Original research on how AI engines cite sources

Why original research matters for AI search

Original research can give publishers a source that is easier to inspect, compare, and cite than an unsupported opinion. Its value depends on a clear protocol, traceable observations, stated limitations, and data that other researchers can evaluate.

SCALZ.AI is building that evidence trail in public. The first baseline names its query set, surfaces, collection date, counting rules, and limitations. Response-level raw files are not yet publicly downloadable, so the current findings should be treated as self-reported observations rather than an independently reproduced benchmark.

Completed July 7 baseline

The completed baseline checked 30 fixed buyer, how-to, and local or vertical queries across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. It produced 150 query-surface observations. Four LLM surfaces returned generated answers for all 30 queries. Google AI Overviews appeared for 20 queries and did not appear for 10, leaving 140 generated-answer observations and 10 no-AI-Overview observations.

Read the completed citation study, its counting method, results, and limitations. The article reports a single-day snapshot. It is not an independent ranking or endorsement.

Collection method and evidence status

The completed baseline used the DataForSEO AI Optimization and SERP APIs to collect four web-search-enabled LLM surfaces plus Google AI Overviews from US desktop search results. The published findings are SCALZ.AI's self-reported analysis of that collection. Response-level raw files are not yet publicly downloadable, and no claim on this page should be read as independently reproduced.

Planned expanded protocol

A separate, expanded protocol proposes 60 queries across healthcare and behavioral health, home services, legal and professional services, AEO topics, and local intent. Running those queries against four LLM surfaces would create 240 query-surface observations. This phase has not been collected, and it must not be combined with the completed 30-query baseline.

Status

The 30-query baseline is complete and published. The expanded 60-query protocol remains planned. A future dataset release should include versioned CSV and JSON files, the locked query list, timestamps, model and surface labels, response-level records, counting rules, limitations, and a changelog before Dataset markup or download claims are added.