AI search is new enough that most of what circulates about it is folklore, and a lot of the popular advice is either wrong or backwards. Over 2025 and 2026, Ahrefs and others ran large studies — millions of prompts, hundreds of thousands of brands and SERPs, controlled before-and-after tests — and the results quietly killed a set of confident claims that GEO consultants still sell. Adding a schema markup barely moved AI citations in a 1,885-page controlled test. Nearly all llms.txt files are never read by an AI at all. Publishing your own "best-of" list does not reliably get your brand recommended; in one controlled experiment the AI named a competitor 43% of the time. Backlinks and domain authority correlate only weakly with AI mentions, while branded web mentions and — most strongly of all — YouTube mentions correlate far more. Traditional search rankings still feed AI answers heavily, yet ranking page one guarantees nothing: most AI-cited URLs no longer rank in the top ten. AI answers churn on the surface but stay stable underneath. And AI search has not replaced Google, which still sends vastly more traffic — even as its own AI Overviews cut clicks to the pages they summarize. This guide takes the nine most common myths one at a time, states what the data actually shows, and draws the through-line: visibility in AI search is earned by genuinely credible, specific content and by earned mentions across the surfaces AI trusts — video especially — not by the technical shortcuts most tools are still selling.
AI search is young enough that the loudest advice about it is mostly untested, and a striking amount of it turns out to be wrong when someone finally runs the numbers. Over 2025 and 2026, Ahrefs and other analysts published large studies — some spanning millions of AI prompts, hundreds of thousands of SERPs, or tens of thousands of brands, several using controlled before-and-after designs — and the results quietly demolished a set of confident claims that GEO tools and consultants still sell as tactics. The pattern is consistent: the technical shortcuts people reach for barely move anything, and the things that actually correlate with getting cited are harder, slower, and less sellable as a product.
This guide takes the nine most common AI search myths one at a time. For each, it states the belief, then what the data actually shows, and it keeps the honesty rule that runs through all credible analysis of this space: correlation is not causation, single-vendor studies are directional not gospel, and the exact percentages will drift as the engines change. What does not drift is the direction, and the direction points somewhere specific — visibility in AI search is earned by genuinely credible, specific content and by earned mentions across the surfaces AI trusts, not by editing a file almost nothing reads. It sits alongside the definitional generative engine optimization entry and the channel-level playbook in AI search visibility; this page is the myth-by-myth counterweight to both.
Most of the strongest evidence here comes from Ahrefs, which sits on a large web index and its own analytics panel and has published a run of AI-search studies through 2026 — a schema before-and-after test, an llms.txt server-log analysis, a 75,000-brand correlation study, and traffic comparisons across tens of thousands of sites. Others corroborate the click-loss findings: Seer Interactive, Authoritas, and independent analysts have measured the same AI Overviews CTR collapse. Where a claim rests on a single study, this guide says so. Two cautions apply throughout. Correlation studies tell you what co-occurs with citation, not what causes it — schema-heavy sites also tend to be better-run in every other way. And any one vendor's counts are a sample, not an audit, so treat exact figures as directional. With those caveats stated once, the myths.
The first cluster is the most expensive to believe, because these are the tactics sold as the quick path to AI visibility. Each is a technical or self-serving move that feels like it should work and, in the data, mostly does not.
The reasoning sounds airtight: AI loves to quote listicles, so publish a "best tools for X" post with yourself at the top and the AI will repeat it. Listicles genuinely are cited heavily — in one analysis of 750 prompts, "best X" pages made up around 44% of ChatGPT's citations. But that is the format being cited, not your self-interested version of it winning. A controlled experiment that tracked self-promotional lists across thousands of AI answers found the model frequently ignored the brand's own ranking and recommended a competitor instead — in one setup, about 43% of the time. AI engines cross-reference; a list that exists mainly to promote its publisher reads as exactly that, and the model reaches for a more neutral-looking source. Being genuinely, verifiably the best answer is what gets quoted, not declaring yourself so on your own page.
llms.txt — a proposed standard file telling AI crawlers what your site is about — got adopted as a checkbox item almost before anyone checked whether it does anything. The server logs say it largely does not. In an Ahrefs analysis, roughly 28% of about 137,000 sites had published an llms.txt file, but around 97% of those files were never fetched by anything at all. Of the small remainder that were read, most of the reads came from non-AI tools — SEO auditors and GEO dashboards checking for the file — rather than from the AI assistants it is meant for. No major assistant has confirmed it uses llms.txt to build answers. It is not harmful to have one, but it is effort spent feeding a file that almost nothing reads, and no substitute for content an AI can actually find and quote.
This one had real-looking evidence behind it — cited pages are markedly more likely to carry JSON-LD schema than uncited ones — which made "add schema" a near-universal GEO recommendation. Then Ahrefs ran the controlled version. They tracked 1,885 pages that added JSON-LD schema, matched them against roughly 4,000 control pages with similar prior citation levels, and measured the change over 30 days. The results: about +2.2% for ChatGPT and +2.4% for Google AI Mode, both close enough to zero to count as noise, and a −4.6% decline for AI Overviews. The correlation was real but spurious — schema lives on better-maintained sites that also publish stronger content, build more authority, and earn citations for those reasons. Adding schema to an otherwise-unchanged page did not move citations. Structured data still has legitimate uses; manufacturing AI citations is not a proven one.
The second cluster is about where AI visibility comes from — and here the myths pull in two opposite, both-wrong directions. One camp says traditional SEO is dead and irrelevant to AI; the other says a page-one ranking is your ticket in. The data rejects both and lands in a more useful middle.
The claim that AI search has decoupled from classic SEO is popular and wrong. When Ahrefs looked at where ChatGPT's citations come from, the large majority — on the order of 88% — traced to pages in the general web search index, the same corpus that classic search ranks, rather than to some separate AI-native source pool. News, community sites, and video make up smaller slices. The practical read: the assistants overwhelmingly quote pages that already exist and are discoverable in ordinary search. If your content is not in that index and not findable through conventional means, it is not in the pool the AI draws from. Traditional search visibility is not obsolete for AI — it is the substrate AI answers are built on.
The mirror-image myth is that if you rank in the top ten, the AI will cite you. It often will not. Analyses of large SERP-and-citation datasets find that only around a third of AI-cited URLs actually rank in the top ten for the same query; a large share rank on pages two through ten of the results, and roughly a third do not rank in the top 100 at all. That overlap has been falling over time, not rising. So rankings feed AI (myth 4) but do not gate it (myth 5): a page can be cited without ranking, and a page-one ranking is no promise of a citation. What the AI selects for is the page that most credibly and cleanly answers the specific question, which is correlated with ranking but far from identical to it. The nuance of that split is the whole subject of SEO in the age of AI Overviews.
This is the load-bearing myth of old-school SEO carried into the new world, and it is the one the data contradicts most sharply. In Ahrefs' study of roughly 75,000 brands, the classic authority signals correlated only weakly with AI mentions — domain-rating and backlink correlations sat in the low tenths, the statistical definition of a weak relationship. What correlated strongly were mentions: branded web mentions came in around 0.66, and mentions on YouTube came in highest of all, near 0.74. In plain terms, how often your brand is talked about across the web — and shown on video — tracks AI visibility far better than how many links point at your domain. Correlation is not causation, and the study's authors say so plainly. But it reframes the work: from accumulating links to becoming genuinely and visibly mentioned, with video as an unusually strong surface. The broader version of this shift is in AI visibility beyond SEO.
The last cluster is about the shape of the channel itself — whether it is trackable, whether it has replaced Google, and whether being in an AI answer helps your traffic. Getting these wrong leads teams to either ignore AI search or wildly overreact to it.
Ask ChatGPT the same question twice and the wording changes, so it is tempting to conclude AI visibility is noise you cannot measure. The surface churns, but the substance is stable. In one study tracking 43,000 keywords over a month with many checks each, the exact wording changed around 70% of the time, the brands named shifted about 46% of the time, and the cited sources swapped roughly 45% of the time — yet the semantic similarity of the answers stayed very high, around 0.95 out of 1.0. The answer is saying nearly the same thing in different words with a rotating cast of citations. That means AI visibility is trackable if you measure the right thing: share of voice and citation rate across many prompts and repeated checks, not whether you appeared in one snapshot. How to do that properly is covered in AI visibility measurement and how AI search visibility metrics are actually calculated.
The breathless version of the AI-search story treats Google as finished. The traffic data says otherwise, emphatically. Across roughly 76,000 sites in late-2025 measurement, Google sent on the order of 190 times more referral traffic than ChatGPT, and ChatGPT's search volume was a small fraction — around a tenth — of Google's. AI search is real, growing, and worth optimizing for, but as a share of how people actually reach content it remains small next to Google. The more important nuance is that the real disruption is happening inside Google, not beside it: its own AI Overviews are changing search behavior at Google's scale, which is a far bigger force than the standalone assistants. Which leads to the last myth.
Some hoped that being cited in an AI Overview would send traffic the way a rich result once did. The opposite is closer to true. Ahrefs measured that AI Overviews cut the click-through rate on the top organic result substantially — a 2025 reading found roughly a 34% reduction, and 2026 measurement put the hit far higher, on the order of a 58% cut to position-one CTR. Independent analyses corroborate a large drop: Seer Interactive reported declines above 65%, Authoritas around 47%, others in between. The exact figure depends on query type and method, but the direction is unanimous. An AI Overview answers the question on the page, so fewer people click through — including to the very sources it cites. This is why being the cited source inside the answer, rather than a blue link beneath it, is now one of the few durable positions; the mechanics are detailed in AI Overviews are reducing organic clicks.
Read together, the debunked myths draw a single clear line. Everything on the losing side is a shortcut: a file you edit, a schema you paste, a list you publish about yourself, a link you buy. Everything on the winning side is earned and hard to fake. The AI quotes pages it can find in the ordinary search index (so a real search presence still matters), it selects the most credibly specific answer rather than the highest-ranked one (so genuine quality beats position), and it leans on how often and how visibly your brand is mentioned across the web — with video the strongest single surface — rather than on your domain's link profile. It answers questions consistently enough to be tracked, it has not replaced Google, and it is actively reducing the clicks that used to be the whole point.
So the honest strategy is unglamorous. Be genuinely and demonstrably good on the specific questions your audience asks, get talked about and shown across many surfaces — especially video — keep a consistent, quotable identity so the engines can corroborate you, and measure the channel properly instead of chasing snapshots. The reason this is hard is not that it is mysterious. It is that it is a lot of production: real content, on many platforms, in many formats, at a cadence, holding one brand line the whole time. That production load is the actual barrier between knowing what works and doing it — and it is the one thing on this page a tool can genuinely help with.
Notice what the surviving levers have in common: they are all output. You cannot fake earned brand mentions, but you can create the content that earns them. You cannot buy your way into AI citations with schema, but you can publish specific, credible answers across the web at a scale that makes being mentioned likely rather than lucky. And the single strongest correlate in the data — mentions on video, YouTube in particular — is the format most brands produce least, precisely because video is the most expensive thing to make by hand. The myths clear away the shortcuts and leave a production problem standing where the tactics used to be. That is the problem Kompozy is built for.
Kompozy is a full AI content generation and multi-platform publishing engine — not a repurposing add-on — which matters here because the winning list is a generation list, not an optimization one. It produces net-new short-form and avatar video that speaks directly to the strongest-correlated surface, alongside image posts, carousels rendered brand-exact through HyperFrames, listicle and naturalistic video, blog articles, and newsletters. Every one of those is a place your brand can be specifically, credibly present and quotably mentioned — the earned-presence side of the data rather than the technical-trick side. And because AI corroborates before it quotes, the fact that Kompozy governs all of it with one Persona Brief — fixing voice, claims, and banned words across every asset — is not a cosmetic feature: it is what keeps your name and facts identical across surfaces so an engine can confidently attribute them to you.
Then autopilot schedules and publishes that spread across the surfaces AI actually mines — eight social platforms plus blog and email — from one queue, behind a per-post review gate so a person signs off before anything ships, and on a recurring cadence, because the myth-busting data on volatility is also a mandate: a single dormant post is invisible, and consistent, current presence is what gets tracked and trusted. The honest framing is the one the whole guide has earned: Kompozy does not sell you a citation, and no tool can, because the myths just showed that shortcuts to citation do not exist. What it removes is the production ceiling that keeps the real, unglamorous strategy — be genuinely good, be mentioned everywhere, especially on video, consistently — from being something only a large team can sustain. It turns the correct answer into an affordable one.
AI search rewards almost the exact opposite of what most AI-search advice tells you to do. The schema, the llms.txt file, the self-promotional list, the backlink campaign — the data shows each barely moves AI citations, and some move them the wrong way. What moves them is being genuinely findable in ordinary search, being the most credibly specific answer to a real question, and being mentioned across the web and shown on video often enough that an engine can corroborate and quote you. AI has not replaced Google, its answers are stable enough to track, and appearing in an AI Overview cuts your clicks rather than growing them — so the position worth holding is the cited source inside the answer, earned the hard way. There is no shortcut hiding in a file or a markup tag. There is only the work of producing credible, specific, on-brand content across the surfaces AI trusts, at a scale most teams cannot reach by hand. Strip away the myths and that is the entire game.
The most common false beliefs, each contradicted by data: that schema markup boosts AI citations, that an llms.txt file is needed, that publishing your own "best-of" list gets you recommended, that backlinks and domain authority drive AI mentions, that ranking page one guarantees AI visibility, that AI answers are too volatile to track, and that AI search has already replaced Google. Ahrefs and other 2025–2026 studies show each of these is either wrong or badly overstated.
The data says barely, if at all. Ahrefs tracked 1,885 pages that added JSON-LD schema against roughly 4,000 matched control pages and measured citation change over 30 days: about +2.2% for ChatGPT and +2.4% for Google AI Mode — close enough to zero to be noise — and a −4.6% decline for AI Overviews. Cited pages do tend to have schema, but that is because well-run sites both add schema and earn citations for other reasons. Adding schema to an otherwise-unchanged page did not move the needle.
There is no evidence they do, and strong evidence most are ignored. In an Ahrefs analysis of server logs, roughly 28% of about 137,000 sites had published an llms.txt file, but around 97% of those files were never read by anything — and of the small share that were, most accesses came from non-AI tools like SEO auditors and GEO platforms, not from AI crawlers. No major AI assistant has confirmed it uses llms.txt to build answers. It is effort spent on a file almost nothing reads.
Three things the data supports. First, genuinely credible, specific content — because AI assembles answers from sources it can confidently quote, and specific, first-hand, well-evidenced material wins over generic coverage. Second, earned brand mentions across the web, which correlate far more strongly with AI mentions than backlinks do — and YouTube mentions correlate strongest of all, making video unusually valuable. Third, a solid traditional search presence, since most AI citations still come from pages in the general search index. Shortcuts lose; earned presence wins.
No. In late-2025 traffic data across roughly 76,000 sites, Google sent on the order of 190 times more referral traffic than ChatGPT, and ChatGPT's search volume was a small fraction of Google's. AI search is real and growing and worth optimizing for, but treating it as a replacement for Google is a myth. The bigger practical shift is that Google's own AI Overviews cut clicks to the pages they summarize, so the change is happening inside Google as much as outside it.
The debunked myths point at levers that are all production problems: earned brand mentions across many surfaces, video presence (the single strongest correlate of AI mentions), and specific, credible content instead of technical tricks. Kompozy is an AI content generation and multi-platform publishing engine that manufactures exactly that — net-new short-form and avatar video, image posts, carousels, blogs, and newsletters, all governed by one Persona Brief so your name and claims read consistently everywhere, published across eight social platforms plus blog and email. It does not sell a shortcut; it makes producing the signals that correlate with citation affordable at scale.
Most AI search advice is folklore the data disproves. Large 2025–2026 studies show schema markup barely moves AI citations, nearly all llms.txt files go unread, self-published "best-of" lists don't earn recommendations, and backlinks correlate only weakly with AI mentions — while branded and especially YouTube mentions correlate strongly. Traditional rankings still feed AI answers, but page-one ranking guarantees nothing, and Google still dwarfs AI search in traffic. Visibility is earned by credible, specific content and earned mentions, not technical shortcuts.
Get started → · ← All guides · Compare Kompozy vs other tools