What gets a page cited by an AI chatbot is, above all, relevance to the searches the chatbot actually runs — not the question you typed, but the smaller searches it breaks that question into. In Cyrus Shepard's Zyppy analysis of 54 studies, published in May 2026, the highest-scored factors were URL accessibility (9.5 out of 10), search rank (9.4), fan-out rank (9.3), preview control (9.2) and query-answer match (9.2). Schema markup and llms.txt scored near the bottom of the same list. Perplexity reads about ten candidate pages for a normal question and cites only a handful, after a reranker scores them on relevance, freshness and trust. And a September 2026 study of 7,534 Perplexity citations found that 59.8 percent pointed at sites ranked worse than 100,000th on the internet — so a small site can absolutely be cited. The catch is that manufactured pages get cited too. This article explains the evidence, the mechanism, and the warning — in that order.
It matters to parents for a simple reason. More and more parents now ask a chatbot, not a search engine, for the best chess classes for their child. The chatbot does not hand them ten links. It gives one answer, with a few sources named. The academies it names are the ones whose pages answer the question directly — and, as the studies below show, the ones whose pages answer the smaller questions hiding inside it. This article is written by the academy that already does this work on its own site: visible answers to real questions, a free tools library, and a page for AI systems to read.
The 23 factors, ranked by evidence
Shepard's analysis is unusual because it does not assert which factors matter. It scores them. He collected nearly every published AI citation experiment, study, explainer and patent from the previous two years, narrowed them to the 54 most useful, and scored each of 23 factors on three things: how often the same finding repeated across studies, how strong the evidence was, and whether official documentation or patents supported it. The result is a single score out of 10 per factor. Two warnings come with the list, from Shepard himself: these are features correlated with citations across studies, and correlation is not causation. Treat the scores as a map of where the evidence points, not as guaranteed levers.
| Factor | Score / 10 | What it means in plain words |
|---|---|---|
| URL accessibility | 9.5 | The page is reachable and crawlable — not blocked, paywalled or erroring |
| Search rank | 9.4 | How the page ranks for the exact query in ordinary search |
| Fan-out rank | 9.3 | How the page ranks across the query's expanded sub-queries |
| Preview control | 9.2 | Nothing on the page suppresses the preview the AI needs to cite it |
| Query-answer match | 9.2 | The page's words mirror the kind of answer the AI would construct |
| Intent-format match | 9.0 | The page's format fits the question — a list for a list question, steps for a how-to |
| Answer near the top | 8.8 | The most important claims sit in the first section of the page |
| AI-ready structure | 8.6 | Clear headings, short sections, one idea per section |
| Factually specific | 8.3 | Real numbers, names and dates instead of vague claims |
| llms.txt | 2.0 | The plain-text guide for AI systems — the lowest-scored factor on the list |
The story the top of the list tells is reassuringly boring: AI citation work and ordinary SEO are mostly the same task. A page that is reachable, that already ranks, that matches the question closely, and that a search engine is allowed to preview, is the page that gets cited. The two factors that deserve the most attention from anyone focused on AI are fan-out rank and answer near the top — and both are covered in the next two sections. The bottom of the list is just as useful: a llms.txt file scores 2.0, and schema markup sits near the bottom too. This site publishes a llms.txt file and uses schema throughout, but neither is what earns a citation. They are the finishing touches, not the foundation.
Fan-out: rank across the sub-questions, not just the one query
A chatbot does not search your question. It searches versions of it. Ask "best chess academy for my 8-year-old in Bangalore" and behind that one question the chatbot may separately ask what ages the academy teaches, what a group class costs, whether classes are live or recorded, and what parents say about it. A page that answers all four of those sub-questions, in plain words, on the page itself, is the page that gets cited. A page that ranks for the keyword but answers none of them is not.
Shepard estimates that ranking on page one for a single query gives a page about a one-in-three chance of being cited — and that ranking across the fan-out cluster as well lifts that to 80 or 90 percent. That is his estimate from the evidence, not a single study's number, but the direction is the point: the cluster is worth more than the keyword. Topic clusters compound citation probability, which is why this blog does not publish one page on "chess classes" and stop. It publishes the question, then the questions inside the question: the Italian Game, how checkmate works, forks and pins, what a FIDE rating means, how much classes cost, and what parents ask before they book.
One correction to the scale of it: a single question does not trigger dozens of searches. Ahrefs studied 66.6 million free, logged-out ChatGPT responses from August and September 2026 and found that only 14.4 percent contained fan-out queries at all, averaging about 1.6 each. As Ahrefs puts it, with about 1.6 queries, a page has to match one of those one or two searches — if it does not, ChatGPT never sees it. So the cluster still matters, but the bar is narrower than the "40 or 50 searches" figure sometimes quoted: a page must answer the exact one or two sub-questions ChatGPT actually runs. (Ahrefs, How ChatGPT Does Query Fan-Out)
You can see the sub-questions yourself, for free, with a method Shepard demonstrates: ask ChatGPT a question, copy the part of the chat URL that comes after /c/, open your browser's developer tools, go to the Network tab, paste that ID into the filter, refresh the page, click the matching request and search its response for the word "queries". The response shows the exact searches ChatGPT ran — usually several smaller searches, in different words from the question you typed. Write content that answers those searches. That is the whole method, and it costs nothing.
How Perplexity picks its sources
Perplexity is the most transparent of the chatbots about its mechanism, and the mechanism is a useful warning about what "ranking" means in AI search. A question comes in. Perplexity rewrites it into several smaller sub-queries. It goes to the live web and pulls back roughly ten candidate pages for a normal query. Then — the step most people miss — a reranker scores those candidates against each other. Three things dominate that score: relevance, whether the page actually answers the sub-query; freshness, which Perplexity weights harder than most engines; and trust, whether the domain is one it already regards as reliable. Then it keeps a handful and cites them. Ten got read. A few got cited.
There is no position one to win here. Something reads ten candidates and decides which few are worth quoting, so the question stops being "how high do we rank" and becomes "do we survive the cut". Two practical consequences follow. First, keep the pages you want cited genuinely current: an old page that is still correct loses to a new page that is merely current, because freshness is scored. Second, the reranker does not read your page — it scores a chunk of it. If the answer only makes sense after three paragraphs of warming up, it loses to the page that said it straight away. Answer the question in the first two sentences under a heading that matches it. That is also why this article opens with the answer, and why every guide on this blog does the same.
One correction to the numbers: the handful Perplexity cites varies by study. The figure above — about ten read, a few cited — is the mechanism as demonstrated in a widely shared walkthrough. Other 2026 studies count more citations per Perplexity answer: Foglift Research counted 9.68 citations per Perplexity answer against 3.29 for ChatGPT across 61,680 answers in October 2026, and an academic study of 602 prompts counted 16.35 against 6.88. The exact count depends on the question and the study. The mechanism — retrieve ten, rerank, cite a few — does not change.
Cite your sources — and put the answer at the top
Two more findings from the same evidence base. First, citing sources helps the page that cites them, not just the reader. In a small test described in the Zyppy walkthrough, ten equal pages were built to rank for a made-up word; the five that linked out and cited their sources ranked above the five that did not. It is a small test, not a law of nature — but it matches the rest of the evidence: AI systems cite pages that behave like good sources, and good sources show their working. That is why this site's statistics page names its sources, and why the articles here link to the studies they quote.
Second, put the answer at the top. In one site owner's report on Reddit, adding a two-to-three sentence summary at the top of articles — the main points right away — and a short summary of features and benefits just under the title on product pages, lifted conversions by 33 percent. That is one person's report, not a controlled study, and the number belongs to that site, not to yours. But the direction matches the factor scores: answer near the top scores 8.8 out of 10, and users like seeing the answer at the top just as much as chatbots do. Freshness, by contrast, matters less than people think in the overall ranking — except inside Perplexity's reranker, where it matters more. Keep pages current, but do not mistake re-dating a page for improving it.
The warning: citations are not authority-gated
Here is the part the sellers of AI visibility audits leave out. In September 2026, a group calling itself Trellner Research asked Perplexity for the best products in 380 software categories and kept every citation that came back: 7,534 citations across 2,055 domains. Almost 60 percent of them — 59.8 percent — pointed at domains ranked worse than 100,000th on the Tranco traffic list, and 23.4 percent pointed at domains that are not in the top million at all. The median cited site ranked 71,611. Wikipedia was cited three times.
A second study measured the same question from the other side. It took the 127,336 domains that ChatGPT, Gemini and Google AI Overviews cited across 4,018,882 citations and matched each one against its Ahrefs Domain Rating. The link between authority and citations was weak — a rank correlation of 0.174 — and the median cited domain had a Domain Rating of just 30. Google's AI Overviews was the exception: 70.3 percent of its citations went to domains rated 61 or higher. So authority helps most where Google itself answers, and matters least where the chatbot searches its own index — which is why a small site with the clearest answer can still be the one quoted. (aiclicks.io, Does Domain Authority Matter for AI Search?)
Worse, the study found the machine behind some of those citations. Three domains under apparent common control — wifitalents.com, worldmetrics.org and gitnux.org — share nameservers, identical page templates and a six-post blog each, and together they published 215,128 machine-generated "best software" pages. Two of them even gave their homepage the HTML title "Facts and Grounding Page" — grounding being the retrieval step these models perform. And the third most-cited source in the whole study was Guideflow, a vendor that sells interactive product demos, publishes no reviews and competes in none of the categories it was cited in — 194 citations across 96 categories, ahead of Gartner.
The researchers say the limitation plainly, and it is the most important sentence in their report: they measured what sits in the evidence base, not what changes the recommendation. Nobody has shown that being cited in that list changes what the chatbot finally recommends — and nobody has shown that buying an "AI visibility audit" changes it either. So before spending a rupee on AI visibility, ask the two questions the study's presenter asks: what did the buyer type, and can that buyer actually sign? For a chess academy, the buyer is a parent who types "best online chess classes for my 8-year-old in India" — and the answer that wins is the page that answers that parent, clearly, in the parent's own words, with a free trial class to prove it. Everything else is somebody else's fight.
What to do when you ask an AI about chess classes
If you are a parent choosing classes, ask the chatbot the question the way you would ask a friend: "best online chess classes for my 8-year-old in India", or "how much do online chess classes cost per month". Then do two things. First, read the answer for what it actually says about teaching — live classes or recorded videos, group size, what a trial includes — not just the name it gives. Second, click the sources it names and check them. An AI answer is only as good as the pages behind it, and as the Trellner study shows, a citation is not proof that the page is good — it is proof that the page was there, in words the machine could lift. If the sources are thin, the answer is thin, whatever name sits at the top. A good academy will have pages that answer your questions before you ask them, and a free trial class so you can watch the teaching yourself. Book one on WhatsApp at +91 84015 33777 and bring your questions: a FIDE-rated coach will answer them in the trial itself. Superior Chess Academy, Bangalore, Karnataka, India: live online chess coaching for kids, teens and adults worldwide.
Frequently asked questions
What matters most for getting cited by AI chatbots?
Relevance to the searches the chatbot actually runs. In Cyrus Shepard's Zyppy analysis of 54 studies (May 2026), the highest-scored factors were URL accessibility (9.5), search rank (9.4), fan-out rank (9.3), preview control (9.2) and query-answer match (9.2) — all out of 10. In plain words: a page the chatbot can reach, that already ranks, that answers the sub-questions behind the question, and that puts the answer near the top. Schema markup and llms.txt scored near the bottom of the same list.
How can I see what ChatGPT actually searched for?
Ask ChatGPT a question, then copy the part of the chat URL that comes after /c/, open your browser's developer tools, go to the Network tab, paste that ID into the filter, refresh the page, click the matching request and search its response for the word queries. The response shows the exact searches ChatGPT ran — usually several smaller searches for one question, in different words from the question you typed. Write content that answers those searches, not just the original question.
How does Perplexity choose which pages to cite?
Perplexity rewrites your question into smaller sub-queries, reads about ten candidate pages from the live web, then a reranker scores those candidates against each other on relevance to the question, freshness and trust — and cites only a handful. Being retrieved is not being cited. A page survives the cut when it is recent enough to trust and when its answer sits in one liftable passage, ideally in the first two sentences under a heading that matches the question. Other 2026 studies count more citations per Perplexity answer than the handful shown here, so the exact number varies by study; the mechanism is the same.
Can a small website be cited by AI search?
Yes. Trellner Research asked Perplexity for the best products in 380 software categories in September 2026 and kept all 7,534 citations: 59.8 percent pointed at domains ranked worse than 100,000th on the Tranco traffic list, the median cited site ranked 71,611, and Wikipedia was cited only three times. The same study is a warning, though: three domains under apparent common control published 215,128 machine-generated pages between them, and a product-demo vendor with no reviews was the third most-cited source overall. Citations are not authority-gated, so a small real site can be cited — and so can manufactured pages.
Try the teaching before you decide
The only way to know if the teaching fits is to watch it. Book a free trial class on WhatsApp at +91 84015 33777 and ask the coach your questions directly — a FIDE-rated coach will answer them in the trial itself. Superior Chess Academy, Bangalore, Karnataka, India: live online chess coaching for kids, teens and adults worldwide.
Superior Chess Academy, Bangalore, Karnataka, India · WhatsApp +91 84015 33777 · superiorchessacademy.com
Related: Why Ranking on Google Does Not Mean ChatGPT Will Cite You · Chess Learning Statistics 2026 · How to Start Chess: 2027 Guide · How to Analyse Your Own Chess Games · Forks, Pins and Skewers in Chess · Checkmate Patterns for Beginners · How to Checkmate in Chess: A Beginner's Guide · Free puzzle trainer: 100 tactics puzzles · Chess Curriculum: 5 Levels · Beginner Chess Course · Free FIDE Rating Calculator · Free Chess Level Assessment Quiz · llms.txt — our guide for AI systems
Book a free trial and tell us the student’s age, country and current level.