Independent publication about AI search

How AI finds information and chooses sources

Research, observations and practical experience about visibility in machine-generated answers.

Latest

Guide28 Oct 20266 min read

Google translated results: checks, reporting, and notranslate

Google may automatically translate a result’s title link and snippet when the original content’s language differs from the query language. Following the link opens a machine-translated page, with the original also available. Monitor impressions and clicks through the Search Appearance filter in Search Console’s Performance report. No opt-in is required, while notranslate opts out of all translation features in Google Search. Russian is absent from the translation-language list checked on 2026-10-27.

GeoAeoAle Editorial · read →

Briefing28 Oct 20263 min read

Search Console’s multimodal filter: exporting image-initiated search

To isolate searches initiated with an image, select the multimodal search type in Search Console’s Search results or Generative AI performance report, then click Export. Google includes Lens, Circle to Search on Android, image uploads to Google Search, and image searches from Chrome’s context menu in this segment.

GeoAeoAle Editorial · read →

Briefing28 Oct 20263 min read

Shareable pages: indexing defaults and spam monitoring

If your platform lets users create shareable pages, first decide whether those pages are intended for search discovery. Google recommends applying noindex or blocking crawling by default for pages created for direct sharing, collaboration, or utility. Separately monitor unexpected growth in indexable publications, indexed pages, and unrelated search queries.

GeoAeoAle Editorial · read →

Guide27 Oct 20265 min read

How to pin Perplexity Agent API effort for repeated evaluations

Choose dynamic effort when you want automatic Perplexity Agent API configuration updates. For an evaluation series with a pinned configuration, send the version returned in the response’s effort field. For explicit settings, copy the model, tools, instructions, and parameters and omit effort. Pinning limits configuration drift, but it does not freeze web sources or guarantee identical answers.

GeoAeoAle Editorial · read →

Briefing27 Oct 20263 min read

UGC Fresh Data Program: is your platform ready to apply to Google?

Consider applying to Google’s UGC Fresh Data Program if your platform has a high volume of user-generated content and a significant audience, dedicated public content pages, public creator profiles, and active moderation. You also need the capacity to transmit structured data and keep engagement signals current. Meeting the published criteria supports an application-readiness assessment; it does not establish acceptance or guarantee appearance in Google Search.

GeoAeoAle Editorial · read →

Briefing27 Oct 20263 min read

VideoObject: identify creators and separate interaction counters

In VideoObject, identify the person or organization that created or published the video through creator or author. For statistics, use InteractionCounter with an action type and its corresponding integer userInteractionCount: WatchAction for views, LikeAction for likes, CommentAction for comments and ShareAction for reshares. This is documented markup support; the source does not establish that adding counters improves rankings.

GeoAeoAle Editorial · read →

Guide8 Oct 20267 min read

How AI search changes publisher referrals: a Google field experiment

In a randomized field experiment with 1,100 participants, forcing Google's generative answer mode reduced external clicks per search by 18.8 percentage points relative to current Google Search. This does not imply the same traffic loss for every site: the study measured outbound clicks from Google, not appearances of a particular brand, visits to a particular page, target actions, or revenue.

GeoAeoAle Editorial · read →

Guide7 Oct 20266 min read

How to audit Fast Search: returned pages, fact use, and visible links are different events

A Fast Search result shows what the application programming interface returned and in what order. It does not prove that the system used every result, carried a fact into the answer, or showed a source link to a reader. Audit four stages separately: request and parameters -> results and order -> answer claims and textual support -> shown link and later action. Keep a numerator and a base count for each stage; when a step cannot be observed, record missing evidence rather than zero.

GeoAeoAle Editorial · read →

Guide6 Oct 20266 min read

How to test whether source identity changes an AI agent's choice

First score which explicit prompt requirements each candidate satisfies. Then show the same descriptions in both orders while changing only the visible URL or source name. If the selection changes systematically while the content stays fixed, that is evidence of a source effect in those test conditions, not a universal ranking of websites.

GeoAeoAle Editorial · read →

Guide5 Oct 20267 min read

How to measure Yandex Search quick AI answers and Alice AI separately

Do not merge these results into one Alice visibility score. Test the quick AI answer under the Yandex Search bar, Alice AI with automatic model selection, and Alice AI with Expert selected manually as separate conditions. For each run, preserve the exact prompt, date, language, region, launch method, complete answer, and visible links; then record the brand name, target-page link, and factual accuracy separately.

GeoAeoAle Editorial · read →

Guide2 Oct 20266 min read

Why a relevant page is not enough: finding where a brand disappears

A page that fits the prompt, an owned domain in the recorded source list, and a brand mention are three different checkpoints. Record each one for every run. If the page fits but the domain is absent from the source record, inspect access, completeness, and public distribution. If the domain is present but the brand is absent, inspect how clearly the page connects the brand with the supported fact. The association reported in one study does not prove causation.

GeoAeoAle Editorial · read →

Briefing30 Sept 20266 min read