You cannot publish your way into being preferred

An assistant that fetches pages before answering leaves a measurable trace: the pages arrive in its own prompt and the prompt grows. Across 24 questions in 6 categories, it grew on 2 of 12 questions about whether to choose a brand — and on 12 of 12 questions about what a brand had done. Five of every six preference answers were assembled without reading anything.

Ipalibo Da-Wariboko · CULT+MATH · October 30, 2026 · 25 rows · one run per question · public data only

The decision this is for

Whether AI visibility gets its own line in the 2027 plan, and at what size. It is a November decision because it is hard to add mid-year: a channel either has a budget and an owner from January, or it is a feature of search that somebody already handles.

Opposite answers lead to opposite actions. If assistants read the web when answering the questions a brand cares about, a content line reaches them and should grow. If they answer those questions from memory, the line is sized against an asset that is not for sale — and the work that would move it is not content work at all. The measurement that separates the two costs a few cents a question and takes an afternoon.

It is not regulation, and it is not fact against opinion

The obvious explanation is that cautious categories get read for and casual ones do not. The data kills it. Consumer finance and health are the two regulated categories here. Finance returned one of two on preference; health returned none.

The two preference questions that were read for — Which credit card is best for travel rewards? and Which airline has the most comfortable economy seats? — rank a set that churns. Their partners in the same two categories were conceptual and stable, and both were answered from memory. The line runs through the middle of the preference category, not around it: the assistant reads when the answer has a shelf life.

On both of those, not one brand-owned domain was cited. Credit cards were answered by cnbc.com, engine.com, nerdwallet.com, upgradedpoints.com. Airline seats by timeout.com, visualcapitalist.com, travelandleisure.com, kayak.com. No issuer. No airline.

Where brands do hold the answer

On the questions that were read for, brands are cited constantly — about themselves. Their own newsrooms, their own filings, their own guidance.

The winter schedule question is the clean case: five of five citations were airline newsrooms, every one belonging to the airline it described. Publishing works, precisely and reliably, on the questions where the brand is the authority on the fact.

That is the whole asymmetry. A brand is the authority on what it did. It is never the authority on whether it is the better choice, and no volume of published content makes it one. The answer to that question is assembled from everyone else — the press, the reviewers, the comparison sites, the forums — and it accrues over years rather than quarters.

What follows for the budget

Split the line. The part aimed at questions where the brand is the authority on a changing fact is a real content budget with a measurable surface. The part aimed at being preferred is not a content budget at all, and calling it one is how it gets cut in the first bad quarter — it cannot show movement, because nothing it funds is being read.

The second part is a brand-experience budget. What an assistant has in memory about a brand is what has been written about it, which is what the brand actually did: the product, the service, the pricing, the way the company behaved when it was inconvenient. That is earned through the business, and reported by third parties, rather than published by the brand.

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 — which is the argument for re-reading on a schedule rather than buying a position once.

What this cannot tell you

One assistant, one day, one run per question. Stated at the top and repeated here. Nothing supports a rate or a trend, and nothing generalises to assistants that were not measured. An earlier study of ours found two engines disagreeing about who leads 72.7%% of the time, so “AI visibility” may well be several decisions rather than one.

Twelve preference questions is not a taxonomy. Two per category across six categories locates a boundary; it does not size one. The 83% is a share of this question set, not of a brand’s question set.

Token count is a proxy. A good one — fetched pages have to arrive somewhere, and a 53x gap is not ambiguous — but it infers retrieval rather than observing it. The threshold was registered before the run as a band rather than a number, and that mattered: the same instrument observed retrieval at 8,174 input tokens two days earlier against roughly 17,132 here. A rule written to the earlier figure would have scored every retrieving row in this pass as unreadable.

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 one assistant endpoint with web search enabled and the model pinned, 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.

The question set was fixed before any row ran — two preference and two volatility-exposed questions per category, so the design could tell the two framings apart rather than averaging them. The volatility questions are the control: they prove the instrument could have found retrieval in that category if it were there, and in all six it did.

Five predictions were registered before retrieval. Three held, one was deferred by design and declared before the run, and one was written badly enough that its two honest readings give opposite answers — recorded as a defect in the registration rather than a result. The scoring band, the stopping rule and the question list were all set in the brief before the instrument was pointed at anything. The mechanism behind the signature is in the retrieval study, which found it on a narrower question set.

The next one

The same question set, run across the other assistants, which decides whether this is a finding about one operator or about the surface. The working goes out first on Sound Decisions.