# AI citations: Earn mentions and answer follow-up questions - Episode: P3E03 (Part 3: AI SEO) - YouTube: https://www.youtube.com/watch?v=VwScl68RIas - Page: https://irongoo.com/videos/ai-citations - Duration: 14:52 - Published: 2026-10-22T09:00:00+05:00 Learn which SEO practices have evidence behind them for AI citations, and which popular tactics showed little measurable benefit. See why different AI engines pick different sources, and how to decide where your own site needs work. ## What this episode covers - How crawl access and readable pages make your content eligible for retrieval. - Why rankings for related searches can matter beyond your main target keyword. - What studies suggest about genuine brand mentions, earned coverage and useful updates. - Why llms.txt, schema markup and content rewriting tricks have weak evidence as citation tactics. - Where correlations, limited tests and changing models leave room for uncertainty. - How to audit the questions your buyers ask, compare cited sources and choose your next SEO priorities. ## Key moments - [00:00 Intro](https://irongoo.com/videos/ai-citations?t=0) - [00:59 How each engine picks sources](https://irongoo.com/videos/ai-citations?t=59) - [03:35 Does ranking still matter?](https://irongoo.com/videos/ai-citations?t=215) - [05:52 What goes with being cited](https://irongoo.com/videos/ai-citations?t=352) - [09:08 Claim 2: llms.txt](https://irongoo.com/videos/ai-citations?t=548) - [10:13 Claim 3: Schema gets you cited](https://irongoo.com/videos/ai-citations?t=613) - [11:12 Claim 4: GEO tricks](https://irongoo.com/videos/ai-citations?t=672) - [12:47 Your citation audit and wrap up](https://irongoo.com/videos/ai-citations?t=767) ## Transcript ### Intro We asked two AI assistants the same ten questions about finding films to watch. Both searched the web, and both cited their sources. On six of the ten questions, they didn't cite a single website in common. So when someone sells you the one trick that gets you cited by AI, the first question is: which AI? Every engine runs its own retrieval and picks sources its own way. A few levers seem to help across all of them. Several of the most popular ones measured at roughly zero. Today we sort them, with a verdict for each, and everything is as of October 2026, because this area moves monthly. Why care about citations at all, if AI answers send so few clicks? Because when an AI answer shows up, being cited in it beats being left out, and episode one has the numbers. The real question is how you get picked. ### How each engine picks sources Start with Google. AI Overviews and AI Mode answer from Google's own index. Google says both may use query fan-out: the model runs a bunch of related searches across subtopics, then writes one answer from what came back. The two features may also use different models, so the links can differ for the same question. To be eligible as a supporting link, Google says a page must be indexed and able to show with a snippet. That is the whole list. In Google's words, there are no additional technical requirements. Remember that sentence, because two of today's myths walk straight into it. Claude has its own crawlers, including a search crawler called Claude-SearchBot, and Anthropic lists Brave Search among the companies that handle its web search. ChatGPT search runs its own crawler, OAI-SearchBot, and OpenAI says it sometimes sends rewritten questions to outside search partners it doesn't name. You'll often read that ChatGPT is just Bing. OpenAI has never said that, so treat it as a guess. Perplexity runs its own crawler and index. Copilot is Microsoft's, and Microsoft now reports how often Copilot cites your pages inside Bing Webmaster Tools, which we use next episode. Five engines, five ways in. Back to our test. We wrote ten questions ReelRadar's readers might ask, from which streaming service has a film to how long films take to reach streaming after the cinema. Each went once to Claude and once to OpenAI's model, both with web search on, on the nineteenth of October 2026. Claude sometimes answered from memory without searching, so we told it to search first. Claude's ten answers cited forty five sources: a third product or company sites, a third news and editorial pages, about a quarter small blogs. OpenAI's model cited twenty nine sources, and twenty eight of them were the official sites of products it named. Same questions, two very different ideas of what a good source looks like. One run is a snapshot, and the next episode shows how much answers wobble between runs. The direction still matches the big studies. You are not optimising for one machine, so any tactic that only fits one of them is a weak bet. One thing every engine shares: it can only cite what its crawler is allowed to read. Block OAI-SearchBot and you're out of ChatGPT's search answers. Anthropic's search crawler follows your robots.txt too. And if your answer only appears after JavaScript runs, don't count on any AI crawler outside Google seeing it. Episode two has the full check. ### Does ranking still matter? So does ranking on Google still matter for AI answers? Yes, but the link between page one and the AI answer has loosened. In March 2026, Ahrefs found that thirty eight percent of the pages cited in AI Overviews also ranked in the top ten. Their mid 2025 study had put it at about three quarters. Read the small print before you panic. Ahrefs says part of that drop comes from its own better detection of citations, and Google moved AI Overviews to a newer model between the two studies. The real change is smaller than it looks, and nobody knows exactly by how much. The more useful finding is about fan-out. Surfer looked at ten thousand keywords and the extra searches Google's model runs for each one. Pages that also ranked for those fan-out searches were one hundred and sixty one percent more likely to be cited than pages that ranked only for the main query. It's a correlation, and it fits how Google describes the system. In practice that means covering the next questions. A ReelRadar page about a new thriller shouldn't stop at where to stream it. Is it on a free tier? Does it need a rental? When does it leave the service? Is there a subtitled version? Each one is a search the model may run, so answer each near the top of its own section, in plain sentences. Where do you find those next questions without guessing? Start with your own Search Console: the longer queries a page already gets impressions for are the sub-questions people actually type. Then ask Claude to list the follow-ups a reader would have, and keep only the ones your data or your support inbox confirms. The model suggests, the data decides. Ranking also wins on the research side. The NeurIPS benchmark we'll meet in a few minutes found that improving a page's real rank in the results the model reads beat every rewriting trick they tested. So the old work still counts: crawlable pages, a clear answer, and being the result worth reading. Quick scorecard before the next chapter. Each engine picks sources its own way. Being crawlable and readable is the entry ticket. And ranking still helps, especially for the follow-up searches, just less automatically than a top ten spot used to promise. Now for the part people actually argue about. ### What goes with being cited Now the big one: what goes with being named in AI answers? The largest study we trust here is Ahrefs, December 2025: seventy five thousand brands, checked against how often ChatGPT, AI Mode and AI Overviews mention them. The strongest factor was YouTube mentions. Branded mentions across the web came a close second. Now look how far down Domain Rating sits, the link score a lot of people treat as the master lever. Links still matter for classic ranking, and ranking still feeds these systems. For being named in an AI answer, though, people talking about you by name tracked far more closely than any link metric. Who does the talking matters too. Muck Rack looked at more than twenty five million links cited by ChatGPT, Claude and Gemini. About eighty four percent came from earned media: journalism, research, government and reference sites. Paid and advertorial content was a rounding error. A University of Toronto paper found the same tilt, plus a head start for big brands. So what does getting talked about look like for a small site like ReelRadar? Give people something worth citing by name. A monthly count of which films left which streaming service is original data that writers and forums can quote. Videos count, given that YouTube mentions led the correlations. None of it can be bought by the link, which is exactly why it carries weight. What doesn't count is manufactured talk. Paying for a wave of forum posts that all happen to mention you, or seeding your name into comment sections, is the kind of thing Google's guide calls inauthentic, and it says chasing it isn't as helpful as it might seem. Real coverage is slower and much harder to fake, which is the whole reason it works. Then freshness. Ahrefs compared about seventeen million cited URLs with ordinary search results. Pages cited by AI assistants were about a quarter fresher on average, and ChatGPT leaned newest of all. Freshness isn't a date you can fake. Bumping the modified date without changing the page tells readers and machines something false. Update when something real has changed: a film moved to a new service, a price changed, a season was added. Then the date is true and the page is genuinely better. For ReelRadar, real freshness is a monthly sweep of the pages whose facts move: where-to-watch pages, rental prices, leaving-soon lists. Pages that rarely change, like a director's filmography, get touched when something actually happens. Put the sweep on a calendar and the dates stay honest on their own. First verdict, then. Being talked about by real people and real publications has the strongest evidence of anything in this episode. We mark it works, with one honest label: it's correlation, and big brands are both talked about more and cited more. Google adds its own warning in its 2026 guide: chasing inauthentic mentions isn't as helpful as it might seem. One habit for reading every study in this chapter. They only see who got cited, never the sites that did the same things and got ignored. And most of them come from companies that sell AI visibility tools. We still use them, named and dated, because they're the best data available, and then we check them against our own. ### Claim 2: llms.txt Claim number two. Add an llms.txt file and AI engines will read it and cite you. The idea is a markdown file at the root of your site that lists your best pages for language models. It's been sold as the robots.txt of AI, which should already make you suspicious. Ahrefs checked server logs across one hundred and thirty seven thousand websites for May 2026. About thirty eight thousand had an llms.txt file, and ninety seven percent of those files got zero requests all month. Not low. Zero. Google's 2026 guide adds that you don't need AI text files or markdown to appear in Google Search. There is one honest use. Among the AI agents that did fetch these files, the second most frequent was Claude Code, a coding tool. That's what the file is actually good for: documentation that developers' coding agents read. If you publish software docs, add one. If you run a film site like ReelRadar, spend the afternoon on something else. ### Claim 3: Schema gets you cited Claim number three. Schema markup gets you cited by AI. It sounds sensible, because structured data describes your page in a format machines read easily. This is also the one most agencies are currently billing for. Ahrefs tested it properly in May 2026. They took about nineteen hundred pages that added JSON-LD and compared them with four thousand similar pages that didn't, thirty days before and after. In ChatGPT and in AI Mode, citations moved by about two percent, which is indistinguishable from zero. In AI Overviews they actually fell a little, and the authors don't blame schema for that. Two caveats, in fairness. The test covered pages AI already cited a lot, so it says less about pages nobody cites yet. And Google says the same thing from its side: there's no special schema markup you need for its AI features. Keep schema for rich results and for describing your entities clearly. Just don't count it as a citation lever. ### Claim 4: GEO tricks Claim number four. GEO tricks lift your AI visibility by forty percent. GEO stands for generative engine optimisation, and the forty percent comes from a real research paper, which is why this one has legs. The paper was accepted at KDD in 2024. The authors rewrote pages in a lab setup by citing sources, adding quotations and adding statistics, and saw visibility gains of up to forty percent in generated answers. Keyword stuffing, for the record, scored below doing nothing at all. Some things never change. Then in 2025 a benchmark called C-SEO Bench, accepted at NeurIPS, tested these methods more broadly. Most were largely ineffective, and some hurt. Improving a page's real rank did far better. And as more sites used the same tricks, the gains shrank. When everyone adds a statistic to every paragraph, nobody stands out. So the forty percent is contested in the lab and a placebo in the wild. Citing real sources and using real numbers is still good writing, and good writing gets read, so do it for the reader. Google's guide says the same about hacks: many of them aren't effective, and there's no need to chop your content into tiny chunks for AI. Google sums up its position in one line, and it's worth reading as written: optimizing for generative AI search is optimizing for the search experience, and thus still SEO. You can disagree with Google about plenty. On this one, the independent studies in this episode point the same way. ### Your citation audit and wrap up Here is how to check all of this in your own niche, and it's what the free audit sheet does. Pick ten questions your buyers really ask, in their own words. Run each through the engines that matter to you and log every cited source. For each source, note what kind of site it is, whether it ranks for that question in Google, and whether the brand behind it gets talked about elsewhere. You don't need a tool subscription to run it. We ran our test from Claude Code: a short script sends each question to each engine with web search on, saves every answer, and writes the cited domains into one table. Then a person fills in the type column, because deciding whether a site is a blog or a publisher is a judgement call, not a regex. Then look for the pattern. In our ReelRadar test, the cited sources were mostly the services everyone already talks about, plus editorial pages that answered one specific question well. Those are the two ways in for a small site: get talked about by name, or become the best answer to a specific question the model goes looking for. The type column tells you where your work is. If the cited sources in your niche are mostly news and reviews, your job is to become worth writing about. If they're pages that answer one narrow question, your job is on your own pages: answer it better and earlier. Most niches need some of both, and the sheet shows you the mix. The episode on one screen. Being talked about works, as far as correlation can tell us. A top ten ranking helps, but less directly than it used to. And the three tactics sold hardest as citation levers, llms.txt, schema and GEO tricks, all came out as placebo for citations. What will change: which sources each engine prefers, and every number in this episode, probably within months. What won't: be worth talking about, and answer the next question before your reader has to ask it. The citation audit sheet is free on the episode page at irongoo.com. Next episode, we measure all of this, layer by layer.