A 60-Answer Claude Test Puts Page Formats in Focus

CueScout’s October 10 report follows 60 Claude answers to software-buying questions, with different page formats dominating each question type. For a small B2B marketing team, it gives the next content brief a concrete check: does the page answer the buyer’s particular question?

The October 9 experiment asked 30 questions of Sonnet 5.5 and Haiku 5.5. Listicles accounted for 77 of 86 citations on best-tool questions. Comparisons accounted for 86 of 99 citations on head-to-head questions. Guides accounted for 94 of 110 citations on how-to questions.

Founder Telman Gadimov described the table in their write-up:

“The diagonal is the finding.”

A category recommendation and a setup problem give the content team different jobs.

The observation supplies a reason to test a more specific brief. It doesn’t establish that a new comparison page will increase a brand’s citations, send visitors, or close a sale. Those outcomes require their own measurement. Our coverage of inconsistent AI recommendations examined why a single answer is a weak basis for a visibility target.

The question supplies the assignment

Consider the work involved in answering a comparison between two named products. The useful page needs to organize evidence about that pair. Pricing, implementation effort, and limitations have to sit where a reader can evaluate them together. A broad product introduction leaves the assistant to assemble the comparison elsewhere.

A setup question needs a different kind of evidence. A guide can give the sequence of actions, prerequisites, and checks that show whether the setup worked. A category list helps someone choose among products. It spends less time resolving the task after they’ve chosen one.

These are editorial distinctions a team can inspect before paying for another draft. The experiment’s pattern suggests a testable content assignment: take a real question from a sales conversation and compare the existing page with the cited pages’ purpose. Record which part of the question each page actually answers. Any gap in the company’s own page is specific enough to put in a brief.

In the recommendation section, Gadimov wrote:

“One page cannot cover all three.”

Treat that as the researcher’s proposed strategy. The test didn’t modify a website and measure a before-and-after result. A comparison page also needs accurate evidence and an honest account of the competing product. Its format alone proves nothing about its quality.

The setup belongs beside the score

Anthropic’s web-search documentation describes controls for search limits and permitted or blocked domains. Citation objects carry a source URL and title. Those settings belong in a measurement record because they define the search the assistant was allowed to conduct.

Changing the setup changes the question the research answers. A tightly limited test can tell a content team what appeared under those limits. A customer’s open-ended conversation can include follow-up questions and a different search path. A tracking report should make that distinction visible before anyone compares its result with what a customer saw. Buyer details belong in the record too, as our coverage of how buyer descriptions change AI answers explores.

CueScout capped each answer at one search and requested 180-250 words. Every prompt described the same hypothetical buyer, a B2B company with 20 employees. There was one answer per question per model. Its author explicitly limits the findings to this small run and cautions that the consumer app can behave differently.

Page labeling also had limits. A second review covered 98 of 212 readable pages; the rest retained provisional model labels. The published CSV lets a reader inspect individual classifications. We reproduced its main format counts, but haven’t independently rerun the experiment. CueScout sells AI visibility monitoring, including a Claude add-on.

Repetition is part of the research

Rand Fishkin, SparkToro’s co-founder and CEO, raised a related measurement problem in research published January 27. That earlier project collected 2,961 answers from 600 volunteers across 12 prompts and found substantial variation in recommendation lists. It also found useful patterns in repeated brand appearances. The research involved Gumshoe, a company selling AI tracking.

Fishkin’s written conclusion included:

“More data is needed.”

That work concerned recommendations, rather than CueScout’s page-format classifications. It supplies a reason to repeat a local test before attaching a budget decision to it. The new study’s narrowness leaves open whether the format pattern persists across fresh runs, other buyers, and longer conversations.

For a company with a small content budget, the next step can fit inside an existing editorial review:

1. Choose a buyer question already in the sales record. Keep its wording and name the page supposed to answer it. Separate product selection from comparison or setup work.
2. Repeat the search and preserve the evidence. Record the model, search settings, date, answer, and cited URLs. Check the actual pages before accepting a format label.
3. Commission the missing answer. Give the writer the unresolved comparison or task. Review citation frequency separately from visits and sales, and require repeated observations before calling an apparent change a result.

The experiment earns a place in the content planning meeting as a prompt for a bounded audit. Fund the page that answers an observed customer question, with evidence good enough for a person to use.