Codebook: Trust in AI Commerce Report, Edition 2

This codebook documents the public dataset for Edition 2 of the Trust in AI Commerce Report, fielded August 27 to September 1, 2026. It lists every published column with its question wording, response options, scale anchors and the respondents who saw it. With this page and the dataset, a reader can reproduce the report’s figures except the open-text coding and the most- and least-trusted assistant grid (see Data notes).

Edition 1 (published June 2026) has its own codebook and dataset. The two files use different column sets and are not interchangeable.

Study summary

For the full methodology, see the Methodology section of the report.

File structure

The dataset is a UTF-8 CSV with 1,493 rows and 135 columns. Each row is one complete response. Disqualified and partial records are not included, so no filter is needed to reach the analysis basis.

Columns removed or recoded before publication

The raw survey export has 152 columns. Before release, every column’s values were profiled, and columns that could identify a person, a device or a session were removed. Open-text answers were removed because a respondent could be identified by what they wrote.

How to read this codebook

Column headers are the question wording as exported from the survey platform, including the question number (for example 12. Have you used an AI assistant…). Question numbers follow the instrument. Gaps in the numbering are instrument items that do not appear as columns in the raw export.

• Matrix questions (grids) have one column per row of the grid. The header is the row label, a colon, then the question. The cell holds the answer chosen for that row.
• Multi-select questions have one column per option. The header is the option, a colon, then the question. The cell holds the option label when selected and is blank otherwise.
• Blank cells in a single-select or grid column mean the question was not shown to that respondent (see Universe), or, for a handful of cells, that it was skipped.
• Characters are exported as typed in the instrument, including curly apostrophes and the mixed dashes in the age bands (18–24, 25–34 use an en dash; 35-44 and older use a hyphen).

Bases

The report uses four bases. Each can be rebuilt from two columns.

Every comparison with Edition 1 uses the trend base, because Edition 1 screened out store-only shoppers. The AI-user blocks (Q13 onward, except Q22 to Q24) were shown only to AI users, so they are blank for the other 943 rows.

Routed sub-bases used in the report: verifiers are AI users with Q13 = “Yes - always” or “Yes - sometimes” (493); error-finders are verifiers whose Q17 answer is not “Never” (443); completers are AI users with Q44 = Yes (119); the located subsample is the 1,008 rows with a Region.

Scale encoding: map the worded anchors before averaging

The survey platform exports most rating scales as a mix of words and numbers: the anchor points carry their label and the points between them carry a number. A mean computed on the numbers alone silently drops every anchor answer. Map each label to its scale point first. Every mean in the report was computed this way.

1-7 grids and scales: anchors at 1, 4 and 7

0-10 items: anchors at 0, 5 and 10

Fully worded 1-7 scales

These export every point as words. Map them in this order, 1 to 7.

Remap in Python

ANCHORS_1_7 = {"Not at all confident": 1, "Do not trust at all": 1, "Neutral": 4,
               "Very confident": 7, "Trust completely": 7, "Extremely confident": 7}
ANCHORS_0_10 = {"Not at all confident": 0, "Neutral": 5, "Extremely confident": 10}

def to_num(value, anchors):
    value = value.strip()
    if value == "":
        return None
    return anchors[value] if value in anchors else float(value)

Apply ANCHORS_1_7 to the Q9, Q11, Q22 and Q24 columns and ANCHORS_0_10 to Q5. For Q33, Q34 and Q42, map each label to its position in the tables above.

Anchor statistics

These are the report’s headline figures, each recomputed from this CSV. A replication that matches them has the bases and the scale mapping right.

Variables

A. Administrative

B. Age (Q1)

Column 4: 1. What is your age?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)
• Notes: Respondents under 18 were screened out. 45-54 corrects Edition 1’s overlapping “44-54” label.

C. Shopping mode (Q2)

Column 5: 2. Where do you make your purchases when you shop for products?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)
• Notes: Defines the trend base: rows other than “I shop only in physical stores”. Edition 1 screened store-only shoppers out; Edition 2 kept them as a segment (246).

D. Categories purchased, past 12 months (Q3)

Question: “Which of the following have you purchased for yourself or your household in the past 12 months? (Select all that apply)”

• Type: Multi-select; the cell holds the option label when selected, blank otherwise
• Universe: All respondents

E. Category of the most recent $50+ online purchase (Q4)

Column 24: 4. Think about the last time you bought a product online that cost $50 or more. From which category was that purchase?

• Type: Single-select
• Universe: All respondents (1,489 non-blank)
• Notes: Not applicable (214) includes 100 store-only shoppers.

F. Last-purchase confidence, 0-10 (Q5)

Column 25: 5. For that purchase, how confident were you that you found the best product for your needs before completing the purchase?

• Type: 0-10 scale, anchors exported as words
• Universe: All respondents (1,487 non-blank)
• Notes: Map Not at all confident = 0, Neutral = 5, Extremely confident = 10. Trend-base mean 7.03 (n=1,243).

G. Sources used to research that purchase (Q7)

Question: “Which of the following did you use to research that purchase? (Select all that apply)”

• Type: Multi-select; the cell holds the option label when selected, blank otherwise
• Universe: All respondents
• Notes: The “Other” write-in text is not published.

H. Category research confidence, PRCI (Q9)

Question: “In general, how confident are you in your ability to find the best product when shopping online in each of the following categories?”

• Type: Grid, one column per row
• Response set: Not at all confident, 2, 3, Neutral, 5, 6, Very confident (1 to 7)
• Universe: All respondents (1,493 per column)

I. Retailer statement, 0-10 (Q10)

Column 53: 10. How much do you agree or disagree with the following statement: 'Online retailers always have my best interest as a customer in mind.'

• Type: 0-10 scale, exported as numbers
• Universe: All respondents (1,493 non-blank)
• Notes: 0 = “Completely disagree”, 10 = “Completely agree”. Net Retailer Trust Score = share answering 9 or 10 minus share answering 0 to 6.

J. Source trust hierarchy (Q11)

Question: “When researching a product you may buy online, how much do you trust information from each of the following sources?”

• Type: Grid, one column per row
• Response set: Do not trust at all, 2, 3, Neutral, 5, 6, Trust completely (1 to 7)
• Universe: All respondents (1,493 per column)

K. What “verified” means (Q11a, open text)

Every respondent was asked: “When a shopping site or an AI assistant says a product claim is ‘verified,’ what do you think that actually means?” The answers are not published (see File structure). Report §9.1 gives the coded shares. Representative answers, quoted as typed, one for each theme:

L. Trust in AI recommendations by category (Q22)

Question: “How much do you trust AI-generated product recommendations in each of the following categories?”

• Type: Grid, one column per row
• Response set: Do not trust at all, 2, 3, Neutral, 5, 6, Trust completely (1 to 7)
• Universe: All respondents (1,493 per column)
• Notes: Asked before the AI-use question so it stays comparable with Edition 1 and can be cut by AI use.

M. Most you would spend on an AI pick without checking (Q23)

Column 79: 23. What is the most you would spend on a product recommended by an AI assistant without checking any other source first?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)
• Notes: Asked of everyone, before the AI-use question.

N. Review trust (Q24)

Column 80: 24. How confident are you that online product reviews accurately reflect product quality?

• Type: 1-7 scale, anchors exported as words
• Universe: All respondents (1,493 non-blank)
• Notes: Map Not at all confident = 1, Neutral = 4, Extremely confident = 7.

O. AI use for product research, past 90 days (Q12)

Column 81: 12. Have you used an AI assistant to help research a product you were considering buying in the past 90 days?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)
• Notes: The gate for every AI-user block. Yes = 550 (517 on the trend base).

P. Verified the AI recommendation (Q13)

Column 82: 13. Did you verify the AI assistant’s recommendation through another source before making a purchase?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

Q. First place checked (Q14)

Column 83: 14. The last time you double-checked an AI assistant’s recommendation, where did you look first?

• Type: Single-select
• Universe: AI users who verify (495 non-blank)
• Notes: Non-blank for the 493 verifiers and 2 respondents who answered No at Q13. The report’s base is the 493 verifiers.

R. How often checking found the AI wrong (Q17)

Column 84: 17. How often has checking shown you that the AI’s answer was wrong or out of date?

• Type: Single-select
• Universe: AI users who verify (495 non-blank)
• Notes: Non-blank for the same 495 rows as Q14. The report’s base is the 493 verifiers.

S. What you did after finding the AI wrong (Q18)

Column 85: 18. The last time you found the AI’s answer was wrong, what did you do next?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)
• Notes: Shown to all AI users. The report’s §2.5 base is the 443 error-finders.

T. Did the AI save you money (Q19)

Column 86: 19. Thinking about your most recent AI-assisted purchase — do you believe the AI saved you money?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

U. Did it find the best available price (Q20)

Column 87: 20. Do you believe it found you the best available price?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

V. Assistants used for product research, past 90 days (Q25)

Question: “Which AI assistants have you used to research a product in the past 90 days? (Select all that apply)”

• Type: Multi-select; the cell holds the option label when selected, blank otherwise
• Universe: AI users (Q12 = Yes)
• Notes: The write-in text is not published. Grok and Meta AI recurred among the write-ins (report §3.1). The most- and least-trusted assistant grid in report §3.2 is not in the raw export and is not in this file.

W. Noticed a sponsored recommendation in an AI answer (Q30)

Column 97: 30. Have you ever noticed a sponsored or paid recommendation inside an AI assistant’s answer?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

X. Trust if a recommendation is clearly labeled sponsored (Q31)

Column 98: 31. If an AI assistant clearly labelled a recommendation as sponsored, how would that change your trust in it?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

Y. Trust if a paid recommendation was not disclosed (Q32)

Column 99: 32. And if you found out afterwards that a recommendation had been paid for, without being told, how would that change your trust?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)
• Notes: Disclosure Premium = share answering “Somewhat less” or “Trust it much less” at Q32 minus the same share at Q31 (277 of 550 minus 219 of 550, or 10.5 points).

Z. Believed accuracy of AI shopping answers (Q33)

Question: “How accurate do you think AI shopping answers usually are about each of these?”

• Type: Grid, one column per row
• Response set: Completely Inaccurate, Mostly Inaccurate, Somewhat Inaccurate, Neutral, Somewhat Accurate, Mostly Accurate, Completely Accurate (1 to 7)
• Universe: AI users (Q12 = Yes) (550 per column)

AA. “AI shopping answers are unbiased” (Q34)

Column 103: 34. How much do you agree with the statement “AI shopping answers are unbiased — not influenced by advertisers or brands.”

• Type: 1-7 scale, every point worded
• Universe: AI users (Q12 = Yes) (550 non-blank)

AB. Knew an AI answer can repeat brand marketing (Q35)

Column 104: 35. Before today, did you know an AI answer can repeat a brand’s own marketing claims as if they were facts?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

AC. Agent tasks allowed without asking first (Q36)

Question: “Would you let an AI shopping agent do each of these without asking you first?”

• Type: Grid, one column per row
• Response set: Yes - without asking, Yes - only if it asks first, No, never
• Universe: AI users (Q12 = Yes) (550 per column)

AD. Most an agent should spend without checking (Q37)

Column 112: 37. What is the most an AI agent should be able to spend without checking with you?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

AE. Give an agent a payment method (Q38)

Column 113: 38. Would you give an AI agent access to a payment method?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

AF. What holds you back (Q39)

Question: “What holds you back? (Select all that apply)”

• Type: Multi-select; the cell holds the option label when selected, blank otherwise
• Universe: AI users (Q12 = Yes)
• Notes: The write-in text is not published.

AG. Who is at fault if an agent buys the wrong thing (Q41)

Column 120: 41. If an AI agent bought the wrong thing with your money, who would be most at fault?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

AH. Comfort with a monthly agent allowance (Q42)

Column 121: 42. Suppose you could give an AI agent a set monthly amount to spend on your behalf. How comfortable would you be doing that?

• Type: 1-7 scale, every point worded
• Universe: AI users (Q12 = Yes) (550 non-blank)

AI. Let an AI assistant complete a purchase (Q44)

Column 122: 44. Have you ever let an AI assistant complete a purchase for you, start to finish?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)
• Notes: Completers are Q44 = Yes (119).

AJ. Assistant that completed the purchase (Q45)

Question: “Which assistant did you let complete a purchase for you?”

• Type: Multi-select; the cell holds the option label when selected, blank otherwise
• Universe: AI users (Q12 = Yes), routed on Q44
• Notes: 124 rows have a selection: 118 of the 119 completers and 6 who answered No at Q44. The report’s shares use the 119 completers (Q44 = Yes).

AK. Already used AI for holiday shopping (Q48)

Column 130: 48. Have you already used an AI assistant for any holiday shopping this year?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

AL. Expect to use AI for holiday shopping this season (Q49)

Column 131: 49. Do you expect to use an AI assistant for holiday shopping this season?

• Type: Single-select
• Universe: AI users (Q12 = Yes) (550 non-blank)

AM. Gender (Q50)

Column 132: 50. What is your gender?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)

AN. Household income (Q51)

Column 133: 51. What is your annual household income before taxes?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)
• Notes: The report groups these into four bands: under $30,000; $30,000-$59,999; $60,000-$100,000; over $100,000.

AO. Ethnicity (Q52)

Column 134: 52. What is your ethnicity?

• Type: Single-select
• Universe: All respondents (1,493 non-blank)
• Notes: The “Other” write-in text is not published.

AP. Census region

Column 135: Region

• Type: Derived
• Derivation: Census region (Northeast, Midwest, South, West) of the U.S. state the survey platform assigned by passive geolocation. Blank when the respondent could not be located in the United States.
• Universe: All respondents (1,008 non-blank)

Data notes

• Open text. The Q11a and Q29 answers and the “Other” write-ins are withheld to protect respondents. The coded shares in report §3.3 and §9.1 cannot be recomputed from this file.
• Most- and least-trusted assistants. Report §3.2 comes from the survey platform’s standard report. That grid is not in the raw export; it will be added to the dataset if supplied.
• Weighting. The data are unweighted, as is the report.
• Location. 98.9% of completes geolocated to the United States. Region is published for the 1,008 respondents located to a U.S. state.

Citation

Product.ai Research. The Trust in AI Commerce Report, Edition 2 [Data set]. Product.ai, 2026. https://product.ai/research/trust-in-ai-commerce-report/. License: CC BY 4.0.