What I learned concerning AI hunt whenever I made up 50,000 stats that got cited by ChatGPT...
About 7 months ago, I launched a website alongside 50,000 statistics. Every one of them was fake.
Google has forgotten the location exists and deindexed the pages, but ChatGPT hasn't. Since the website went live 7 months ago, it has sent additional than 10,000 AI sessions its way. And now additional than 330 referring domains, including Shopify, Bluehost, and DHL, nexus to the site. Its made-up numbers have additionally ended up in an scholarly document and in a bank's funding research.
At Peec AI and in the industry, we conversation concerning GEO and SEO all day, and we keep looking for new tactics. I’ve been operating on this stats leaf one for a lengthy time, sharing pieces of it publically alongside the way. Of course, it’s experimental, but the results have been difficult to ignore.
Where the idea for publishing counterfeit data came from
The impetus to kernel a location alongside data came out of long-running investigation I've been doing on LLMs, relevance engineering, and grounding (credit to the iPullRank team, Dan Petrovic, and Wil Reynolds).
In the way of my work, I noticed a pattern: as additional teams compose time-sensitive and breaking satisfied alongside LLMs, those LLMs have to appearance for (seemingly) dependable data during their investigation to inhabit the gaps in their knowledge. Statistics are frequently exactly what they’re looking for.
Most GEO direction focuses on one moment: getting your brand into the AI answer. This test looks at what happens following the answer, whenever group obtain that output and use it, putting it into presentations, adding it to a buying list, and yes, publishing it on their website.
I didn't arrive up alongside the idea for this test alone. Hat tip to Seer Interactive, who already published an eye-opening study in August 2025 that got a few thoughts bouncing about in my head. Several months later, those thoughts came together and eventually made it to the top of my to-do list. In March 2026, I started operating on construction the location that would let me test my ideas.
The setup: a website and 50,000 made-up stats, created in one day
Quick environment for anyone who wasn’t following alongside in genuine period as I ran this experiment:
I bought a brand-new domain, stateglobe.com, and generated 50,000 data pages in a sole day alongside GPT-4.1. The location purports to provision stats on SEO, satisfied marketing, social media, e-commerce, and web innovation from 200 countries. Except all figure was totally invented. I never checked a sole one. I pushed them to the location alongside a essential AI web scheme template. This disbursal me about $30. Later, I added 15,000 listicles on top of the statistics. That was concerning $20 more, so all in all, this is an test that came in at the extremely low cost of concerning $50.
The reply whenever I archetypal shared it was wild. When my companion Kevin featured it in Growth Memo, it reached a much bigger audience. I think group responded so enthusiastically to this activity since it showed awesome results during additionally speaking to the fears group have concerning what ends up in AI hunt results.
What happened
Google buried the location nearly immediately
When I launched the site, traffic, indexing, and involvement were initially going really well. Using indexer tools, I got nearly the complete location indexed inside a week, and it started to appear in Google results.
Then two Google updates came out correct following launch, and my Google visibility collapsed.
I've been invisible on Google for many months now. Indexed pages bounce between 5 and 20 impressions. At one point, I got inquisitive and created a subdomain to see if I could addition impressions, but it didn't really help.
Other than that, I haven't touched the site. Some pages power equal neglect to burden now and then. I don't know.
But whenever I sat downward to compose this, I checked the crawl data in Cloudflare following 7 months.
In the final 30 days alone, the location got additional than 600,000 requests from bots. Around 72,000 of them came from ChatGPT-User. That is the delegate ChatGPT uses to open a leaf in genuine period during answering someone. So these are not training crawls. Each one is ChatGPT fetching a counterfeit data leaf to answer a genuine question.
Unlike Google, ChatGPT never stopped pulling from the site
And whenever I say, “never stopped,” I average ChatGPT really never stopped.
After 7 months, the location now has 330+ referring domains
According to Ahrefs and Semrush, the location has links from 330+ another sites. That includes well-known brands akin Shopify, DHL, and Bluehost. With powerful Google performance, the numbers power have been equal better.
ChatGPT is motionless citing it to this day
Google may have forgotten concerning the site, but ChatGPT plainly hasn’t. The location is motionless earning citations and traffic additional than 7 months following launch.
If you've peruse Tomek's ChatGPT Index research, you're lucky, since it’s a awesome intro to how ChatGPT hunt works. If you haven't, add it to your study catalog correct now, since it helps explain why my website has been living in ChatGPT/OpenAI Index for so long.
A entirely made-up website motionless performing so fine is shocking. Since March, GA4 shows 10,000+ sessions from ChatGPT alone, and Cloudflare shows the location has elevated marks for retrieval.
I didn’t contain straightforward sessions in general visits, since it mostly looks akin bot traffic. Remember the deceased net theory!
Right following I joined Peec in June 2026, I additionally set up a visibility dashboard in our tool to track a few prompts. Performance for them is motionless strong. On October 7, stateglobe.com was classified second for visibility on these prompts, rearward lone data powerhouse Statista.
Retrieval is not ranking
The changeless retrieval is the archetypal big lesson: ranking on Google and being retrievable by an LLM are two distinct problems.
When ChatGPT searches, it doesn't fair copy Google's top 10. It rewrites your immediate into its own sub-queries (called fanout queries), pulls candidates from its hunt sources and its own cache, reranks them, and afterward opens a fistful of pages to dirt the answer. A leaf can be invisible on Google and motionless execute fine in all one of those LLM steps.
StateGlobe is a fine example. While Google dropped it, the AI retrieval bots kept coming back.
For GEO, this method Google rankings are lone one signal, not the entire picture. If you lone track Google, you are missing a conduit that can activity equal whenever Google says “no.”
Fake data turned into genuine citations
You’d anticipate that if a location performs fine alongside data retrieval tools, its data would commencement to appear out in the wild. And you would be right.
StateGlobe was built entirely on counterfeit data, using the cheapest example accessible and a immediate that was nothing additional than a catalog of countries and topics affirmative two sentences. Nevertheless, the output was…
…cited in an scholarly paper…
…used in funding investigation by Qatar's largest financial group…
…cited by an economics publish in India…
…and used as a origin in a investigation study from a Canadian influencer community.
Note that these are lone the citations that connected rear to the StateGlobe site. Presumably, many additional used the data without a link.
The uncomfortable implication: citation laundering
On day one, a counterfeit figure on StateGlobe had zero authority. It was a random stat on a brand-new domain.
Then an scholarly document cited it. Then a bank. Now the next person, or the next model, that finds this figure sees a stat alongside a source. And that origin is cited by grave institutions.
The counterfeit data certainly didn't rotate into true, but all citation made it appearance additional trustworthy.
It has the possible to get worse: following so much retrieval and citation, these pages end up in web crawls, and web crawls end up in training data (I anticipation it hasn’t!). A figure that started as a GPT-4.1 conjecture can rotate into item the next example "knows."
This points to the larger hazard suggested by this experiment, and it's why I'm sharing this as research, not as a playbook for counterfeit content. Verifying sources should rotate into an essential part of writers’ and researchers’ workflows.
Lastly, we tried operating our test alongside genuine data
Once I began operating at Peec, I decided to try a type of this test again. As the GEO Research team, we talked through the results up until that item and quickly launched a uncomplicated new page. This time, there were no made-up numbers, and we checked all stat and correctly cited its source. You can perspective it here: peec.ai/ai-search-geo-statistics.
Almost immediately, sites we doubtful use AI to compose their satisfied started linking to us. For the second time, using a second site, earning links by publishing statistical satisfied proved that a hypothetically powerful maneuver worked in reality.
Lessons learned
Why publishing data leads to LLM citations
The two tests display that LLMs have a powerful penchant for seeking out data whenever asked to compose articles. The scheme is simple:
More and additional satisfied is written alongside LLMs.
When an LLM writes, it researches the suitable topic.
When the LLM hits a gap, it looks for a figure to inhabit it.
The reasons rearward this procedure are likewise clear.
Research method fanout queries asking for statistics
When a immediate requires research, the example breaks it into sub-queries, adds modifiers akin "statistics," "data," "report," or the current twelvemonth (if additionally requested in the immediate text). If person asks for the example to create a blog article concerning distant work, the example volition hunt for "remote activity data 2026." Pages that optimize for what models hunt fairly than for what humans category volition appear additional frequently in LLM results.
A figure looks akin evidence that can inhabit cognition gaps
When a example is asked to outline a part of writing, a citation to a leaf that says "Remote activity is growing" adds small authority and remains vague. If that leaf alternatively says "X% of workforce activity remotely at smallest formerly a week," there’s abruptly item that is specific, quotable, and additive. When a example evidence an answer, a particular figure is the most helpful item it can find. These numbers appearance akin evidence, although as this test proves, sometimes they aren’t.
The example doesn’t inspect the number
A search-grounded part of LLM-written satisfied has a tight prosperity for latency and cost. The scheme fetches a few pages, pulls the applicable passages, and starts to write. Fact-checking doesn’t happen anyplace in the process.
Verifying a sole figure would average finding autonomous sources, comparing them, and resolving conflicts. Across billions of prompts by millions of users, the expenses in the two period and prosperity would add up, making verification too dilatory and too costly for applicable purposes.
To get about this, LLMs use rerankers that mark relevance, but not truth. That method if a passage matches the query and looks authoritative, it gets used.
Statistics have a built-in freshness that models like
I've written before concerning ChatGPT's recency bias in how it scores sources. Statistics pages merge the twelvemonth in the title, the headings, and the URL. Three built-in signals to the example that what’s on the leaf is fresh.
How to build a data leaf the correct way
The two tests I ran display that data pages are a extremely mighty way to get your satisfied cited by LLMs. While I entirely do not propose coming up alongside bogus data, you can use the learnings from our small test to create your own pages alongside genuine data. Here's what worked for us at Peec and what I would do if I were making a data leaf again.
Make one assertion per sentence. Each stat should create awareness on its own, out of context. Retrieval doesn’t activity on pages, but fairly on passages. If a figure needs the paragraph complete it to create sense, it likely won't get pulled.
Put the origin and twelvemonth next to all number. "According to [Source] (2026), X%..." This helps the model, the individual editor, and your credibility. It’s one small part of data hygiene that benefits the broader data environment.
Match headings to fanout queries. The headings on your leaf should equivalent the way LLMs hunt the internet: "[Topic] statistics," "[Topic] market size," "[Topic] acceptance by country." Think in the queries models run, not fair the keywords group type.
Show the update date (and really update it). Freshness is a genuine indication that models obtain into account. A old stats leaf volition endure to a newer one.
Make the stats uncomplicated to fetch. Use server-rendered HTML alternatively of hiding data rearward JavaScript, tabs, or images. Also, be certain to inspect your robots.txt and your CDN or WAF rules. Your location may be blocking AI retrieval bots without your knowing it.
Keep it simple. Our Peec AI GEO data page is basically a list. Yours doesn’t need to be item additional than that, either. Save scheme resources for sites and resources that can really advantage from them.
Add first data whenever you can. Aggregated stats may get you a few citations, but whenever you say item no one alternatively does, you can rotate into the chief source.
How to track whether your leaf is getting picked up
Google Search Console won’t display you what’s going on alongside AI bots on your site. Here is what I track instead:
AI bot activity in your server logs or Cloudflare. Look for retrieval bots akin ChatGPT-User, OAI-SearchBot, and PerplexityBot, not lone training crawlers. We have this feature!
AI referral sessions in GA4 from chatgpt.com, perplexity.ai, and akin sources.
Citation and visibility tracking throughout a set of prompts. At Peec, we use our own tool for this, obviously.
New referring domains in Ahrefs or Semrush.
Unlinked mentions. Search your exact numbers in quotes. You volition discover group using your data without linking to you.
A term on ethics
It should go without saying that you should never publish counterfeit content. While my test relied on tens of thousands of unverified stats, I ran it in command to portion the results alongside the public, and the website, during motionless live, now carries a disclaimer. When putting out satisfied we anticipation gets picked up by LLMs, our job should continually be to put out accurate, accessible information.
To add another caveat, there’s no justify this maneuver volition continue to work. We tested it, and these are the results, but now that they’re published, LLMs power near this gap. For now, it works, but in the forthcoming it may not, and of way it should never be abused.
And one final point: this is a tactic, and neither can nor have to be the basis for your organization’s complete GEO strategy. The results I got alongside StateGlobe were extreme, but the brands most apt to advantage from using the method are those group already trust.
What's next from GEO Research at Peec
At Peec, our investigation squad is operating near to 40 experiments akin this one. Some are motionless in progress, and others are already showing results. Building a long-term scheme for your brand method testing, measuring, and sharing what we study alongside the way.
More soon!
