How we do research

Where the data comes from

Our studies use Leewou's proprietary creator data, built on public profiles. Follower counts and posting data describe those public profiles directly. Audience measures, including credibility and follower types, are modelled estimates rather than direct observations, so treat them as estimates. We use only public data, and we publish aggregates only.

The Instagram purge study: sample and collection

• Population: the 11,440,763 Instagram accounts with 10,000 or more followers in the index on October 30, 2026. Of these, 10,961,081 (96%) have a known country.
• Reports: 24,501 analytics reports collected October 30 and 31, 2026: 19,803 Instagram, 3,120 TikTok and 1,578 YouTube. Each holds monthly readings from April to October 2026.
• Sampling: random draws within country and follower-band cells for Instagram, allocated in proportion to the square root of cell size so small countries have enough accounts to report, plus a smaller global draw. TikTok and YouTube are global draws by follower band.
• In the weighted figures: 19,624 Instagram accounts with a known country. Two rows under 10,000 followers and 177 with no usable country are excluded. Accounts younger than a given reading are excluded from that month only, so April to May uses 19,103.
• Non-response: about 1.9% of attempted Instagram profiles never returned a usable report. Of the 2,395 that first asked us to retry, 2,205 eventually did.

Weighting and uncertainty

We post-stratify by country and October follower band, using known counts per cell, so each cell counts for its real share of the population. Population in the 0.6% of cells with no sampled account is spread across the rest of that band. The effective sample size after weighting is about 7,550. TikTok and YouTube are weighted by follower band only, and YouTube's bands below 25,000 are merged because the platform rounds its subscriber counts.

Confidence intervals are 95% intervals from 1,000 bootstrap replicates, redoing the weighting in each. Medians and percentiles are weighted. Cuts with fewer than 30 accounts are suppressed, and countries are ranked only with at least 50 accounts and an effective sample of at least 30.

The difference-in-differences compares the April to May share losing more than 2% with the average of June to July, July to August and August to September, for each platform. It assumes that without the purge the April to May share would have looked like those months, and that Instagram, TikTok and YouTube would otherwise have moved alike. May to June is left out of the baseline because it still shows elevated losses.

Definitions

• Follower change: an account's follower count at one monthly reading against the previous one. It is net of new followers, unfollows and removals.
• Audience credibility: a modelled score from 0 to 1 of how likely an account's followers are to be real and active. We report it for Instagram only. It was measured in October 2026, after the purge. It is not a fake follower percentage.
• Suspicious followers: the share of followers that a model places in a suspicious category. It is a separate measure and doesn't map onto credibility: one public account we checked scores 0.78 on credibility, which would imply 22% if read as a fake share, but has 12.5% suspicious followers. We never use the two interchangeably.
• Engagement rate: on this page, average likes per post divided by followers, from the monthly history. Reported engagement rates in the index are per follower on Instagram but per view on TikTok and YouTube, so we never compare them across platforms.
• Follower bands: set by October 2026 follower count and aligned to the index's filter grid. We do not band by April size, because accounts that fell below 10,000 aren't in the sample, which would make the lowest April band look healthier than it was.
• Followers removed: the purge month's net change across all accounts minus the average net change of June to September. It counts follows, so one removed account can count many times.

How we validate a study

Known limits

• Survivorship. We sample accounts that exist with 10,000+ followers in October. Deleted accounts and those that fell under 10,000 are missing.
• Timing. Monthly readings don't carry an exact date. Comparing closed and open months suggests around the 22nd, with a rough range of the 17th to the 28th, based on 64 steadily growing accounts. The latest month in a report is the live count at report time, not a monthly reading.
• Cold profiles. About 11% of reports had to be computed from scratch. Those accounts are almost never verified, and they show less loss in May and more later. Their history doesn't change on re-fetch, and the mechanism remains unclear, so we publish results with and without them.
• Band position. Our sampling tool returns results sorted by followers, with a cap on how deep it can page, so samples in large cells lean toward the top of a band. Within-band position has a small effect, and we report the sensitivity.
• Control quality. About a third of TikTok and YouTube monthly readings show exactly zero change, so their history is coarser than Instagram's. Compare them using the share losing more than 2%, not medians.
• Regional models. Audience estimates may work less well in some regions and languages. We don't present any country as having fake audiences.
• Accounts, not people. Accounts can't be de-duplicated within or across platforms.

What we will not publish

We don't name individual accounts as fake or inflated, and we don't publish modelled pricing as observed fact. Data is never gated behind a form.

Correction policy

If a published number changes, we fix it, keep the URL and record the old value, new value and date in the study's corrections log. To report a problem, write to support@leewou.com. We publish recurring studies at one URL and keep earlier editions available so old citations still resolve.

Change log