Publish stats, get cited: The tactic that put my information everywhere in the net

What I realized about AI search after I made up 50,000 stats that received cited by ChatGPT…

About 7 months in the past, I launched a web site with 50,000 statistics. Every considered one of them was pretend.

Google has forgotten the positioning exists and deindexed the pages, however ChatGPT hasn’t. Since the web site went reside 7 months in the past, it has despatched greater than 10,000 AI classes its method. And now greater than 330 referring domains, together with Shopify, Bluehost, and DHL, hyperlink to the positioning. Its made-up numbers have additionally ended up in an instructional paper and in a financial institution’s asset placement analysis.

At Peec AI and within the business, we discuss GEO and search engine marketing daily, and we maintain searching for new ways. I’ve been engaged on this stats web page one for a very long time, sharing items of it publicly alongside the best way. Of course, it’s experimental, however the outcomes have been onerous to disregard. 

Where the concept for publishing pretend statistics got here from

The impetus to seed a website with statistics got here out of long-running analysis I’ve been doing on LLMs, relevance engineering, and grounding (credit score to the iPullRank workforce, Dan Petrovic, and Wil Reynolds).

In the course of my work, I observed a sample: as extra groups write time-sensitive and breaking content material with LLMs, these LLMs must search for (seemingly) dependable information throughout their analysis to fill the gaps of their data. Statistics are sometimes precisely what they’re searching for.

Most GEO recommendation focuses on one second: getting your model into the AI reply. This experiment seems at what occurs after the reply, when folks take that output and use it, placing it into shows, including it to a procuring record, and sure, publishing it on their web site.

I did not provide you with the concept for this experiment alone. Hat tip to Seer Interactive, who already revealed an eye-opening study in August 2025 that received just a few ideas bouncing round in my head. Several months later, these ideas got here collectively and at last made it to the highest of my to-do record. In March 2026, I began engaged on constructing the positioning that may let me take a look at my concepts.

The setup: a web site and 50,000 made-up stats, created in in the future

Quick context for anybody who wasn’t following alongside in actual time as I ran this experiment:

I purchased a brand-new area, stateglobe.com, and generated 50,000 statistics pages in a single day with GPT-4.1. The website purports to offer stats on search engine marketing, content material advertising and marketing, social media, e-commerce, and net know-how from 200 international locations. Except each quantity was completely invented. I by no means checked a single one. I pushed them to the positioning with a primary AI net design template. This value me round $30. Later, I added 15,000 listicles on prime of the statistics. That was about $20 extra, so all in all, that is an experiment that got here in on the very low value of about $50.

The response after I first shared it was wild. When my buddy Kevin featured it in Growth Memo, it reached a a lot greater viewers. I believe folks responded so enthusiastically to this work as a result of it confirmed nice outcomes whereas additionally talking to the fears folks have about what leads to AI search outcomes.

What occurred

1. Google buried the positioning nearly instantly

When I launched the positioning, site visitors, indexing, and engagement had been initially going rather well. Using indexer instruments, I received nearly all the website listed inside per week, and it began to seem in Google outcomes.

Then two Google updates got here out proper after launch, and my Google visibility collapsed.

I’ve been invisible on Google for a lot of months now. Indexed pages bounce between 5 and 20 impressions. At one level, I received curious and created a subdomain to see if I may improve impressions, nevertheless it did not actually assist.

Other than that, I have never touched the positioning. Some pages would possibly even fail to load from time to time. I do not know.

But after I sat down to put in writing this, I checked the crawl information in Cloudflare after 7 months.

In the final 30 days alone, the positioning received greater than 600,000 requests from bots. Around 72,000 of them got here from ChatGPT-User. That is the agent ChatGPT makes use of to open a web page in actual time whereas answering somebody. So these usually are not coaching crawls. Each one is ChatGPT fetching a pretend statistics web page to reply an actual query.

2. Unlike Google, ChatGPT by no means stopped pulling from the positioning

And after I say, “by no means stopped,” I imply ChatGPT actually by no means stopped.

After 7 months, the positioning now has 330+ referring domains

According to Ahrefs and Semrush, the positioning has hyperlinks from 330+ different websites. That contains well-known manufacturers like Shopify, DHL, and Bluehost. With sturdy Google efficiency, the numbers might need been even higher.

ChatGPT remains to be citing it to this present day

Google could have forgotten in regards to the website, however ChatGPT clearly hasn’t. The website remains to be incomes citations and site visitors greater than 7 months after launch.

If you have learn Tomek’s ChatGPT Index research, you are fortunate, as a result of it’s an amazing intro to how ChatGPT search works. If you have not, add it to your studying record proper now, as a result of it helps clarify why my web site has been dwelling in ChatGPT/OpenAI Index for therefore lengthy. 

A very made-up web site nonetheless performing so properly is surprising. Since March, GA4 reveals 10,000+ classes from ChatGPT alone, and Cloudflare reveals the positioning has excessive marks for retrieval.

I didn’t embrace direct classes in general visits, as a result of it principally seems like bot site visitors. Remember the useless web principle!

Right after I joined Peec in June 2026, I additionally arrange a visibility dashboard in our device to trace just a few prompts. Performance for them remains to be sturdy. On October 7, stateglobe.com was ranked second for visibility on these prompts, behind solely information powerhouse Statista.

Retrieval is just not rating

The fixed retrieval is the primary massive lesson: rating on Google and being retrievable by an LLM are two completely different issues.

When ChatGPT searches, it would not simply copy Google’s prime 10. It rewrites your immediate into its personal sub-queries (known as fanout queries), pulls candidates from its search sources and its personal cache, reranks them, after which opens a handful of pages to floor the reply. A web page will be invisible on Google and nonetheless carry out properly in each a kind of LLM steps.

StateGlobe is an efficient instance. While Google dropped it, the AI retrieval bots stored coming again.

For GEO, this implies Google rankings are just one sign, not the entire image. If you solely monitor Google, you might be lacking a channel that may work even when Google says “no.”

3. Fake information was actual citations

You’d anticipate that if a website performs properly with info retrieval instruments, its info would start to seem out within the wild. And you’ll be proper. 

StateGlobe was constructed completely on pretend information, utilizing the most cost effective mannequin out there and a immediate that was nothing greater than a listing of nations and matters plus two sentences. Nevertheless, the output was…

…cited in an instructional paper…

…utilized in asset placement analysis by Qatar’s largest monetary group…

…cited by an economics publication in India…

…and used as a supply in a analysis report from a Canadian influencer neighborhood.

Note that these are solely the citations that linked again to the StateGlobe website. Presumably, many extra used the info with out a hyperlink.

The uncomfortable implication: quotation laundering

On day one, a pretend quantity on StateGlobe had zero authority. It was a random stat on a brand-new area.

Then an instructional paper cited it. Then a financial institution. Now the subsequent individual, or the subsequent mannequin, that finds this quantity sees a stat with a supply. And that supply is cited by critical establishments.

The pretend information actually did not develop into true, however each quotation made it look extra reliable.

It has the potential to worsen: after a lot retrieval and quotation, these pages find yourself in net crawls, and net crawls find yourself in coaching information (I hope it hasn’t!). A quantity that began as a GPT-4.1 guess can develop into one thing the subsequent mannequin “is aware of.”

This factors to the bigger danger steered by this experiment, and it is why I’m sharing this as analysis, not as a playbook for pretend content material. Verifying sources ought to develop into a vital a part of writers’ and researchers’ workflows.

4. Lastly, we tried operating our experiment with actual information

Once I started working at Peec, I made a decision to attempt a model of this take a look at once more. As the GEO Research workforce, we talked by means of the outcomes up till that time and rapidly launched a easy new web page. This time, there have been no made-up numbers, and we checked each stat and correctly cited its supply. You can view it right here: peec.ai/ai-search-geo-statistics.

Almost instantly, websites we suspect use AI to put in writing their content material began linking to us. For the second time, utilizing a second website, incomes hyperlinks by publishing statistical content material proved {that a} hypothetically sturdy tactic labored in actuality.

Lessons realized

Why publishing statistics results in LLM citations

The two checks present that LLMs have a robust desire for searching for out statistics when requested to put in writing articles. The mechanism is straightforward:

1. More and extra content material is written with LLMs.

2. When an LLM writes, it researches the suitable matter.

3. When the LLM hits a niche, it seems for a quantity to fill it.

The causes behind this course of are equally clear.

Research means fanout queries asking for statistics

When a immediate requires analysis, the mannequin breaks it into sub-queries, provides modifiers like “statistics,” “information,” “report,” or the present 12 months (if additionally requested within the immediate textual content). If somebody asks for the mannequin to compose a weblog publish about distant work, the mannequin will seek for “distant work statistics 2026.” Pages that optimize for what fashions search reasonably than for what people kind will seem extra usually in LLM outcomes.

A quantity seems like proof that may fill data gaps

When a mannequin is requested to draft an article, a quotation to a web page that claims “Remote work is rising” provides little authority and stays obscure. If that web page as an alternative says “X% of workers work remotely no less than as soon as per week,” there’s out of the blue one thing that’s particular, quotable, and additive. When a mannequin grounds a solution, a particular quantity is essentially the most helpful factor it could discover. These numbers seem like proof, although as this experiment proves, generally they aren’t.

The mannequin doesn’t verify the quantity

A search-grounded piece of LLM-written content material has a good price range for latency and value. The system fetches just a few pages, pulls the related passages, and begins to put in writing. Fact-checking doesn’t occur anyplace within the course of.

Verifying a single statistic would imply discovering unbiased sources, evaluating them, and resolving conflicts. Across billions of prompts by thousands and thousands of customers, the prices in each money and time would add up, making verification too gradual and too costly for sensible functions.

To get round this, LLMs use rerankers that rating relevance, however not reality. That means if a passage matches the question and appears authoritative, it will get used.

Statistics have a built-in freshness that fashions like

I’ve written earlier than about ChatGPT’s recency bias in the way it scores sources. Statistics pages include the 12 months within the title, the headings, and the URL. Three built-in alerts to the mannequin that what’s on the web page is contemporary.

How to construct a statistics web page the proper method

The two checks I ran present that statistics pages are a really highly effective technique to get your content material cited by LLMs. While I completely don’t suggest developing with bogus information, you need to use the learnings from our small take a look at to create your personal pages with actual information. Here’s what labored for us at Peec and what I might do if I had been making a statistics web page once more.

• Make one declare per sentence. Each stat ought to make sense by itself, out of context. Retrieval doesn’t work on pages, however reasonably on passages. If a quantity wants the paragraph above it to make sense, it most likely will not get pulled.

• Put the supply and 12 months subsequent to each quantity. “According to [Source] (2026), X%…” This helps the mannequin, the human editor, and your credibility. It’s one small piece of data hygiene that advantages the broader info surroundings.

• Match headings to fanout queries. The headings in your web page ought to match the best way LLMs search the web: “[Topic] statistics,” “[Topic] market dimension,” “[Topic] adoption by nation.” Think within the queries fashions run, not simply the key phrases folks kind.

• Show the replace date (and truly replace it). Freshness is an actual sign that fashions bear in mind. A stale stats web page will lose to a more moderen one.

• Make the stats straightforward to fetch. Use server-rendered HTML as an alternative of hiding information behind JavaScript, tabs, or photographs. Also, you’ll want to verify your robots.txt and your CDN or WAF guidelines. Your website could also be blocking AI retrieval bots with out your figuring out it.

• Keep it easy. Our Peec AI GEO statistics page is principally a listing. Yours doesn’t must be something greater than that, both. Save design sources for websites and property that may truly profit from them.

• Add unique information when you possibly can. Aggregated stats could get you just a few citations, however whenever you say one thing nobody else does, you possibly can develop into the first supply.

How to trace whether or not your web page is getting picked up

Google Search Console gained’t present you what’s occurring with AI bots in your website. Here is what I monitor as an alternative:

• AI bot exercise in your server logs or Cloudflare. Look for retrieval bots like ChatGPT-User, OAI-SearchBot, and PerplexityBot, not solely coaching crawlers. We have this characteristic!

• AI referral classes in GA4 from chatgpt.com, perplexity.ai, and related sources.

• Citation and visibility monitoring throughout a set of prompts. At Peec, we use our personal device for this, clearly.

• New referring domains in Ahrefs or Semrush.

• Unlinked mentions. Search your precise numbers in quotes. You will discover folks utilizing your information with out linking to you.

A phrase on ethics

It ought to go with out saying that you must by no means publish pretend content material. While my experiment relied on tens of 1000’s of unverified stats, I ran it to be able to share the outcomes with the general public, and the web site, whereas nonetheless reside, now carries a disclaimer. When placing out content material we hope will get picked up by LLMs, our job ought to all the time be to place out correct, accessible info.

To add one other caveat, there’s no assure this tactic will proceed to work. We examined it, and these are the outcomes, however now that they’re revealed, LLMs would possibly shut this hole. For now, it really works, however sooner or later it might not, and naturally it ought to by no means be abused.

And one ultimate level: it is a tactic, and neither can nor must be the premise on your group’s complete GEO technique. The outcomes I received with StateGlobe had been excessive, however the manufacturers most certainly to learn from utilizing the method are these folks already belief.

What’s subsequent from GEO Research at Peec

At Peec, our analysis workforce is operating near 40 experiments like this one. Some are nonetheless in progress, and others are already displaying outcomes. Building a long-term technique on your model means testing, measuring, and sharing what we be taught alongside the best way.

More quickly!