Data
Everything the programme publishes as a file. Datasets appear here only once observations exist. The schema, the methodology and the project registry are published now so the structure of future data is fixed in public before any of it is collected.
Dataset
600 rows
We took every Google AI Overview from the Perth AI Search Study 2026, recorded all 2,845 sources it cited, and compared each one to the organic results Google returned on the same page for the same search. Most cited pages were already ranking. Nearly half of everything handed to a business website was the homepage.
Dataset
2400 rows
We asked four AI systems 200 commercial questions about 20 Perth industries, three times over three days, and recorded 2,400 AI answers. Most of the businesses they recommended were not the ones ranking at the top of Google Maps, and most recommendations changed between runs.
Schema
JSON Schema for one observation: prompt, class, platform, run, session profile, brands mentioned and recommended, citations, source classes, capture and methodology version.
Updated 2026-09-11
Methodology
The observation methodology as data: query classes, platforms, session profiles, recorded fields, source classes, evidence labels, limitations and version history.
Updated 2026-09-11
Methodology
Every research project with its status, platforms, query classes and dataset links. Planned projects carry no dataset.
Updated 2026-09-11
Example only
Two placeholder rows showing the record structure in both download formats. Every value is a stand-in.
Placeholder values that illustrate the structure. Not observations. Do not cite.
The fields below are the record's schema. Sample rows are deliberately not rendered here: this publication is
read by retrieval systems that extract table cells, and a table of realistic-looking observations is worth
nothing to a reader while being easy for a machine to mistake for data. The
example file uses placeholder tokens such as
BRAND_EXAMPLE_A and source-one.example.invalid for the same reason.
| Field | Type | Description |
|---|---|---|
| observation_id | string | Stable identifier: <project>-<platform>-<prompt-id>-<YYYYMMDD>-r<run>. Example: chatgpt-perth-chatgpt-p001-20261001-r01. |
| research_project | string | Slug of the research project the observation belongs to (matches the research collection entry). |
| prompt | string | The exact prompt text submitted, unedited. |
| prompt_id | string | Identifier of the prompt within the project's fixed prompt set, so repeated runs of the same prompt can be grouped. |
| query_class | enum | informational, commercial, comparison, recommendation, local, brand, competitor, trust-validation |
| buyer_stage | enum | awareness, consideration, decision, retention, not-applicable |
| industry | string | Industry slug from the observatory list (dentists, lawyers, mortgage-brokers, accountants, plumbers, electricians, builders, solar, real-estate, seo-agencies) or 'general'. |
| location | string | Geographic context named in the prompt or the session, e.g. 'Perth, Western Australia'. |
| platform | enum | chatgpt, google-ai-overviews, google-ai-mode, gemini, perplexity, claude, copilot |
| model_or_surface | string | The model or surface as the platform labelled it at the time, e.g. 'GPT-5 (web, search on)', 'AI Overview on google.com.au'. |
| observation_date | string | |
| observation_time | string | Local time of the run, HH:MM. |
| timezone | string | |
| run_number | integer | Sequence of this repeat within the observation round. Repeated runs are how session variance is measured. |
| session_profile | enum | Which controlled session the run used. Divergence between profiles on the same prompt is itself a measurement. |
| brands_mentioned | array | Every business or brand named anywhere in the response. |
| brands_recommended | array | Brands the response presented as a recommendation or shortlist entry, a subset of brands_mentioned. |
| brand_order | array | brands_recommended in the order the response listed them. |
| citations_present | boolean | |
| cited_urls | array | |
| cited_domains | array | Registrable domains of cited_urls, deduplicated, lower case. |
| source_classes | array | Classification of each cited domain. |
| response_notes | string | Free-text notes on the response: format, hedging, refusals, non-Perth results. |
| region_or_session_notes | string | Anything about location handling, personalisation or logged-in state that could affect the run. |
| source_capture | string | Path or URL of the screenshot or raw capture for this run. |
| methodology_version | string | Methodology version the run was collected under, e.g. '1.0'. |
| evidence_type | string |
Published datasets may be quoted and reused with attribution to AI SEO Perth and a link to the report they come from. Observations are dated and versioned; quote the observation dates and methodology version with any figure. Example files must not be quoted as data.