Search engine optimisation was built to influence rankings. A newer, less visible discipline has emerged alongside it: optimising content not to rank highly for a human reader, but to be retrieved and repeated by a language model. When that discipline is used to inject a false or self-serving narrative rather than to genuinely inform, it is worth naming for what it is — poisoned SEO.
What "poisoned SEO" means
The mechanism is straightforward once stated plainly: AI assistants answer questions using content indexed from the open web, whether through live retrieval, a search-augmented pipeline, or a training corpus built from crawled pages. Anyone who can get content indexed can, in principle, influence what a model repeats. Poisoned SEO is the deliberate production of content engineered for that purpose — written to be picked up by crawlers and cited or paraphrased by a model, rather than written to genuinely inform a human reader.
The label is not a judgment on search optimisation itself, which remains a legitimate discipline. It applies specifically to content whose primary purpose is to distort what a model says about an entity — a competitor, a public figure, an organisation — rather than to earn visibility on its own merits.
The tactic itself is not new — planted stories and paid placements predate generative AI by decades — but the payoff has changed. A planted story once needed to fool a specific journalist or editor into running it; today it only needs to accumulate enough indexed volume to shift a statistical average across a retrieval layer, which is a substantially lower bar to clear.
The supply chain a model inherits
A model has no independent way to verify what it retrieves; it inherits the trust decisions made by whatever crawler, search index, or dataset curator sits upstream of it. If that upstream layer indexes a page, the model treats its content as available material, largely without knowing how that page came to exist, who funded it, or whether it was written to be read by people or specifically to be read by machines. This is not a flaw unique to any one vendor — it is a structural property of retrieval-plus-generation systems built on an open, adversarial web.
This upstream inheritance also explains why a single successful injection can surface across otherwise unrelated products: a search engine's index, a retrieval-augmented assistant, and a separately trained model can all draw on an overlapping pool of crawled pages, so content seeded once can end up repeated by several tools that share no other infrastructure in common.
Four techniques we see used to inject a narrative
- Volume seeding — publishing the same claim, reworded, across dozens of low-cost domains so that it appears to be independently corroborated when it is really a single source multiplied.
- Authority mimicry — structuring a page to look like a credible reference source (dated bylines, footnotes, an "About" page, structured data markup) regardless of whether the underlying content has been verified.
- Review and forum manufacturing — seeding review platforms, Q&A sites and forums, which models frequently retrieve for "real-world" sentiment, with coordinated but inauthentic posts.
- Machine-oriented phrasing — writing content in the declarative, list-friendly, FAQ-style format that retrieval systems tend to extract cleanly, increasing the odds a specific sentence gets lifted verbatim into an answer.
None of these techniques requires much technical sophistication on its own. What makes a campaign effective is combining several of them at once, so that domain-level, content-level and engagement-level signals all point the same direction simultaneously.
Why volume can beat authority
Retrieval systems generally weigh some notion of source quality, but "quality" signals — domain age, backlink profiles, engagement metrics — are themselves gameable, and none of them directly measure factual accuracy. A coordinated campaign that produces enough distinct-looking domains, each individually unremarkable, can accumulate more aggregate retrieval weight than a single authoritative source that a model's index happens to crawl less frequently. This is the same volume dynamic behind how a false claim gets laundered into an assistant's answer — poisoned SEO is simply that process run deliberately, at scale, with intent.
Don't know what AI says about you today?
A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.
Detecting an injected narrative
A few patterns tend to distinguish manufactured content from organic coverage:
- Domain clustering — a set of unrelated-looking domains registered around the same time, hosted on similar infrastructure, publishing on the same narrow topic.
- Timing correlation — a burst of near-simultaneous publication around a specific date, rather than the organic, staggered pace of independent coverage.
- Absence of a real audience — no social engagement, no backlinks from unrelated sites, no evidence anyone other than a crawler ever read the page.
- A single narrative thread — every piece of "independent" coverage happens to converge on the same specific framing, figure, or accusation.
- Interchangeable boilerplate — author bios, disclaimers or site descriptions that are identical, word for word, across supposedly unrelated domains.
No single signal is conclusive on its own — legitimate coverage can occasionally cluster in time around a real news event, and small independent publishers can legitimately share hosting infrastructure. It is the combination of several of these signals converging on the same narrow claim that distinguishes manufactured corroboration from an ordinary news cycle.
Where legitimate optimisation ends and poisoning begins
Structuring your own content clearly, publishing accurate FAQs, and making factual material easy for a model to retrieve is simply good practice — the discipline sometimes called generative engine optimisation. The line is crossed when the content being optimised is not accurate, not disclosed as promotional or adversarial, or not about an entity that consented to being described. Optimising the discoverability of true, sourced, attributable information is not the problem this article describes; manufacturing the appearance of independent corroboration for an unverified or false claim is.
This distinction also matters for how an organisation should respond. Aggressively suppressing or reporting content simply because it competes for the same search terms risks silencing legitimate commentary and inviting exactly the kind of backlash a bad-faith campaign would want to provoke. A sound response targets the specific markers of manufacture described above, not the mere existence of unflattering coverage.
Defending your presence in the index
An organisation cannot police the entire open web, but it can reduce how much room a poisoning campaign has to work with. The approach we recommend starts with maintaining a small number of clearly authoritative, well-structured, frequently updated source pages that directly address the topics most likely to be targeted — because a model that can retrieve a strong, specific, on-topic source has less need to fall back on the noisier alternative. From there, it becomes a monitoring problem: the same multi-model auditing used to catch hallucinations and laundered claims also surfaces the early signs of a coordinated injection campaign, usually before it has fully propagated.
Speed matters more than perfection here. A campaign caught and countered within its first few weeks rarely accumulates enough indexed volume to meaningfully shift what a model retrieves; the same campaign left unaddressed for months can become difficult to fully displace, for the same reason a laundered claim grows harder to correct the longer it circulates unchallenged. Our audits typically flag domain-clustering and timing anomalies before a client's own communications team would have noticed the underlying narrative at all. See our public methodology for how the two checks work together.