The new link-building: fake stats, laundered into LLM answers

What happened

A growth marketer named Jon posted a frank writeup on the Peec AI blog describing a tactic he calls "publish stats, get cited." The method is embarrassingly simple: he wrote a blog post full of specific-sounding statistics about his industry — percentages, adoption rates, year-over-year growth figures — and made no attempt to source them because there was nothing to source. He made them up.

He then did standard outreach: pitched the "study" to SEO-focused bloggers and roundup writers who are perpetually hungry for a quotable number to pad a listicle. Within weeks, his fabricated figures were appearing on third-party sites, each one dutifully linking back to his original post as "the source." Once a stat has been repeated across a few dozen domains, it stops looking like one person's claim and starts looking like consensus — which is exactly how retrieval-augmented LLMs decide what is true.

The payoff came when he started asking ChatGPT, Perplexity, and Google's AI Overviews questions in his niche. The models cheerfully returned his invented numbers, sometimes attributed to his site, sometimes attributed to nobody, sometimes attributed to one of the downstream blogs that had copied him. The fabrications had been laundered into the training and retrieval corpus. The post landed on Hacker News and raced to a score of 95, because every engineer who has ever shipped a RAG system felt a cold prickle reading it.

Why it matters

This is not a new category of attack — SEO spam has existed as long as PageRank — but the economics have shifted sharply in the attacker's favor. In the old world, a bogus stat that ranked well on Google still had to survive the reader clicking through, scanning the page, and deciding whether to trust it. In the new world, the LLM reads the page for the user and returns a confident one-sentence summary. The human-in-the-loop friction that used to catch obvious nonsense has been engineered out of the funnel in the name of "answers, not links."

The underlying failure mode is well understood by anyone who has built a retrieval pipeline. Vector search and keyword retrieval both reward recurrence: a claim that appears in many documents gets more chances to be retrieved, scored higher by rerankers that use document frequency as a prior, and treated by the generator model as a safe statement to repeat. None of these systems have any mechanism for distinguishing "many independent sources confirm this" from "one source got copied thirty times." Provenance collapses the moment content is scraped, chunked, and re-embedded; the LLM sees thirty chunks saying the same thing and infers consensus where there is only replication.

The community response on HN split into two camps. The first is the obvious one — this is fraud, Google and the LLM providers should penalize it, Jon is a bad actor for publishing a how-to guide. The second, more interesting response came from practitioners who pointed out that Jon's tactic is just the aggressive version of what every "state of X" report from a vendor already does: invent a methodology, publish a plausible number, and ride the citation flywheel. The difference is one of degree, not kind. Vendor-authored stats have been poisoning B2B content for a decade. LLMs just removed the last layer of skepticism between the stat and the reader.

There is also a darker read. If a solo marketer can do this to game AI answers for ad clicks, a state actor or a well-funded competitor can do it to shape consensus on consequential questions — drug efficacy, election procedures, which cloud provider has the fewest outages. The barrier to entry is a domain, a blog, and enough patience to seed the citations. We have built an information retrieval system with no immune response to coordinated inauthentic authorship, and we are plugging it into every product that answers a user question.

What this means for your stack

If you ship a product that uses retrieval over the open web — a research assistant, a sales-intelligence tool, a "ask our docs plus the internet" chatbot — you now have a concrete threat model to design against. Three practical implications:

First, treat statistics as a special content class in your pipeline. A number without a primary source is not data; it is a rumor with decimal places. Extract numeric claims during ingestion, try to resolve them to a canonical origin, and down-weight or flag any claim whose chain of citations loops back on itself. This is annoying engineering work and it will catch perhaps 60% of garbage, but 60% is better than the current 0%.

Second, reconsider the domains you trust. Most RAG stacks either trust everything they scrape equally or lean on a crude domain allowlist. Neither works against this tactic — the fabrications live on real-looking SEO blogs with real-looking author bios. Consider pulling in signals the SEO industry already uses against itself: domain age, backlink diversity, whether a site publishes anything other than listicles, whether its "authors" have provable identities. Pair that with primary-source preference: when a .gov, a peer-reviewed paper, or an SEC filing is available, retrieve it instead of the twelfth blog that summarizes it.

Third, if your product makes authoritative claims to users, start logging the retrieved chunks that produced each answer and sampling them for review. You will not catch everything, but you will build the dataset you need the first time a customer complains that your chatbot told them something plainly wrong. The incident report that says "we traced it to a single fabricated source that had been replicated across 34 downstream domains" is a dramatically better story than "we don't know where the model got that."

Looking ahead

The honest answer is that the open web is becoming a less reliable substrate for truthful retrieval, and the LLM providers are not going to fix this for you — their incentive is to answer confidently, not to say "I don't know who to trust here." Expect a wave of middleware whose entire pitch is provenance: tools that take a model's answer, trace each factual claim back to its earliest documented appearance, and show you the citation graph. Expect, also, more Jons. The tactic works, he published the recipe, and the next quarter of SEO conference talks will be about how to do it without getting caught. The practitioners who come out ahead are the ones who stop treating "the model said so" as the end of the research and start treating it as the beginning.