We Tested 10 AI SEO Myths on 100,000 Sites (Here's What Actually Works)

Right now, you might be doing dozens of things to rank in AI search that do absolutely nothing. We tested 10 of the most common AI visibility claims across 100,000 sites and about half are wrong. Just a quick example. Take that LLMs.txt file everyone keeps arguing about. Are we adding it or what? We checked 300,000 domains. Barely 10% of them use it and it has zero connection with you getting cited. However, this story isn't even fully settled because of a twist Google recently announced. We'll come back to that. Some of the other claims we looked at are things you're probably doing right now, thinking it helps, when the data actually says it doesn’t. We'll go through 10 popular claims, but stay for the last one, because it's the single biggest thing that decides whether AI even cites you. Now, back to that LLMs.txt story because it's a weird one. If you haven't seen it, LLMs.txt is a little text file you drop at the root of your site. The idea is

that it hands AI models a map of your best pages so they cite you more. Sounds smart. We tested it across 300,000 domains. Two things came back. First, almost nobody uses it. Only 10% of sites have one. Here's the part that should stop you. We broke it down by traffic, and the biggest, highest-traffic sites use it less than mid-tier ones. If this were a real edge, the big players would be all over it. They're not. We even built a machine learning model to predict which sites get cited. When we pulled LLMs.txt out of the model, the predictions got more accurate. The file was adding noise. It literally made the picture worse. In fact, Google put out its own guide on showing up in AI search, and LLMs.txt is right there on the list of things you don't need. Case closed, right? Except a few weeks later, Google's own Chrome tool, Lighthouse, added a check that looks for an LLMs.txt file. One part of Google checks for the file,

another part says skip it. The catch is that the Lighthouse check is about AI agents browsing your site, not about getting cited. Two completely different jobs. Now in case you already have LLMs.txt, that’s fine, leave it, it costs nothing. Just don't expect a single extra citation from it today, and definitely don't spend another minute optimizing it. Claim two. FAQ schema markup. As in add FAQ schema markup to your pages, and AI cites you more. True or false? Mostly false. Just look at the numbers first. In ChatGPT, pages with FAQ schema averaged at 3.6 citations. Pages without it sit at 4.2. In Google's AI Mode, we see roughly the same picture. This means that as a citation lever, the code does practically nothing. Here's the irony. Google has spent years walking away from FAQ snippets, and people keep pushing it as an AI tactic anyway. The visible payoff it used to earn those expandable FAQ dropdowns under

your search results got cut back to government and health sites in 2023, vanished from search completely in May, and on June 15th Google deleted the documentation for it. The code may still work and Google can still be reading it, but the reward people were chasing is gone, and it was never moving AI citations to begin with. The FAQ idea isn't worthless, though. It's the schema code that does little, not the FAQ itself. The real answers on the page are what get picked up. In AI Mode, pages with genuine FAQ sections landed 4.9 citations versus 4.4 without them. In ChatGPT it's more situational, FAQ sections there even score a touch lower on average, but mostly because they tend to live on thin support pages. On a strong page, they help. The takeaway? Treat FAQ schema as nice to have, not must have. On a page that already has strong content and real authority, adding clean markup can give a modest extra lift,

but it won't rescue a thin page, and on its own it won't get you cited. If you do want to use it, build dedicated FAQ pages around one clear topic, with full, real answers. That way, AI can map questions to solid responses. Claim three. Pack your main keyword into your URL and title, and AI cites you more. The oldest reflex in SEO. True or false? False. Pages with broad, topic-clear URLs averaged at 6.4 citations, keyword-stuffed ones at 2.7. Same with titles. More than double the gap, both times. Unfortunately, the more you optimize the way we were taught, the worse you do. AI doesn't read a URL like an old search engine counting keywords. A slug that describes the topic tells the model what the page is about. One stuffed with repeated keywords reads like spam, and it gets treated like spam. The fix costs nothing. Write your URLs and titles for a human, describe the topic in plain words, and stop. Less is more in this case.

Claim four. AI search is stable. Run the same prompt twice and you get the same sources back. True or false? Obviously false. The problem is that people try to apply traditional rank tracking to AI tools but it doesn't work like that. We ran the same local searches through Google's AI Mode over and over. On average, only 35% of the sites showed up again on a repeat run. Two-thirds got swapped out. Same prompt, different sources, minutes apart. Two things worth noting here. First, it's not pure chaos. Vague queries bounced around wildly, but when we named a specific place, consistency almost doubled. The more specific the prompt, the more stable the result. Second, even when the same sites came back, AI Mode often cited different pages on those sites. The domains repeat more than the exact URLs do. If you check your AI visibility once and write it down, you're recording a coin flip, not a result.

Prompt tracking is its own discipline. You can't treat it like keyword rankings. That's why, when we track AI results on our platform, we run every prompt several times and look at what actually holds. We've also got a full guide on our blog with a practical framework for finding, filtering, and prioritizing the prompts actually worth tracking. The link's in the description. That's four myths down. Now the other half, what actually works. Each one's true, that's not the interesting part. The interesting part is how much it moves the needle, and the catch that flips the advice you've probably heard. First up. ChatGPT sends people to dead pages more often than Google does. The why is a little unsettling, because ChatGPT is inventing URLs that were never real. We found it on our own site first. Back in March, our Head of SEO noticed ChatGPT was sending real traffic to pages that didn't exist. Not old pages that broke. URLs we never created.

That’s over three months and about 70 of these fabricated pages got traffic. Some pulled over 20 sessions. Around 20 of them had picked up backlinks. That’s other sites linking to pages that were never real. Then we checked the scale. Of the 145,000 URLs ChatGPT cited, 1.22% returned a 404. That’s higher than Google organic search, AI Overviews, and AI Mode. On paper, 1% looks tiny, but ChatGPT handles around 2.5 billion requests a day. That's still millions of broken experiences, every single day. The takeaway is simple. Basic technical SEO still matters. Keep your site clean of broken pages, handle your redirects properly, and keep your URLs consistent. AI traffic needs its own monitoring now. You want to track where that traffic actually lands, and whether those pages even exist. Claim six. Longer content gets cited more by AI. True, you probably knew that, but word count isn't really the point, and that trips people up.

The numbers show that in ChatGPT, articles over 2,900 words averaged at 5.1 citations. Pages under 800 words, at 3.2. AI Mode shows almost the same climb. The sweet spot starts at around 1,500 words, and you go longer when the topic actually needs it. However, length actually means depth here. ChatGPT rewards pages that cover a topic fully with all the context, the nuances and related questions. Padding to hit a number won't do it and two things back that up. In ChatGPT, pages with expert quotes get cited noticeably more than those without. Pages packed with real data, which is 19 or more stats — pull well ahead of thin ones. Now, here's the part that matters most for smaller sites. Length pulls about 65% more weight for small domains than for big ones. If you can't compete on backlinks or traffic yet,

going deep on a topic is one of the few levers that actually works in your favor. Again, covering it once isn't enough. A page that crushed it last year can quietly stop getting cited, and most people never notice why. Claim seven. Refreshing old content gets it cited more by AI. True and it's one of the biggest points on the whole list, but the catch is brutal, because most people do the version that doesn't work. The data says that pages updated in the last three months pull nearly double the citations of stale ones. AI Mode shows the same pattern. I want to point out that new isn't the same as fresh. Brand-new pages barely outperform years-old ones. Publishing more doesn't help, and neither does changing the date and calling it updated. The lift comes from a real refresh with new stats, new examples, a section that didn't exist before. Something that actually makes the page better than it was. The move is simple. Take your pages that already get cited, the ones that matter,

and genuinely update them every few months with new data, current examples, and anything that has changed since you wrote it. One warning though before you go reaching for the fastest way to refresh at scale. A lot of people are about to think, "I'll just have AI rewrite all my old posts." We actually tested that, published AI-generated content and tracked what happened. The result surprised us, and it's a whole video on its own. (POINT TO TOP LEFT) The link is here and in the description. Claim eight. Profiles on review sites such as Trustpilot, G2, and Capterra, get you cited more by AI. True and it’s more significant than you may think. Domains with profiles across multiple review platforms pull around three times the citations of sites without any. ChatGPT reads them as trust signals, but the real surprise is what happens the moment someone searches your brand by name. Ask AI a general question and review sites show up in about 7% of the answers. Now, ask it about a specific company and that jumps to 27%.

Four times more. In US data, Yelp was the most-cited site that wasn't somebody's own brand. When someone asks AI "is this company any good," part of that answer is coming straight off of your review profiles, whether you've claimed them or not. The verdict is that you must claim your profiles on the platforms that matter in your market, and get real reviews on them. Reviews are one place AI looks for outside proof. The other is the one everyone hypes — Reddit and Quora. The advice there is half right. Claim nine is that getting your brand mentioned on Reddit and Quora gets you cited more by AI. This is true. Go from barely mentioned to heavily discussed on Quora, and citations jump roughly four times. Reddit has the same story. AI treats these as digital word of mouth, real people talking about you, not you talking about yourself. Here's where the usual advice falls apart. Everyone says "go build a presence on Reddit and Quora."

That's right, until someone searches your brand by name. Then the platforms AI trusts completely flip. Reddit and Quora drop. Facebook, Instagram, and LinkedIn take over, with LinkedIn nearly doubling its share. If you've poured everything into Reddit and Quora and left your LinkedIn and Facebook half-dead, you've optimized for the general query and abandoned the one where someone's already asking about you by name. Why not do both? Earn the organic Reddit and Quora mentions for discovery, and keep your owned socials, LinkedIn especially, active and real, because that's what AI reaches for when the question is you. Keeping every social channel active and on-message is a real job, especially now that AI is reading all of them. We use Planable for this, and it does a great job at helping us plan and publish content across all socials so nothing goes stale or off-brand. The tenth and final claim is the big reveal I asked you to stick around for.

Your regular organic SEO, as in your Google rankings, is the single biggest thing driving the AI’s decision to cite you. We measured the relationship between a site's organic traffic and its AI traffic across the whole dataset. The correlation came back at 0.72. On a scale where 1 is a perfect match, 0.72 is huge. It was the strongest signal in the entire study. Everything else we've covered — the content, the freshness, the reviews — all of it matters, but nothing came close to good old SEO. Sites that do well in organic search get more AI traffic. Moreover, this relationship is getting stronger. Early last year, the correlation was weak at around 0.19, but it nearly tripled by the end of the year. We’re talking 12 months. In ChatGPT, the higher on average a page sits in Google, the more it gets cited. In AI Mode, pages with more than four keywords ranking in Google’s top 10,

get about 11% more citations than pages with weaker keyword coverage. On top of that, across the major AI platforms, a big share of what they cite is already ranking in Google's top 20 for that same search. Here are the three things you should definitely take away from this. One. Keep doing SEO. AI search isn't a new game with new rules. SEO isn't being replaced by it. SEO is how you get into it. The technical work, the content, the authority you've been building for years — that's the same thing powering your AI visibility right now. Two. Manage everything AI already owns about you. When someone asks AI about your brand, most of what it pulls comes from the sources you control. Your site, your Maps listing, your socials, your reviews. Your move is to get them current, active, and claimed. Three. Build your own way to measure this. Don't wait for your analytics to catch up. Ask new customers how they found you. The businesses that can see

their AI traffic right now are the ones who set up the feedback loop themselves. That's the playbook. If you want to go deeper on any of this, the videos on screen are your next step. Be sure to check them out. Now go optimize!