4 AI Search Myths That Cost You Money (These Don't Matter)
The most repeated piece of AI search advice on the internet stopped being true back in 2024, and Google's own AI Overview was still handing it to me last week.
It told me to go optimize for Bing. That one is the bonus at the end. First, four AI search tactics that fall apart the moment you check them, all of which someone is actively selling you right now.
One quick note on where this came from. Dan Petrovic of DEJAN AI went on the Ahrefs Podcast in August and took apart several of these live. He is one of the people worth following on AI search, and I verified his claims independently rather than repeating them. Most held up. The one that did not is the same line Google is repeating.
Fair warning on my bias. All four already sit on the do-not-chase list inside my AI SEO System, so I am not arriving at this neutrally. Where a primary source settles a question I quote it directly, and where the evidence is one practitioner's own testing I label it as that. Every source is linked so you can check the reasoning instead of taking mine.
Myth 1: llms.txt Gets You Into AI Answers
There are generators, plugins, agency line items, and consultants selling you a plain text file.
Start with Google, since it is the only player here that states a position in writing. Its documentation on AI features says this, word for word:
"You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."
That is Google Search Central, not a blogger. Same page also states there are no additional requirements to appear in AI Overviews or AI Mode.
Then there is the question of whether anything reads the file at all. Ahrefs studied 137,000 domains and found 28 percent publish an llms.txt. Of the roughly 38,000 that had a valid one, 97 percent received zero requests for it in May 2026. Their words: no bots, no humans, nothing.
A separate study from SE Ranking ran 300,000 domains through correlation and machine learning and found the file did not make a domain more likely to be cited. Google's John Mueller compared it to the old keywords meta tag, and noted you can tell from your server logs that the AI services do not even check for it.
Now Here Is My Own Contradiction
Go type ryandoser.com/llms.txt into your browser. There is a file sitting right there.
Petrovic has one too, and his reasoning on the podcast was blunt. He said he put it on his site because it is so easy to do and it does not harm anything, then added that some fringe models use it and nothing important does.
That is the honest position rather than a clean one. The file costs fifteen minutes and zero dollars, so calling it a scam overstates things. Calling it a strategy overstates them a lot more.
It does carry one cost people skip past. It is a document you wrote about yourself that something might read, so a stale price or a dropped capability in there shows outdated information to anything that opens it. Keep the facts current, then stop thinking about it, and do not pay anyone a monthly fee to maintain one.
Myth 2: You Can Look Up Prompt Volume
Several tools now sell you something labeled prompt volume, presented the way keyword volume is presented, as though a number of people per month typed a given prompt.
Petrovic's response to this on the podcast is the most quotable thing in the whole hour. His words: "for everyone who's been hypnotized by the prompt search volumes, look at me in the eyes. There is no prompt volume."
His argument is mechanical rather than cynical. Outside a few short common phrasings, the odds of two people typing the same prompt the same way are small. People also do not send one prompt and stop, they have multi-turn conversations, and personalization layers on top so location, history and logged-in state all bend the answer.
Tim Soulo, Ahrefs' CMO, backed this up from their own sampling of clickstream providers. Real prompts run long, multi-topic, and often follow on from something else.
There is a defensible version of the measurement, and it is genuinely useful. Take one prompt, pull the fan-out queries the search backend actually fired, and look at real search volume for each. Petrovic's condition is the whole point. What you are looking at there is the volume of the queries the backend triggered, so call it that, and since those queries overlap, treat any sum of them as a rough ceiling rather than a headcount.
That distinction matters because the fan-out is real and documented. Google's own AI features page confirms both AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics. So the sub-queries exist and have measurable volume. The prompt on top of them does not.
Myth 3: You Can Rank in ChatGPT
Search how to rank in ChatGPT and you will find guides promising a number one position, some in 48 hours.
There is no fixed position to hold. Answers are generated per request, so while a brand can certainly be named first, that order is not a standing placement anyone assigns. Ask the same question twice and you can get different companies in a different sequence, from the same model, on the same day.
What does exist is a rate. Across many asks, phrasings and engines, you get named some percentage of the time. That number is real, it moves with your work, and it behaves nothing like a ranking you can hold.
The distinction changes what you should buy. If a vendor uses the word ranking as shorthand for a metric they will show you, fine. If they cannot say what position means or how they measure it, that is the moment to slow down, because everything else they tell you sits on top of it.
Same logic on the dashboards. If a tool scores your brand out of 100, ask what it measures, how many prompts it samples, which engines, and whether it can tell being recommended apart from being cited in an answer that sells somebody else. Those two are not the same, and I dug into why grounding, citation and mention are three different outcomes separately. Any of these numbers will move on their own between runs, which is exactly why you want to see the underlying answers rather than the score.
Myth 4: You Can Spam Your Way Into the Training Data
The theory goes that if you publish enough content saying your brand is the best, the next model learns it.
The labs have moved toward curated and filtered datasets, and the filtering step is cheap enough that volume alone is a bad bet.
Petrovic showed how cheap one kind of detection gets. He built a text classifier on gzip compression alone, no machine learning involved, and reports 90 to 99 percent accuracy separating quality writing from filler on his own test set. I covered how that classifier works in my roundup of AI search experts.
Be precise about what that shows. A quality filter is a different tool from a brand-manipulation detector, and this was his own unpublished test without a stated sample size or false-positive rate. He tells you to go replicate it, which is the right way to offer a claim like that.
What it does establish is the economics. If sorting real writing from generated filler costs almost nothing, then the cheapest thing you can do at scale is also the easiest thing to screen out. Soulo drew that conclusion on the episode, and it is the part of the argument that holds regardless of the exact accuracy figure.
The Bonus Myth: Rank in Bing and ChatGPT Will Find You
The other four all have somebody selling them. This one is different, and it is the reason I wanted to write this post at all.
When I searched "how to rank in chatgpt" on August 28, Google's own AI Overview listed optimizing for Bing indexing as step two, on the reasoning that ChatGPT leans on Bing's index. Petrovic gave a version of the same answer on the podcast, with the caveat that he was simplifying. Run that search yourself, since AI Overviews are personalized and change constantly.
It has gone stale. Your eligibility to appear in ChatGPT now runs through OpenAI's own crawler rather than through Bing, and OpenAI's crawler documentation says so plainly. I laid out the full evidence, including what the retrieval stack looks like by tier, in my breakdown of which search index each AI assistant reads.
What matters here is who was repeating it. Petrovic is a credible practitioner with fifteen years in and first-party research behind his claims, and he hedged the line as he said it. Google's own AI feature repeated it too. Nobody was selling anything.
That is the actual lesson. In a field moving this fast, most bad advice starts out as a true statement that quietly goes stale while everyone keeps repeating it, including the people who got it right the first time, and including Google.
Final Thoughts from Ryan
Here is the test to run on anything you get pitched from here on. Ask what it measures in one plain sentence. Ask to see the actual AI answers behind the number. Ask what happens to the score if you change nothing for a month.
What survives that test is usually boring. Can AI crawlers reach your pages. Are those pages indexed. Does the page answer the question in the first 40 words instead of the fifth paragraph. Is there a comparison table a model can lift cleanly. Does the web mention you anywhere you do not own. None of that photographs well in a dashboard, which is roughly why nobody is selling it to you.
AI search is new enough that nobody has to lie to sell you something useless. As the bonus item shows, a description that was accurate two years ago does the job on its own, even coming from Google.
Almost everything above traces to a document you can open in ten minutes, and where it traces to one guy's unpublished test instead, I said so.
Go check my llms.txt while you are at it.
AI Search Myths FAQs
Does llms.txt actually do anything for AI search?
There is no demonstrated citation benefit. Google's documentation states you do not need machine readable files or AI text files to appear in its AI features, and of roughly 38,000 domains Ahrefs found with a valid llms.txt, 97 percent received zero requests for it in May 2026. Some tools and smaller models do read the file, so this is not the same as saying nothing touches it. Keeping an accurate one costs little. Building a strategy on it is the mistake.
Is prompt volume a real metric?
Not as it is usually sold. Outside a few short common phrasings, identical prompts are improbable, most real usage is multi-turn, and personalization changes the result per person. The defensible version is to pull the fan-out queries a prompt actually triggers and look at search volume for those, treating any total as a ceiling rather than a headcount, since the sub-queries overlap.
Should I pay for an AI visibility tracking tool?
Only if it shows you the underlying answers. A tool that surfaces the actual prompts and responses is doing work you could not do by hand at volume. A score with no visible evidence behind it cannot be audited, and it drifts between runs regardless of what you changed, so you cannot tell a real move from noise.
How I sourced this: Google's position is quoted from Search Central's AI features documentation and OpenAI's from its published crawler docs, both read directly on August 28, 2026. The llms.txt figures come from Ahrefs' 137,000-domain analysis and SE Ranking's 300,000-domain analysis, both linked above and run on different samples, so their adoption rates are not directly comparable. Dan Petrovic's claims are quoted from the Ahrefs Podcast episode published August 18, 2026, and are his own first-party testing rather than published studies, which is how I have labeled them throughout. The AI Overview I describe was what I saw on August 28 and is personalized, so yours may differ. I sell an AI SEO product, which is linked in this post. Full disclosure.