Weekly GEO research: user choice, a new model, and measurement errors

This passport covers August 10–24, 2026. It includes Google documentation, an xAI release, a preprint on GEO editing, an industry graph about Reddit citations, and a small field test of seven AI search systems. They cannot be combined into one ranking because they concern different pipeline layers. Official documents establish a feature or rule. A research paper establishes an effect only in its stated task. A field graph is an observation, not a mechanism. Editors should ask which artifact changed instead of calling every update a new factor.

On August 20 Google published a standard JavaScript button, a custom option, and a deeplink that let a reader add a domain or subdomain to Preferred Sources. For a user who made that choice, the source may be more likely to appear in Top Stories and be marked in AI Mode or AI Overviews where available. The document does not call it a general ranking signal and does not promise a citation, click, or conversion; it also says the button is not required for selection. It is therefore an audience-retention mechanism, not a mass visibility tactic.

xAI announced Grok 4.6 on August 12 and described additional training, new SFT trajectories, and agentic RL. This establishes the release and developer-reported properties, but not a change in search backend, URL candidate set, or citation ranking. A weekly monitor comparing Grok before and after should record the exact alias, version, access mode, region, and date on every row. Otherwise a mention change may come from the model, connector, personalization, or query mix. A developer benchmark is not a GEO metric and cannot replace a matched test.

The GEO-Flag v2 preprint used 3,200 examples, 400 queries, four domains, and eight optimizer families; its field audit retained 10,095 accessible pages. The authors report an F1 of 0.944 and estimate that 8.90% of pages contained the studied GEO signals. These numbers belong to that benchmark and URL set. A label does not prove human or machine authorship, page falsity, or causal citation. Verifiability classification means whether a URL can be opened and checked under the authors’ rules, not whether a claim is true. A detector cannot be used as an editorial filter without review.

Promptwatch reported a decline in Reddit’s average share of ChatGPT Search citations from 3.83% to 0.52% across its compared periods. The public report does not disclose daily denominators, prompt mix, geography, personalization, or raw data, so it is an early signal, not evidence of a penalty or algorithm change. A separate Habr test of seven systems found few shared sources, but a common search plugin and parser errors could imitate system differences. Before publishing a strong conclusion, validate the connector, repeat queries, and compare control domains.

The database should version the model, surface, connector, and parser beside every answer; store a numerator and denominator for every rate; and label evidence as official, primary-research, industry-study, or field-report. Preferred Sources should be documented as a user feature. GEO-Flag is an experimental signal, not advice to write AI text. Reddit and the seven-system test need replication with a disclosed sample. The next check is four matched repeats per surface, control domains, and preserved full answers.

Use observed for an official feature and cite the exact documentation section. For a preprint, state its method, benchmark, and lack of peer review where relevant. For an industry report, preserve author, window, denominator, and unknown parameters. For a field test, separate the useful idea from the sample size. These labels do not make a weak source useless; they show what to verify next. An editor may publish an early signal when its limits are explicit, but should not turn it into advice for every site or a hidden ranking factor.

Four weeks should be connected through stable entities rather than turned into many thin pages. Each research record keeps its date, version, surface, source, claim, and next review. A new article appears only for new evidence or a material data change. Weekly research then becomes part of an encyclopedia: a weak signal gets an honest status, a strong study gets a full analysis, and readers can see how a hypothesis changed over time. Internal observations must pass a public-safety review before publication.

Preferred Sources describes a user action; Grok 4.6 a model version; GEO-Flag an experimental classification; Promptwatch an industry aggregate; and Habr a small field test. Each signal has its own denominator and next verification step. Connect them through source_id and evidence_type rather than one generic visibility tag. If a run does not preserve model, connector, region, and query set, it cannot be compared honestly with the previous week. This discipline keeps an early signal useful without turning it into a false ranking factor.