Can Fake Reviews Influence AI Recommendations?

Yes. Reviews are part of the published record that AI systems read, retrieve, and quote when they describe and recommend businesses — so fake reviews can influence what an AI system says, including which businesses it names. That does not make them a strategy: fake reviews are illegal under U.S. Federal Trade Commission rules, they violate the policies of every major review platform, and because AI systems repeat the descriptive language they find, planted reviews leave a permanent, discoverable liability in the record a business depends on.

Why reviews carry weight with AI systems

When an AI system answers “what’s a good plumber near me?” or “is this company reliable?”, it is working from information it can find — and reviews are some of the most abundant, most specific business information on the public web. A business’s own website says what the business claims. Reviews, in aggregate, say what customers reportedly experienced: response times, pricing behavior, quality, complaints, specialties. Review text is also written in exactly the language customers use, which makes it highly retrievable when a customer’s question uses the same words.

Two consequences follow. First, AI answers about a business frequently paraphrase or quote review content — the praise, the complaints, and the specific details. Second, the volume and consistency of the review record helps an AI system treat a business as established and describable. A business with a deep, coherent body of reviews is simply easier for a system to talk about confidently than a business with almost no public record.

This is also why the fake version works — briefly, and at a price. Planted reviews enter the same record, in the same language, through the same channels.

Fake reviews are a form of poisoning

Researchers have a name for this, and it predates modern AI. In the recommender-systems literature, injecting fake profiles and ratings to push an item up (a “push” attack) or drag a competitor down (a “nuke” attack) is called a shilling attack — documented in the academic literature since at least 2004, when researchers demonstrated that fabricated reviewer profiles could shift what recommendation systems surfaced. The mechanism has not changed; the reader has. What used to manipulate a star ranking now feeds systems that generate prose answers, quote the planted language back as if it were customer experience, and carry it into contexts the planter never sees.

In the terms this pillar uses: fake reviews are poisoning of the recommendation record. They are false information, published in sources AI systems rely on, intended to change what those systems say. Reviews planted against a competitor — fake one-star attacks — are the same act with the polarity reversed.

The legal and platform line

This is not a gray area. In 2024, the FTC finalized a rule on consumer reviews and testimonials that bans fake and false reviews — including reviews written by people who never used the product, reviews generated to misrepresent experience, and reviews bought to express a predetermined sentiment. The rule also covers paying for indicators of social proof and requires disclosure of material connections, such as reviews written by a company’s own officers or employees. Civil penalties apply per violation.

Separately, every major platform — Google, Yelp, TripAdvisor, Amazon, and the rest — prohibits fake reviews in its terms and runs detection systems that remove them in bulk. Enforcement is real but incomplete: platforms remove large numbers of fakes, and some still get through, at least for a while. “Some get through” is a statement about enforcement gaps, not about legality or safety. A review that survives a filter today sits in the corpus that AI systems read tomorrow.

A concrete example

Consider a small HVAC company competing in one metro area. A marketing vendor offers “reputation enhancement”: forty five-star reviews over two months, each mentioning specific services — “same-day AC repair,” “honest pricing on furnace replacement,” “technician explained everything.” The reviews are written to sound local and varied.

For a period, the plan appears to work. The company’s rating climbs. AI answers about HVAC services in that area begin naming the company, sometimes echoing the planted phrases — because those phrases are now part of the retrievable record, repeated across the company’s profile. Then the platform’s periodic enforcement sweep flags the pattern: reviews from accounts with no other history, posted in bursts, sharing phrasing. The reviews are removed and the profile may be flagged or suspended. But the removal does not rewind everything. Pages that aggregated the reviews, screenshots in directories, and AI answers generated during the window have already propagated the claims. The business now has a public record that includes a fraud episode — and when the inflated language disappears from its profiles, the descriptions AI systems produce change again, this time downward, with the business unable to explain the gap.

The competitor who spent the same two months earning real reviews — slowly, with ordinary language and a few imperfect ratings — ends with a smaller number and a record that never has to be defended.

The strategic point: language is the liability

The deepest problem with fake reviews is not the star rating; it is the words. AI systems repeat descriptive language: “known for same-day service,” “premium pricing,” “family-owned since 1985.” If that language was planted, the business’s AI presence is built on statements it cannot substantiate — and every future comparison a system makes, every answer a customer reads, inherits them. When the underlying reviews are removed or exposed, the business does not return to neutral. It returns with a documented discrepancy between what the record claimed and what it could support.

There is also a direction of risk most buyers of fake reviews never consider: the vendor. The business has handed a third party proof of a deliberate deception, payable and documented. Disputes, leaks, and competitor complaints to platforms and regulators are all standing possibilities for as long as the reviews exist.

Limitations and misconceptions

“Platforms catch everything.” They do not. Detection is statistical and periodic, aimed at scale. But incomplete enforcement is not the same as safety — it means exposure is delayed and unpredictable, not eliminated.

“A few fakes among real reviews can’t matter.” For AI answers, wording matters as much as counts. A handful of detailed fakes can supply the specific descriptive claims a system repeats, which is exactly the mechanism that makes them risky.

“Everyone does it, so it’s just marketing.” The FTC rule treats bought sentiment as deception regardless of how common it is, and “everyone does it” has never been a defense in a platform enforcement action.

“Negative fake reviews about us are just part of the game.” Fake reviews aimed at a competitor are equally prohibited. If your business is hit by them, the remedy is documentation and platform reporting — the same dated-record discipline described in How Can Businesses Protect Their AI Reputation?, not retaliation in kind.

What we cannot claim: no platform publishes how much weight reviews carry in its AI answers, and we do not claim a specific effect size. The influence described here is the documented mechanism — reviews are read, retrieved, and quoted — not a measured percentage.