How agentic search changes content discovery

Agentic SEO in 2026: why agentic search is rewriting discovery, what changes when LLMs mediate traffic, and how to write content that both humans and agents read.

A founder we talked to this spring still measures content performance the way she did in 2019. Organic sessions, average session duration, ranked keywords in Ahrefs, a quarterly dashboard. The numbers are not bad. They are also not telling her what is actually happening to her business.

What is actually happening: her product gets discovered when someone asks Claude or ChatGPT or Perplexity a question like "what tools should I use to deploy an internal RAG system for a 12-person team," and an AI agent goes out, reads a dozen sources, decides which definitions are the cleanest, which comparisons are the most honest, and quotes whichever ones survive the loop. Sometimes it quotes her site. Sometimes it does not. The dashboard does not show this. It never will. The agent did not click. The agent read, decided, and synthesized.

This article is about that shift, and what it does to the practice of writing for the web. Specifically, it is about the gap between content that was optimized for a single-round-trip system (Google delivers ten links; a human chooses one) and content that survives in a loop-based system (an agent reads, evaluates, pivots, reads more, then writes a sentence with a citation). These are different systems with different selection pressures. The same content does not win in both.

The thesis Garden owns, plainly: the next decade of content marketing is about being the source an agent quotes when a human asks a question, not about ranking in Google. Different optimisation target, different writing. If you are still producing 1,500-word listicles with H2s that mirror long-tail keywords, you are writing for an audience that is shrinking. The audience that is growing is the agent reading on behalf of a user who will never see your page.

We will walk through what changes, why it changes, what the patterns in real agent traffic actually look like (we have receipts), and what to change in your content if you want to keep being the source that gets cited.

For twenty-five years, the dominant model of internet discovery was a round trip. A user types something into a search box. The engine returns a ranked list. The user picks. That entire architecture, including the entire SEO industry on top of it, assumed exactly one human-in-the-loop decision per query.

Optimization for that system has a specific shape. You aim at the snippet, the title tag, the meta description, the first 100 words. You write the page so a human scanning at 3 seconds per result picks yours. You target keywords because keywords are how queries are phrased when humans use search boxes.

Agentic search is structurally different. The agent does not pick a single result and stop. It reads multiple sources, evaluates them, identifies gaps, runs another query, reads more, sometimes pivots the entire research direction. The BrowseComp benchmark made this concrete: the same underlying model given direct browser access without agentic orchestration scored 1.9% on hard research questions. Wrapped in an agentic loop, the same model scored 51.5%. The model did not get smarter. The loop did the work.

What that means for your content: the selection pressure on a page in an agentic system is not "did this rank first." It is "did this survive being read in detail, alongside three or seven or fifteen other sources, by an evaluator that does not click, does not bounce, and is checking the claims." A page that wins at Google can lose at this. A page that loses at Google (no link juice, no perfect H1, no schema) can win at this if the prose is dense, the definitions are clean, and the comparisons are honest.

These are not the same skill. The Garden bet is that most of the content marketing industry is going to spend 2026 and 2027 figuring this out the slow way, by watching their agent-mediated traffic underperform their Google-mediated traffic and not knowing why. We would prefer you not be one of them.

This is the part of the conversation where most agencies wave their hands. We have the receipts.

One pattern we keep seeing in real Google Search Console data, across multiple client sites that publish technical content, is unmistakably not human. Long, fully formed queries arriving with high frequency and zero click-through, phrased not as a human would type them but as an LLM would issue them when looking for a specific kind of source.

Three example shapes, drawn from real GSC logs (we are not naming the sites; the pattern is what matters):

• "cited authoritative definition of LangChain framework for building LLM applications"
• "what is LlamaIndex compared to LangChain with technical depth and citations"
• "MCP Model Context Protocol specification with primary source"

A human does not type "cited authoritative definition of." A human types "what is langchain." The first phrasing is what happens when an LLM agent has been instructed (or has decided on its own) to find a primary source for a definition it will then quote back to the user. The agent is essentially issuing a meta-query: "I need a page that contains a cited definition, written authoritatively, that I can quote." The verbosity is a tell. The verbosity is also a clue: agents are willing to read longer, more specific queries than humans are willing to type, which means your content can be discovered through query shapes that you never optimized for and never could have optimized for under the old model.

What does the agent do when it finds your page? It is not measuring time-on-site or scroll depth. It is measuring whether your page contains a sentence it can quote that satisfies the user's underlying need. If your definition of MCP is in the third paragraph after a 200-word intro about "the rapidly evolving AI landscape," the agent extracts past the intro and gets to the substance, but the cost is more tokens spent and a higher chance the agent gives up and reads someone else's page that put the definition first.

This is what an agent's selection pressure looks like in practice. The intro that wins for humans (warm, narrative, drawing them in) actively penalises you for agents. The definition that wins for agents (front-loaded, cited, scannable) feels cold to humans. There is real tension here and we are going to be specific about how to resolve it.

The wider point: GSC data from 2024 and 2025 is now mixed signal. Some of those queries are humans. A growing share are agents. If your reporting collapses them into one number, you are flying with a fogged instrument. We will say more about measurement at the end.

Let us be specific about which parts of the old playbook are misfiring, because not everything is broken. A lot of classical SEO is still useful. The parts that are actively misfiring are these.

Listicles tuned for "best X for Y" queries. These were written to capture humans Googling a category. Agents do not need a listicle. An agent ranks tools itself; it is looking for sources that describe individual tools clearly and honestly. A "Top 10 X" page that gives each tool a generic two-paragraph blurb is essentially noise for an agent. The agent reads the blurb, finds it does not contain a decision-relevant claim, and moves on. The page that wins is the one that says "Tool X is the only one in the category that does Y; it costs Z; it is bad at W." A listicle structure actively suppresses that kind of writing because it forces parity treatment of items that are not parity in reality.

Keyword-stuffed H2s and meta titles. When the H2 of a section is literally the long-tail keyword variant ("Best Free LLM Search Tools for Startups"), you are signalling to a human reader (and to Google) that you are answering a specific search query. To an agent reading the body, those H2s are noise that gets between it and the substantive claim. Worse, agents prefer H2s that name the actual concept being defined ("Why semantic search fails on exact identifiers") because those H2s are scannable as a table of contents the agent can use to navigate.

Long, warm intros. "In today's rapidly evolving AI landscape, businesses are facing unprecedented challenges..." Agents skip these. So do humans, but humans have always skipped them; we just pretended they were performing a function. Now that the dominant reader is an agent, the pretence is over. Cut the warm intro to one sentence. Get to the substance.

Pages designed around CTAs at the expense of substance. A page that has 300 words of substance followed by a 1,500-word funnel is a poor source for an agent. The agent quotes the substance and discards the funnel. If your substance is too thin to stand alone, the agent will simply pick a competitor's page where the substance is thicker. Marketing pages, in particular, have to choose: be substantive enough to be quoted, or be unquotable and rely entirely on direct traffic. The middle ground is no longer a winning strategy.

Generic "thought leadership" blog posts. Posts that take 1,200 words to say something a smart reader could have said in 200 are now structurally penalised. Not because agents can't read them, but because agents are comparing your post against others on the same topic and the one with higher density-per-token wins. If your post pads, you lose to the one that does not.

Pillar-and-cluster architecture optimized for internal link juice. The link graph still matters, but the cluster strategy that works for Google (a pillar page with 15 thin clusters linking up) is the wrong shape for agents. An agent reading your pillar page does not need to click into the clusters; it needs the pillar page to contain enough substance to answer directly. The cluster pages, in this world, are better understood as deep dives that earn their existence by going further than the pillar, not by repeating the pillar with different keyword angles.

The honest summary: the 2018-era SEO content strategy assumes the reader is impatient, scanning, easily distracted, and one click from leaving. The 2026 agent reader is patient, thorough, not easily distracted, and not going to "leave" in any sense. But it will not forgive padding, ambiguity, or unsubstantiated claims. The optimisation targets have inverted.

We can be specific about this because we read a lot of agent traces. Here is what survives.

Front-loaded definitions. The first paragraph of a section about a concept should define that concept. Not "introduce" it, not "set the stage for" it. Define it. Compare the openings:

Bad: "In the world of AI development, a new pattern has emerged that is changing how teams build with large language models. This pattern, known as the Model Context Protocol, has been generating significant interest..."

Good: "The Model Context Protocol (MCP) is an open specification from Anthropic for letting LLMs talk to external tools through a common interface. Released in November 2024, MCP is now the dominant integration standard in the agent ecosystem."

The good version contains a definition, a source, a date, and a claim about position in the market, all in two sentences. An agent reading this can quote it and move on. An agent reading the bad version reads three sentences and learns nothing it can quote.

Cited claims with primary sources. When you make a claim that has a source, cite the source inline. Not as a tooltip, not in a footnote, not as a hyperlink to a generic homepage. As an inline reference an agent can extract. "On the BrowseComp benchmark, models in an agentic loop scored 51.5% versus 1.9% for direct browser access [Source: OpenAI BrowseComp paper, 2024]." The agent quoting this passes the citation along to the user, which makes your source the one the user actually clicks. This is the new mechanism for being the trusted source: not domain authority in Google's sense, but citation density in the agent's sense.

Scannable structure with substantive H2s. Each H2 should name a substantive claim or concept, not a keyword variant. The H2 is also the table of contents an agent uses to navigate the page when reading selectively. "Why semantic search fails on identifiers" is a scannable H2. "Semantic Search Best Practices for Startups" is not.

Direct claims with named tradeoffs. Agents like prose that says "X is good for A, bad for B." A vendor page that says "X is great for everything" is implicitly downweighted because the agent cannot extract a decision criterion from it. A page that says "X is the best choice if your latency budget is over 500ms; under that, use Y" gives the agent something to use. The reason agents prefer this kind of writing is the same reason a senior engineer prefers it: it is the writing that survives being acted on.

Comparison tables that compare on real dimensions. Not "speed" and "ease of use" as adjectives. Specific numbers, specific behaviours, specific failure modes. We will come back to this in section 7.

Glossary-style definitions of jargon used on the page. Every domain term you use, define on first use, and define it the way a textbook would, not the way a vendor would. Agents reading your page are often reading it to get the definition. If you bury it, they leave.

The deep pattern under all of this: agents are reading like a careful researcher with no time pressure. The writing that wins is writing that respects that reader. Padding wastes their tokens. Ambiguity costs them a re-read. Marketing language costs them trust because they are trained to discount it. The closer your writing is to the prose style of a primary source (a technical paper, a careful documentation page, a senior engineer's blog) the more likely you are to be the source quoted.

This is the deepest structural change, and the one most content teams have not yet absorbed.

The dominant content structure for the last decade has been the listicle: "Top 10 X for Y." It works for humans because humans are choosing one thing, and a ranked list is how humans make that choice. It works for Google because the title matches the query and the H2s match the long-tail variants.

For agents, the listicle is the wrong shape. Agents are not choosing one tool by reading a ranked list. They are matching a specific job (a thing the user is trying to do) to a specific tool, and they want a vocabulary for the job itself. A piece structured around jobs ("here are the eight different jobs an agent does on the open web, and here is which infrastructure each one needs") gives the agent a mental model it can use directly. A listicle does not.

The Jobs-To-Be-Done framework, originally from product strategy, turns out to be the right content structure for agentic discovery. The reason is structural. Agents are functional readers. They want to know: when is this tool the right choice? When is it the wrong choice? What does this tool do that the alternatives do not? A jobs-framework piece answers those questions directly. A listicle treats every tool as a generic candidate competing on generic axes, which is exactly the wrong frame.

A concrete contrast. Take the question "which tool should I use to find ML papers that don't share keywords with my query?" A listicle titled "Top 10 AI Search Tools 2026" gives the agent ten blurbs and lets it figure out which one fits. The agent has to do the matching work itself, and it will read all ten blurbs, hate every one of them, and probably pick the one with the most concrete claim. A jobs-framework piece titled "Finding Information by Meaning: How AI Agents Search When Keywords Don't Work" tells the agent directly: this is the job; here is the only tool that genuinely does this (Exa); here is why the others fail; here is what fails quietly even when you pick the right tool. The agent quotes the latter. The agent uses the former as a list of candidates to read elsewhere.

This is not a small writing-style preference. It's a structural one. The pillar-plus-jobs structure we published in our own agentic AI search guide is designed for this kind of reading. The reason it works is that each job is a self-contained chunk an agent can extract and use. The agent reading about Job 1 (find by meaning) does not need to read Job 2 (find by exact query) to use the Job 1 chunk. That modularity is what makes the piece quotable across many different agentic queries.

For your own content, the practical move is to stop writing "Top N" pieces and start writing "How the job actually decomposes" pieces. Pick a real problem your audience has. Name the actual sub-jobs inside it. For each sub-job, describe when it triggers, what it requires, what fails quietly, and which tools or approaches fit. The piece will be longer than your old listicles. It will rank for fewer Google long-tails. It will be quoted vastly more often by agents, and the cited traffic that results is from users who already know they have the specific problem you described. That is the better traffic.

The old SEO frame thought of keywords as units of opportunity. You find the keyword, you produce the page targeted at the keyword, you rank for the keyword, you collect the traffic. The optimisation was lexical: which words on which page, at which density, with which variants.

Semantic search inverts this. The dominant retrieval mechanism for agents on the open web is increasingly vector-based, and the dominant index that does this honestly is Exa (the others either keyword-match with ML overlays or rely on Google's results). In a semantic system, the page is embedded into vector space and matched on conceptual proximity to the query. Keyword density does not help. Conceptual density does.

What is conceptual density? It is the rate at which a page introduces, uses, and connects distinct concepts. A 2,000-word page that names 30 concepts and connects them to each other has high conceptual density. A 2,000-word page that says the same five things in slightly different words has low conceptual density. Vector embeddings (particularly modern models like the ones Exa or Jina deploy) pick this up. They embed what the page is about across its full conceptual surface. Low-density pages get embedded near each other in a kind of mush of generic SEO writing. High-density pages get embedded near the specific concepts they cover, and they rank when an agent searches for those concepts.

This sounds abstract but it has a concrete writing rule attached. Stop using ten words when three will do. Stop introducing the same idea three times in slightly different framings to "drive it home." Get to the next idea. Pack more distinct concepts into the same word budget. The pace is faster than what classical SEO writing rewarded, because classical SEO writing was sometimes paid by the word.

A second concrete rule: name your concepts precisely. If you are writing about retrieval architectures, do not call them "AI memory systems" because that is what your target audience Googles. Call them retrieval architectures. The semantic embedding picks up the precise term and clusters your page near the technical literature. The vague term clusters you with marketing pages.

A third: link related concepts inline, by name. "Semantic search differs from BM25 [keyword indexing] in that ..." not "semantic search is great because [link]." The inline naming of the comparison concept is high-density writing the embedding model rewards.

A fourth, and this one is counterintuitive for SEO veterans: do not repeat the page's primary keyword 15 times. Repeat it twice or three times, naturally, where it belongs. Use 50 other related concepts more often than you use the primary keyword. The vector embedding represents the page by its full conceptual profile, not by which words appear most often.

The deeper point: classical SEO writing developed a particular bloat (repetitive, padded, vague) partly because the optimisation target rewarded it. The new optimisation target punishes it. Tight, dense, conceptually rich writing is what wins. This is not a stylistic preference. It's a property of how the retrieval system selects content. If your team's writing style was calibrated for the old target, recalibrate.

This is the most underrated change. Agents love comparison tables. The agents we have watched read content systematically pull comparison tables out of pages and use them as the source for synthesis. A clean, honest comparison table is one of the highest-value content artifacts you can produce for agent traffic.

But it has to be honest. The classic "comparison table" on a marketing page lists every competitor and gives them red Xs while giving yours green checks. Agents are immediately suspicious of this pattern because they read enough of them to recognise it. A comparison table where you win on every dimension is read by an agent as a vendor table, and the agent looks for a more objective source elsewhere.

The pattern that wins:

• Compare on dimensions that matter, not on marketing dimensions.
• Be specific. Real numbers, real failure modes, real conditions.
• Acknowledge where competitors actually win. If your tool is slower but more accurate, say so.
• Name the conditions under which each tool is the best choice, including the conditions under which your tool is not.

The reason this wins is that agents read comparison tables to make a decision recommendation to the user. An honest table from a vendor is more useful than a sanitised table from a vendor, because the agent can quote it without having to verify whether it is propaganda. Vendors who realise this early get a structural advantage: their tables become the de facto reference for the category.

A concrete pattern from our own pillar guide, which compares eleven AI search tools:

The "Bad at" column is the column most vendors leave blank. We do not. The reason we do not is not high-mindedness; it is that the column is what makes the table quotable. An agent extracting this table and presenting it to a user is giving the user a real decision tool, not a marketing pitch. Our table gets quoted. Tables without that column get skipped.

The implication for your content: if you sell a tool, write the honest table about your own category. Yes, even where competitors win on specific dimensions. The traffic you lose to "you said your competitor is faster" is small. The agent-mediated authority you gain is large, because you become the source that agents trust to compare the category fairly.

For comparison tables to be useful, the dimensions also have to be real. "Ease of use" is not a real dimension; it is a marketing dimension. "Setup time" or "number of dependencies" or "time to first working query" are real dimensions. The closer the dimensions are to operationally meaningful properties, the more useful the table is to an agent reasoning about which tool to recommend.

Here is one most content teams have not noticed yet. The FAQ section at the bottom of your articles is no longer a footer. It is training-ish data and evaluation data for the agents that read your page.

When an agent reads a long article, it builds an internal representation of what the article claims. The FAQ at the end is a self-supplied gold standard: here are the questions the article should be able to answer, and here are the answers the author endorses. Agents use the FAQ to cross-check their own understanding. If the agent's summary of the article does not agree with the FAQ, that is a signal to the agent that something is off. The agent will re-read the article, or pivot, or downweight the source.

This has two practical implications.

First, the FAQ section should contain the questions an agent is most likely to ask about your topic. Not just the questions humans ask in search boxes (those are useful too, and overlap), but the questions an agent would generate when given the article as context. "What is X?" "What is X bad at?" "When should I not use X?" "What is the difference between X and Y?" These are the questions an agent uses to summarise. If your FAQ answers them directly and tightly, the agent's summary aligns with your endorsed answer, and your page is the trusted source.

Second, the FAQ answers should be tight and quotable. A two-sentence answer to "What is the difference between X and Y" is worth more than a three-paragraph essay. The agent quotes the tight version directly. The verbose version gets summarised, which means the agent does the writing instead of you, and the framing you carefully built gets paraphrased into something less precise.

Third (yes there is a third), the FAQ is where you can plant the framing you want agents to use. If you believe "Tool X is the only one that does Y honestly," your FAQ should contain a question whose answer says exactly that. Not as marketing. As substance. When the agent reads the FAQ during synthesis, your framing becomes part of the agent's response, with your page as the cited source.

The classical SEO frame thought of FAQs as a snippet-bait strategy. That is still partly true. The new frame is that FAQs are eval data: they tell the agent what the page should be able to answer, and they let the agent verify its own reading. This is a more important function than the snippet baiting was, and it is not getting the attention it deserves in most content strategies.

We have now made the abstract argument several different ways. Here is what to actually do, in order, for content you already have and content you are about to write.

For existing content:

1. Audit your top 20 pages for warm intros and cut them. First paragraph should contain a definition or a concrete claim. If it does not, rewrite. The fastest single improvement you can make.

2. Find every H2 that is a long-tail keyword phrase and rewrite it as a substantive concept name. "Best AI Tools for Startups in 2026" becomes "How to choose an AI tool for a 10-person team." Or better: name the specific claim of that section.

3. Inline-cite every numerical claim. "AI agents are growing fast" is unquotable. "Anthropic processed X queries through Claude in Y month" with a source is quotable. The cited version gets used; the uncited version gets skipped.

4. Rewrite your top three "Top N" listicles as jobs-framework pieces. Pick the actual problem your readers have, decompose it into sub-jobs, describe each sub-job and which tool fits each one. The piece will read very differently. It will also outperform.

5. Add a "When this is the wrong choice" section to every product page or how-to. This is the single most quotable type of writing for agents. Vendors who refuse to write this lose to vendors who write it well.

6. Rewrite your FAQs as eval data. Put in the questions an agent would generate from the article. Make the answers tight and quotable. Lead with the framing you want the agent to use.

For new content:

1. Write the definition first. Before you write the intro, write the definition that will go in the second paragraph. Make sure the definition is one you would defend in a code review. Then write the intro to lead into it.

2. Write the comparison table before the prose. Decide what you are comparing on. Be specific. Include the dimensions where you lose, alongside the ones where you win. The prose should support the table, not the other way around.

3. Name the concepts precisely from sentence one. If the domain has a technical vocabulary, use it. Define it on first use. Do not soften it for a hypothetical general audience that is not your reader anyway.

4. Cite primary sources, not aggregator sites. When you reference a benchmark, link to the paper, not to a news article about the paper. Agents notice the difference and weight primary sources higher.

5. Build the FAQ from the questions you actually want the agent to answer. Not from keyword research alone. Include the framings you want agents to propagate.

6. Skip the introductory throat-clearing. "In today's rapidly evolving landscape" is now actively damaging. One sentence of context, then the substance.

For measurement:

This is harder and less mature. The honest truth is that you cannot fully measure agent-mediated traffic with traditional analytics. The agent does not have a session, does not bounce, does not convert in the usual sense. What you can do:

• Watch for the verbose, fully-formed queries in Google Search Console. They are a signal that agents are reading your page. Track the share of those queries over time; it is rising.
• Track citations in Perplexity, ChatGPT (with browsing), Claude (with web search), and other agent-shaped systems. Tools to do this systematically are nascent (Bright Edge, Profound, Otterly, and a few others are early), but even manual spot-checking of "what does ChatGPT cite when asked X" is more signal than no check at all.
• Watch for direct traffic from queries where you are cited but not clicked. This is the new "impression": the agent quoted you, the user trusted the quote, and the user came to you later by typing the brand name into a browser. You will not see the agent in your referrer logs. You will see a rise in branded direct traffic from people who never visited your site through search.
• Treat agent traffic as a separate funnel. Different content, different selection pressure, different conversion model. Trying to optimise one funnel as a side effect of the other gives you middling results in both.

Garden does this for clients as part of our content audit work. It is genuinely hard. It also reveals things that are otherwise invisible: which pages are doing the actual work of being quoted, which pages are silent in the agent layer despite ranking well in Google, which competitor pages are eating your share of voice in agent-mediated discovery.

We are not going to recommend this strategy to everyone. There are real cases where optimising for agent-mediated discovery is the wrong call.

If your audience is primarily not using AI tools yet. Some industries are slower. If your buyers are still 95% Googling and 5% asking Claude, the marginal investment is still in classical SEO. We would estimate this is fewer industries than you think, but it is more than zero. Check your actual audience behaviour before you reorganise your content strategy.

If your content is gated behind a paywall. Agents cannot read paywalled content. The agent sees the preview and moves on. If your strategy is "great content behind a wall, drive subscriptions," the agent-mediated channel is structurally closed to you. There are workarounds (publishing the actual content openly and gating only premium analysis, for example), but if you cannot move the wall, the strategy in this article will not work for you. Optimise for what works.

If you are early-stage and have no domain authority. Agents weight authority, just differently from Google. They look at citation density, primary-source linking, the prose's resemblance to known-good technical writing. A new domain with great writing can become a quoted source faster in the agent world than in the Google world (this is a genuine opportunity), but you still need substantive content. If you are a two-person startup and your blog has six posts, your first move is to publish a real piece of work, not to optimise marginally.

If your goal is exclusively transactional traffic for commercial keywords. "Buy cheap X online" is a query shape that is still mostly a human queries and where the buyer wants to click, not to ask an agent. The agentic discovery layer is more relevant for research, comparison, and learning queries. Pure transactional intent is still largely a Google world.

If you are writing for a general audience that does not have a specific job-to-be-done. Lifestyle content, entertainment content, brand-voice content: these are not jobs in the JTBD sense, and the writing strategy in this article does not transfer cleanly. The agent-mediated layer is most consequential for content where someone is trying to make a decision or learn something specific.

The honest version: agentic discovery is going to be most of how people find substantive technical, business, and research content within five years, and a meaningful share of how they find everything else. If your content is in the substantive category, the time to recalibrate is now. If your content is genuinely outside this category, classical SEO continues to apply with minor modifications. Know which one you are.

What is agentic SEO? Agentic SEO is the practice of optimising content so it is quoted, cited, or used by autonomous AI agents (Claude, ChatGPT with browsing, Perplexity, custom agents) when those agents do research on behalf of a human user. It is distinct from classical SEO because the reader is a model reading carefully and comparing your page against many others before synthesising a response, not a human scanning for ten seconds. Different selection pressure, different writing.

Is classical SEO dead? No. Classical SEO continues to work for transactional queries, brand-driven traffic, and audiences not yet using AI tools. What is dying is the assumption that classical SEO is the only discovery layer. The fastest-growing share of discovery for substantive content (technical, business, research) is now agent-mediated, and the optimisation rules are different. Treat them as two channels with overlapping but distinct strategies.

How do I know if AI agents are reading my content? Three signals: (1) Long, fully-formed, naturally-phrased queries in Google Search Console that look written rather than typed ("cited authoritative definition of X" rather than "what is X"); (2) Direct citations in tools like Perplexity, ChatGPT, and Claude when you ask them questions in your domain; (3) Rising direct traffic from users who appear to know your brand without having ever clicked through search. None of these are perfect signals. Together they are reasonable evidence.

What is the single highest-leverage change I can make today? Cut the warm intro on your top 20 pages and replace each with a substantive opening that contains a definition, a cited claim, or a specific concrete fact in the first paragraph. This single change tends to produce a measurable lift in agent-mediated discovery within weeks. It is the lowest effort, highest ROI move.

How long is the right length for an agentic-search piece? There is no single right length. The right length is the length at which conceptual density stops being maintained. Some topics are 1,200 words; some are 6,000. The wrong move is to pad to a target word count. Agents punish padding. Write until the substance ends, then stop.

Should I publish my content openly even if I usually paywall it? For most knowledge businesses, yes, for the content you want agents to quote. The reasoning is straightforward: a paywalled piece cannot be cited by agents, which means it does not contribute to your authority in the agent-mediated layer. You can still gate premium analysis, proprietary data, or interactive tools. But the foundational explanatory content that establishes you as the source is more valuable open than paywalled, in 2026 and after.

The honest version: we got into this by accident. We started building agentic systems (a Telegram bot for Berlin culture with multi-agent search, a stem-cell research feed, a research-question dataset for an academic thesis) and we kept noticing how differently content needed to be written for the agents inside those systems versus for the humans outside them. The patterns in this article are not theoretical. They came from watching real agents read real content and choose what to quote.

We do content audits for AI-readability. We read your top pages the way an agent would, score them against the patterns above, identify the highest-leverage changes, and put them in priority order. We do this for teams of around 5 to 20 people because that is the team size where the founder or operator is still close enough to the content to act on the audit findings without it going to die in a committee. We are not a generic SEO agency. We do not produce roadmaps or 90-slide decks. We do small, real, measurable interventions on content you actually have.

If you read this article and recognised your own content in some of the patterns we said are misfiring, that is the conversation we have in a Garden content audit. Different optimisation target, different writing. But you have to know what you are actually optimising for. Email a@gardenresearch.eu if it is useful.

This article is part of Garden Research's investigation into how AI agents are reshaping discovery. The complete agentic AI search guide covers the underlying infrastructure (eleven tools, eight jobs); the semantic search guide covers the retrieval mechanism that does the heaviest lifting in agent-mediated discovery. This piece is about what that infrastructure means for the people who produce the content the agents are reading.