The Next Generation of SEO Is AI Visibility Engineering

Last spring, I sat across from a client who had every reason to be happy. Their flagship service page ranked #1 for the exact term they had hired us to win three years earlier. Position one, featured snippet, the whole trophy case. And their inbound leads were down almost 30% year over year.

So I did something that has become a reflex for me: I opened ChatGPT, Claude, and Gemini and asked each one the question their customers actually ask. Who should I hire for this service, in this market, for this kind of problem?

Their company didn’t come up once. Not in a single answer, across three different AI assistants. Two of their competitors, both ranking below them in Google, got named repeatedly.

That meeting changed how I explain my job to people. Here’s what I’ve learned: ranking #1 is no longer the finish line. It’s not even the same race. The finish line moved, and most of the industry is still sprinting toward the old one.

The Scoreboard Broke Before Most of Us Noticed

I don’t want to hand-wave at “changing user behavior,” because the data here is specific and it is brutal.

Rand Fishkin, co-founder of SparkToro, has been measuring zero-click search behavior since 2019, when roughly half of Google searches ended without a click. His June 2026 study with Similarweb’s clickstream panel found that 68% of US Google searches now end without a click to anywhere. Not without a click to your site. Without a click, period. For every 1,000 US searches, only 276 clicks reach the open web, down from 374 just two years earlier. That’s the open web losing roughly a quarter of its clicks per thousand searches in 24 months.

Two details in that study matter more than the headline number. First, AI Overviews now appear on more than 20% of searches, and when they show up, click-through rates drop by roughly 60%. Second, AI Mode accounted for only about a third of one percent of searches in the study window, yet Google announced at I/O 2026 that it had passed a billion monthly users with query volume more than doubling each quarter. The feature that will drive the next decade of this trend has barely started driving it.

Fishkin’s own conclusion is one I’d frame this way: you cannot fight this by simply getting better at traditional SEO, because SEO still matters as much as ever, it just no longer pays out in traffic the way it once did. The value moved from the click to the citation. Your content’s new job is to be the source the machines trust enough to name.

And in classic form, the clearest confirmation isn’t in anything Google said on a stage. It’s in what Google built. AI Overviews, AI Mode, the steady expansion of machine-readable everything. Watch what they build, not what they say. What they built is an answer engine, and their earnings suggest it’s working fine for them.

Ranking is a position. Visibility is a presence. AI search rewards the second one.

How AI Search Actually Retrieves (and Why Your Rank Tracker Can’t See It)

Before we can optimize for these systems, we have to be honest about how they work, because most of the industry is still optimizing against a mental model that’s a decade out of date.

Mike King, founder and CEO of iPullRank and the practitioner who coined the term relevance engineering, has done the most thorough public teardown of how AI Mode and AI Overviews function. The core mechanic is something Google calls query fan-out. When a user enters a query, the system doesn’t just run that query. It hands it to Gemini, which expands it into a set of synthetic queries, pulls related data from the Knowledge Graph, runs those searches in the background, extracts relevant passages from the results, and then synthesizes an answer from those chunks.

Sit with the implications of that for a minute, because they rewrite several assumptions at once.

First, you’re no longer ranking pages for queries. You’re competing at the passage level for queries you can’t see. King made this point in an interview with Danny Goodwin at Search Engine Land: when the industry saw pages ranked #52 appearing in AI Overviews, the early read was that AI Overviews ignore rankings. What was actually happening is that the document ranked #3 for one of the expanded synthetic queries that nobody’s rank tracker was monitoring. We were out of phase with the system we thought we were measuring.

Second, retrieval happens on chunks, not pages. These systems decompose your content into passages, embed them as vectors, and score them for semantic relevance against the expanded query set. A 3,000-word page with one brilliant paragraph buried at word 2,400 competes passage against passage with a competitor who put their answer in a clean, self-contained block near the top. This is why semantic chunking, one idea per section, direct answers stated early, and clear semantic triples (subject, predicate, object statements a machine can extract as facts) have moved from nice-to-have writing hygiene to retrieval strategy.

Third, information gain matters. If your passage says exactly what fifty other passages say, there is no reason to retrieve yours. Original claims, proprietary data, and named expertise are what make a chunk worth pulling. I’ve been calling this Catalyst Content for years; the retrieval layer just made the case for me.

Entities: The Decade-Old Idea That Just Became the Gating Factor

Let me be clear about something, because I’ve watched too many “AI SEO” pitches present entity optimization as a shiny new discovery: entities are not new. Anyone framing them that way is telling you they weren’t paying attention for the past decade.

Google announced the Knowledge Graph in May 2012 with Amit Singhal’s famous “things, not strings” post. Hummingbird rebuilt the core algorithm around semantic understanding in 2013, documented at the time in Danny Sullivan’s definitive FAQ at Search Engine Land. The late, great Bill Slawski spent years dissecting Google’s semantic and entity patents at SEO by the Sea and telling anyone who would listen where this was headed. Practitioners like Dixon Jones at InLinks built entire platforms on entity SEO while most of the industry was still counting keyword density. The SEO community has been discussing, debating, and implementing entity optimization for well over ten years.

So what actually changed? The stakes.

For the past decade, entity clarity was a competitive advantage. You could rank respectably with sloppy entity signals because lexical relevance, links, and content quality could carry you. In retrieval-augmented systems, entity resolution happens before and during retrieval. When a model fans out a query about your category, it’s resolving which entities are relevant, pulling Knowledge Graph data about them, and deciding whose passages to even consider. An ambiguous entity doesn’t rank lower in that process. It gets filtered out of consideration entirely, or worse, conflated with someone else and misrepresented in the answer.

Entity clarity went from edge to prerequisite. That’s the shift, and it’s why work many of us have been doing for years, consistent naming across every surface, Organization and Person schema, sameAs corroboration to authoritative sources, internal architecture that reinforces entity relationships, suddenly has a much higher ceiling and a much harder floor. I build this out as an internal knowledge graph, which deserves and will get its own deep-dive article. For now: the machines can’t cite what they can’t confidently disambiguate.

The Crawler Reality: Your Site Is Probably Half-Invisible

Here’s the section where I get to ruin somebody’s week, because through my testing this is the single most common failure I find, and the data backs up how widespread it is.

Vercel and MERJ published a large-scale analysis of AI crawler behavior across Vercel’s network, covering hundreds of millions of requests. In a single month, GPTBot generated 569 million fetches and Anthropic’s crawler another 370 million, together about 20% of Googlebot’s 4.5 billion. These crawlers are hitting your site at serious scale. And here is the finding that should reorganize your technical priorities: they fetch JavaScript files, GPTBot in about 11.5% of requests, ClaudeBot in about 24%, but they do not execute them. None of the major AI crawlers render JavaScript. The one exception is Gemini, which inherits Googlebot’s rendering infrastructure.

Chris Long of Go Fish Digital flagged this study when it dropped, and the practical translation is stark: a React or Vue site with client-side rendered content can sit at position one in Google, because Googlebot renders JavaScript, while being a literal blank page to the crawlers behind ChatGPT, Claude, and Perplexity. Same URL, two readers, opposite outcomes. If your key content, pricing, or schema only exists after JavaScript runs, it does not exist for most AI systems. The test takes thirty seconds: view source, not DevTools, and search for a sentence of your main content. If it’s not in the raw HTML, you have a rendering gap.

The second layer of this problem is robots.txt, and it’s more nuanced than “don’t block the bots.” The major AI companies now run distinct crawlers for distinct jobs. OpenAI operates GPTBot for training, OAI-SearchBot for its search index, and ChatGPT-User for live fetches when a user asks about a URL. Anthropic mirrors this with ClaudeBot and Claude-SearchBot. That separation means you can make deliberate policy choices, like allowing the retrieval and search crawlers so you appear in AI answers while restricting the training crawlers, instead of the blunt block-everything rules I keep finding in audits, half of them added by a plugin update nobody reviewed.

The Machine-Readable Declaration Layer

On top of crawlability sits an emerging layer of files that describe your site to AI systems directly, and I want to give you the honest version of this, because it gets oversold constantly.

The llms.txt proposal came from Jeremy Howard, co-founder of Answer.AI, in September 2024: a curated markdown file at your domain root giving AI systems a token-efficient map of your most important content, skipping the navigation chrome and script noise that makes HTML expensive for a context window. Since then, adoption has spread across thousands of sites, including Anthropic, Cloudflare, and Vercel, and adjacent proposals have emerged, including Google’s Open Knowledge Format, which I’ve covered separately.

The honest caveats: llms.txt is a proposed standard, not a ratified one. No major crawler operator has formally committed to consuming it, and anyone selling it as a guaranteed ranking factor is ahead of the evidence. I’ve written at length about where these standards stand and where I think they’re going. My position: these files are cheap, low-risk infrastructure for an agentic web that is clearly coming, and early field reports of AI systems citing them directly are starting to appear. I implement them, but I just refuse to overpromise on them.

The genuine complication is maintenance. Sites change weekly, and a hand-maintained declaration file drifts out of truth fast, which is arguably worse than not having one. That decay problem is why I built LLMS Amplifier, a WordPress plugin that generates and updates these files automatically. Whether you use my tool or a cron job and discipline, the principle holds: a declaration to machines has to stay true to be worth anything.

So What Is AI Visibility Engineering?

Let me define the term I’ve been circling.

AI Visibility Engineering is the discipline of making a business identifiable, retrievable, trustworthy, and citable to AI systems. It’s what SEO becomes when the surface you’re optimizing stops being a results page and starts being a set of retrieval pipelines that answer questions directly. Mike King calls his framing relevance engineering, and the fact that serious practitioners are independently converging on engineering language is itself the tell: this work now requires understanding information retrieval, not just following best-practice checklists.

I’ve watched this industry declare SEO dead roughly every eighteen months since I started, and SEO has attended all of its own funerals in good health. This is not a replacement story. It’s a re-pricing. The fundamentals that made sites rank are the same fundamentals that make sites citable; what changed is the surface area, the unit of competition (passages, not pages), and the standard of proof.

The way I structure the discipline, it rests on four pillars:

Machine readability. Can AI systems crawl, parse, and extract your content? Server-rendered HTML, clean structure, structured data, deliberate crawler policy, and machine-readable declarations.

Entity clarity. Can machines resolve exactly who you are, what you do, and how your pieces relate, without ambiguity? Schema, consistent naming, sameAs corroboration, and knowledge graph architecture.

Trust and evidence. Can machines verify you’re real, credentialed, and corroborated? Author entities, third-party validation, and increasingly, provenance signals like verified authorship.

Citability. Does your content contain passages worth retrieving? Direct answers, original claims, real data, information gain over the rest of the corpus.

Each pillar gets its own deep treatment in the articles that follow this one. Consider this the map; the territory is coming.

What I’d Do First If I Were Starting Today

If you’re reading this with a sinking feeling, the useful news is that the gap between where most sites sit and where they need to be is made of fixable problems, and almost nobody is fixing them yet. Based on my experience running these audits, here’s the sequence.

Start by measuring actual AI visibility. Ask ChatGPT, Claude, Gemini, and Perplexity the questions your customers ask, in a few phrasings. Record who gets named, what gets said about you, and what gets said that’s wrong. Inaccuracy is its own emergency; a model confidently misdescribing your business is worse than one ignoring you.

Then verify readability at the raw HTML level. View source on your key pages and confirm the content is actually there. Pull your server logs and look for GPTBot, OAI-SearchBot, ClaudeBot, and PerplexityBot: are they getting 200s on content-rich pages, or crawling empty shells? Audit robots.txt for blanket blocks you didn’t consciously choose.

Then think in fan-outs, not keywords. For each priority topic, map the cluster of sub-questions a query expansion would generate, and make sure you have clean, self-contained passages answering each one, with the answer stated early, not buried under eight hundred words of preamble.

Then establish entity clarity and the trust layer: Organization and Person schema, consistent naming across every platform you control, author bios with depth, corroborating profiles that agree with each other.

The advanced work, declaration files, provenance signals, knowledge graph refinement, belongs on top of that foundation. I’ve watched people implement llms.txt on sites that block GPTBot, which is like printing a beautiful menu for a restaurant with the doors locked.

The Citation Layer Is Still Up for Grabs

Twenty years in this industry has taught me that the biggest advantages go to practitioners who take a shift seriously about two years before it becomes conventional wisdom. Mobile-first rewarded the early movers. Helpful Content punished the late ones. The data says this shift is bigger than either, and moving faster.

Right now the citation layer, the set of sources AI systems trust enough to name, is still forming. Businesses that make themselves readable, resolvable, and verifiable over the next year or two get cited first, and citation compounds, because machines, like people, keep recommending what they already trust.

That’s the work I’m doing, testing, and documenting here. Next in this series: why most websites will never be recommended by AI, how digital trust is starting to outweigh backlinks, and the full checklist I use to audit AI readiness. If you’ve run your own AI visibility experiments, I’d genuinely love to hear what you’re seeing. The practitioners comparing notes right now are the ones writing the next playbook.

Written by Brian Winum

Digital marketing veteran, partner at MAXBURST and MAXPlaces, and creator of LLMS Amplifier and Authority Amplifier.

About Brian Work with me