MAKE YOUR SERVICE THE ANSWER.Discuss your service 
SAG TECHNOLOGY FRONTIER
We design how search, answers and generative AI understand brands, and explain the architecture behind SAG in clear technical terms.
SAG / ARCHITECTURE NOTES
From discovery and evidence to data boundaries, reliable execution and delivery—see how each engineering choice works.
75 articles · 2 / 8
An observation cohort is a comparison unit that groups observations with the same question, engine and model, region, language, and device conditions. The same sentence can receive different answers when the model and market change. An average that removes these conditions may be easy to calculate, but it is hard to explain what changed.
A model that connects observation conditions and the original response in one verifiable record. If you store only rankings in a table, it is difficult to determine why those numbers appeared after search results change. Even if a source link is still live, there is no guarantee it shows the same answer that was observed at the time.
An operational structure that defines the report’s publication date separately from the period being evaluated. Mixing this month’s in-progress data with last month’s finalized data makes it difficult to interpret month-over-month changes and the effects of assigned work. Describing an unfinished period as finalized results can cause even greater misunderstanding.
First check the research conditions, then validate separately against customer questions and channels. Effects may differ when the paper’s models, markets, or metrics differ. Copying a research improvement rate as an expected customer outcome goes beyond the evidence.
A way to manage images and text/DOM as distinct evidence types. A screen may show a product description, but it does not show meta tags or canonical tags. Treating a screenshot as equivalent evidence to an HTML diagnosis leads to conclusions about technical items that have not been verified.
This approach classifies sources as official or external based on the domain relationships of registered brands. Mistaking similarly named sites and media citations for official specifications changes accountability. String similarity is not domain ownership.
This is the boundary that converts external service inputs, errors, and source data into common observations. Replacing an API failure with an example and marking it as success can be mistaken for actual collection. Recovering a call and substituting a result are different.
The principle for calculating an aggregate metric while preserving the numerator and denominator of each ratio. A simple average of ratios from engines with different answer counts gives the same weight to small and large samples. This can misrepresent the actual set of answers.
A data model that stores content from the same address as records tied to each collection time. If specifications or policies change while the address stays the same, the evidence behind an earlier report may differ from what is currently displayed. Storing only the latest content makes it difficult to reproduce past decisions.
Evaluation data for comparing the effects of analysis changes by fixing reviewed questions, evidence, and expected judgments. If you change a model or extractor and check only for polished answers, past errors may return. Search, generation, and metric errors must be measured separately.