Data quality
Most of what Poseidon shows is read straight from Meta. A few labels are written by AI. Here is how often each one is right, measured on ads people checked by hand, and what we hold back until it is good enough.
Read straight from Meta, never written by AI
- The ad's start and stop dates, and how many days it ran
- Meta's EU reach (the number of people Meta showed the ad to)
- The country, age and gender shares of Meta's own breakdown
- The ad's text and its link in the Meta Ad Library
Every number shows the day it was read.
Labels written by AI, and how often they're right
Each label is checked against ads a person labelled by hand, ads the AI had never seen. The dot is how often the AI was right; the line is the likely range. A label is served only when it is right at least 90% of the time and the low end of its likely range is 85% or more.
Hook type
What kind of opening line an ad uses: an offer, the product, a story, a pain, or a claim.
Served for 1 of 5 types on image ads: Pain. The other types are held until they pass the rule.
Image ads
- Claim: 31 of 33 right, likely between 80% and 98%. Held: its likely range goes down to 80.4% (the rule is 85% or more).
- Story: 29 of 32 right, likely between 76% and 97%. Held: its likely range goes down to 75.8% (the rule is 85% or more).
- Pain: 24 of 24 right, likely between 86% and 100%. Served.
- Product: 20 of 21 right, likely between 77% and 99%. Held: its likely range goes down to 77.3% (the rule is 85% or more).
- Offer: 10 of 11 right, likely between 62% and 98%. Held: its likely range goes down to 62.3% (the rule is 85% or more).
Video ads
- Claim: 11 of 16 right, likely between 44% and 86%. Held: right only 68.8% of the time (the rule is 90%); its likely range goes down to 44.4% (the rule is 85% or more).
Too few ads checked by hand to judge yet (under 10): Pain (9), Product (6), Story (6), Offer (5).
Measured on ads checked by hand that the AI had not seen, 13 Nov 2026. As served by Poseidon's API.
Creative type
What kind of creative it is: UGC, a video sales letter, a demo, a product photo, a before and after…
Not served for now: no type passes the rule yet.
Image ads
- Product photo: 38 of 40 right, likely between 84% and 99%. Held: its likely range goes down to 83.5% (the rule is 85% or more).
- Lifestyle photo: 18 of 28 right, likely between 46% and 79%. Held: right only 64.3% of the time (the rule is 90%); its likely range goes down to 45.8% (the rule is 85% or more).
- Native photo: 12 of 12 right, likely between 76% and 100%. Held: its likely range goes down to 75.7% (the rule is 85% or more).
Too few ads checked by hand to judge yet (under 10): Before / after (6), Text only (5), Illustration (3), Screenshot (2), Comparison (1).
Video ads
- UGC: 26 of 29 right, likely between 74% and 96%. Held: right only 89.7% of the time (the rule is 90%); its likely range goes down to 73.6% (the rule is 85% or more).
- 3D: 10 of 11 right, likely between 62% and 98%. Held: its likely range goes down to 62.3% (the rule is 85% or more).
- VSL: 10 of 11 right, likely between 62% and 98%. Held: its likely range goes down to 62.3% (the rule is 85% or more).
Too few ads checked by hand to judge yet (under 10): Text on music (2), Demo (1), Podcast (1).
Measured on ads checked by hand that the AI had not seen, 13 Nov 2026. As served by Poseidon's API.
Held back, and why
- Script sections (hook, problem, proof, offer…)Shown in full ads, but No accuracy figure is published for them yet. It will appear here when it does.
- Weights by hook typeHeld until they are recomputed on the 5 new hook types.