Why Faked "Personal Review" Posts Never Spark AI Citations — The Double Failure of Fake Sponsored Posts on PTT / Dcard / Threads / Xiaohongshu

The Old Playbook Got Crushed, So Black Hats Pivoted to Social

The previous post, “The Double Death of the Thatched Hut and the Sandcastle,” dissected the double-death logic of on-site structural black hat tactics (PBN / hidden text / cloaking / templated mass-produced sites). When the combined Google + AI crackdown flattened the ROI of the PBN / content-farm route, cheap vendors don’t simply pack up and quit — they go hunting for the next “path that can still rise.”

Watching the Traditional-Chinese market from 2024 to 2026, that path is faking “user review sharing” on social platforms:

On channels like PTT (the Gossiping board and various topic boards — the highest registration barrier, where netizens call out “yè-pèi” sponsored posts in real time, yet also the high-authority corpus ghostwriters covet most), Dcard, Threads, Xiaohongshu (targeting the Chinese market), Instagram (some sponsorships), and Mobile01 (for cars / consumer electronics / real estate), they use fake accounts or ghostwriter accounts to post things like:

“I used X myself for a few months and it really worked, so I’m putting together a summary for friends with similar struggles.” “I stumbled on an underrated service by accident — not a sponsorship, please, just a genuine share.” “Everyone around me is using X, so I’ve compiled the pros and cons for everyone.”

These sponsored review posts that mimic real users look at first glance no different from an ordinary netizen’s Dcard review post — and that disguise is exactly what counts as “success.”

This post breaks things into three layers: (1) how this industry actually operates internally, (2) why AI engines structurally refuse to buy it, and (3) how owners should spot it and pull out.

Part One: Industry Insider View — The Ghostwriter Account Lifecycle

To understand why this path is structurally doomed, you first have to see how fake-review sponsorship actually operates.

1.1 The “Account Seasoning” Phase — The First Overlooked Cost

Every social platform has a “new-account credibility threshold”: posting a sponsored piece right after registering gets your reach throttled by the algorithm, or even classified outright as spam. So a ghostwriter account needs to be “seasoned” for 3-6 months first:

The “X0,000/month, 300,000 reach guaranteed” quote the owner sees never tells you that this X0,000 includes the amortized fixed cost of the vendor seasoning accounts. So a meaningful portion of your sponsorship fee is subsidizing the historical upkeep of ghostwriter accounts.

More crucially: this whole account-seasoning logic only works while the platform anti-spam hasn’t yet seen through it. Once the platform’s model catches the pattern of “credibility looks high, but the account suddenly started heavily pushing one brand recently,” the entire batch of accounts gets banned in bulk — and the seasoning costs the owner already paid go straight to zero.

1.2 The Pricing Structure of “KOLs” at Different Tiers

The current gray-market sponsorship pricing tiers in the Traditional-Chinese market (from publicly available vendor rate cards and industry estimates — not the prices vendors actually collect from clients):

What cheap vendors mostly sell is the “ordinary-person + micro-influencer bundle”: 50-100 posts pushing simultaneously in a single month, made to look like a “natural word-of-mouth explosion.”

This tactic worked well in 2020, but by 2025-2026 it has reached close to its point of diminishing returns — because: - Platform-side anti-spam has accumulated enough capability to catch the cluster of “100 posts in the same period concentrating on a single brand” - User-side ability to recognize “what a sponsored post looks like” is improving (Dcard has dedicated boards hunting sponsored fake accounts) - AI engine-side recognition of “same-source bulk content” is being reinforced

1.3 The 5-Part Template of the Sponsored Post

The content architecture of fake-review sponsored posts is highly templated. Cheap vendors typically reuse the same outline across 5-10 ghostwriter accounts, changing only the details and wording:

1. Personal-story opening (300-500 words) — tell a life struggle, build an emotional connection, get the reader to project themselves into it
2. The search for a solution (200-400 words) — “I tried A, B, and C and none of them worked” (conveniently name-checking competitors)
3. The chance discovery (200-300 words) — “a friend / colleague / my mom recommended I try X” (the owner’s brand makes its entrance)
4. The usage experience (500-1,000 words) — detailed product / service description (the vendor provides a “ghostwriter copy brief” for this section)
5. Conclusion + closing disclaimer (200 words) — “just sharing / no sponsorship pitches please / everyone can judge for themselves” (the classic adversarial declaration)

This 5-part structure is itself a detector signal: when you vectorize a batch of Dcard sponsored posts with an LLM, content with extremely high structural similarity clumps into a cluster. The AI model doesn’t need a human to catch it at all — the vector space sorts itself into groups.

1.4 Fake Comments and the Boost Mechanism

Once a sponsored post goes live, the first wave of 30-100 comments is usually also bundled in by the vendor:

• Opening comments: “Have you actually used it, OP?” “Roughly how much does it cost?” (propping up the vibe of “someone genuinely cares”)
• Mid-thread comments: “My friend used it too, it really is good” “I had OOO before, and later X also saved me” (multiple accounts corroborating one another)
• Closing comments: “OP please DM me where to buy it” (manufacturing an “urgent demand” signal)

When the Dcard / Threads algorithm sees “30+ high-engagement comments within 30 minutes of publishing,” it judges the post as trending → recommends it to the homepage. That is the technical mechanism behind the “reach guarantee.”

But in the eyes of an AI engine, this whole boost is an entirely different matter — which the sections below will explain.

Part Two: The AI Engine’s Technical Perspective — Why Anonymous UGC Scores Low on Trust

The question owners get bluffed on most often by vendors: “This post went viral and hit the homepage — surely ChatGPT will see it?“

It will see it, but it won’t cite it. Being seen versus being cited are two completely different things.

2.1 The Quality Filter on LLM Training Corpora

Mainstream LLMs (the OpenAI GPT family, Gemini, and other leading models) don’t just dump the entire internet in for training — the data passes through multiple layers of quality filters:

• Duplicate content removal (deduplication): highly similar content gets aggregated and weighted, not each piece gets +1 vote. The same 5-part template posted across 8 accounts = the LLM may count it as only 1 vote during training
• Source domain trust score (domain authority): Wikipedia, well-known media, academic-institution sites, and government open data carry markedly higher weight than anonymous forum posts
• Content objectivity heuristic: content that is “all praise + zero downsides” has been publicly noted in multiple labs’ training documentation as getting down-weighted (it doesn’t match the statistical distribution of real usage experiences)
• Author entity verification: content with a real name, a historically stable identity, and a consistent cross-platform record carries higher weight than anonymous UGC

When your 100 Dcard sponsored posts enter Common Crawl (one of the LLMs’ main web sources for training): - The templated structure triggers deduplication → many posts become one vote - Dcard is an anonymous platform → its domain trust leans medium in the “lifestyle sharing” category and low in the “authoritative source” category - “All praise + no sponsorship pitches please” triggers the objectivity heuristic → the whole batch gets down-weighted - No author entity verification → it can’t get into the “citation priority pool“

Result: the cost paid for 100 posts → after LLM training, the increase in brand authority score is close to zero.

2.2 The Citation Ranking of Live Web Search (RAG)

When ChatGPT, Perplexity, and others enable “live search,” the candidate results pulled back from the web pass through a reranker before it decides which few to cite. The reranker mainly weighs:

Fake-review Dcard sponsored posts score low on all five of these dimensions: anonymous source, all praise, rambling-chronicle structure, posts that easily disappear, no cross-source corroboration. The reranker drops them to the bottom of the candidate list, and the citation probability approaches zero.

Want to understand exactly which steps AI uses to filter out sources like these during live citation? See: How Do AI Search Engines Pick Their Citations? The 4 Key Steps (VIP).

Owners assume “as long as the AI sees it, there’s a chance” — but the AI first looks at who is authoritative, and only then decides whom to cite.

2.3 Same-Source Cluster Detection in Vector Space

The deepest problem: ghostwriter bulk content is an extremely conspicuous group in vector space.

After converting a batch of fake-review posts into embeddings: - The content vectors clump into a tight cluster (identical structural template) - The author embeddings across content show the anomaly pattern of “short-term, dense pushing of one brand” - The temporal burst signal (50-100 posts on the same topic in one month) is utterly unlike the long-tail distribution of natural word-of-mouth

For a human reader: you might have to read 50 posts + cross-check account IDs to notice. For an LLM / embedding model: it’s a cluster you catch in a single vectorization pass.

Want to understand how AI distinguishes “written by a real person” from “bulk-spammed” at the language level? See: The 7 Sub-Metrics of Language Naturalness (VIP).

This is not speculation — it’s mature technology that’s been around in the information-retrieval field for 20+ years. An owner buys fake-review sponsorship → gets seen through at both the AI-training and live-citation levels → and the whole budget amounts to money burned for nothing.

2.4 The Three-Layer Difference Between “Being Seen” and “Being Cited”

Putting the above three threads together, owners must distinguish:

The owner’s money buys the first layer (the crawler sees it) — but what GEO actually wants is the fourth layer (being cited). The two middle layers wring the entire funnel dry.

Part Three: The Evolution of Platform-Side Anti-Spam

Another structural risk of fake-review sponsorship: the platforms themselves are hunting it too.

3.1 The Evolution of Dcard’s Anti-Spam Measures

Since 2022, Dcard has progressively strengthened its anti-fake-sponsorship mechanisms: - New-account restrictions: daily caps on posts / comments within 30 days of registration - Credibility-rating system: cumulative point deductions for the number of reports / board-rule violations - Bulk bans: discovering “multiple accounts on the same IP / same device fingerprint” gets the whole batch banned - AI-assisted content recognition: machine-learning detection of “sponsored-template signatures” (the specific model isn’t public, but the behavior is observable) - Sponsorship disclosure mechanism: lawful sponsorship requires clear labeling, and Dcard’s ToS explicitly prohibits undisclosed sponsorship

Over the past few years, incidents of ghostwriter accounts being banned in bulk have happened repeatedly, with the seasoning costs the owner previously paid going entirely to zero.

3.2 The Inherited Advantage of Threads (the Meta family)

Threads is a Meta product, so it directly inherits the anti-spam models Instagram / Facebook have accumulated over more than a decade:

• IP / device fingerprint detection
• Cross-platform account linkage (recognizing one device operating multiple accounts)
• The anomaly pattern of “a new account suddenly with high engagement“
• Detection of AI-generated content (including LLM-written sponsored posts)

For cheap vendors, Threads’ anti-spam bar is an order of magnitude higher than Dcard’s — but this also makes sponsored posts on Threads look more “normal” (the survivors are the ones that passed Meta’s filter), so owners misjudge that “Threads sponsorship works better.” In reality it’s survivorship bias: you see the batch that hasn’t been caught yet, and you don’t see the batch that got banned and never delivered.

3.3 Xiaohongshu and Regulation Within China

Xiaohongshu is constrained by KOL-sponsorship regulations within China (amended several times since 2017): - Explicit requirements for sponsorship labeling - Clear personal legal liability for KOLs who post undisclosed sponsorships - The platform has been penalized by regulators multiple times

Traditional-Chinese owners who want to use the strategy of “hit the Chinese market with Xiaohongshu + incidentally influence Taiwanese search” face far greater legal exposure than they imagine.

3.4 PTT: The Double Squeeze of a High-Authority Corpus × Netizen Self-Policing

PTT (the largest Taiwanese BBS) is a special case in the Traditional-Chinese market — one cheap vendors both covet and fear:

• The highest account-seasoning barrier: PTT account registration has long required verification and was even closed off for periods, so mass-producing ghostwriter accounts costs far more than on Dcard / Threads — you basically can’t season 50 disposable accounts to push one brand at once.
• Real-time netizen self-policing: PTT’s reply (推文) culture means call-outs like “yè-pèi” (a homophone for “sponsorship”), “paid shill / 工讀生,” and “this is a sponsored post, right?” happen on the very first page of replies. A fake-review post often gets caught before it can even go viral.
• And ironically, its corpus authority is relatively high: PTT is heavily crawled, frequently referenced by news and AI, and is a relatively high-domain-trust source in Traditional Chinese — which is exactly why ghostwriters covet it.

But the latter two stack up to bite the owner back: once a ghostwriter finally lands a fake review on PTT, the replies “this is obviously yè-pèi” / “brand X’s paid shills are back again” enter that high-authority corpus alongside the original post. The result isn’t “the brand endorsed by a high-authority source,” but “brand + yè-pèi / paid shill” permanently recorded by a high-authority source.

In other words: faking it on a low-trust platform (anonymous UGC) at worst means “money wasted, never cited”; getting caught faking it on a high-trust platform (PTT) means “paying to buy a negative label that the AI will see — and cite” — worse than getting down-weighted on Dcard.

3.5 Why New Platforms Will Inevitably Follow Google’s Anti-Spam Trajectory

Google went through more than a decade of anti-spam evolution — Penguin (2012), Panda (2011), Helpful Content (2022), SpamBrain (ongoing), and more. Each generation knocked the ROI of black hat tactics down a notch.

Dcard, Threads, and other new platforms essentially have to walk the same evolutionary path once, just with the timeline compressed into 5-10 years — because: - Machine-learning tools are already mature (no need to build detectors from scratch) - Google’s spam-detection papers / techniques are already public - Commercial pressure on platforms (advertisers demanding a clean environment)

What cheap vendors are profiting from is the time gap of “the platform hasn’t caught it yet + the AI isn’t ready yet.” That gap will keep closing — it won’t open back up.

Part Four: The Owner’s Decision Framework — How to Spot It, How to Pull Out

4.1 Five Contract / Proposal Red Flags

If you’re currently in talks with an SEO / marketing vendor, the moment the other party’s proposal includes any one of the following, you are being sold a fake-review sponsorship plan:

1. “Guaranteed N0,000 reach“: the cost structure of a reach guarantee can only come from buying traffic / platform boosts / ghostwriter accounts cross-promoting one another — natural word-of-mouth can never be guaranteed
2. “Ordinary-person recommendations / micro-influencer bundle” with per-post quotes below NT$3,000: any “KOL” below this price is basically a ghostwriter account / seasoned account
3. “Synchronized across Dcard / Threads / Xiaohongshu” + “50+ posts a month“: natural word-of-mouth never moves at this cadence — this is a vendor’s bulk production line
4. Requiring the owner to provide a “sponsored-post copy brief“: it means the vendor will hand ghostwriters a script to “fake a real review per the brief”
5. The contract has no sponsorship-disclosure clause: the statutory labeling of undisclosed sponsorship is the owner’s legal risk, and a legitimate vendor protects you in the contract

Any one trigger warrants pressing for details; multiple triggers mean what the vendor is selling is fake-review sponsorship.

4.2 Real Content-Marketing ROI vs. Fake-Review ROI

What the owner actually wants is not “cheap reach” but “cumulative brand authority.” A comparison of the long-term ROI of the two routes:

The owner’s thinking is “prop things up with cheap tactics first, then go legit later” — but this strategy has two failing assumptions: 1. It assumes “the downside of cheap tactics won’t spread” — but if caught, the brand name + “bought Dcard sponsorship” will live permanently in search results 2. It assumes “you can just switch over later” — but AI engines don’t “re-evaluate” signals that the quality filter has already down-weighted in the past

4.3 How to Transition from the Current State to Legitimate Channels

If you’re already running fake-review sponsorship, stop immediately + switch to the lowest-cost approach:

1. Take inventory of your existing fake Dcard / Threads sponsorships: make a list so you know your exposure when the platform catches it
2. Gradually switch to genuine third-party reviews: find industry KOLs / media willing to give a real usage experience (including pros and cons)
3. Get your official website’s GEO fundamentals in order: a 12-dimension health check + structured data + brand entity building
4. Real customer testimonials with identity: customer testimonials with full company name, real name, and real use cases
5. Long-term goal: inclusion in Wikipedia / industry wikis / public databases — structured and cross-verifiable

All four of these are slow, expensive, and don’t guarantee short-term reach — but only these four get cited by AI, and only these four hold up for two or three years.

Part Five: Legal and Brand Risk

Many owners assume “it’s just buying a sponsored post — what legal problem could there be?” In fact, Taiwan’s regulation of undisclosed sponsorship is relatively strict:

5.1 The Fair Trade Commission’s Undisclosed-Sponsorship Rules

• Fair Trade Act §21 (false advertising)
• Since 2017, the Fair Trade Commission has explicitly classified “undisclosed sponsorship” as a form of “false or misleading advertising“
• The multimedia sponsorship-disclosure guidelines explicitly state that “the owner / agency / KOL all three bear a disclosure obligation“
• Violators can be penalized

If the owner merely “buys sponsorship through a vendor,” the legal liability does not therefore transfer to the vendor. When a report is filed, the owner themselves is the subject of the penalty.

5.2 Damage Assessment When a Report Is Filed

The common consequences of a fake-review sponsorship being reported: - Platform level: accounts banned / posts removed / the platform throttling the brand’s reach - Legal level: Fair Trade Commission penalties / litigation (if a competitor sues) - Media level: “XX brand caught buying Dcard sponsorship” hits PTT / Threads — permanent negative SEO - AI level: the quality-filter down-weighting + brand-entity contamination mentioned in the previous section

Any one of these four layers of damage runs 10-100 times higher than the sponsorship fee the owner originally paid.

5.3 The Boundary of the Vendor’s Responsibility

A legitimate vendor writes the following clearly into the contract: - “All sponsored content is disclosed per the Fair Trade Commission’s rules” - “If the owner is penalized due to incomplete disclosure, the vendor is jointly liable for damages” - “The KOL’s real identity is verifiable“

Cheap vendors leave these clauses blank or write them vaguely. If you can’t find these three clauses in the contract, you should change vendors.

Part Six: The Owner’s Self-Check List — 12 Fake Signals

If you want to judge for yourself whether a Dcard / Threads post is a fake-review sponsorship, the 12 signals below mean it’s highly suspicious if any 3 or more hit:

1. The opening story is overly personalized, with heavy emotional build-up (300+ words before getting to the point)
2. It uses the 5-part “pain point → trial and error → chance discovery → experience → recommendation” structure
3. The ending features adversarial declarations like “no sponsorship pitches please / just sharing / judge for yourselves” (a genuine review usually doesn’t proactively disavow it)
4. Within 1-2 weeks, 5+ structurally near-identical “reviews” on the same topic appear across different Dcard boards
5. The account’s posting history has an obvious break of “the first 80% genuine life, the last 1-2 months started pushing a certain brand“
6. Among the first 30 comments on the post, 5+ accounts are of similar account age and similar posting frequency
7. The brand name turns up suspicious discussion in Google searches for “brand name + sponsorship,” “brand name + Dcard,” “brand name + PTT,” or “brand name + yè-pèi“
8. The brand’s official-site link inside the post uses a ref / utm carrying a KOL tracking code
9. The opening or ending lacks a per-platform disclosure of “ad / collaboration / sponsored“
10. The account avatar is a generic image findable online (verifiable via reverse image search)
11. The account bio / school / job is filled in too perfectly but with no verifiable links
12. The recommended “friend / colleague / mom” is always generic, with no name / no detail

An owner can judge by catching 1-2 themselves; catching the pattern across an entire industry requires tooling.

Part Seven: Why This Path Is Structurally Hopeless in 2026-2027

Integrating the trends across the three layers:

All four curves converge toward “fake-review sponsorship becomes less and less worthwhile.” What cheap vendors are selling now is money made off a structural window of opportunity — in two or three years that window will close, while the negative SEO on the owner’s brand will remain.

A One-Sentence Closing

Fake-review posts are reach made for the algorithm to see, not a brand signal made for the AI to see. What the owner buys is the firework that “trends on Dcard for 24 hours,” not the asset that “gets continuously cited by ChatGPT for two years.”

Look back in two years and the former leaves only platform-report records, negative brand SEO, and legal exposure; the latter is still compounding.

The price of cheap is making the AI remember forever that you shouldn’t be cited.

Further reading (go deeper)

Rather than burning budget on reach that’s destined to be down-weighted, put it into the signals AI actually trusts. To dig deeper:

• How Do AI Search Engines Pick Their Citations? The 4 Key Steps (VIP) — understand it from the flip side: what kind of source actually makes it into the citation pool.
• Why E-E-A-T Matters More in the AI Era Than Ever Before (VIP) — why “a real name + a verifiable identity” is something anonymous UGC can never make up for.
• The 7 Sub-Metrics of Language Naturalness (VIP) — the technical detail of how bulk ghostwriter content gives itself away at the language level.

This post is part of the GeoWeb blog’s “Off-Site Signals” series. Further reading: “The Thatched Hut and the Sandcastle: The Double Death of On-Site Structural Black Hat” · “Off-Site Visibility and AI Citation Authority” · “Why Wikipedia Is Critical for GEO”