AI Search

What user ratings can tell you about AI answer quality

Read Pew’s chatbot health-information ratings with complete response tables, survey-form bases and a transparent reconciliation of reported totals and rounded cells.

Separate speech-bubble, magnifier and evidence-folder motifs on a calm editorial desk

Key takeaways

  • Pew’s survey measures how users perceive chatbot health information. It does not test whether individual answers are correct. [1] [2]
  • Convenience, accuracy, personalization and clarity use separate questions, with two different survey-form bases. [2]
  • Keep Pew’s reported 18% high-accuracy total alongside the 17% sum of its rounded response cells, with each clearly labeled. [1] [2]
  • Review what a saved answer says and whether its claims have suitable support separately from how readable or convenient it seems.

A user rating describes a perception of information; it cannot become a factual-accuracy score for an AI answer. Marketing and SEO teams need that distinction when explaining what AI visibility evidence can support. An answer may be easy to understand, a source may be cited, and a factual claim may still require verification. Each observation answers a different question.

Pew Research Center’s health-information survey offers a concrete case. It asked U.S. adults about several qualities of information they get from AI chatbots. Fieldwork ran from October 20 to 26, 2025, and the report was published on April 7, 2026. These historical findings concern health-information users in the relevant survey forms, including people who get that information rarely. They are not current ratings across every industry or a benchmark of individual chatbot products. [1] [2]

Four qualities, four different questions

The chart shows the report’s published shares rating chatbot health information extremely or very highly on each quality. Accuracy concerns users’ assessment of accuracy; convenience concerns getting the information; personalization concerns how tailored it feels; clarity here means ease of understanding. Keeping the labels intact prevents a positive rating on one quality from becoming an unsupported claim about another.

How chatbot health-information users rate four qualities

Pew Research Center, U.S., October 20–26, 2025. Published extremely/very totals; subjective ratings, not tested answer correctness.

483624120AccuracyConven.PersonalClarityQuality rated
  • Published high rating
Full data table
How chatbot health-information users rate four qualities · % of eligible users in each form
Quality ratedPublished high rating
Accuracy18
Conven.48
Personal23
Clarity41

Accuracy and convenience: Form 1, unweighted N=1,057. Personalization and clarity: Form 2, N=1,127. Eligible users include rarely-users; percentages are weighted. Clarity means easy to understand. Accuracy is the report’s 18% net; rounded Extremely and Very cells sum to 17%, shown separately in the full tables. Different forms do not support individual-level comparisons. Source: Pew Research Center, Washington, D.C., April 7, 2026. © 2026 Pew Research Center. Original Ansvra graphic. Conven. = convenience; Personal = personalization; Clarity = easy to understand. [1] [2] [3]

The source reports 18% as highly accurate, 48% as highly convenient, 23% as highly personalized and 41% as extremely or very easy to understand. Accuracy and convenience use Form 1, with an unweighted item base of 1,057 chatbot health-information users. Personalization and clarity use Form 2, with a base of 1,127. The percentages are weighted, so the bases are not counts to multiply by the displayed percentages to reconstruct respondent totals. [1] [2]

The chart preserves the reported accuracy total even though adding the rounded Extremely and Very cells below gives a different integer. That difference is disclosed in the reconciliation section. None of these bars represents the proportion of answers that passed a factual check.

The complete response tables

A high-rating total compresses a distribution. The following tables retain every printed response cell for the four selected chatbot questions, including Somewhat, the two lower ratings and No answer. The response words take the meaning of the quality being asked: Extremely in the accuracy table means extremely accurate, for example. All values are weighted percentages within the stated form and source-user base. [2]

Accurate

Form 1; unweighted item base 1,057. Topline page 9, HLTHACC, item g. [2]

ResponsePercent
Extremely4%
Very13%
Somewhat58%
Not too18%
Not at all5%
No answer1%

Convenient to get

Form 1; unweighted item base 1,057. Topline page 11, HLTHCONV, item g. [2]

ResponsePercent
Extremely22%
Very26%
Somewhat31%
Not too13%
Not at all7%
No answer1%

Personalized

Form 2; unweighted item base 1,127. Topline page 13, HLTHPERS, item g. [2]

ResponsePercent
Extremely9%
Very14%
Somewhat29%
Not too27%
Not at all20%
No answer1%

Easy to understand

Form 2; unweighted item base 1,127. Topline page 15, HLTHUNDSTND, item g. [2]

ResponsePercent
Extremely13%
Very28%
Somewhat40%
Not too12%
Not at all5%
No answer1%

In this instrument, No answer includes web respondents who skipped the question and telephone respondents who refused or did not know how to answer. Keep that category visible rather than recoding it as a negative rating. The six displayed values total 99% for accuracy and clarity, and 100% for convenience and personalization. The topline warns that rounded rows may not total 100%. We preserve the printed figures without rescaling them. [2]

Why the reported accuracy total is 18%, while 4 + 13 is 17%

Pew’s analysis reports 18% of chatbot health-information users rating the information highly accurate. Its detailed topline prints 4% Extremely accurate and 13% Very accurate. Adding those displayed cells gives 17%. Both values belong in a transparent account: one is the publisher’s reported combined total, and the other is Ansvra’s arithmetic from rounded components. [1] [2]

The table below reconciles every selected quality. Reported is Pew’s published extremely/very total. Cells is the sum of the two printed categories. Gap is Reported minus Cells, in percentage points. Row total adds all six printed response categories. These local calculations do not recover unrounded weighted data.

QualityReportedCellsGap (pp)Row total
Accurate18%17%+199%
Convenient to get48%48%0100%
Personalized23%23%0100%
Easy to understand41%41%099%

The one-point difference is compatible with rounding: independently rounded components need not sum to the rounded combined estimate. The public cells do not reveal the exact underlying weighted total. We therefore keep 18% in the reported-total chart, show 17% as the rounded-cell sum, and do not change a source cell or assert that the publisher made an error.

This is also why an automated report should retain the source location and calculation alongside a number. A seemingly simple correction can quietly replace a reported result with a different measure. When a net and its displayed components differ, label the distinction before deciding what the chart should show.

What the survey design permits you to say

The research comes from Wave 182 of Pew’s American Trends Panel. The overall survey included 5,111 U.S. adults, with SSRS conducting online and live telephone interviews in English and Spanish. The methodology describes weighting for selection and nonresponse and a target population of noninstitutionalized U.S. adults aged 18 or older. The selected rating questions have the smaller form-specific bases shown above. [3]

Eligibility matters as much as the sample size. The rating questions include respondents who get health information from chatbots rarely, as well as those who do so more often. They do not describe only frequent users, all U.S. adults or everyone who has used a chatbot for any purpose. And the two form groups must not be pooled as though each person answered all four questions. [2]

The selected public summaries contain no respondent-level joint table. They cannot tell us how many people simultaneously rated information convenient and accurate, or whether finding it convenient changed someone’s trust. They also contain no tested answer corpus, expert correctness assessment or clinical outcome. An 18% high-accuracy rating cannot be interpreted as 18% of answers being correct; its complement cannot be interpreted as an error rate.

Pew reports a margin of error for the full survey, but that whole-sample figure must not be attached to these smaller subgroups. This analysis does not calculate subgroup confidence intervals or claim statistically significant gaps between qualities. The study is useful for understanding how the questions and ratings differ, without turning those differences into a causal claim.

Pew produced the research, with SSRS conducting fieldwork. The report acknowledges support from The Pew Charitable Trusts, Pew’s parent and primary funder. The selected findings come from Pew’s original survey and public instrument. Our analysis retains the study’s health context; it does not extend these percentages to other sectors or offer a recommendation about health care. [1] [3]

Two hands compare an abstract answer card with a separate source page and clipped evidence through a magnifier
A clear answer and an available reference still leave a claim to check.

Keep perception, answer content and claim support distinct

For a brand team, the practical lesson is about labeling evidence. A satisfaction survey could tell you how a group evaluates an information experience. A saved answer shows what appeared for a particular question. Reviewing a material claim against a suitable source can establish what support is available for that claim. Combining those into a single quality label hides what was actually checked.

Consider a fictional question about a product’s return conditions. An answer might express the conditions in clear language. It might also include a reference to a policy page. Those observations leave a specific verification: does the referenced policy support the stated condition, for the relevant product and period? Clarity and citation presence do not complete that check. This example illustrates the evidence distinction without inventing a measured result.

The next diagram is Ansvra’s suggested reporting sequence. Start by naming the measure and its base, preserve the source value, then open the relevant evidence and state the remaining limit. It is an editorial decision aid, not a validated accuracy-scoring method.

Keep perception, answer content and claim support distinct

Ansvra’s suggested evidence-labeling workflow. This is editorial interpretation, not a validated medical or factual-accuracy scoring method.

  1. Label measure

    Identify whether a number describes user perception, saved answer content or checked claim support. The selected survey measures perceptions. [1] [2]

    Next: Record base

  2. Record base

    Keep the fieldwork, health-information context, form and source-user base with each percentage. Rarely-users are included. [2] [3]

    Next: Preserve source

  3. Preserve source

    Retain the reported net and all rounded cells. Explain 18% versus 17% without inventing unrounded values or correcting the publisher. [1] [2]

    Next: Open evidence

  4. Open evidence

    For an actual brand question, open the saved answer and its available references. A readable answer or a citation is not a correctness certificate. [1]

    Next: State limit

  5. State limit

    Describe what was observed and what remains unverified. Do not turn subjective ratings into the percentage of answers that are factually correct. [1] [2]

Pew supports the survey interpretation and source limits. The five-step review sequence is original Ansvra editorial guidance; it does not establish causal effects or certify answers. [1] [2] [3]

Use the references to make the next check concrete

AI Citations brings available cited URLs together with the saved answer and question context. That gives the team a page to investigate. A cited page is evidence of a reference appearing with that answer; its presence does not certify every surrounding statement.

Citation Source Analysis helps organize recurring pages and websites across the tracked questions. Use those connections to choose which material deserves attention, then inspect the source and the claim in context. Recurrence is a research lead, rather than proof of truth or an explanation of why a platform chose a source.

A review note can keep this boundary visible without becoming complicated. Record the question and saved answer, describe the relevant statement, identify the reference available, and say whether that reference supports the statement or leaves an unresolved question. When evidence is unavailable, retain that status rather than replacing it with a reassuring overall impression.

When presenting a survey finding alongside this work, keep its population, fieldwork and wording attached. Report a perception as a perception, an observed citation as a citation and a verified claim as a checked claim. That gives stakeholders a useful account of what the team knows and a precise next step for what it still needs to investigate.

Summary

User ratings can describe an information experience when their questions, bases and dates stay visible. They do not establish factual correctness for an AI answer. Preserve Pew’s reported totals and rounded response cells separately, then examine saved-answer claims and their available source support on their own terms. Use the evidence to define the next check before making a broader quality claim.

Sources

  1. Users of social media and AI chatbots for health information are more likely to say they are convenient than accurateAccessed
  2. ATP Wave 182: Health and Wellness Survey — toplineAccessed
  3. Health information: survey methodologyAccessed

Discussion

Join the discussion

Start the conversation

Share a question or tell us what worked for you.

Comments appear after review. How we handle comments

The Ansvra journal

Keep reading

Browse all articles

Measure AI shopping by task

Separate discovery, purchasing and service questions using Salesforce’s historical shopper survey, complete response tables and a practical AI visibility review workflow.

Ansvra

A clearer next step.

MEET YOUR AI GUIDE

A little clarity.
A clearer next move.

Ask about Ansvra, AI visibility, or your next step with content.

Opening chat…