Skip to content
AboutJoin the waitlist

Research · Method

How we built the Baltic LinkedIn Substance Index

Every number on the index page comes from one fixed process. This article walks through it step by step: the company list, the posts, the blind panel of five models, the checks we ran and the rules we publish under.

Companies checked
111
Posts read
2,638
Model labs
5
Panel agreement
alpha 0.74

Published 2 November 2026 by Byro · Read the index

The process at a glance

The diagram follows 7 milestones from 111 company pages to a ranked table with 4 tiers. Likes and follower counts never enter a score. The method is explained below, with the numbers from the Q4 2026 edition.

  1. 111Companies checkedEstonia, Latvia and Lithuania
  2. 2,638Posts collected90 days to 26 October 2026
  3. 540Set asideHiring and culture posts, shown separately
  4. 5Models score blindFive labs, company names masked
  5. MedianOne score per postMiddle of five, scaled 0 to 100
  6. 10Pull toward averageCompany average drawn toward the mean
  7. 4Tiers18 Leaders named
Figure 1. From company list to tiers. 111 companies were checked and 91 ranked. Only the 18 companies in the Leaders tier are named. Every other company is published as a count.

Step 1. The company list

We started with 111 Baltic business-to-business companies founded in Estonia, Latvia or Lithuania, each with a LinkedIn company page. The list is our selection. It is not a census, and some sectors hold only a few companies.

We recorded each company's headquarters and the industry shown on its LinkedIn page. The listed industry sets the sector. Where it was missing or unclear, we assigned the sector by hand. Country is the country of founding, so a company that moved its headquarters abroad still counts for the country it started in.

Step 2. Collecting the posts

The window is 90 days to 26 October 2026. Only public company-page posts enter a company's score. We collected profile posts separately for the leaders analysis.

Public company-page and profile posts collected through third-party LinkedIn data providers (accessed via an API marketplace), not through Byro accounts.

Some posts may be missing from a provider's copy, and like and comment counts keep changing after collection. The counts on the index page are the counts as collected.

Step 3. Sorting

Not every post tries to say something about the company's field. Job ads, team events and employer branding are a different kind of post. The panel labelled them hiring and culture, and we set them aside: 540 of the 2,638 posts, 20%. They are shown separately on the index page and never scored.

A company needs at least 5 non-hiring posts in the window to be ranked. 91 companies cleared that bar. 20 did not, and they appear on the index page only as a count.

Posts2,638
Companies111
Figure 2. What was set aside. Hiring and culture posts are shown on the index page but do not affect any score. Companies with fewer than 5 non-hiring posts are counted but not ranked.

Step 4. Blind scoring by a panel of five

Posts admitted to scoring were assessed by five language models from five different labs. We masked company names in the text of company-page posts and leader names on leaders' own posts before scoring. The short image descriptions were not masked. In the rerun, 45% of posts had one that named the company, and within each company those posts scored no differently. Each model scored on its own and never saw another model's score.

The models scored three criteria from 0 to 3 each. S is specificity and evidence. E is expertise. G is generic style, inverted, so a distinctive voice scores high and template text scores low. A post's score is the median of the five totals, scaled from 0 to 9 to 0 to 100. With a median, one unusual reading cannot move a post on its own.

A company pageIllustrative post, written by us

We moved document checks for new freight customers to the first call. Onboarding time fell from 14 days to 6 this quarter. The checklist is in the comments.

  1. Specificity and evidence2 of 30 slogans, nothing concrete3 claims a reader can check
  2. Expertise2 of 30 unrelated to the field3 know-how only an expert would write
  3. Generic style, inverted2 of 30 template or filler3 a voice that would be missed

Total6 of 9Post score67 of 100

Figure 3. The rubric applied to an illustrative post we wrote for this article. It is not a post from the index. Each criterion runs from 0 to 3. A total of 6 scales to 67 of 100.
  1. ClaudeClaude Sonnet 5.5 scores 6 of 9
  2. GeminiGemini 3.5 Flash scores 7 of 9
  3. gpt-ossgpt-oss-120b scores 6 of 9
  4. GLMGLM 5.3 scores 5 of 9
  5. KimiKimi K2.7 scores 6 of 9

Panel median6 of 9scaled to 67 of 100

Figure 4. How five blind scores become one. The illustrative totals 5, 6, 6, 6, 7 have a median of 6, which scales to 67. The panel also labelled each post with one tone, used for the findings on the index page.

The model panel

Five models, five labs. We chose models from different labs so that one lab's habits could not set the index. One of the five is an open-weights model.

  • Claude Sonnet 5.5Anthropic
  • Gemini 3.5 FlashGoogle
  • gpt-oss-120bOpenAIopen weights
  • GLM 5.3Zhipu (Z.ai)
  • Kimi K2.7Moonshot AI

Logos identify the models we used. No lab sponsored or reviewed this research.

Gemini 3.5 Flash also did the comment classification described under Engagement.

Step 5. Checking the panel

Agreement across the five models was Krippendorff's alpha 0.74. We then compared the panel with two outside readings. Byro's founder scored 10 posts blind, without seeing the model scores, and agreed with the panel at r = 0.93. That is a small sample from one reviewer, so it is a sanity check, not proof. A separate AI judge that was not on the panel scored 50 posts blind and agreed at r = 0.80.

Panel agreement
α 0.74Krippendorff's alpha across five models
Founder check
r 0.9310 posts, one reviewer, scored blind
Separate AI judge
r 0.8050 posts, scored blind
Figure 5. Three agreement figures, each on a 0 to 1 scale. Alpha measures agreement among the five models. The two r values compare the panel median with an outside reader. The founder sample is 10 posts from one reviewer.

Post length

Longer posts scored higher in the founder sample and the model panel. The correlation between post length and panel score is r = 0.56. Long posts scored 38 points higher than short posts in the founder sample and 34 points higher in the panel. We kept length in the score, because a longer post has more room for specifics, and we computed a length-adjusted rank as a check. 15 of the 18 Leaders are in the top 18 on that rank, and none is lower than 24th.

Step 6. From post scores to a company score

A company's score is the average of its non-hiring post scores, pulled toward the overall average. The overall average is weighted as if it were 10 posts. A company with 5 strong posts moves only part of the way from the average. A company with 30 posts keeps most of its own average. This stops a short, lucky run from topping the table.

  1. Small company5 posts
  2. Large company30 posts
Figure 6. Small-sample adjustment with illustrative numbers. Both companies average 85 on their own posts. With the overall average set at 60 for this example, the company with 5 posts lands at 68 and the company with 30 posts at 79.

Tiers

Ranked companies are split by score. Leaders are the top 20%, Strong the next 30%, Average the next 30% and Generic the bottom 20%. In this edition that is Leaders 18, Strong 27, Average 27, Generic 19.

Ranks close to a tier line are within run-to-run noise. When we reran the whole panel on 2 November, 15 of the 18 Leaders stayed in the tier. The other three placed 19th to 22nd, each within 1.2 points of the line.

Engagement, reported beside the score

Likes, comments and follower counts never enter a score. We report them beside it, because the gap between the two is one of the main findings. The correlation between a post's likes and its substance score is 0.02.

For comments, we pulled every comment on scored posts that had at least one: 2,872 comments on 1,188 posts. A model classified each commenter from their public headline, such as own staff, decision-maker elsewhere, practitioner, vendor or recruiter, and each comment by kind, such as substantive, social, self-promotion or link only. This classification is automated and not hand-checked.

Of the classified comments, 48% were social and 32% substantive. At the median company, 19% came from its own staff. Expert posts received 1.0 comments per post from outside the company, compared with 2.3 for celebratory posts.

About one in five commenters were decision-makers at every substance level. More substantive posts did not draw a larger share of decision-makers in this sample.

Engagement rates are given per 1,000 followers, so a small page and a large page can be compared.

Leaders' own posts

We also looked at what company leaders post themselves. For each company we looked for up to three leaders, the CEO or founder, the CTO or CPO and the head of marketing, through LinkedIn people search and web search. An AI check of the profile and the search evidence confirmed each one, and roles were left empty where the evidence was unclear. 97 leaders were confirmed; 53 did not post in the 90 days and 22 posted five or more times.

Leaders' posts received a median of 49 likes, against 23 for company-page posts. Expert-tone leaders' posts scored 81 on average, against 75 for expert company-page posts. Only 17 companies had enough posts on both sides to compare directly; company pages scored higher in 13. That sample is too small for a general claim about which side posts more substantively.

Their names were masked before scoring, and their posts are reported only in aggregate. No individual is named on the index page, in the report or here.

What we publish, and what we do not

  • Only the 18 companies in the Leaders tier are named, with their scores.
  • Every other ranked company appears as a count in its tier. Any ranked company can ask for its private scorecard. Scorecards are not published.
  • Companies that were checked but had too few posts are a count only.
  • No individual is named anywhere.
  • The index measures what a post says. It does not measure what the post did for the company.
  • Corrections are requested and logged on the index page, and we publish every correction we make.

Read the indexRequest a private scorecard