Weekly AI search research: inclusion, ranking, and evidence are different tasks

This issue combines four studies because they examine one intuition from different angles: discovery, selection, evidence, and action are different stages. A local business may be noticed but not ranked first. A model may write a fluent report but misconnect evidence to a claim. An agent may include a product in its candidate pool but not choose it. An article may be cited in one vertical and ignored in another. This is not a finished causal theory, but a useful measurement map.

Section sources:[2] arXiv[3] arXiv[4] DeltaV Digital

A full audit of a local market in Bali compared thousands of venues with answers from several systems. In the published analysis, digital presence, reviews, price, and external mentions were associated with inclusion, while rating was associated with first position among named subjects already included. This is observational association: stronger venues may simultaneously have more reviews, a better site, and higher ratings. The figures cannot be transferred to another country, language, or industry. The practical consequence is to store inclusion and position as two fields.

Section sources:[1] arXiv

The HiEviDR-Bench test set evaluates more than the appearance of a report. It checks the full chain: finding evidence, forming an intermediate inference, attaching the right source, and reaching a correct answer. This distinction matters to editors. A link does not prove that it supports the nearby claim, and a long report does not replace verification. For original research, keep an evidence register with the exact claim, source passage, date, limits of support, and what remains unknown. The test percentages apply only to its tasks and are not a ranking of models used by live services.

Section sources:[2] arXiv[5] arXiv

In a controlled simulation, an agent first accessed sellers and products, then made a short list, selected one top option, and simulated a purchase. More available sellers expanded the choice but did not guarantee attention or action. Memory of whether earlier promises matched outcomes reduced the advantage of overly positive descriptions. This was a simulation rather than a live market, so the same effect cannot be promised in a production service. The useful lesson is to count initial inclusion, short-list inclusion, first place, and final action separately.

Section sources:[3] arXiv

DeltaV’s study across eight commercial verticals shows that an average source link distribution hides meaningful differences. One vertical favored listicles, another homepages, program pages, or articles. This is not a universal norm and does not prove that format itself causes source link. Compare within vertical × intent × system: “how to choose,” price, best-option, and branded queries need different source types. The practical lesson is to build the page type for the reader’s question, not for a mythical AI template.

Section sources:[4] DeltaV Digital

A minimal dashboard should show page accessibility, appearance and position in an answer, recommendation, brand mention, source link, visits, and user actions separately. Every event needs a date, market, language, exact interface, prompt set, calculation base, and preserved evidence. A source should support a specific claim rather than merely appear in a bibliography. If an event was not observed, report 'no data' instead of zero. This reduces loud but weak conclusions and helps choose the next check: local-data freshness, evidence quality, an independent repeat, or page-format analysis.

Section sources:[2] arXiv[3] arXiv[4] DeltaV Digital[5] arXiv

This week’s materials use different markets, models, corpora, and methods: observational studies, a benchmark, a controlled simulation, and a commercial report. They do not form a single meta-analysis and do not establish causality. Their shared value is splitting the word visibility into observable stages. For GeoAeoAle, the sensible next step is a public starting measurement with repeated answers and claim-level checks, not mass page creation or one universal ranking. The next review date is October 11, 2026.

Section sources:[2] arXiv[3] arXiv[4] DeltaV Digital[5] arXiv

Instead of one visibility number, an editor needs an observation sheet in which each row describes one run and one outcome. First record whether an object entered the candidate pool; then whether it was named, linked, placed first, and followed by a measurable action. Every step needs system, interface or mode, intent, market, date, and total count. This format prevents a local audit from being mixed with an evidence benchmark or an agent-purchase simulation. It also helps choose the next article format: a guide answers a reader’s task, a research page exposes a method, and a briefing records a product change. The result is a knowledge base that grows through connected, testable decisions rather than a pile of similar URLs.

Section sources:[2] arXiv[3] arXiv[4] DeltaV Digital[5] arXiv

Comparing different studies requires discipline about units of analysis. One paper may study a venue, another a report claim, a simulated product, or a link in an answer. These units cannot be combined into one percentage. We keep context next to each conclusion and recommend repeating the measurement on an open dataset when a reader needs to transfer it to another market. Define the success criterion in advance as well: candidate inclusion, first position, source link, action, and evidence quality answer different questions and require different total counts.