Seven AI Visibility Myths, Tested Against the Evidence

Every gold rush breeds a map-selling industry, and the rush to appear in AI answers is no different. Some of the maps are drawn from real terrain. Others are drawn from confidence. So let us do the unfashionable thing and check seven popular claims against the studies that actually exist.

Myth 1: Add llms.txt and the citations follow

The claim has a seductive shape: robots.txt tamed crawlers, sitemaps tamed indexing, so surely llms.txt, a tidy file summarising your site for language models, is the new secret handshake. It costs ten minutes, which is precisely why it spread.

Then someone actually measured it. SE Ranking ran a statistical study comparing sites with and without the file and found no measurable effect of llms.txt on citation frequency. None of the major engines has committed to reading it, and the study's null result matches that indifference. We should be honest where the folklore is not: as of mid-2026, llms.txt is a hopeful convention, not a lever.

Should you rip yours out? No. It is harmless, cheap, and might matter someday. But every hour spent polishing it is an hour taken from publishing the kind of citable work that measurably does move answers. File under tidy, not transformative.

Myth 2: A big following means AI will name you

This one comforts a decade of audience-building, which is exactly why it needs testing. The mechanism fails on contact: engines assemble recommendations from published, crawlable, corroborated evidence, and follower counts are none of those things. Much of the feed content behind a large following is unciteable by design: ephemeral, gated inside platforms, stripped of the claims and specifics that survive extraction.

We dedicated a full essay to this in why a big following won't make AI name you, and the pattern since has only hardened. The quiet expert with twelve deep, quotable articles keeps beating the influencer with six figures of followers and nothing a machine can lift. An audience is a fine asset for humans. It is simply not the currency the engines count.

The half-truth worth rescuing: a following is a distribution engine for the assets that do count. If your audience sends readers to a published essay, earns you podcast invitations, or prompts a trade publication to quote you, the following has converted itself into the machine's currency. The error is not having an audience; it is mistaking the audience for the record.

Myth 3: You can pay your way into ChatGPT answers

A persistent rumour, usually offered by someone selling the alleged plumbing. The reality: the organic answer, the part where an assistant says "consider this person", is not a placement product on any major engine. Advertising exists and will grow around AI experiences, labelled as advertising. The recommendation itself is synthesised from the model's training and retrieval, which is why nobody can invoice you for it.

We unpacked the full question, including what money can legitimately buy, in can you pay to appear in ChatGPT. The short version: you cannot buy the answer, but you can afford the evidence the answer is built from. That is either frustrating or wonderful news, depending on whether your budget was for shortcuts or for substance.

Myth 4: AI traffic is too small to bother with

Here the mythmakers have a real number, just the wrong one. AI referrals are indeed a trickle by volume, around 1.08% of website traffic across ten industries. Stop there and the case for waiting looks respectable.

But volume was never the point of a referral channel; intent is. Conductor's 2026 benchmarks, collected by SEO Sherpa, show ChatGPT referrals converting at 14.2 to 15.9% and Claude referrals at up to 16.8%, against roughly 1.76% for Google organic. A visitor arriving from an AI answer arrives pre-persuaded, because the machine already made the introduction and the case. And the upstream behaviour is moving fast: a G2 buyer-behaviour survey from March 2026 found 51% of B2B buyers now start research in an AI chatbot more often than in Google, up from 29% in April 2025. Small stream, extraordinarily valuable water, rising.

Myth 5: AI visibility is just SEO with a new name

The one myth with real substance inside it. The overlap is genuine: crawlability, structured pages, authority signals, quality content, all still load-bearing. If you did SEO well, you are not starting from zero, and anyone declaring SEO irrelevant is selling something.

But the unit of competition has changed, and that is not cosmetic. SEO competes for position on a results page the user reads; AI visibility competes for presence inside an answer the machine writes, judged at the level of passages, entities and corroborated facts. The scoreboard changed too: zero-click searches grew from 56% to 69% in the year after AI Overviews launched, and AI Overviews now appear on roughly a quarter of searches. Winning position ten on a page nobody scrolls is a trophy in a drained pool. The disciplines are cousins, not synonyms; we drew the full boundary in our plain-language guide to answer engine optimization.

Practically, the overlap means sequencing rather than choosing. Sound technical foundations and crawlable pages remain the entry fee; the new work sits on top, shaping passages a model can lift, consolidating your identity into one entity, and earning the third-party corroboration answers are assembled from. Teams that frame it as either-or usually end up playing the old game well and the new one not at all.

Myth 6: Once AI cites you, you are set

A pleasant belief for anyone who has been cited once, and the data is unsentimental about it: tracking by the Semrush AI Visibility Index shows 40 to 60% of sources cited in AI answers rotate month over month. Models update, indices refresh, retrieval re-ranks, and fresher or better-structured evidence displaces the incumbent.

The correct reading is that AI visibility is a practice, not a plaque. The people who hold their place keep publishing, keep their facts current, and keep their identity consistent, so that every rerun of the contest finds them slightly stronger. The consolation cuts the other way too: whoever is cited above you today is defending the spot monthly, whether they know it or not.

Rotation also explains why one-off audits mislead. A snapshot taken in March can flatter you into complacency or panic you unnecessarily; only a monthly series shows whether your presence in answers is a trend or a fluke, and the series costs nothing but the discipline of asking.

Myth 7: It is too early, wait for the space to settle

The most expensive myth on the list, because it sounds like prudence. The behaviour has already shifted: ChatGPT reached roughly 900 million weekly users by February 2026, up from around 400 million a year before, and the money has noticed, with the US market for generative engine optimization services projected at around $365 million in 2026, growing at roughly 42.9% a year.

Meanwhile the mechanics reward early movers structurally. Training data accumulates; corroboration compounds; a consistent entity builds machine familiarity the way interest builds capital. Waiting for the field to settle means arriving after the settlement, when the names the engines reach for are already habitual. The tools will keep changing. The advantage of having five years of citable, corroborated work will not.

Prudence also misreads what settling would even mean here. The interfaces will keep churning, but the underlying selection logic, published evidence plus consistent identity plus independent corroboration, has held stable across every model generation so far. You are not being asked to bet on a tool. You are being asked to start a record that any future tool will read.

Bonus myth: any AI mention is good, no matter what it says

This one sounds almost too obvious to need busting, but it quietly drives a lot of bad decisions. The assumption is that more mentions, in any context, with any framing, must be moving you in the right direction. It doesn't hold up. An AI answer that mentions your name alongside the wrong credentials, attaches you to a claim you never made, or surfaces you in a context that undersells what you actually do can do real damage, and it does that damage with the same authority-borrowing effect that makes a correct mention so valuable in the first place. A reader trusts the machine's summary; if the summary is wrong, they trust the wrong thing just as confidently as they would have trusted a right one.

We've covered the mechanics of this directly in what to do when AI gets facts wrong about you and in when AI invents your credentials, and the short version is that accuracy is not a side concern here, it is the entire currency. A hundred vague, half-wrong mentions are worth less than five precise, correctly-sourced ones, because the goal was never volume. It was ever being the trusted, accurately-described answer.

Bonus myth: blocking AI crawlers keeps you safe with no visibility cost

The instinct is understandable. Some publishers feel uneasy about a language model reading their work without payment or a click in return, and robots.txt looks like a clean, reversible way to opt out. The myth is believing this is a free, safe default with no downside. It isn't. Blocking a crawler doesn't just withhold your content from a model's training run. It removes you from the pool of pages an engine like Perplexity or an AI Overview searches through the moment someone asks a live question. You cannot be the cited, named answer to a question if the tool answering it was never allowed to read what you wrote in the first place. Blocking is not neutral. It is an active decision to be unfindable by exactly the systems this whole field is about being found by.

That doesn't mean blocking is always the wrong call. There are legitimate reasons some publishers choose it, and the tradeoffs deserve a full look rather than a slogan either way, which is exactly what our deeper piece on whether you should block AI bots walks through. The point here is narrower: treating a block as costless is the myth, not the decision to block itself.

All nine myths, side by side

Here is the full scorecard in one place, myth against the honest reality, including the two bonus ones above.

Table: all nine myths from this piece, side by side with the reality that survives contact with the evidence.

So what actually works?

Strip the folklore away and the surviving programme is almost embarrassingly plain: publish real, specific, liftable work under a consistent name; earn independent corroboration; keep your facts current everywhere machines read; and measure by asking the engines the questions your buyers ask. No magic file, no purchased answer, no follower arithmetic. The myths persist because each promises a shortcut around the same unavoidable fact: the engines reward accumulated, verifiable work, and accumulation cannot be retrofitted. That is bad news for hacks and rather good news for anyone willing to be patient for a year. If you want your own scorecard run honestly, myth-free, that is what our services are for, and the rest of the journal tests the folklore one claim at a time.