Free tool · measured October 28, 2026 · no sign-up

Does an AI assistant read anything before it answers?

When ChatGPT searches the web, the pages it reads arrive in its prompt, so the prompt size shows whether it went looking. Reading costs about 8,174 input tokens; answering from memory costs about 320. We put the eight biggest US dinner questions to it: 2,365,000 searches a month, and it read nothing on any of them.

Read this first. One run per question, on one day, on one assistant. It shows whether the web was read, not how often. A brand can only win an answer the assistant reads for, so this is the check to run before paying for content aimed at AI answers.

Ipalibo Da-Wariboko · CULT+MATH · measured October 28, 2026 · ChatGPT (gpt-4.1-mini, web search on) · public data only, nothing commissioned

Eight questions, nothing read

ChatGPT on the eight highest-demand questions, October 28, 2026
QuestionUS monthly searchesInput tokensRead the web?
dinner ideas823,000318no
mac and cheese recipe450,000321no
easy dinner ideas450,000319no
mac and cheese246,000319no
healthy dinner ideas135,000319no
homemade mac and cheese110,000321no
easy chicken recipes90,500319no
baked mac and cheese60,500321no

Every row came in at 318 to 321 input tokens, the size of the question alone. When the same assistant reads the web, the count jumps to about 8,174. Nothing in between.

What decides it is time, not truth. A recipe is a fact and it was not read. The same assistant did read for a live price, this week’s recalls and a current stock quote: answers that change. Answers that stay the same come from memory, however factual. The mechanism has its own study.

Who holds each question on Google instead

The same day, the Google organic results for nine of the category’s questions, seven of them from the table above. Open one to see who holds it. These are search results, not AI citations: they show who a person finds when they search. Retailers are marked.

dinner ideas
  • redgoldtomatoes
  • pinchofyum
  • recipetineats
  • thepioneerwoman
  • dillons · retailer
  • delish
mac and cheese recipe
  • thechunkychef
  • allrecipes
  • joyfoodsunshine
  • momontimeout
  • thepioneerwoman
  • youtube
easy dinner ideas
  • easydinnerideas
  • pinchofyum
  • recipetineats
  • seriouseats
  • foodnetwork
healthy dinner ideas
  • healthyfood
  • honestandtruly
  • facebook
  • womenshealthmag
  • therecipecritic
baked mac and cheese
  • foodnetwork
  • youtube
  • smittenkitchen
  • cheflindseyfarr
  • southernbite
homemade mac and cheese
  • feastandfarm
  • youtube
  • cooking.nytimes
  • modernhoney
  • seriouseats
easy chicken recipes
  • food52
  • damndelicious
  • foodnetwork
  • youtube
  • allrecipes
  • sidechef
grocery list
  • nutrition.gov
  • kroger · retailer
  • smithsfoodanddrug · retailer
  • menumagic.ai
  • grocerylists.org
lunch ideas for kids
  • wholefoodsmarket · retailer
  • facebook
  • oxo
  • theleangreenbean
  • tasteofhome
  • yummytoddlerfood

Independent publishers hold the cooking questions outright; one holds “easy dinner ideas” with a domain that is the question. Retailers appear on 3 of the 9, two of them the shopping questions (the grocery list and kids’ lunches). A tenth question failed on a network error and is not shown.

What we predicted, written before the data

Seven predictions registered before the run
PredictionResultWhat happened
ChatGPT reads the web on fewer than half of the eight questionsHeldIt read on none of the eight.
The brand studied is cited on none of the eightHeldHeld, but not as predicted: nothing was cited by anyone, so there was no contest to lose.
Publishers hold the Google answer on at least eight of tenHeldPublishers led all nine that returned.
The three recipe questions get read; “dinner ideas” does notMissedNone was read. The miss produced the rule above: reading tracks whether the answer changes, not whether it is a fact.
Retailers appear in two or fewer of tenMissedThree of nine, two of them on the shopping questions.
The brand appears less on the scratch questions than on the branded oneUnscoreableIt appears nowhere, so there is no difference to read.
At least half the sources for “lunch ideas for kids” are written for parents of young childrenUnscoreableThe classification was not written down before the sources were read, so scoring it now would follow the answer. Our process failure, recorded as one.

3 held, 2 missed, 2 could not be scored. The misses are shown at the same weight as the hits.

What this cannot tell you

One run per question, one day, one assistant. Nothing here supports a rate or a trend.

Other assistants behave differently. In the same week Perplexity read the web on both questions put to it, and Anthropic’s model read on a branded product comparison. The lever exists; this assistant, on these questions, did not offer it.

Google AI Overviews are not measured here.

Whether an assistant reads is a provider setting. It can change without notice, which is why every figure carries its date.

It does not run on your brand. This page is one category, measured once. Yours is below.

Run it on your own category

The full write-up of this category, with the assistants compared, is Who answers dinner. An AI agent can run the check on any question itself, paid per run, through CULT+MATH connected to its assistant.

Which of your questions an assistant reads is a budget decision

Whether an assistant reads your category’s questions decides which of them to fund content for, and which to price as ordinary search. The AI visibility brief works through that decision and the ones around it.

See the AI visibility decisions

$495 standard, $995 custom. Pre-made and delivered on payment. If the decision turns out not to be decidable from evidence, it is refunded.

CULT+MATH LLC · Houston, Texas · Ipalibo Da-Wariboko
Standing research. Figures are re-measured on the schedule stated on each page and the page is corrected in place, dated. Corrections are published, not applied quietly.
Cite as: CULT+MATH LLC, Ipalibo Da-Wariboko. Attribution to cultmath.com is requested.