Kraft Heinz is not in the answer to its own category’s questions.
The questions were not written by us. 254 candidate phrases were pulled from measured US search demand and collapsed to 21 distinct concepts, the largest of them “dinner ideas” at 823,000 searches a month. Put through four AI assistants, Kraft Heinz held none of the 9 organic answers.
Read this first. This is one run per question, on one day. It is a screen, not a study: enough to show that a brand is absent and which operators can be reached at all, not enough to put a number on how often. A funded version runs repeats and reports variance. Google AI Overviews are not measured here, and they are read by more US adults than any chat assistant.
Nine answers to the category’s own questions. The brand held zero of them, and a home-decor site answered one.
On the one question that asked which of the brand's own two products is better for a child's lunch — the question where the brand has the strongest possible claim to the answer — an assistant did go and read the web, 11,359 tokens of it. It cited yahoo.com · eatthis.com · wokwellesley.com · decorwithstyle.com · saladproguide.com. None of the brand’s own pages.
The decision this is for
A content and renovation budget, and whether the part of it aimed at category questions is reaching anyone. The buyer is a Senior Brand Manager or a Head of Content. They are funding occasion and recipe content on the assumption that publishing a good answer puts the brand in front of the household asking.
Opposite answers lead to opposite actions. If that assumption holds, the spend continues and spreads evenly across the category’s questions. If it does not, owned-answer investment concentrates on the questions an assistant demonstrably reads, and the rest is priced as human-facing search and social — a different budget, with a different success measure. Today they share one.
Where the questions came from, and why that is the finding
The first version of this question list was written by hand, from category judgment. Checked against measured demand before anything expensive ran, 1 of its 10 questions survived. The list was rebuilt from the data and the hand-written one was discarded.
| Question concept | Monthly searches |
|---|---|
| dinner ideas | 823,000 |
| mac and cheese recipe | 450,000 |
| easy dinner ideas | 301,000 |
| lunch ideas for kids | 135,000 |
The same check was run on a second category, with a second hand-written list. Zero of ten appeared. Across 150 candidate phrases in that category, not one carried a condition, an age or a use-case word — the axis the whole list had been built on. Two question lists written from expertise, two lists the demand data did not recognise.
Neither list had shipped, so neither is a correction. What they produced is the rule the rest of this study runs on: before an expensive pass, run the cheapest row that could prove its premise wrong. Four such checks ran today and each overturned the thing before it.
The four checks
| Check | The belief | What it returned |
|---|---|---|
| The question list | Ten family-meal questions written from category judgment | One of the ten appeared in measured demand. The list was rebuilt from the data. |
| The second question list | Ten running-shoe questions written the same way, for a different category | Zero of ten appeared. Across 150 candidates, no term carried a condition, age or use word. |
| The absence | A brand read as missing from its own category's candidate pool | The pool caps at fifty per seed, sorted by volume, and the cut sat at 135,000. Absent from the head is not absent. |
| The metric | A citation count read as a measure of brand position | On a question whose answer does not move, the count measures name ambiguity rather than standing. |
The third is the one worth dwelling on. A brand read as absent from its own category’s candidate pool looks like a finding until you read the instrument: the pool returns fifty suggestions per seed, sorted by volume, and the cut sat at 135,000 searches a month. Absent from the head of a category is not absent from the category, and the difference between those two sentences is the difference between a finding and an error.
Which assistants can be reached at all
| Assistant | Does it read the web? | What happened |
|---|---|---|
| Perplexity | Reads every question | Cited the brand's own domain at position one, on two models |
| Anthropic | Reads some questions | Read 11,359 tokens on a branded product comparison and cited five sites, none the brand's |
| OpenAI | Read nothing on 8 of 8 | 318 to 321 input tokens across 2,404,500 monthly searches |
| Unreadable through this instrument | Citations return as an opaque redirect; AI Overviews are measured a different way and were not measured here |
A brand can hold an AI answer. Perplexity cited the brand’s own domain at position one, and did it again on a second model. So the lever exists. Which assistant, and which question, decides whether it is reachable at all — and that is answerable for about three cents a question, before any content budget moves.
Retailer domains held 3 of the 9 answers, and only on shopping tasks rather than cooking ones. The cooking occasion is still open and nobody has claimed it.
And when the assistant buys
One AI shopping agent, asked to describe its own grocery ordering rules, said this:
“Muse asks before any swap that changes an allergen, an essential ingredient for the meal, or the budget materially; trivial brand swaps are simply noted.”
Allergens, essential ingredients and the budget are above the line and a human is asked. The brand is below it. The same agent places a grocery order at one store rather than splitting it, so a brand not stocked there is not in the basket and the substitution happens without anyone being asked.
That is a different problem in a different seat. Whether an agent can discover, evaluate and complete a purchase is a product-feed and checkout question, owned by digital commerce and funded from a different line than the one this study is about. Two problems, two buyers, one account.
What this cannot tell you
One run per question, on one day. Stated at the top and repeated here because it is the limit that matters most. Nothing here supports a rate, a frequency or a trend.
No revenue effect. The data that would connect an AI citation to a purchase is held by the retailer and the assistant provider, and is not obtainable from outside at any tier. This study does not claim one and neither can anyone else.
Google AI Overviews are not measured. They are read by more US adults than any chat assistant and they are built from the ranked web, so retrieval there is guaranteed by construction. It is the largest unmeasured surface in this work and the next run scheduled.
Whether an assistant retrieves is a provider setting. It can change in either direction without notice. Every figure here carries its date for that reason, and the study is re-run rather than cited forward.
The branded comparison is one observation. On the single branded product-comparison question measured, retrieval happened and the brand lost the citation. Whether that generalises across branded comparisons is the next thing to measure, not something this study establishes.
How it was built
Questions first, from the data. 254 candidate phrases were pulled from measured search demand across six seeds, then collapsed to 21 distinct concepts by removing word-order duplicates — a step that matters because the same four words in two orders returned two different volumes, and carrying one across the other would have inflated a concept by more than double.
Each question then went to four assistants with web search enabled, and three things were recorded: whether the assistant read anything before answering, which sources it cited if it did, and whether the brand appeared. Retrieval is readable without trusting the assistant’s own account of itself — fetched pages arrive in the prompt, so the prompt size says whether it went looking. That mechanism is the subject of its own study.
Kraft Heinz was chosen because it is large, well known, and not a client. Nothing was commissioned and the company was not contacted. Every figure is from public data and is reproducible.
The brief, its priors and its falsifiers were written before the data they describe. Both question lists that failed are named above rather than quietly replaced.
The next one
This study is re-run when an operator changes how it retrieves, and the AI Overviews measurement publishes when it runs — including the version where it disagrees with this one.
Who answers the question is one of six decisions
This study measures one question category. The CPG Decision Brief takes the six decisions a brand manager owns and answers each from filings, with the seat that contests it named and the limits of the evidence stated.
See the six 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.