All of it is necessary.

Fourth, most of the "AEO and GEO hacks" circulating online are unsupported by how Google's systems actually work. No llms.txt file. No special "AI markup." No required chunking of content into bite-sized pieces. No need to rewrite content in a specific machine-friendly cadence. No benefit from buying inauthentic mentions or coordinated brand chatter.

Google also nods at agentic experiences, the world where AI agents browse sites on behalf of users to complete tasks. They link to an emerging best-practices guide for agent-friendly sites and to a Universal Commerce Protocol that some search agents will use to transact.

That is the official guide in a paragraph. Useful, calm, intentionally boring. Now the parts that matter.

How the retrieval layer actually works

If you only understand one thing about generative AI search, understand this: the answer the user sees was written by a model, but the facts in that answer were retrieved from a small set of pages the system found relevant to the query. The model is not making things up from training data. It is summarizing what it pulled.

This pattern is called retrieval-augmented generation, or RAG. Google calls it grounding. The mechanics are roughly:

The user asks a question.
The system uses the existing search ranking infrastructure to find the most relevant pages in the index.
For complex questions, the system generates additional related queries on the fly, called fan-out queries, and retrieves additional pages for each.
The system selects passages from those retrieved pages.
A generative model synthesizes those passages into an answer.
The system cites the sources it pulled from.

The fan-out part is the new and underappreciated piece. If the user asks "how do I fix a lawn full of weeds," the system might silently ask itself "best herbicides for residential lawns," "how to remove weeds without chemicals," "how to prevent weeds from returning," and pull pages for each one in parallel. The final answer is a synthesis of all of them.

This has three consequences.

The first is that pages which answer narrow, specific sub-questions extremely well are disproportionately valuable. The fan-out queries are exactly the kind of long-tail, intent-specific queries that ranking systems already reward. A page that nails the answer to "how often should I dethatch my lawn in zone 6" is much more likely to be retrieved than a generic "complete guide to lawn care" page that mentions dethatching in passing.

The second is that the page itself does not need to match the original query. If the original query is broad and the fan-out queries are narrow, the page that gets cited might never have ranked for the original query directly. It got pulled in because it was the best answer to one of the sub-questions the system decided to ask.

The third is that this entire architecture rewards depth and specificity in a way the old ranking-only model did not. A long, comprehensive page that covers a topic well is more likely to contain a passage that answers one of the fan-out queries. A short, generic page is less likely to surface even if it is technically relevant.

We have seen this in practice with clients. Content that ranked at position eight for a head term suddenly started showing up as the cited source inside AI Overviews for a dozen long-tail variations of that term. The ranking did not change. The visibility did, dramatically. Same page, new economy.

The unit of optimization is the passage, not the page

Here is where Google's guide is technically correct but practically incomplete. The guide tells you not to "chunk" your content into tiny pieces for AI. That is good advice for the surface action: do not break your article into thirty single-sentence pages and call it AI optimization. The systems can read normal prose. They do not need help with parsing.

But the underlying truth is more subtle. The retrieval system is, in fact, extracting passages from your pages and feeding those passages into the model. The page is the document. The passage is the unit of evidence. When you write a page, you should think about which passages on it could be lifted out and used as a standalone answer.

This is not a request to write in bullet points or to break every paragraph into a heading. It is a request to write in a way that makes individual sections self-contained enough to stand on their own.

The test we use internally is what we call the lift test. Take any paragraph in the middle of your article. Could a stranger read just that paragraph, with no other context from the rest of the article, and walk away with a complete and accurate answer to a specific question? If yes, that paragraph is liftable. The system can cite it. If no, the paragraph requires the rest of the article as scaffolding, which means the system either has to cite the whole article (rare) or skip the page entirely (common).

Most blog content fails the lift test. Most reference content passes it. This is one of the reasons documentation, encyclopedic content, and tightly structured guides tend to show up disproportionately in AI citations. It is not that the systems prefer them stylistically. It is that they contain more liftable passages.

You do not have to rewrite your entire site into reference content. But the next time you draft a piece, ask yourself: does each section answer a specific question completely, or does each section just continue the narrative? Both have their place. Only one gets cited.

Non-commodity content is now a technical requirement

Google's guide spends more space on content quality than on anything else, and they are right to. We want to extend their framing.

Commodity content is content that any reasonably competent writer could produce by reading two existing articles on the topic and rephrasing them. It is structurally indistinguishable from the median of what already exists on the web. AI systems are trained on the web. They have already absorbed the median.

When a generative system has the entire commodity layer in its training data, it does not need to retrieve commodity content from your site. It can produce commodity content directly from its own weights. There is no reason to cite you. The system will only retrieve from external sources when those sources contain something the model could not produce on its own.

That "something" usually falls into one of four categories.

First-hand experience. A specific story, project, decision, or outcome that only you lived. Google's "sewer line" example is exactly this. A first-person account of what happened during a real inspection is something no general model can fabricate without hallucinating.

Proprietary data. Numbers, benchmarks, survey results, internal metrics, or measurements that originated with you. If you ran a study, published the results, and structured them well, your data is one of the few things the model genuinely needs.

Synthesized expertise. A novel framework, a contrarian argument, or a way of organizing existing information that did not exist on the web before you wrote it. This is harder than it sounds because most "novel frameworks" are repackaged versions of existing thinking. But genuine synthesis, when it happens, is highly citable.

Recent specificity. Information that is too new or too specific for the model's training data to cover. A new product spec, a recent policy change, a specific local detail, an update to a known process. The model has to retrieve this from somewhere. It is going to retrieve it from whoever published it most clearly.

What does not work is what most agencies still publish at scale. Generic top-of-funnel content covering well-known territory. "What is X" pages with no point of view. Pillar articles that string together five obvious subtopics. Listicles that recycle existing best-of roundups. The systems have all of this in their training data already. There is nothing to cite.

We covered the deeper version of this argument in A Keyword Isn't a Strategy. The principle is the same: when you optimize for what is searchable rather than what is true and useful, you produce content that is technically findable but functionally invisible.

Entities, not keywords

Modern search systems do not really operate on keywords. They operate on entities, which are the underlying concepts those keywords refer to. An entity is the thing your content is actually about: a company, a person, a product, a method, a place, an idea. The keyword is just one of many surface expressions of that entity.

This matters because the model behind AI Overviews and similar features is fluent in entity-level understanding. It does not need you to repeat a keyword phrase fourteen times to understand what your page covers. It needs to know, with high confidence, which entity your page is the authoritative source for.

The implication is that being the canonical source for a specific entity is more valuable than being one of many pages ranking for a specific keyword. Canonical sources get cited. Other pages get crowded out.

How do you become a canonical source? Roughly four things, in order of importance.

Coverage depth. The canonical source covers the entity exhaustively, not just at a surface level. It is the page somebody who genuinely needs to understand the topic would bookmark.

Internal coherence. All of your other pages about adjacent entities link to the canonical page. The canonical page links to your other pages in a way that builds a coherent web of related topics. Search systems use this internal link graph as a structural signal. We will get into the mechanics of internal linking later in this guide.

External validation. Other reputable sites link to or reference your canonical page. Not because you bought the links, but because they are the natural source to cite. This is what links have always meant. Generative AI search has not changed this.

Structured data. When appropriate, structured data tells the system explicitly what the page is about, what entity it represents, and how it relates to other entities. Google's guide is careful to say structured data is not required. That is true. It is also true that structured data is a high-leverage way to remove ambiguity, and removing ambiguity is one of the few things you can do that genuinely helps.

We see this play out repeatedly in audits. A client will have ten thin pages each targeting a slight variation of the same entity. None of them are canonical for that entity. None of them rank well. None of them get cited. The fix is consolidation: take the ten thin pages, merge them into one comprehensive resource, redirect the old URLs, and let the consolidated page absorb the equity. This is one of the most effective interventions we run, and it works because it converts a sprawl of non-canonical content into a single canonical source.

We document the methodology for this in our technical SEO and content strategy practice. The pattern is the same regardless of vertical.

The technical foundation, with no compromise

Google's guide lists a handful of technical requirements: crawlable, indexable, snippet-eligible, semantically structured, fast enough to deliver good page experience, free of duplicate content. We will not repeat the list. We will name the parts that operators most commonly get wrong.

JavaScript rendering. Google can process JavaScript-rendered content. The qualifier is "as long as it is not blocked." In practice, JavaScript-rendered sites still create more retrieval risk than statically rendered or server-rendered sites. The model is being fed text. If your text is buried inside a hydration step that the rendering pipeline did not complete, the retrieval system sees an empty page. We have audited dozens of Single Page Application sites where critical content was technically present but functionally invisible. If you are running a JavaScript framework, server-render or pre-render your content. Our Webflow and headless build practice defaults to this for exactly this reason.

Indexation hygiene. Every page that should be in the index needs to be allowed to be in the index, and every page that should not be there needs to be excluded. Most enterprise sites have a long tail of accidental indexation: faceted search URLs, parameter variations, internal search result pages, staging environments leaked into production. This wastes crawl budget and dilutes the system's understanding of which page is canonical for which entity. A periodic indexation audit is one of the highest-return technical activities you can run.

Crawl budget. For sites with more than a few thousand URLs, crawl budget becomes a real constraint. Generative AI features pull from the index, but the index has to be kept fresh by the crawler. If your crawl budget is consumed by low-value URLs, your high-value URLs get crawled less often. They go stale. The model retrieves stale content. Use server logs, Search Console crawl stats, and a clean robots.txt to direct crawl budget toward the URLs that matter.

Core Web Vitals. Page experience is a real signal and a real user concern. We treat Largest Contentful Paint, Cumulative Layout Shift, and Interaction to Next Paint as production-line metrics, not as a one-time pre-launch checklist. Our performance engineering practice exists because most performance regressions happen post-launch, slowly, in ways nobody catches until a quarterly audit. By that point, you have months of degraded page experience baked in.

Semantic HTML. Google's guide says you do not need perfect semantic HTML. They are right. They are also right that using semantic elements where it is reasonable is a good idea, because it helps screen readers, parsers, and any system that needs to understand the structure of your page. A <header> should be a <header>. An <article> should be an <article>. A heading hierarchy should be coherent. Two <h1> elements on the same page is not the end of the world but it is also not a habit worth keeping.

Duplicate content. Generative AI systems do not need help deciding which of three near-identical pages is the canonical one. They will pick one, and they will probably pick the one that is the most authoritative or most linked, which may not be the one you wanted. Canonical tags, redirects, and content consolidation are the tools. The principle is one entity, one canonical URL.

We treat all of this as part of a search analytics and attribution practice rather than a one-time audit. The audit finds the issues. The practice keeps them from coming back.

Structured data, demystified

Google's official position on structured data is that it is not required for generative AI search and there is no special schema you need to add. They also recommend continuing to use it for the existing rich result programs in Search.

We agree with the official position. We want to add a layer.

Structured data is not magic, and it does not move rankings directly. What it does is remove ambiguity. When you mark up a product page with Product schema, an Article page with Article schema, or a local business page with LocalBusiness schema, you are giving the indexing system an unambiguous description of what the page represents. The system already knows. The schema confirms.

Where structured data earns its keep is in three places.

Eligibility for rich results in conventional search. This is the original use case and still the most measurable one. Recipe cards, product carousels, FAQ rich results, event listings, video carousels. These are real visibility surfaces with real click-through impact. If your category has a rich result type, implement the corresponding schema.

Entity disambiguation. When you have a name that could refer to multiple things, structured data tells the system which thing you mean. A page about "Mercury" the planet versus Mercury the bank versus Mercury the chemical element will benefit from explicit entity signals. For most clients this is not a concern, but for clients with ambiguous brand names or industry-shared terminology, it is high leverage.

Structured fact assertion. When you have specific data, attributes, or relationships you want the system to understand precisely, schema provides a vocabulary. A SaaS product with feature lists, integration partners, supported platforms, and price tiers can encode all of that in structured form. The model can then pull it directly.

What we do not recommend is over-engineering the schema layer. We see clients with twelve nested schema types on a single page, half of them duplicating information from the others, all of it generated by a plugin nobody on the team understands. Pick the schema types that match the rich result programs you want to be eligible for. Implement them cleanly. Validate them. Move on.

Our GEO and AEO optimization practice treats structured data as one input among many. It is not the centerpiece. It is a tool used in service of clarity.

Internal linking as a structural asset

Internal linking is the most underused intervention in the GEO playbook. Google's guide does not mention it explicitly. It is implicit in everything they say about site structure and discoverability.

Internal links do four things at once.

They shape crawl paths. A page that is linked from many other pages on your site will be crawled more often. A page that is only linked from the sitemap and nowhere else will be crawled rarely.

They distribute link equity. Each link passes a small amount of authority from the linking page to the linked page. This is not as significant as it was in the link-graph era, but it is still real. Concentrating links on your canonical pages strengthens them.

They define topical clusters. A coherent web of internal links signals to the indexing system which pages are related to each other, which pages are the parent for a topic, and which pages are supporting subtopics. This is one of the strongest signals you have for entity-level understanding.

They distribute fan-out retrieval. This is the new behavior. When the system runs fan-out queries to gather additional context for an AI answer, the related pages on your site become candidates for retrieval. A well-linked topical cluster gives the system multiple plausible passages to pull from. A disconnected page sits alone.

The principle we use is the topical hub model. For every important topic, identify the canonical hub page. Build supporting pages that each cover a specific subtopic in depth. Link every supporting page back to the hub. Link the hub to every supporting page. Cross-link supporting pages where they naturally relate to each other.

The anti-pattern is the orphan page. We see this constantly. A high-value resource gets published, sits in the sitemap, never gets linked from anywhere else on the site, and slowly fades from the index. The fix is mechanical: build the link graph deliberately, not by hoping someone on the content team remembers to link to old posts.

This is also why long-form pillar resources like this one matter beyond their own ranking. A 6,000-word resource that links naturally to twelve related service and content pages does work for the whole site. The links you are reading right now are not decoration. They are structural.

What to ignore

Google's "mythbusting" section is the most direct part of the official guide. It is worth restating their list, with our own commentary on each.

The llms.txt file. A proposed standard for a machine-readable file at the root of your domain that tells AI systems how to consume your content. Google explicitly says it does not use this. Other major systems have made similar statements. The file does nothing for visibility. If it makes you feel better to publish one, fine. If you are spending engineering hours on it, stop.

Content "chunking." The idea that you should manually break your content into AI-friendly chunks. The systems already do this for you. They can read normal prose. Breaking your article into thirty single-sentence pages is a bad idea that some agencies are still selling.

Rewriting content for AI. There is no special "AI voice" you need to write in. The model understands synonyms, near-synonyms, and conceptual meaning. Writing for humans is writing for the model. Writing in stilted, keyword-stuffed prose is writing for neither.

Inauthentic mentions. Buying or coordinating brand mentions across the web to influence what AI systems say about you. This is unsupported by the core ranking systems and risks running afoul of spam policies. The model is trained on a lot of text. It is reasonably good at distinguishing genuine industry discussion from manufactured chatter.

Schema overuse. As covered above. More schema is not better schema. Cleanly implemented, validated schema for the rich result types you actually qualify for is the goal.

We will add three from our own audits.

Keyword density as an optimization target. The idea that a page needs to contain a target keyword some specific number of times to rank well. This has not been true for a long time. The model understands what your page is about from the entirety of its content. Forcing a keyword into the text past the point of natural use makes the page worse, not better.

Word count as a quality proxy. The idea that longer pages rank better. They sometimes do, because longer pages contain more passages that can answer more queries. But adding fluff to hit a word count target produces a worse page, not a better one. The right length is the length the topic genuinely requires.

Content velocity as a primary signal. The idea that publishing frequently is itself a ranking factor. It is not. Publishing frequently can help if every piece is good. Publishing frequently when every piece is mediocre is a way to dilute your site's overall topical authority and waste crawl budget.

Most of the optimization industry sells motion. The systems reward substance. The two are not the same.

The agent layer

Google's guide ends with a section on agentic experiences. It is short, deliberately so, because the field is moving fast. We will say more about it than they did.

An agent is a system that performs tasks on behalf of a user. A browser agent visits sites, reads them, fills in forms, completes transactions. A purchasing agent shops for the user across multiple sites. A research agent gathers information for a report. These agents are real, they are improving rapidly, and they are starting to make purchasing decisions on behalf of real customers.

The implications are not yet fully understood, but a few things are already clear.

Agents read content the same way models do. They pull text from the rendered page, parse the structure, and act on what they find. A site that is hard to parse for a model is hard to parse for an agent. The technical foundation that helps your AI search visibility also helps your agent compatibility.

Agents need clear functional affordances. A "Buy Now" button that is a div with an onclick handler is hostile to agents. A semantic <button> element with a clear label is friendly. The same is true for forms, navigation, error states, and confirmation flows. Building for accessibility, which we have always done, turns out to be building for agents.

Agents do not click ads in the way humans do. They have a goal, they execute the goal, and they leave. This has implications for paid traffic that nobody has fully worked out yet. The trend, however, is clear: paid acquisition is going to get noisier and less efficient as more traffic becomes agent-mediated.

Conversions will increasingly be machine-mediated. When an agent fills out your contact form, completes your checkout, or signs up for your newsletter, the data you get looks different. The user is real. The agent is the proximate actor. Your analytics and attribution stack needs to be ready for this. We covered the framework in Perfect Attribution Is a Myth. Agent traffic makes the imperfection worse, not better.

For most clients, building specifically for agents is not the priority yet. The priority is building well in general, because the same things that work for human visitors and AI retrieval also work for agents. When the agent layer matures, the sites that are already well-built will be well-positioned. The sites that need to be retrofitted will discover the cost is non-trivial.

If you want a deeper look at how this connects to our broader thinking on the topic, The Wrong People Are Worried About AI covers our position on which parts of the agency model survive automation and which do not. Spoiler: the parts that survive are the parts that compound, and compounding requires the kind of substantive content and infrastructure work this guide describes.

Measurement in the citation economy

The hardest part of the shift is not the content work or the technical work. It is the measurement work, because the metrics most teams report on are the ones losing meaning fastest.

Rankings are still useful but increasingly partial. A page can rank at position three and lose half its traffic because the AI Overview above it answered the user's question without sending the click. Or, conversely, a page can rank at position eight and gain traffic because it became the cited source inside an AI Overview that appears on a much wider set of fan-out queries than it ever ranked for organically.

Sessions and pageviews are still useful but increasingly noisy. AI-mediated traffic patterns are different from human patterns. Direct traffic is rising as branded queries get answered in AI tools and users navigate to the brand without going through search. Referral traffic from AI tools is appearing as a real channel, often miscategorized in default analytics setups.

Conversions still matter, of course. They always will. But the source attribution is harder.

Here is the measurement frame we recommend.

Track citation frequency. For your target queries, measure how often your site is cited in AI responses across the major surfaces: Google AI Overviews, ChatGPT, Perplexity, Claude, Gemini. There are tools that automate this and there is no shame in spot-checking manually for high-value queries. The trend over time is more informative than the absolute number.

Track answer inclusion. When your site is cited, what passage was cited? Is it the passage you wanted? Is it accurate? Is it framed in a way that represents your brand correctly? This is the new version of the meta description audit. The citation is your meta description now.

Track source positioning. When multiple sources are cited in a single AI response, where are you in the list? First-cited sources receive disproportionate trust and disproportionate click-through. Being cited is good. Being cited first is better.

Track AI referral traffic explicitly. Build a custom channel grouping in your analytics that segregates AI tool traffic from generic referral or direct. We documented our approach in our attribution playbook and the patterns are stable across most clients we work with.

Track downstream conversion behavior. AI-referred traffic tends to convert at higher rates than generic organic, because the user already received an answer they trusted and arrived with intent. Measure this. If you are not breaking out AI traffic from organic, you are averaging two very different behaviors into one number.

Reframe the rankings dashboard. Rankings are still a leading indicator for visibility, but they need to be interpreted alongside SERP feature presence. A page ranking at position five on a query with an AI Overview is in a different situation than a page ranking at position five on a query without one. Most rankings tools now report SERP feature data. Use it.

The teams that get this right will see clearly that visibility is increasing even as some traditional metrics flatten or decline. The teams that get this wrong will spend the next two years optimizing for a measurement model that no longer reflects reality.

Workflow: how to actually do this

We promised an operating manual, so here is the workflow we run for clients moving from old-school SEO into the citation economy. It applies whether you are starting from scratch or retrofitting an existing site.

Step 1: Inventory. Pull every URL on your site. Map each one to a primary entity it covers. Identify which entities have multiple competing URLs (consolidation candidates), which entities have no canonical page (gaps), and which entities have a canonical page that is thin (depth candidates).

Step 2: Audit the technical foundation. Crawlability, indexability, page experience, JavaScript rendering, duplicate content, internal link structure. This is the boring layer and it is non-negotiable. Most sites have at least three meaningful issues at this layer that are quietly suppressing visibility.

Step 3: Consolidate. For every cluster of competing URLs targeting the same entity, decide on the canonical, merge content, set up 301 redirects, and update internal links. This is the highest-return content intervention available. A site that goes from twelve thin pages on a topic to one strong page on the same topic typically sees that one page outperform all twelve combined within a quarter.

Step 4: Deepen canonical pages. For each canonical page, run the lift test. Every section should answer a specific question completely. Add proprietary data, first-hand examples, specific numbers, and recent updates. Remove generic filler. The page should read as the resource somebody would bookmark.

Step 5: Build the link graph. Every canonical page should be linked from every plausibly related page on your site. Every supporting page should link back to its parent canonical. Anchor text should be descriptive but natural. This is mechanical work and most teams skip it.

Step 6: Implement structured data cleanly. Identify the rich result types you qualify for. Implement the corresponding schema. Validate it. Stop there. Resist the urge to add more.

Step 7: Set up the new measurement layer. Citation tracking, AI referral segmentation, SERP feature presence, conversion behavior by source. Build it once. Watch the trends.

Step 8: Publish new content with intent. Only new content that meets the non-commodity bar. First-hand experience, proprietary data, synthesized expertise, or recent specificity. If a piece does not have at least one of these, do not publish it. The cost of mediocre content is no longer just opportunity cost. It is dilution of your topical authority.

Step 9: Refresh on a cycle. The model retrieves fresh content preferentially for time-sensitive queries. Pages that go stale lose visibility. Build a refresh cycle into the calendar. Update statistics, examples, and date references at least quarterly for high-priority pages.

Step 10: Treat the system as a system. This is not a campaign. It is a permanent operational layer. The teams that win are the ones who run it continuously, not the ones who do a quarterly project.

If this list looks like more work than the average "10 GEO hacks" article suggests, that is because it is. The shortcuts do not exist. The work compounds.

What this means for the agency model

We run an agency. We have a position on this shift, which we have written about in pieces like Strategy Decks Lie and The Handoff Tax. The short version: the agency services that survive AI are the ones that produce compounding infrastructure, not the ones that produce isolated deliverables.

A monthly content quota is a deliverable. A topical hub model with consolidated canonical pages and a maintained link graph is infrastructure. The first one is increasingly easy to replicate with AI tools and increasingly commoditized in the market. The second one requires judgment, taste, and continuity that AI tools alone do not yet provide.

We talk about this internally as the difference between making widgets and tending a garden. Most agencies still make widgets. The structural work this guide describes is gardening. It pays back over years, not weeks.

If you are evaluating agency partners, here is the question we would ask. Show me a client where, twelve months ago, you implemented a content and technical SEO strategy. Where is that client today? Are the canonical pages compounding traffic and citations? Is the link graph getting stronger or weaker? Is the technical foundation maintained or has it drifted? If the answer is "we shipped a project and moved on," the agency is making widgets. If the answer is "we are still running it, here are the trend lines," the agency is gardening.

This is one of the reasons our retainer engagements look different from our project engagements. The retainer model is the one that produces gardens. The project model is the one that produces widgets. Both have their place. Only one wins in the citation economy.

A note on AI-generated content

Google's guide addresses this explicitly. Their policy is consistent: content created with AI assistance is not penalized by virtue of being AI-assisted. Content that meets the quality bar is fine. Content that is mass-produced for the purpose of manipulating rankings is in violation of their scaled content abuse policy regardless of whether a human or an AI produced it.

We agree. We use AI tools heavily in our own production process. We do not use them to produce content. We use them to accelerate the work of producing content: research synthesis, first-draft outlines, fact-checking, formatting, internal-link suggestion, schema generation, quality audit. The judgment, the point of view, and the proprietary inputs remain human. The output is reviewed and edited by humans before it ships.

The shorthand we use is: AI can shorten the distance between a strong input and a publishable output. It cannot replace the input. The teams that try to use AI to bypass the input layer produce exactly the commodity content the systems are now penalizing. The teams that use AI to operationalize their strongest thinking produce more of what works, faster.

Putting it together

Generative AI search is not a separate channel. It is the next phase of search itself. Treating it as a side project to be optimized for in addition to "real SEO" is the wrong mental model. It is real SEO, with the surface area of visibility broadened to include AI-generated answers alongside ranked links.

The principles that apply to one apply to the other. Be the canonical source for the entities you care about. Build a technical foundation that does not get in the way. Write content that contains something a model could not produce from its training data. Link your pages into a coherent topical structure. Measure the right things. Run the system continuously, not as a project.

What you should ignore: the special markup files, the chunking tricks, the AI-only content rewrites, the inauthentic mention services, the keyword density spreadsheets, the word count targets, the publishing cadence as proxy for quality. Most of the optimization industry is still selling motion. The systems reward substance.

What you should focus on: depth over breadth, specificity over generality, structure over volume, infrastructure over campaigns, continuity over projects. The shift is real, the work is concrete, the timeline is now.

If you are evaluating where you stand, the questions are simple. For each of your most important entities, are you the canonical source? Does your site contain liftable passages a model would prefer to cite over its own training data? Is your technical foundation strong enough that those passages are reliably retrieved? Is your internal link graph helping or hurting? Is your measurement layer keeping up with how visibility actually works now?

If you are honest about the answers, the work plan writes itself.

The Citation Economy: An Evergreen Guide to Being Found, Read, and Trusted by AI Search