Skip to content
Get in touch

An industry poll found 27% on the side of the restrictions. The open conversation found 57%.

Most of the gap comes from how the poll worded its questions. Thinkwell measured it in about 45 minutes, and every number on this page can be checked.

Data storyNovember 1, 2026Measured with Thinkwell
  • 57%of people who took a side online favored the restrictions, against 27% in the poll
  • 19 of 22poll questions are worded toward the industry; where the wording is neutral, the gap closes
  • 95%agreement when each reading was checked blind
  • 45 minfrom setup to confirmatory results

The issue, from the start.

In 2026, Congress has been weighing legislation that would put new rules on large online platforms and marketplaces, including limits on how those companies can favor their own products and services. Supporters say it would increase competition. Opponents say it would cost consumers features they use every day.

An industry group commissioned a poll on the bill. It asked 1,999 respondents 22 questions about it: how concerned they would be about its consequences, whether they would support it, and what they thought would happen if it passed. Among respondents who gave an answer, 27% came down on the side of the restrictions.

We asked a different question: when people talk about this bill in public, unprompted, which side are they on? We read Reddit, Bluesky and news coverage from August 1 to October 31, 2026, matched what people said to the same 22 poll questions, and counted only clear, explicit stances. Of 10,751 people in on-topic threads, 243 took a side, across 140 threads.

What this is, and is not

  • Not a voter sample. The conversation is the people who chose to speak. Every comparison here is benchmarked against the poll, never weighted to it.
  • Locked first. The hypotheses and their tests were written down and locked before any result was read.
  • Anonymous accounts. No one in the conversation was identity-checked. Authors are kept as keyed hashes; no comment text, usernames or handles appear here.
  • Every number checkable. Each figure on this page is read from one results file and carries its key in the page source.

Step 1 of 9

57% vs 27%

the open conversation against the poll, on the side of the restrictions

57% of people who took a side online favored the restrictions. The poll found 27%.

Among the 243 people who took a clear side in public, 57% were on the side of the restrictions, most likely between 50% and 65%. The poll found 27% of its decided respondents there. The gap is far larger than chance would explain (p = 0.0002).

BuiltBuilt and checked; not yet a product feature.
Share on the side of the restrictions, from 0 to 100%. The ring is the poll; the dot is the open conversation, with its 95% interval.
Show the numbers
SourceOn the side of the restrictions95% interval
The poll27%not reported
The open conversation57%50% to 65%
Why it matters

The poll asked everyone every question. The conversation records only what people chose to raise, so the two are different kinds of number: one is answers to a prompt, the other is unprompted stances. A gap this size says the poll is not describing the public conversation.

How it was made

Each person counts once per question they took a side on. Intervals come from resampling whole threads, so a busy thread cannot narrow them; that leaves an effective sample of about 162 people rather than 243.

Step 2 of 9

95%

of 161 pairs agreed on a blind second reading

A second, blind reading agreed with the first on 95% of 161 checks

An AI judge read every comment and decided whether it affirms, denies or qualifies each poll question, or takes no stance. To check it, a second reading of 161 comment and question pairs was drawn after the first and done blind. The two agreed on 153 of 161. Every disagreement is a stance against no stance: the readings never flip affirms into denies, the one error that would bias a share.

LiveRunning in the product today.
Each of 161 pairs read twice, blind. Rows are the first reading, columns the second; the diagonal is agreement.
Show the numbers
First readingAffirmsDeniesQualifiesNo stance
Affirms19003
Denies02802
Qualifies0020
No stance210104
Why it matters

Everything after this rests on reading stances correctly. A judge that confused sides would move the headline number; one that sometimes misses a stance only makes the sample smaller.

How it was made

The judge passed a fixed set of test cases before the run. The check pairs were weighted toward the pairs the first reading called a stance, where errors would matter most. A cheaper judge was also tried and did not agree well enough to be used; it stays a comparison, never the judge of record.

Step 3 of 9

7.8%

of people who took a side came from the busiest thread

The busiest thread holds 7.8% of the people, so no single crowd sets the result

The 243 people who took a side came from 140 threads. The busiest supplied 19 of them, and the top five together 22%. The study set a limit on any one thread before it ran, and this is well inside it.

BuiltBuilt and checked; not yet a product feature.
The five threads that supplied the most people who took a side, then the other 135 together, as a share of all 243.
Show the numbers
ThreadsPeople who took a sideShare
Thread 1197.8%
Thread 2145.8%
Thread 372.9%
Thread 472.9%
Thread 562.5%
Other 13519078.2%
Why it matters

If one thread dominated, its crowd could set the answer, and a single viral post would look like public opinion.

How it was made

Threads are counted by the people in them who took a side. The tests resample whole threads rather than people, which is why the interval in step 1 is wider than a simple count would give.

Step 4 of 9

3 of 3

families of questions lean further toward the restrictions than the poll

The conversation leans further toward the restrictions than the poll in all 3 families of questions

The poll's questions fall into three families: consequences of the bill, how it is framed, and whether to pass or block it. In each, the conversation sits well to the side of the restrictions compared with the poll, and each result holds after correcting for testing several at once (Benjamini-Hochberg). The two sides also argue differently: people against the restrictions argue mostly with facts, people for them mostly with policy (p below 0.0001).

BuiltBuilt and checked; not yet a product feature.
Each line is a 95% interval from resampling whole threads; the dot is the conversation and the tick is the poll's share for the same questions.
Show the numbers
HypothesisConversation95% intervalPollPeople
H1 All 22 questions57%50% to 65%27%243
H2 Pass or block78%66% to 90%33%54
H3 Consequences43%35% to 52%27%156
H4 Framing78%67% to 88%20%58
Why it matters

A gap in one family could be a quirk of a few questions. A gap in all three says the lean is general.

How it was made

Thread-cluster bootstrap, explicit stances only, with the hypotheses and corrections locked before results.

Step 5 of 9

rho 0.07

rank agreement between the conversation and the poll, question by question

Item by item, the conversation does not rank the questions the way the poll does

If the poll and the conversation were measuring the same thing, the questions the poll finds most popular would also be the most popular online. They are not: across the 11 questions with enough people, the rank correlation is 0.07 (p = 0.42), no better than chance. This is the one locked test that did not hold.

BuiltBuilt and checked; not yet a product feature.
Each row is one poll question, sorted by the poll's share. The ring is the poll; the dot is the conversation, larger with more people. Faint dots rest on fewer than 10 people.
Show the numbers
QuestionFamilyPollConversationPeople
How convenient these online services are to useConsequences53%29%7
How well your personal data and payments are protected from fraud and scamsConsequences52%38%29
How affordable online products and services are for youConsequences49%67%6
Industry innovation or government action: which comes closer to your viewFraming42%79%52
Support or oppose Congress passing the legislationPass or block42%88%43
Adopt rules like Europe's, even if it meant losing featuresPass or block41%30%10
Making digital advertising less effective or more expensive for small businessesConsequences32%none0
Made these services less convenient to useConsequences28%29%7
Features like maps and directions no longer appearing directly inside search and other apps, because companies could not combine their own servicesConsequences27%0%2
Reduced safety protections for children onlineConsequences25%42%31
Hurt small businesses that rely on these platformsConsequences24%40%5
Weakened security or fraud protectionConsequences23%38%29
Having to pay separately for perks that are now bundled together, such as free shipping, streaming, and storageConsequences22%none0
Fewer free or low-cost features than you have todayConsequences22%none0
Raised the prices you payConsequences20%62%16
Making it harder to keep children and teens safe onlineConsequences20%42%31
No laws that risk disruption without clear evidencePass or block17%100%1
A greater risk of scams, malware, or fraud from unverified outside apps being allowed onto smartphonesConsequences15%38%29
Higher prices on the products and services you buy onlineConsequences14%62%16
Weaker protection for your personal data and privacyConsequences14%41%75
Laws should apply equally to all companiesFraming11%67%6
Congress should focus on everyday costs, not new restrictionsFraming9%100%1
Why it matters

A shift in level with the same order would mean the poll is tilted. A different order means the poll and the conversation are answering different questions. Step 6 shows why.

How it was made

Some questions share one underlying claim, so they show the same conversation share; there are fewer independent points than rows. Questions with no stances online show the poll's ring only.

Step 6 of 9

19 of 22

poll questions are worded toward the industry

19 of the poll's 22 questions are worded toward the industry, and that is where the gap is

Two blind coders read every poll question and agreed on the framing of all of them (100% agreement): 19 lean toward the industry, 3 are neutral, and none lean toward the restrictions. Every consequence question opens with "impose new restrictions"; every support question opens with "harming the people who rely on these online services". Where the poll asks neutrally, the conversation and the poll roughly agree (-7 points). Where it leads, they part by 26 points.

BuiltBuilt and checked; not yet a product feature.
The average gap between the conversation and the poll, in percentage points, with its 95% interval. Zero means they agree.
Show the numbers
WordingQuestionsAverage gap95% interval
Toward the industry19+26 points+16 points to +36 points
Neutral3-7 points-24 points to +17 points

Read the questions, five ways

Coming soon

For every poll question we wrote a neutral version, an opposite-framed version with matching loading, and two close paraphrases. Highlighted words are the ones both coders flagged as loaded. These versions are written and coded; putting them in front of real people, at random, is the next wave.

  1. As the poll asked it

    Congress is considering legislation that would impose new restrictions on online platforms and marketplaces, including limits on how companies can offer their own products and services online. These proposed rules could cause consumers to lose easy access to convenient features and services they currently use. Here are some potential consequences of this legislation. How concerned are you, if at all, about each?
    Higher prices on the products and services you buy online

    Both coders: worded toward the industry
  2. Neutral rewrite

    Congress is considering legislation that would place new rules on large online platforms and marketplaces, including limits on how companies can offer their own products and services online. Supporters say the rules could increase competition and consumer choice; opponents say they could change or remove some features and services consumers use today. Here are some possible effects of this legislation. How concerned are you, if at all, about each?
    Possible changes in the prices of the products and services you buy online

  3. Opposite framing

    Congress is considering legislation that would impose new limits on the largest online platforms and marketplaces, including limits on how they can favor their own products and services over competitors’. These proposed rules could give consumers more choice and fairer prices by ending the unchecked power of a few giant companies. Here are some potential consequences of leaving these companies unchecked. How concerned are you, if at all, about each?
    Higher prices on the products and services you buy online because a few giant platforms face little competition

  4. Close paraphrase

    Congress is weighing legislation that would place new restrictions on online platforms and marketplaces, among them limits on how companies may offer their own products and services online. Under these proposed rules, consumers could lose easy access to convenient features and services they use today. Below are some potential consequences of this legislation. How concerned are you, if at all, about each?
    Higher prices for the products and services you purchase online

  5. Second paraphrase

    New legislation before Congress would add restrictions on online platforms and marketplaces, including limits on how companies can present their own products and services online. These proposed rules might leave consumers without easy access to convenient features and services they currently rely on. Here are some of the potential consequences of this legislation. How concerned are you, if at all, about each?
    Having to pay more for the products and services you buy online

Why it matters

This is the reason the study matters. The poll's {pct(h1['null'])} is partly an answer to its own wording. Read the questions below and judge for yourself.

How it was made

Gaps are averaged over questions with enough stances in the conversation, with intervals from resampling questions. Only 3 questions are neutral, so this comparison is exploratory, not confirmatory. The lead-in also pushes answers to the ends: the strongest answer category takes 52% of decided answers on the bill questions against 35% on the poll's everyday-use questions.

Step 7 of 9

52% vs 49%

strong answers in the poll and in the conversation

Strong answers are about as common online as in the poll. The side is what differs.

A poll answer records a side and a strength, such as "very concerned". Text carries both too. Each of 330 stances was coded as strong or moderate from its language, and a second AI model, coding blind, agreed on 91% of 120. Strong answers make up 52% of the poll's decided answers and 49% of the conversation's. The strength is the same; the side is not.

BuiltBuilt and checked; not yet a product feature.
Share of decided answers. Strength is about the same in the poll and the conversation; the side is not.
Show the numbers
MeasureThe pollThe conversation
Strong answers52%49%
On the side of the restrictions27%57%
Why it matters

This rules out a simple story where online talk is just angrier. The conversation is as strong as the poll and points the other way.

How it was made

Strength is coded from the language of each stance (all caps, exclamation marks and intensifiers are much more common in strong ones) on the 18 questions that have a strength scale.

Step 8 of 9

69%

of top-level comments posted in the first 15 minutes reached their thread's top quarter

69% of top-level comments posted in the first 15 minutes reached their thread's top quarter. After 3 to 12 hours, 22% did.

When to speak matters as much as what to say. Top-level comments written in the first 15 minutes after a post landed in their thread's top quarter by score 69% of the time; those written 3 to 12 hours in, 22%. Compared within the same thread, each doubling of delay costs about 23% of a comment's score, across 283 threads.

BuiltBuilt and checked; not yet a product feature.
Share of top-level comments that ended in their thread's top quarter by score, by how long after the post they were written.
Show the numbers
Posted after the postTop-level commentsReached the top quarter
0 min to 15 min74769%
15 min to 1 h1,14850%
1 h to 3 h1,25429%
3 h to 12 h1,40822%
12 h to 24 h50432%
24 h to 72 h26840%
After 72 h11532%
Why it matters

The comments that take a side tend to arrive hours in, after the window where reach is decided. Anyone who wants to be heard on an issue like this has minutes, not hours.

How it was made

Score was read once, when the data was collected, so this shows where reach ended up, not how it built. Within-thread comparison holds the topic and the crowd fixed.

You might wonder next

Step 9 of 9

45 min

from setup to confirmatory results

The study took about 45 minutes from setup to confirmatory results

A study like this usually takes a research team weeks. This one went from setup to locked, confirmatory results in about 45 minutes, on top of collecting the data earlier the same day. The compute was AI model tokens; no cloud functions ran.

BuiltBuilt and checked; not yet a product feature.
StageTokens
Stance judge, first reading1.18M
Stance judge, blind second reading64k
Cross-check by a second AI model75k
News full text, web search384k
A second reading drawn too early, not used69k
AI model tokens used at each stage of the run.
Why it matters

Speed changes what is worth asking. When a check like this takes an hour, it can run before a decision instead of after it.

How it was made

Token counts are the totals the session reported for each stage. Time is measured from attaching the data to the confirmatory results.

The method

How the personas stay honest: real people check them, every wave.

Thinkwell's AI personas predict how people would answer a question. A smaller group of verified people answer for real. The method combines the two, so the result is sharper than the verified group alone, and its error bars hold for the sample you actually have, not only in theory for very large samples. It is a mixed-subjects design, after van Loon and Kanopka (2026).

It works under three conditions: the verified people are picked at random from the group you want to describe, which of them answer is decided at random, and the personas' predictions are fixed before the answers come in.

The rolling benefit. Every new wave of verified people re-measures how closely the personas track real people. When the personas are good, each verified person counts for more, so the same confidence costs fewer people. When the personas drift, the error bars widen on their own and the estimate stays unbiased: a wrong persona cannot quietly make the number wrong. The cost of a decision-grade answer falls as the personas improve, and drift shows up instead of hiding.

In simulation, the calibrated estimators cut sampling noise by up to 35%, with zero bias, against a plain estimate from verified people alone.

BuiltBuilt and tested in simulation. No live panel of verified people has run yet; the first will be the Pulp team.

Every number can also carry how its people were verified: real ID, video, phone, email, public figure, attested, or anonymous. That labeling is Built. The conversation on this page is anonymous.

What a verified person is worth

Illustration, not a result

At this setting, about 64 verified people give the same error bar as 100 without personas.

A teaching sketch: verified people needed for the same error bar, out of 100, using the textbook rule that a predictor correlated r with the answer removes r squared of the noise. It ignores real-world costs and is not a study result.

Measured with Thinkwell.

Thinkwell is Pulp's product for reading public conversation the way a poll reads a sample: who took a side, how strongly, on which claim, and how sure we can be. The short version of this study, for teams who make decisions on issues like this one, is on thinkwell.cloud.

How every number was made

Sample
Reddit, Bluesky and news headlines, August 1 to October 31, 2026. Not a voter sample.
Identity
Anonymous: social accounts, no identity check.
Status
Confirmatory for the locked hypotheses; the wording analysis is exploratory.
Locks
Hypotheses, tests and corrections written down before any result was read.
Privacy
No comment text, usernames or handles. Authors are keyed hashes.
Not tested
Social against news: only 6 news outlets took a stance, too few to test.

Want an issue of yours measured like this?

Tell us the question. We bring the data and the method.

Get in touch

Pulp

Color mode

Roll into a random page on Pulp. A few hidden pieces can only be accessed by taking a chance.

Get in touch

Color mode

Kept on this device only.

What are you trying to understand?

So we know who to write to.
An address we can reply to, like name@example.com.
A question, a dataset, an argument that keeps going in circles.

Get in touch
Roll into a random page on Pulp. A few hidden pieces can only be accessed by taking a chance.