Back-to-School 2026: America's First Agentic Shopping Season
Published July 7, 2026 · 6,500 simulated respondents · fielded in 9 minutes on production infrastructure
Every July since 2003, the National Retail Federation and Prosper Insights & Analytics have surveyed thousands of American families about back-to-school spending. This year we ran the question on 6,500 simulated Americans first — and froze our numbers before theirs come out, so the two can be compared in public. We also asked what no annual tracker fully covers yet: whether Americans are ready to let AI shop for them. This is a synthetic panel, not a probability sample, and no humans were interviewed.
Headline findings · frozen predictions
$857–870
estimated mean K-12 back-to-school spend — Method A $856.83, Method B $870.37 — pre-registered Prediction 1, frozen July 7.
92–93%
of parents in our frozen panel plan to shop mid-July deal events like Prime Day for school items — pre-registered Prediction 3, frozen July 7.
92%
of US adults say an ad made mostly by AI would make them trust it less — stable across both runs.
63–70%
would nonetheless let an AI assistant complete a purchase for them within a spending limit — the agentic paradox.
+53 pts
how much one sentence — telling the persona what day it is — moved the 'started shopping' result. The full A/B is published below.
The scorecard · appended July 14, 2026
Results vs. NRF benchmark
NRF and Prosper published their 2026 back-to-school survey on July 14, 2026 (July 1–8, 2026 · n=7,677). Predictions below were frozen July 7 and are unchanged. P1 is a clean grade. P2 and P3 are labeled instrument mismatch because our frozen questions do not match what NRF published — numbers are shown for transparency, not as wins.
P1
GRADED — WIN (Method B)Mean K-12 household back-to-school spend (band midpoints, excl. Not sure)
Method A
$856.83
Method B
$870.37
NRF actual
$863.86
Error
A: 0.8% ($7.03)
B: 0.75% ($6.51)
P2
INSTRUMENT MISMATCH% of K-12 parents who started shopping by early July
Method A
32.0%
Method B
84.7%
NRF actual
62%
Error
A: 30.0 pp
B: 22.7 pp
P3
INSTRUMENT MISMATCH% of K-12 parents planning to shop mid-July deal events (Prime Day, etc.)
Method A
92.1%
Method B
93.3%
NRF actual
54%
Error
A: 38.1 pp (not gradeable)
B: 39.3 pp (not gradeable)
P1 · GRADED — WIN (Method B)
Clean grade. Both frozen runs within 1% of NRF; Method B landed closer by $0.52.
P2 · INSTRUMENT MISMATCH
Our pre-registered item is a timing grid — "When did you start, or when do you plan to start…" — not NRF's binary "already started" question. We cannot claim a clean score. Directional numbers shown for transparency: Method A (undated persona) read 32.0%; Method B (dated persona) read 84.7%. Method B's NRF-matched amendment (P2b: "Have you begun purchasing…" yes/no) read 81.3% — still 19.3 pp high vs NRF 62%, and post-hoc.
What NRF measured: NRF headline: 62% had already started back-to-school shopping by early July (Prosper survey fielded July 1–8, 2026).
P3 · INSTRUMENT MISMATCH
Our frozen question asks whether parents plan to shop mid-July deal events. NRF's July 14 release reports June sales participation (54%), a different construct. NRF reported 85% on mid-July planning wording in 2024–2025 but did not publish a comparable 2026 figure. Numbers shown for context only — not a gradeable match.
What NRF measured: NRF 2026 published 54% who already shopped major June sales events (Prime Day, etc.) for school items — a past-behavior construct, not mid-July planning intent.
NRF release · July 14, 2026 · July 1–8, 2026 · n=7,677
NRF press release →Amendment — Method B, same day, still pre-benchmark
We found a design flaw an hour after publishing. Here's the A/B.
Method A's personas were never told what day it is. A human survey respondent knows it's early July; our simulated respondents answered “when did you start shopping” with no idea what “now” means — which is why Method A's raw share for “started before July” was 0.1%. So at 10:05 AM ET the same morning — still before NRF's release — we re-fielded the entire study with the date on the persona, plus one NRF-matched question. Method A stands unedited above, and both methods are scored in the scorecard above.
Started shopping by early July (Q3)
The undated model put 'started before July' at a raw 0.1% — it did not know July had begun.
A
32.0%
B
84.7%
Used AI for a shopping task, past 3 months (Q7)
'Past three months' is also a temporal anchor; the undated model answers from around its training cutoff.
A
20.0%
B
44.6%
Comfortable delegating a purchase to AI (Q9)
Moderate shift, same direction as Q7.
A
63.3%
B
69.7%
All non-temporal questions (spend bands, channels, trust ranking, AI-ad trust)
Effectively unchanged — the dateline moved only the questions that reference time.
A
—
B
±0–3 pts
The methodological takeaway we're publishing against ourselves: synthetic panels are highly sensitive to temporal grounding on time-referenced questions, and stable without it on attitudinal ones. Every question that references a date or window moved; the trust ranking, the AI-ad reaction, spend bands, and channel choices barely moved at all. Method B (dated persona) becomes our standard from Issue 02 — and if NRF's data says the undated run was closer, we'll print that too.
Module 1 · The spend picture
Parents of K-12 children (n=1,500)
Full results below are Method A (the original pre-registration). Method B's full distributions are in the materials available by request; where the two diverge materially it is flagged on the question.
Question 1 · Parents of K-12 children · n=1,500
“Do you expect your household's total back-to-school spending this year to be higher, lower, or about the same as last year?”
Much higher
3.1%raw 0.0%
Somewhat higher
54.8%raw 55.1%
About the same
15.5%raw 17.8%
Somewhat lower
26.2%raw 26.5%
Much lower
0.4%raw 0.5%
Net: 57.9% expect to spend more; 26.6% expect to spend less.
Question 2 · Parents of K-12 children · n=1,150 of 1,500 (350 non-responses dropped, disclosed)
“Approximately how much do you expect your household to spend in total on back-to-school items this year, including clothing, shoes, school supplies, and electronics?”
Under $250
15.8%raw 16.3%
$250 to $499
18.7%raw 19.3%
$500 to $749
17.1%raw 17.7%
$750 to $999
7.9%raw 8.1%
$1,000 to $1,499
18.2%raw 18.8%
$1,500 or more
19.3%raw 19.9%
Not sure
2.9%raw 0.0%
Band-midpoint estimated mean: $856.83 — the pre-registered Prediction 1.
Notable splits (directional)
- Spending separates hard by income: under-$50k households concentrate below $500; $100k+ households concentrate at $1,000+ (directional subgroup cells).
Question 3 · Parents of K-12 children · n=1,500
“When did you start, or when do you plan to start, back-to-school shopping this year?”
I started before July
9.0%raw 0.1%
Early July
23.0%raw 19.6%
Late July
57.5%raw 56.3%
August or later
1.4%raw 1.2%
I buy things as needs come up
9.1%raw 22.8%
Started by early July: 32.0% — the pre-registered Prediction 2. Comparability caveat: NRF's instrument differs, and NRF reported 67% by early July in 2025. This is the most date-sensitive question in the study: Method B (persona told the date) moved it to 84.7%, and the NRF-matched instrument (P2b) reads 81.3%. See the Method B amendment above.
Question 4 · Parents of K-12 children · n=1,500
“Which of the following best describes your main approach to back-to-school shopping this year?”
Choosing lower-priced or store brands
49.4%raw 45.9%
Waiting for sales or deal events
20.4%raw 29.4%
None of these — shopping as usual
20.3%raw 22.6%
Buying fewer items overall
4.5%raw 0.0%
Spreading purchases over more weeks
2.8%raw 2.1%
Reusing items from last year
2.6%raw 0.0%
Options shown ranked; respondents saw them in rotated order. ~80% report some value behavior.
Notable splits (directional)
- The value squeeze is an income story: under-$50k parents overwhelmingly trade down to store brands; $100k+ parents mostly deal-hunt or shop as usual (directional).
Question 5 · Parents of K-12 children · n=1,500
“Where do you expect to do the largest share of your back-to-school shopping this year?”
Mass merchants like Walmart or Target
66.3%raw 66.4%
Online marketplaces like Amazon
21.6%raw 21.5%
Off-price or discount stores
4.6%raw 9.6%
Dollar stores
3.1%raw 2.1%
Department stores
2.2%raw 0.5%
Office-supply or specialty stores
2.2%raw 0.0%
Question 6 · Parents of K-12 children · n=1,500
“Do you plan to shop mid-July deal events, such as Prime Day or similar retailer sales, for back-to-school items this year?”
Yes
92.1%raw 91.6%
No
7.3%raw 8.4%
Not sure
0.6%raw 0.0%
Pre-registered Prediction 3. NRF's human benchmark reported 85% in both 2024 and 2025.
Notable splits (directional)
- Deal-event intent is near-universal above $50k income (94%); 74% among under-$50k households.
Module 2 · Enter the AI shopper
All US adults (n=5,000)
Question 7 · All US adults · n=5,000
“In the past three months, have you used an AI tool, such as a chatbot or an AI assistant built into a website or app, for any shopping-related task?”
Yes
20.0%raw 16.5%
No
78.7%raw 83.5%
Not sure
1.3%raw 0.0%
Date-sensitive: Method B (persona told the date) reads 44.6% yes — 'past three months' is meaningless to a model that doesn't know when now is. Both runs are frozen; treat the true value as likely between them.
Notable splits (directional)
- Adoption is a young, affluent phenomenon: 33% of 18-34s and 33% of $100k+ households have used AI for shopping vs. 3% of 55+ and 3% of under-$50k households (Method A).
Question 8 · All US adults · n=5,000
“Which of the following shopping tasks have you used an AI tool for most often, if any?”
Comparing products or prices
8.9%raw 8.3%
Finding product ideas
4.5%raw 1.2%
Finding deals or coupons
2.6%raw 0.0%
Completing a purchase for me
2.0%raw 0.0%
Making a shopping list or plan
1.6%raw 0.0%
I have not used AI for shopping tasks
80.4%raw 90.5%
Among tasks, comparison shopping leads — the research phase, not the purchase.
Question 9 · All US adults · n=5,000
“How comfortable would you be allowing an AI assistant to complete a purchase on your behalf, within a spending limit you set?”
Very comfortable
7.4%raw 1.0%
Somewhat comfortable
55.9%raw 48.7%
Not very comfortable
36.2%raw 47.4%
Not at all comfortable
0.5%raw 2.9%
63.3% at least somewhat comfortable — three times the share who have actually used AI for any shopping task.
Notable splits (directional)
- The delegation gap is income-driven: 82% of $100k+ households are comfortable vs. 12% of under-$50k households (directional).
Question 10 · All US adults · n=5,000
“Which of the following would you trust most for product recommendations?”
Friends and family
69.3%raw 74.7%
Online customer reviews
24.2%raw 25.0%
A search engine
2.9%raw 0.2%
A retailer's website
2.3%raw 0.0%
An AI chatbot or assistant
1.3%raw 0.1%
AI ranks last — a 68-point trust gap behind friends and family.
Question 11 · All US adults · n=5,000
“If you learned that an advertisement was created mostly by AI, would that make you more or less likely to trust the ad?”
Somewhat less likely
91.4%raw 93.8%
Much more likely
3.0%raw 0.0%
Somewhat more likely
2.9%raw 0.0%
No difference
1.9%raw 0.6%
Much less likely
0.8%raw 5.6%
92.2% say less likely overall. Near-unanimous results should be read as a strong directional signal, not a precise point estimate.
Question 12 · All US adults · n=5,000
“If a retailer offered an AI shopping assistant on its website or app, would that make you more or less likely to shop there?”
More likely
55.9%raw 57.3%
No difference
35.0%raw 35.9%
Less likely
6.2%raw 6.3%
Not sure
2.9%raw 0.5%
Notable splits (directional)
- Sharp generational split: 81% of 18-34s say more likely vs. 28% of 55+ — the same public that rejects AI-made ads welcomes AI shopping help.
What it means
For CMOs and retailers
- The agentic paradox is the finding: consumers punish AI on the message (92% distrust AI-made ads) and reward it on the service (56% more likely to shop where an AI assistant helps). Build AI into utility, not into creative claims.
- Deal-event intent in our frozen panel reads 92–93%, but P3 is not gradeable against NRF 2026 — NRF published June sales participation (54%), not mid-July planning. See the scorecard.
- The value squeeze is bimodal — under-$50k households trade down to store brands almost uniformly, while $100k+ households keep spending. One campaign will not fit both.
For insights teams
- The comfort-to-usage gap (63–70% comfortable delegating purchases vs. 20–45% adoption across methods) is a leading indicator worth tracking quarterly — intent is ahead of behavior in both runs.
- This entire study fielded in 9 minutes for the cost of GPU time. The right use of synthetic panels is not replacing your tracker — it's pre-testing instruments and getting a directional read the week the question first matters.
- Judge the method by the scorecard above, not by our marketing. That's what it's for.
Methodology
How this data was made — and how wrong it might be
The findings in this report do not come from human respondents. They come from Lewsearch's synthetic panel: simulated American consumers grounded in US Census (ACS 2023) demographic distributions. Both processed and raw shares are shown for every option above, and in the materials available by request.
Cohorts. Adults: n=5,000 generated from the national ACS profile. Parents: n=1,500 with the age distribution re-weighted to a modeled prior for parents of school-age children (concentrated at 25-54; labeled modeled, not census) and the K-12-parent household context injected into each persona. Parent incidence is imposed, not sampled — a modeling choice disclosed in the data file.
Two methods, both frozen pre-benchmark. Method A was the first freeze. Method B, fielded the same morning, dated the personas and added one NRF-matched question. The divergence on time-referenced questions is disclosed above. Both get scored.
Known limitations of this run. Q2 dropped 350 of 1,500 responses as unparseable (7-option format; drop rate disclosed, non-responses excluded, not imputed). Small subgroup shares should be read as directional. The 500-question benchmark showcase records mean error of 3.72 points per answer in the September 1, 2026 historical evaluation. These consumer questions are unbenchmarked, which is why the NRF comparison exists.
If you cite this report, cite it as what it is: a calibrated synthetic panel with published error rates, not a traditional human survey. Attribution: “The Lewsearch Report” with a link to this page.
Join our email list
New issues and scorecards from The Lewsearch Report, straight to your inbox.