Platforms analysed
Instagram, TikTok and YouTube. One creator profile URL per report. Each platform has its own structural model because the metrics that matter differ between short-form feeds and long-form video.
Methodology
Platform evidence first. AI interpretation second.
This page explains how an influencer vetting report is produced: what data is collected, how the structural scores are calculated, where the AI is allowed to operate, and what the product does not claim to assess.
The simple overview
- 1.Public platform data is collected for the single creator you enter.
- 2.Deterministic structural scores are calculated from that data on the server.
- 3.Your brand context is extracted from the brand website and anything extra you provide.
- 4.An AI model interprets the evidence and is constrained within a band around the structural baseline.
- 5.Confidence and forecast limits are applied before the report is written.
Technical detail
Content sample sizes
The report samples the creator's recent public posts or uploads available from the platform at the time of the check. Sample size is reported alongside the analysis, and it directly affects the confidence level attached to the verdict.
Where the sample is too small to support a conclusion, the report says so rather than estimating.
Engagement calculations
Raw engagement rate is calculated from the sampled posts, then normalised into a percentile band so that creators are not penalised or rewarded purely for audience size. The bands used are: very low (under 1%), low (1–2%), average (2–4%), good (4–7%) and excellent (7% and above).
Audience-size tiers (nano, micro, mid-tier, macro, mega) are used to contextualise the normalised score, because a strong engagement rate at scale is a different signal from the same rate on a small account.
Posting and growth signals
Posting cadence is scored from the observed frequency in the sample — higher and more consistent cadence scores higher, occasional posting scores lower.
Growth is derived by comparing recent performance against the creator's earlier baseline in the sample and resolves to a growing, stable or declining signal. Where no baseline can be established, the signal is treated as stable rather than assumed positive.
Structural scoring
Each platform's structural composite is a fixed weighted blend of engagement, safety, growth and posting. YouTube weights engagement 40%, safety 28%, growth 16% and posting 16%. Instagram weights engagement 44%, safety 26%, growth 16% and posting 14%. TikTok weights engagement 45%, safety 26%, growth 16% and posting 13%.
These calculations are deterministic: the same inputs produce the same structural scores every time, independently of the AI layer.
Sponsorship detection and paid resilience
The report detects sponsored content signals within the sampled posts and separates sponsored from organic where enough of both exist.
Paid resilience compares sponsored performance against the organic baseline. Where the sponsored sample is too small, the resilience read is marked directional or withheld rather than presented as a firm conclusion.
Brand-context extraction
The brand website is analysed for positioning, category and public voice. Any additional brand or campaign context you provide is used alongside it. Richer brand context raises the confidence attached to the brand-fit read; thin context lowers it.
Constrained AI interpretation
The AI receives the structural scores and the sampled evidence and produces the written interpretation: the fit thesis, reasons to proceed, reasons for hesitation, campaign shape and recommended next move.
Its fit score is clamped by the server to within 15 points of the structural baseline. The model is told the band it will be clamped into, so its reasoning and its score stay consistent. The AI cannot invent the underlying metrics, and it cannot reduce an existing structural safety concern.
Safety scoring and limitations
Sampled content is checked against a risk-term list covering areas such as controversy, legal issues, explicit content, gambling, drugs, violence, hate, misinformation and drama-driven content. Matches are weighted by how prevalent they are across the sample rather than counted once.
Context matters, so an allowlist prevents common false positives (for example, 'exposed brick', 'beef stew' or 'fake plant'), and clearly humorous usage is discounted.
This is signal detection, not complete reputational due diligence. It reviews the content sampled, not the creator's entire history, off-platform behaviour, imagery or press coverage. Every flag is a prompt for human review.
Confidence derivation
Decision confidence is high, medium or low, and is driven by how much usable evidence exists: the number of posts sampled, the number of sponsored posts detected, and the richness of the brand context. The report states the reason for the confidence level it gives.
Forecast limits
Performance ranges are bounded by the creator's demonstrated historical distribution in the sample. Estimates that fall outside that observed range are pulled back to it, and where the evidence cannot support a range, no figure is shown.
Forecasts describe likely content performance ranges only. They are not predictions of sales, revenue or conversions.
Missing-data handling
When platform data is unavailable or incomplete, the report is marked low confidence and falls back to conservative, rules-based reads of the data that was retrieved. Missing evidence is reported as missing rather than filled in.
What the product does not claim to assess
Audience demographics or audience overlap. Guaranteed brand safety. Complete reputational due diligence. Sales, revenue or conversion outcomes. Contractual availability, rate negotiation or exclusivity status.
Check My Influence supports a human decision. It does not automate approval, and it does not replace influencer-marketing expertise.