Most answers are not researched. They are remembered.
When an assistant fetches pages before answering, those pages arrive in its own prompt and the prompt grows — about 8,174 tokens. When it answers from memory, the prompt stays near 318. Across 8 questions carrying 2,404,500 monthly searches, the largest assistant returned between 318 and 321 every time. Nothing published about those subjects was in the room.
Read this first. This describes four named endpoints on one day, one run per question. It is not a finding about “AI”, and whether an endpoint retrieves is a provider setting that can change without notice. Google AI Overviews are not measured here and are read by more US adults than any chat assistant.
Retrieval tracks whether the answer moves — not whether it is true, and not whether it is a fact.
Prices move. Recalls move. A live score moves. A recipe does not, and a recipe was not read for — which means the rule is not fact-versus-opinion, the shape most people assume. An assistant answering a stable question answers from memory, and memory is not a surface that publishing can reach.
The decision this is for
A content budget carrying an AI-visibility justification, and whether that justification survives. The buyer is a CMO, a VP of Brand or a Head of Content who has been told — often by an agency, sometimes by a board — that publishing authoritative answers is how a brand stays visible as search becomes answering.
Opposite answers lead to opposite actions. If assistants read the web when answering the questions a brand cares about, that spend is reaching them and should grow. If they answer those questions from memory, the spend is doing something else — possibly something valuable, for human readers — and the AI-visibility line in its business case is unfunded. The measurement that separates the two costs about three cents a question.
How the rule was arrived at
Three rules were written and tested against the same runs. The first two were wrong, and were discarded before anything was published.
| Candidate rule | Result | What decided it |
|---|---|---|
| The operator never retrieves | Rejected | It retrieved on questions whose answers change hourly, in the same session. |
| It retrieves for facts and answers preference from memory | Rejected | A recipe is a fact and it was not read for. |
| It retrieves when the answer moves | Survived | Prices, recalls and live events were read for. Recipes, occasion ideas and product comparisons were not. |
The second is the instructive one. “It reads for facts and answers preference from memory” is intuitive, it fits most of the data, and it is wrong: a recipe is a fact and it was not read for. A rule that fits most of the rows is the most expensive kind of wrong, because nothing in the data complains.
What follows for a brand
The addressable set is narrower than the category of questions a brand wants to own. It is the questions whose answers move. For a food brand that is recalls, availability, price and reformulation — not the recipe, not the occasion, not the comparison, which is most of what gets published.
The lever exists, on the operators that read every time. Perplexity retrieved on every question measured and cited a brand’s own domain at position one. So this is not a story about an unreachable surface. It is a story about which surface a given question can be reached on, and that is knowable in advance for a few cents rather than discovered after a quarter of spend.
And it is a setting, not a property of the technology. An operator could begin retrieving on stable answers tomorrow and the addressable set would widen overnight. That is the argument for re-reading on a schedule rather than buying a position once.
And it is only half the problem
What an assistant says about a brand and whether an agent can buy from one are routinely sold as one programme. They are not one problem, and the clearest evidence is organisational: a portfolio CMO can own every lever in the left column and none in the right.
| Answering | Acting | |
|---|---|---|
| The question | What does an assistant say when asked about us? | Can an agent discover, evaluate and buy from us? |
| The seat | VP Brand, Head of Content, CMO | VP Ecommerce, Chief Digital Officer, Head of Digital Product |
| The budget | Marketing | Digital commerce, or IT |
| The levers | Owned pages, third-party citations, structured data, reference sources | Product feeds, pricing APIs, machine-readable inventory, checkout, returns data |
| The failure | The brand is absent, or present and wrong | The agent recommends you and cannot complete the purchase |
| Measured in | Share of answers, citation share, position | Completion rate, agent-attributable revenue |
Different seats, different budgets, different failure modes, different numbers. A proposal that covers both is a proposal to two buyers who do not share a line. The measurement on this page belongs entirely to the left column.
What this cannot tell you
Four endpoints, one day, one run per question. Stated at the top and repeated here. Nothing supports a rate or a trend, and nothing here generalises to assistants that were not measured.
Token count is a proxy. It is a good one — fetched pages have to arrive somewhere, and a 25x gap is not ambiguous — but it infers retrieval rather than observing it directly. An operator could change how it accounts for input and this instrument would need rebuilding.
Google AI Overviews are not measured. They are built from the ranked web, so retrieval there is guaranteed by construction, and they reach more US adults than any chat assistant. Their absence is the largest hole in this page and the next run scheduled.
No revenue effect is claimed. Nothing here connects a citation to a purchase, and the data that would is held by the operators.
How it was built
Each question was sent to four assistant endpoints with web search enabled, and the input-token count returned with the answer was recorded alongside the cited sources. The count is reported by the operator and is not something the assistant can be asked about, which is the point: it does not depend on a model’s account of its own behaviour.
Questions were drawn from measured US search demand rather than written by hand, and the category application of that question set is in the CPG study, which carries the brand-level findings.
Three candidate rules were written down and two were discarded against the data before publication. Both are named above rather than quietly replaced.
Run this on your own category
7 signs, scored on what you already know. No data, no tools, about two minutes.
Every box you leave unticked is a question you cannot currently answer about your own category.
A Category Decision Brief answers all seven for your category. Twenty questions chosen by measured demand, across four assistants and the AI Overview, with who holds each answer today and what was predicted before anyone looked.
See what a Decision Brief covers$495 standard, $995 custom. One category, measured, with the misses published alongside the hits.
The next one
Re-run when an operator changes an endpoint. The AI Overviews measurement has since run, and it disagrees with this one in the way that matters: on that surface, retrieval is the architecture. The working goes out first on Sound Decisions.