I Scraped the Top 50 SEO Sites to Learn How to Get Cited by LLMs

Shane HellmrichFounder, PropeloSEO. Reviewed September 2026. Sources cited where a claim needs one.

What I did

I put 10 questions into Google, things like how to get cited by LLMs and how do LLMs choose sources to cite. That logged 427 results.

I took the 50 strongest domains ranking for those questions and had a script read every page. 45 came back readable, three blocked the request and two handed back almost nothing.

Then the script counted 16 different recommendations across all of them.

What they agree on

Schema won by a mile. That is the hidden code that tells a machine what is on your page, and 36 of the 45 tell you to add it.

After that it is off-site work. Backlinks on 32 pages, third-party mentions on 30, Reddit and review sites on 27 each.

The tidy formatting advice sits lower than you would expect. Answer-first on 22 pages, chunking on 12, FAQ headings on 8, and an llms.txt file on 7.

One thing to be straight about. The script counts that a page raises a tactic, not that it swears by it, so I read the close calls by hand.

• Add schema or structured data: 36 of 45
• Build backlinks and domain authority: 32 of 45
• Earn third-party mentions or digital PR: 30 of 45
• Format as lists and tables: 27 of 45
• Get into Reddit and community platforms: 27 of 45
• Get on review sites and listings: 27 of 45
• Publish original data or research: 23 of 45
• Show author credentials: 23 of 45
• Keep content fresh and dated: 23 of 45
• Put the answer first: 22 of 45
• Build a consistent entity: 17 of 45
• Let the AI crawlers in: 14 of 45
• Write in self-contained chunks: 12 of 45
• Target conversational queries: 11 of 45
• Use question headings and FAQs: 8 of 45
• Publish an llms.txt file: 7 of 45

The bit that stopped me

Seven pages bring up llms.txt. Only one of them tells you not to bother, and it is Google's own documentation.

Google's guide has a list of things you can ignore. It names llms.txt directly and says Google Search ignores the file.

That same list says there is no requirement to chunk your content into tiny pieces, while eleven other pages in the sample raise chunking as a tactic. It also says you do not need to rewrite content just for AI systems, which is most of the formatting advice gone in one line.

Two separate studies back Google up on llms.txt. SE Ranking checked around 300,000 domains and found their model performed better without it, and ALLMO looked at 94,614 cited URLs and found exactly one citation pointing at an llms.txt file.

What the count actually proves

Not much on its own, and that is the honest read. A tally tells you what people publish, not what works.

36 pages saying add schema is 36 people agreeing, not 36 tests. A separate experiment published deliberately invalid schema markup, with an invented type and zero valid Schema.org properties, and ChatGPT and Perplexity both returned the fake data anyway.

A schema-aware parser would have rejected it. The models are reading the text, not parsing the structure.

Strip out the advice that is contested and you are left with backlinks, brand mentions, community presence and reviews. All of that happens somewhere other than your website.

Most of what gets sold as AI search work is on-page work aimed at an off-site problem.

The four-false-proofs test

Before accepting that any tactic influences AI visibility, check whether the evidence would happen regardless of whether the claim were true. Four observations prove nothing, because they happen to any text file on the open web.

If a proposed tactic's evidence rests only on these four bars, it is conjecture. Price it accordingly.

• AI bots crawl it. Fetching things is a crawler's entire job.
• Google indexed it. Google has indexed text files for decades.
• An LLM returned a fact found only in that file. That is ordinary retrieval, not a special standard being honoured.
• ChatGPT says it helps. A model telling you something works is evidence that lots of text on the internet says it works.

What to do first

Rank for the question. The models search, so page one is still the door.

Check a bot can fetch the page. If it cannot read you, nothing else on this list matters.

Go get mentioned by other people. Podcasts, roundups, review sites, real threads where your buyer already is.

Publish something only you have, because your own numbers beat a rewritten summary of someone else's. Add schema because it is cheap and tidy, not because you think it is the lever.

Method

10 Google queries pulled 2 August 2026, United States, desktop, top 50 results each, with 427 organic results logged. 50 publisher domains selected by rank across queries, with forums, video and social excluded.

45 pages read and 5 blocked the request. Counts are per page, by pattern match, then hand-checked where the reading was close.