Getting cited by AI assistants is a supply problem, not an optimization one

The model names your competitor because a corroborated body of work about them exists across the web. About you, it does not. Their markup is not better. The shortlist forms inside the assistant now, before a human loads a single page.

It feels mystical, like a switch a rival flipped. It is more ordinary than that. Gartner's 2025 survey of about 646 buyers found 45 percent used generative AI during a recent purchase. Forrester's 2026 survey of roughly 18,000 business buyers found the overwhelming majority now research this way. Absent from the model, you are absent from the room where buying starts.

So the useful question is not what your competitor optimized. It is what the web already says about them that it cannot yet say about you. The model read independent work about their product, their results, their arguments. About you it found a homepage and little else. That is a supply gap, not an optimization gap. The whole GEO industry sells the second story, because only the second one has a product attached.

Each assistant retrieves from the search index behind it and favors sources the wider web already corroborates. The surfaces differ; the trust signal does not. The model does not know your category from memory. It runs a search, pulls a handful of pages, and writes an answer over what came back.

Two acronyms get sold around this. GEO, generative engine optimization, means shaping content so a generative engine quotes it. AEO, answer engine optimization, aims the same idea at the direct-answer box. Both name a goal. Neither is a separate machine.

The engines differ only at the surface. ChatGPT's web search leans on Microsoft's Bing index, plus OpenAI's own crawler. Google's AI Overviews are, per Google's documentation, rooted in its core Search ranking systems, pulling pages from the Search index. Perplexity runs its own crawler and index and, per independent studies, leans notably on Reddit. They read different indexes and reward the same thing. That last part is the load-bearing one.

It is people-first SEO done to a higher bar, and Google says so in a page of its own. The AI features run on the same core ranking systems, so there is no separate GEO discipline to buy. Read any ranking guide and you get the opposite premise: a new job, with its own checklist, its own retainer.

Google's Search Central guidance frames the whole effort in one line.¹ Optimizing for AI search is still SEO. Not a cousin of it, not a successor. The same systems that decide organic rank decide what the AI features surface. If the feature runs on the ranking system, a second discipline with a separate bag of tricks does nothing the first one does not. The GEO-checklist industry is selling a second job, and the maker of the first job published, for free, the page that says the second one is unnecessary.

No. Llms.txt has no effect on Google Search or its generative AI features, structured data is not required, and the guides' own cited data undercuts the hacks they sell. This is the part the checklist economy would rather you skip.

On June 15, 2026, Google updated its guidance under the heading “Clarifying guidance on llms.txt files.”²Owners do not need llms.txt or other machine-readable files to appear in Google Search, including its generative AI capabilities. Such files “will neither harm nor help your” visibility, because Google Search ignores them.

On markup the same doc is blunt.¹ Structured data is not required for generative AI search, and there is no special schema you need to add. It also says to skip content chunking and AI-only rewriting, and warns that inauthentic mentions help less than they seem. Five hacks, one debunk, from the company that publishes it for free.

The vendors' own numbers agree, quietly. Ahrefs tracked 1,885 pages that added schema, each matched to three controls. Cited pages were almost three times more likely to already carry it, a correlation the guides quote endlessly. But adding schema to pages already cited moved nothing: AI Mode up 2.4 percent, ChatGPT up 2.2 percent, both indistinguishable from zero.

One machine-readable thing is not optional, though. The crawlers have to reach you. Block OpenAI's OAI-SearchBotin robots.txt and you vanish from ChatGPT's answers. Block PerplexityBot and you vanish from Perplexity. For Google's AI features the crawler is ordinary Googlebot, so normal indexing is enough; Google-Extended only governs Gemini training, not whether you appear in AI Overviews. The file that promises everything does nothing. The access nobody sells you is the thing that matters.

We run the exact pipeline every guide describes secondhand. Every article we publish emits registry-driven JSON-LD, submits to Search Console and Bing, and pings IndexNow as it goes out. It is clean, correct, and worth doing, because a crawler should never be the reason you are missed. It is also not a lever. It is hygiene: the table stakes that let the real work be found, never the thing that gets it cited.

One honest distinction, because it is where the confusion lives. The peer-reviewed GEO paper, Princeton-led with co-authors at Georgia Tech, IIT Delhi, and the Allen Institute for AI, found that adding citations, quotations, and statistics lifted generative visibility by up to 40 percent across roughly 10,000 queries.³ Read what that measured. It is enriching the content the engine already retrieves, with real citations and real numbers on your own page. That is doing the work better, not marking it up harder.

Being cited is being linked as a source; being recommended is being named in the answer. Chase links on the engines that link, and chase being named everywhere else. They do not even show up at the same rate. Per AI-visibility vendor Spotlight, across 2.4 million results, Perplexity and Copilot link out in over 77 percent of answers. ChatGPT links in about 31 percent. Claude does not link at all.

Both wins run on one driver, and it is not markup. It is information asymmetry. Aligned models drift toward the typical, a mode collapse driven by a bias toward familiar text in their training.⁴ AI-assisted writing grows more alike across people even as each piece reads as more creative.⁵ It is the same pull to the middle we traced in why all AI marketing content looks the same.

We read those findings one way: a model reaches outside itself only for what it cannot generate from that center. A primary number. A first-hand result. A stance it cannot infer. An answer capsule or a schema block hands it content it already holds, which is why they change nothing.

The only durable lever is producing enough verifiable, first-hand work that independent sources corroborate you, because a citation is a supply problem, not an optimization one. It is work, not a toggle, which is why the listicles skip it. Five moves, in order.

Every move reduces to one bottleneck, and it is not that nothing is optimized. It is that nothing is published. The position, the data, the result sit in the founder's head, where no crawler can read them. That is the founder-as-bottleneck problem, and keeping that published voice yours as the volume grows is its own discipline.

An assistant only cites what it cannot already generate. So the real question was never how to get cited. It was what do we know that a model can't.

You cannot track it like a keyword rank, because the same prompt returns the same brands under one time in a hundred. So treat a citation as a feeder and a differentiation bet, not a KPI. There is no Search-Console-for-AI, and the reason is arithmetic.

Per SparkToro, across 12 prompts run 2,961 times on ChatGPT, Claude, and Google's AI, the same question asked again returns the same list of brands under one time in a hundred. Identical lists in the same order show up about one time in a thousand. You cannot rank-track an output that inconsistent. A dashboard built on it is measuring noise, which is exactly why a tool vendor can bill you for watching it.

A share-of-model dashboard is a vanity metric you can watch, but never move.

Being cited is also a different event from ranking, so your rank tools miss it by design. Only about 38 percent of AI Overview citations also rank in the top 10, down from about 76 percent a year earlier. A Surfer study of 10,000 keywords found 67.82 percent did not rank in the top 10 at all. What you can see is the referral: a click from ChatGPT or Perplexity lands in your analytics as a normal referrer, even though the citation behind it does not. So watch the inputs. The body of published work going up. The referral assist behind it.

The questions founders ask most about getting cited by AI, answered straight.

The citation is a byproduct of being the single most useful, most verifiable answer. The only input you control is the work, and the naming follows. There is no metric to optimize here. There is a body of first-hand work to build.

The moat the listicles cannot copy is boring on purpose. One named human verifies every statistic back to its primary source before it goes out, because a single mis-cited number is the fastest way to get dropped from an AI answer for good. That is a standard, not a hack, and a standard is the one thing a checklist cannot sell you. If you want to know what is already published about you, and what is missing, that gap is what our audit reads.