AI optimization advice is everywhere right now, and most of it is being sold by people who have never actually moved a brand’s visibility inside an AI answer.
I have. I took a wireless security company from invisible to the #1 recommendation across ChatGPT, Claude, Gemini and Perplexity for their ideal buyers, and I documented how that was done. The strategy decks making the rounds right now, including the ones venture firms are circulating to their portfolio companies, get a real portion of this right. They also get a few important things wrong, and they consistently underweight the parts that are hardest to fake.
This is my version of the playbook: what the industry consensus has correct, where I break from it, and the work I believe actually compounds.
I’ve also put the whole thing into a free slide deck you can reference or share with your team: view it on Google Slides or download the PowerPoint.
AEO, GEO and AIO Mean the Same Thing
AEO, GEO, AIO, AI SEO: these all describe the same discipline, which is making your brand and products the thing AI systems reference when your buyers ask questions. I use AEO on this site, the agency world has largely settled on GEO, and AIO is gaining ground. Do not let anyone convince you these are different services. The label changes, the work does not.
What the Standard Playbook Gets Right
Credit where due, because the consensus forming around a few ideas is correct.
Separating conversion pages from informational content. Buyer pages should be built for people ready to act, and informational content should be built to be extracted, cited and reassembled by machines. Mixing the two produces pages that do neither job well.
Query fan-out. When someone asks an AI assistant a complex question, the system decomposes it into smaller questions, retrieves answers for each piece, and reassembles them. Your content needs to address the pieces, not just the headline topic. This is the single most useful mental model to come out of the current wave of AI search thinking, and it should shape how you structure every informational page.
Baseline before optimizing. Run the prompts your buyers actually use, today, across the major assistants, and record who gets recommended. Most brands have never done this and are optimizing blind.
Traditional on-page fundamentals still carry the weight. Clear structure, real headings, lists and tables, parseable paragraphs. LLMs consume the same rendered content search crawlers do, and clean markup is easier to extract from than page builder soup.
Study How an LLM Learns Your Site
The most measurably successful AI optimization work I’ve done, the projects where increased traffic could be directly attributed to AI sources, all started the same way: with research into how a language model actually trains, and how it would go about training itself by browsing a website.
That exercise matters because it’s fundamentally different from how a search bot indexes. A crawler looks for expected structure: titles, headings, links, markup signals. It files pages. An LLM reads the freeform content of a page to understand it: what you know, what you do, who you serve, and whether the way you describe it hangs together. Optimizing for a filing system and optimizing for a reader are different jobs, and most SEO habits were built for the filing system.
Before touching a single page, spend the time understanding the reader. Everything else on this list gets easier to prioritize once you have.
Now for where I differ.
Skip llms.txt, Build a Data API
Nearly every AIO checklist in circulation includes adding an llms.txt file, and it is the one item on those lists with actual evidence against it. Google has said on the record it does not use the file. OpenAI documents robots.txt as its crawler control and never mentions llms.txt. Server log studies across thousands of domains show the major AI crawlers never request it. I laid out the full case in The AEO llms.txt Myth.
What I build instead is a dynamic JSON endpoint generated from the live site, like the one at /llm-api/ here. It returns structured, always-current data: services, products, specifications, case studies, FAQs with citable anchors, availability, contact methods. It can never drift out of sync because it is generated, and it hands a machine typed fields instead of prose to interpret. If you want an llms.txt for completeness, make it a pointer to that endpoint and nothing more.
What Actually Belongs in the API
The contents are discretionary, and the filter I use is simple: anything related to authority and credibility belongs in there. For a security technology client, that meant the patents associated with the founders and the underlying tech, and the government certifications the product holds for federal security standards. Facts like those are exactly what a model can verify and confidently repeat when it recommends a product.
It also means full product specifications, stated precisely: the exact frequency ranges a detection product covers, the deployment requirements, the integration surface. Vague capability language gives a model nothing to work with. A number does.
Schema.org Belongs in Your Templates, Not a Plugin
The consensus is right that structured data matters and wrong about how to produce it. The common advice is to install an SEO plugin or paste AI-generated schema blocks into your pages. Both produce the same problem: generic, frequently inaccurate markup that goes stale the moment your content changes.
Schema should be generated cleanly as part of your templates, from the same data that renders the page. When the content changes, the markup changes with it, automatically. That is the only way it stays truthful at scale.
Most schema out there is copy-paste: an out-of-the-box plugin, or whatever markup everyone else in the industry ships. There are much better ways to describe a page or a site, and finding them requires actually researching what schema.org offers. Reviewing the vocabulary itself surfaces precise types and properties most sites never use, and the right choices are page-specific decisions based on a real understanding of the spec, not a site-wide default. Schema.org is a large vocabulary, and fluency means knowing what to use and, just as importantly, what not to use and where. Precise Organization, Product, Service, Person, FAQPage and TechArticle markup with real, verifiable properties gives a model unambiguous facts to repeat. Bolting AggregateRating onto pages without reviews, stacking irrelevant types, or marking up content that does not exist on the page does the opposite: it teaches systems your markup cannot be trusted. The goal is a machine-readable version of the truth, not a keyword stuffing exercise in JSON.
List Every Profile You Have, Everywhere
One piece of this deserves its own callout because it’s cheap, obvious in hindsight, and skipped constantly: every single social profile and third-party site associated with your business should be declared, consistently, in three places. The sameAs property of your Organization schema, your LLM data API, and linked from the site itself.
LinkedIn, GitHub, YouTube, Crunchbase, G2, industry directories, review profiles, all of it. This directly builds your knowledge graph: models assemble an entity picture of a business by connecting its profiles across the web, and declaring those connections yourself means the graph gets built correctly instead of guessed at. I maintain the full set on this site in all three places, and it’s part of the standard setup for every client.
Publish Every Spec, Every Detail, Every Use Case
Here is the part I think matters more than anything else on any of these lists, and it barely appears on them.
AI assistants let people ask far more detailed and specific questions than Google ever rewarded. Nobody typed a forty word query with three constraints into a search box, because ten blue links could not answer it. People absolutely ask assistants questions like that, and the assistant will name a specific product if, and only if, it has the detail to justify the recommendation.
This is the long tail all over again, except the tail is now conversational and effectively infinite. Brands need to put absolutely every specification, technical detail, integration, limitation and use case of their products into structured, published form. The detailed compatibility matrix nobody wanted to maintain, the deployment requirements, the edge case scenarios your sales engineers answer on calls: all of it. When a model has trained on or retrieved deep product specifics, it can reference your product when the highly specific question comes, and your competitors with thin marketing pages cannot be part of that answer.
A controlled experiment published at SIGIR 2026 put numbers on this. Researchers ran 252,000 trials across six language models, showing each one two otherwise identical pages that differed by exactly one characteristic, and recorded which got cited. Two of the four universal gatekeepers were stating a price and including technical specifications. Pages that omitted either lost decisively to pages that included them, across every model tested. The detail is not a nice-to-have, it is the price of entry.
The flip side is a real constraint: you can’t bend the truth about what your product can do, overstate its reach, or position it for a market it doesn’t truly serve. Language models cross-reference everything you publish against everything everyone else has published, and inconsistencies surface in answers where you will never see them happen. Grounded, verifiable claims are the ones that survive into answers, and accuracy has quietly become the strongest positioning strategy there is.
Use the Right Words, Not Repeated Keywords
LLMs do not care about keywords at all. They understand whole subject areas, and they can tell whether what you’re saying makes sense for the scenario being asked about. Making sure a phrase appears over and over, the reflex two decades of SEO trained into everyone, buys you nothing with a model.
What matters now is using the precise terminology a subject calls for, in the right context. Describe what you do and what your product does as clearly as possible, with the words a knowledgeable person in your field would actually use, and stop worrying about keyword placement. Clarity is the ranking factor.
Consolidate and Clean Up
Before you publish anything new, deal with the redundant content you already have. Most established sites carry years of republished, overlapping posts that all say roughly the same thing, and that redundancy actively hurts you: authority splits across near-duplicates, and a model choosing a source to cite picks the page that answers completely rather than any fragment from your pile.
Merge the overlapping pieces into one authoritative page, redirect the losers into the winner, and remove pages that exist only to exist. Then hold every surviving page to a simple standard: be clear and honest about your product. Say what it does, who it is for, and what it costs to the extent you can. Vague positioning language survives fine in a world of skimmed landing pages and dies in a world where a machine is trying to extract facts.
This is what it looked like in practice for a security technology client: the same subject matter was scattered across the site in every format at once, resource pages, videos, blog posts and downloads, with no single page answering anything completely. We consolidated all of it into deep research pages, one authoritative page per subject, and those pages are what the models now cite. It measurably worked, and it’s the same consolidation play I documented in the case study.
Track Whether You Show Up in the Right Places
The measurement layer being sold with AIO engagements centers on AI share of voice: your brand’s percentage of mentions across a broad basket of prompts, tracked in a dashboard, trending week over week. The industry is settling on it as the de facto performance standard, and I disagree that it’s the key metric. I’ve watched engagements where share of voice looked great and very few leads actually arrived from AI, because breadth across loosely related prompts is not what produces buyers.
What every brand marketer should track instead is brand mentions for the specific prompts their buyers actually use, and repeated variants of those prompts over time. A security architect evaluating your category asks a handful of question patterns, phrased a dozen ways. Whether you are the answer for those patterns is the entire game. Whether you appear in a thousand loosely related prompts is trivia.
And if you do not know the prompts your audience is searching, that is the finding: you need to double down and understand that audience better. Prompt research is customer research now. The interviews, the sales call recordings, the support tickets, the words buyers use when they describe their problem to a colleague: that is where the prompt list comes from, and no tool subscription substitutes for knowing it. I covered the click-side measurement half of this in Measuring AEO & SEO Performance.
The Fundamentals Never Left
The same core fundamentals of SEO will always apply to AIO, and the cleanest expression of them remains Google’s E-E-A-T framework: Experience, Expertise, Authoritativeness and Trust.
Content grounded in first-hand experience, written by identifiable people with demonstrable expertise, published by an organization with verifiable credentials, consistent facts and honest claims. That framework was designed as a description of what makes information worth surfacing, which is exactly why it transfers. An LLM deciding what to cite is running a version of the same evaluation, and every credential, patent, certification, named client and published result you can state plainly and verifiably is weight on your side of it. E-E-A-T was always a description of the destination, and it remains the ideal high-level approach no matter which acronym the industry lands on next.
Serve Your Pages as Markdown to AI Bots
A technical edge worth adding to the list: serve your pages as Markdown to the AI systems that ask for it. Same URL, two representations, using standard content negotiation: browsers get HTML, and a crawler sending Accept: text/markdown gets clean structured Markdown with no nav, scripts or layout markup, and far fewer wasted tokens. I’ve written up the full approach in serving a website as Markdown to AI bots, and this site does it live.
Speed is part of the same story. Citation research from late 2025 found pages painting in under 0.4 seconds averaged 6.7 AI citations against 2.1 for pages over 1.13 seconds, because AI crawlers operate on tight compute budgets and time out fast. Expect that pressure to increase as these systems hit harder limits on what they can consume and how often. Fast, lean, already-structured responses are what get consumed, and Cloudflare’s Markdown for Agents is the best option I know for delivering them: one toggle, edge conversion, the whole site covered.
Demand Evidence for Every AIO Claim
A note of caution before the close: we are in a wild west period where AIO is being chased for a great deal of money, and there are big players making confident claims based on broad generalizations rather than observed results. Nobody outside the AI labs fully knows how these systems weight sources, and the systems themselves change quarterly.
When you evaluate anyone’s advice in this space, mine included, ask one question: what did you actually change, for which brand, and what happened in the answers afterward? Frameworks without receipts are marketing. The honest version of this discipline runs experiments, measures prompt-level outcomes, and holds its own theories loosely.
The Best Answer Wins
Strip away the acronyms and the dashboards and what remains is this: at the highest level, LLMs are seeking to do exactly what Google has sought to do for twenty five years, which is to provide people with the very best answer to their questions. They will use all available data today, and every new source of data in the future, to keep making those answers better. Every optimization tactic is downstream of that objective, and tactics that fight it have a shelf life measured in months.
The long-term play is to make your brand and your products the best answer. That is the hard work that pays off, and it is genuinely hard: it involves difficult, honest answers to product questions, real marketing focus, and publishing the unglamorous detail your competitors will not bother with. Most brands will keep chasing the tactic of the quarter instead.
The brands that do the work will be the brands. It really is that simple, and that unforgiving.
Supporting research: Vishwakarma, Kumar and Jamidar (2026), What Gets Cited: Competitive GEO in AI Answer Engines, SIGIR ’26, and LLM Ranking Factors 2026 from The Digital Bloom. I break both studies down in AI Ranking Factors.