SocialToAI states a figure about itself only with a record of how it was measured. This page prints those records: the date, the sample, the method and the version of the system that was measured.
Every offered interface with a recorded real answer: one answer each, the platform's raw JSON kept as it came.
Method
The same code that serves calls turned each recorded answer into SocialToAI's answer, offline: the JSON an HTTP call and an MCP tool both send, billing and notes included. Both were counted with the o200k tokenizer. No call was made.
System version
SocialToAI as served on 2026-11-06
All of them
8,357,994 tokens raw, 878,622 from SocialToAI
Platform
Interfaces
Raw JSON (tokens)
SocialToAI (tokens)
Fewer
TikTok
18
2,286,105
56,180
97%
Instagram
8
544,022
38,520
92%
X
21
247,096
88,867
64%
Reddit
11
169,102
27,114
83%
YouTube
10
52,538
31,967
39%
Facebook
4
16,144
4,526
71%
LinkedIn
7
36,700
21,249
42%
Threads
4
349,990
19,992
94%
Telegram
3
18,151
5,456
69%
Pinterest
2
43,466
4,677
89%
Xiaohongshu
21
479,829
80,827
83%
Douyin
33
1,836,257
99,034
94%
Weibo
10
543,363
25,361
95%
WeChat Articles
12
51,639
19,824
61%
Bilibili
8
274,162
26,909
90%
Zhihu
16
711,609
189,099
73%
Kuaishou
8
271,296
18,647
93%
WeChat Channels
11
182,863
31,561
82%
Meta Ad Library
5
236,356
79,078
66%
Google Ads Transparency
3
7,306
9,734
-34%
Where an interface gives only a few numbers, the notes every answer carries (what it cost, the balance, the fields it cannot give) can make it larger than the raw answer; the table counts those too.
One keyword search on a short-video platform, the same query sent to a raw platform-data API and to SocialToAI.
Method
Both responses were saved as returned and counted in model tokens: about 465,000 for the raw response, about 3,600 for SocialToAI's. 465,000 ÷ 3,600 ≈ 129.
7 research questions about public social media, each answered by the same model connected to one product at a time.
Method
Same model, same prompt and the same budget for every product, web search off. Each answer was scored from 1 to 10 by a judge model that saw no product names; the figure is the mean over the 7 questions. The other four products are not named.