How Scout works

Most people now say it is getting harder to tell real from fake online [14]. For a small business the harm is concrete: a burst of one-star reviews, a copycat website, a manufactured pile-on. The tools built for this problem were built for platforms and governments. The ones that reached consumers mostly depended on platform data access that has since been withdrawn [13], and the platforms themselves have scaled back enforcement. Since October 2024 a federal rule also makes a business liable for fake reviews bought on its behalf, even unknowingly, with civil penalties per violation [10].

Scout is the owner's own instrument. It works from data the owner is entitled to see, so no platform can switch it off.

Scout takes three kinds of input: a review export or upload from any platform, the business's website domain, and (soon) public mentions found by search. It never scrapes behind a login. The choice is deliberate: when access to platform data closed in 2023 and 2024, the research tools that depended on it went dark within a year [13]. Owner-authorized data is the one source that cannot be revoked.

Timing bursts. Scout compares review volume in any 72-hour window to the business's own long-run rate. Coordination shows up first in timing, and this is one of the few signals that still works now that generative AI writes fluent text: a 2023 study found a 1,140-account botnet on X that content classifiers could not separate from humans, and it was caught by its posting pattern instead [5]. Meta's own enforcement teams report the same thing, that behavior rather than content is what gives operations away [2][3].

Shared wording. Scout measures how much reviews overlap at the phrase level and keeps only tight clusters. Templated or lightly paraphrased copy is a classic tell of a coordinated push [2]. Independent customers of the same shop also reuse stock phrases, so Scout requires a cluster to be tight across every pair, not merely chained through one common phrase [4].

Account age and history. Where a platform exposes it, Scout checks whether reviews come from very new or single-review accounts and whether those accounts arrived in the same week. Creating fresh accounts for a campaign is the most common technique in the public catalog of influence-operation methods [12]. When the platform hides this, Scout says so rather than guessing.

Rating shift. A large swing in the last 30 days against the prior six months. Scout treats this as context only; it is just as consistent with a real service problem.

Look-alike domains. Scout generates typo, homoglyph, hyphen and top-level-domain variants of the business's website and checks which ones are live. In a systematic review of all 63 national elections held in 2024, fake or captured websites were the second most-used interference vector, present in 13 of the 19 elections with documented interference [6]. Impersonation sites are also the commonest way a small business is attacked, and this surface is fully visible without any platform's cooperation.

Scout has two layers. The first is deterministic code that computes the signals above and sets a ceiling on how strong the overall verdict may be. The second is a language model that reads those findings and writes the plain-language assessment and the action list. The model may say organic or can't tell whenever it judges that right; it may say likely coordinated only when the deterministic layer has already found enough. The reason is evidence from the research: language models used as fact-checkers score roughly 65 to 80 percent on binary verdicts and lean toward calling claims false [8], and automated bot detectors produce false positives at rates that make them unsafe as sole judges [9]. Scout therefore uses the model for explanation, not for judgment.

Every report also lists what Scout could not see. Deepfake and synthetic-media detectors, for example, perform near chance on real-world material [7], so Scout does not claim that capability. "Can't tell" is a real answer, and the report says what would settle it.

Scout's language rules are product constraints, not style. It reports counts and patterns ("17 reviews in three days, 16 from accounts under two weeks old") and always offers the ordinary explanation alongside the concerning one. It never calls a review fake, a person a bot, or a claim false, and it never speculates about who is behind a pattern. The reasoning is practical: tools that issued truth or authenticity verdicts have drawn lawsuits and political attack, and the most thorough review of counter-disinformation measures to date concludes that any approach requiring an arbiter of truth inherits that exposure [1]. Pattern description with stated confidence leaves the judgment where it belongs, with the owner and, when needed, a lawyer.

The one piece of evidence on what makes people willing to pay for protection against online manipulation is that they pay after seeing their own exposure, not before [11]. So the snapshot costs nothing and asks for no account. Monitoring, which re-runs the checks on a schedule and alerts on change, is the paid product.

Scout stores each assessment with its evidence so the owner can find the report again. With the owner's consent, an anonymized copy enters a research dataset on how small businesses are targeted. The field's own reviews name the lack of such ground truth as its deepest gap [1][4]. Consent is a checkbox, off by default.