Ask a model the same question twenty times and it can give you several different answers. If 14 of those 20 pick the same option, that is 70%, and the honest range around it runs from 48% to 85%. Everything on this page follows from taking that seriously.
A Mind Polling result describes the replies selected models gave through their APIs, to a recorded question, under recorded settings, on stated dates. This page explains each part of that sentence, so you can decide how far to trust a result and how to describe it in print. How it works is the short version.
"AI polling" often means something else: prompting a model with a made-up person ("a 45-year-old nurse in Ohio") and treating its answer as what such people think. Mind Polling does not do that. It asks each model as itself and reports what the model says.
You can write a wording as a kind of asker, such as "As a small business owner in Ohio, ...". That tests how a model answers that asker. It is never shown, here or in an export, as what small business owners think.
Each model receives your wording and one instruction for the kind of answer you asked for, and nothing else: no system prompt, no conversation, no persona of ours. For a multiple-choice question:
Which ocean is larger?
Options:
A. The Atlantic Ocean
B. The Pacific Ocean
Answer with the letter of exactly one option, and nothing else.The options are shuffled for every call, and the order shown is recorded with the answer, so the first option never wins for being first. A scale question ends like this:
How much do you agree: social media has been good for democracy.
Answer with a single whole number from 1 (Strongly disagree) to 5 (Strongly agree), and nothing else.A number question names its unit and never shows any limits you set, so the limits cannot nudge the answer:
What will the official 2027 Alaska Permanent Fund Dividend amount be?
Answer with a single number in US dollars, and nothing else.A written answer is asked for in at most 150 words:
What should a small city spend its transport budget on first, and why?
Answer in at most 150 words.These are prompt template v1. If these instructions ever change, the version changes with them, and results from different versions are never combined.
The chat apps people use add their own instructions, memory and often web search to what the model sees. Mind Polling measures the model as its maker offers it to developers, under settings recorded with every answer. The two can differ, and the difference is on the record:
Reading the app instead would not show "what people see" either: an automated session is logged out, with no memory and no history. Each method measures a stated stand-in. This one's procedure can be repeated exactly, though the replies themselves will vary, which is the point.
Web search. You can turn search on for any model that offers it, with up to 1, 2 or 3 searches per answer. A model with search on reads current pages before it answers; one with search off answers from what it learned in training. The sources a model cites are kept with its answer when the provider returns them, and a missing list is a gap in what the provider reported, not proof nothing was read. Keep search the same across runs you mean to compare.
A model asked the same question twice can answer differently. So each question is asked several times per model (you choose how many), and every share comes with its sample size and a 95% range, a Wilson score interval. A narrow view, one wording and one model, rests on fewer answers, so its range is wider: the same 70% from 200 answers runs from 63% to 76%.
A 2026 study of AI search answers concluded that results "must be reported with uncertainty estimates", because many apparent differences fall inside the noise (Sielinski, arXiv 2603.08924).
What the range does not cover: it measures sampling, under its own assumptions. It does not include wording effects, a provider's unannounced changes, or the fact that two samples from one model are not independent people. More samples narrow it; they do not repair a badly framed question. And it is not the chance that an answer is true.
Number answers show a mean with a range from the spread of the answers. A small sample or one odd answer can mislead it, and a single number answer has no spread at all: read a range of zero width as missing, never as certainty.
Every call ends as one of five outcomes. A share is computed from the Answered replies only; the other four are counted and shown beside it, never quietly assigned to an option.
"Answered" means the reply was in the form asked for. It does not mean the answer is correct.
A run that did not happen (the worker was down, or the balance could not cover it) is a gap on the chart, never a zero.
Every answer records the model the provider says actually replied, beside the one asked for. Providers sometimes route one model's name to another; when they differ, the answer is flagged, not quietly merged. Models reached through OpenRouter also record the host that served them. What a provider reports is all anyone outside can see: a change it makes without changing the model's name cannot be detected this way, which is one more reason to compare runs rather than trust one.
Models do change: a 36-week study of political answers found a drift that often coincided with new version releases (Nogueira et al., "Changing Tunes", ICML 2026).
Once a poll has run, changing its wording, its options, its models or its answer type makes a new version (before its first run, an edit simply changes it). The new version starts a new series, shown with a break on the chart, so a change you made is never read as a change in the models.
The answer type counts. One study found the best systems answered news questions correctly over 90% of the time as multiple choice, and lost 11 to 13% under free-response evaluation, in the paper's own terms (Suzgun et al., arXiv 2605.22785).
A poll runs once, daily or weekly, at a time of day in its own time zone, between a start date and an end date at most 90 days apart. It stops on its own after its last run. You can also run it once at any time, which does not change its schedule.
| The mistake | The rule that blocks it |
|---|---|
| A change of wording read as a change in the models | Every change starts a new version, never combined with the old |
| The first option winning because it is first | Options are shuffled on every call, and the order is kept |
| One lucky answer read as a finding | Several samples, and a 95% range on every share |
| A model quietly replaced by another | The model that answered is recorded and flagged when it differs |
| A missed day read as a collapse | A missing run is a gap, never a zero |
| A failed call counted as an answer | Errors are their own outcome, never counted and never charged |
These are the rules the software follows today, not plans. None of them makes a result true; they make it checkable.
Say what you measured, in the terms you measured it: "Asked 20 times on Nov. 3 through its API, Model X chose B 14 times (70%, range 48% to 85%)", not "Model X prefers B". Give the question and its wordings, the options, the models, the dates, the settings, the sample sizes and what was left out (refusals, unreadable replies). Name the poll version. A change between runs is an observation; its cause needs more evidence. And keep the export behind the claim.
Every answer downloads as CSV or JSON, with the prompt, its hash, the settings sent, the option order, the raw reply, the model that answered, the tokens and the charge: enough to run the same procedure again and to check every number on the page.
The studies cited above motivate the method; none of them validates Mind Polling. Two more are worth reading: Chen, Zaharia and Zou measured GPT-3.5 and GPT-4 changing between March and June 2023 (arXiv 2307.09009), and Polo and colleagues study how results move across prompt templates (arXiv 2405.17202).
Is a result a forecast? No. It records what models say. A share of answers is not a probability that something will happen.
Does it show what ChatGPT tells people? Not exactly. It shows what the model behind it answers through the API, under stated settings. See Through the API, not the chat app.
If every model agrees, is the answer right? No. Models trained on similar text can share the same mistake. Agreement is consistency, not truth; check facts against a record.
Are two runs different if their ranges do not overlap? That is a useful first look, not a full test, especially when you compare many things at once. Read the answers behind both runs.
Can I cite a result? Yes. Cite the poll, its version and the dates of the runs, and keep the export so a reader can check it.
Other questions, about the API, public polls and payment, are on the FAQ.