How AI SEO Perth observes AI search

Research objectives

1. Record which businesses, brands and sources AI-assisted search systems name, recommend and cite for Perth prompts.
2. Record which domains and URLs are cited, and classify each source.
3. Measure how answers vary across repeated runs, session profiles and platforms.
4. Track change over time on a fixed prompt set.

The programme does not try to explain why a system chose a source. It records what appeared, how often, on which surface, and how that changed. Explanations, where offered, are labelled as inference.

Query and prompt universe

A study starts from a written prompt set, not from whatever comes to mind on the day. Each prompt is assigned one query class and one buyer stage before any run, and the set is frozen for the round. Prompts are written the way a person would type them into an assistant, including the place name where the question is local.

Buyer stages: awareness, consideration, decision, retention, not-applicable. A prompt set for an industry study covers every class that applies to that industry, so a report can say which class a pattern belongs to rather than averaging across intents that behave differently.

Platforms

These are the surfaces a study may observe. A study names which of them it covered; not every platform is available, stable or reproducible for every round, and a platform missing from a report means it was not observed, not that it returned nothing.

The model or surface label the platform showed at the time is recorded per run, because platforms change models without notice and two runs a month apart may not be comparable.

Repeated runs

Every prompt is run at least 3 times per platform per round. Outputs vary between sessions on the same prompt. A single run is an anecdote. Repeated runs let the publication report how often a brand or domain appeared, not whether it appeared once.

Outputs vary for reasons the platforms have documented in part and left undocumented in part: sampling in the model, retrieval that can return different documents, personalisation, location inference and ongoing model changes. The programme does not try to remove that variance. It measures it, and reports frequencies with the number of runs behind them.

Geography and session handling

What can be controlled: the account used, its stated location where the platform accepts one, the time of day, and whether search or browsing is switched on. What cannot be controlled: how a platform infers location from the network, what it remembers across sessions, and what it changed overnight. Runs are therefore made from defined session profiles, and divergence between profiles on the same prompt is recorded as a measurement rather than treated as noise.

Location is named in the prompt where the question is local ("in Perth"), because that is how people ask, and it removes one dependency on the platform's location inference. Where a platform allows a location setting it is set to Perth, Western Australia, and the setting is noted in the record.

What gets recorded

One record per prompt per platform per run. The record contains:

• prompt (verbatim)
• query_class
• buyer_stage
• industry
• location
• platform
• model_or_surface
• observation_date, observation_time, timezone
• run_number
• session_profile
• brands_mentioned
• brands_recommended
• brand_order
• citations_present
• cited_urls
• cited_domains
• source_classes
• response_notes
• region_or_session_notes
• source_capture
• methodology_version

Mentions and recommendations are recorded separately. A mention is any business named in the answer; a recommendation is a business the answer presented as a shortlist entry or suggested option. Order is kept. Citations are recorded as shown; where a surface displays no citation the record says so, which is not evidence that no source was used.

Source classification

Each cited domain is classified into one of the following classes at the time of the observation.

A domain can belong to different classes in different observations: a business's own site is owned when that business is named in the answer and third-party when it is cited for a question about someone else.

Evidence labels

Every non-obvious claim published on this site, in reports and reference pages alike, carries one of five labels.

An inference is never presented as a fact. "X appeared in nine of twelve runs" is an observation. "The system prefers X" is an inference, and it is written that way or not at all. "The system trusts X" is not a claim this publication makes.

Limitations

• Session variance: the same prompt returns different answers across runs.
• Personalisation: account history, saved preferences and prior chats can change results.
• Location: platforms infer location differently and not all of them let it be set.
• Model updates: platforms change models and retrieval behaviour without notice, so observations are dated and versioned.
• Logged-in state: consumer products behave differently logged in, logged out and through an API.
• Citation availability: some surfaces show citations inconsistently or not at all, so absence of a citation is not evidence a source was unused.
• Sampling: a fixed prompt set is a sample of the questions people ask, not the population.
• No universal score: this publication does not compute a single ranking or 'GEO score'. Frequencies are reported per prompt set, platform and period.

The last point matters most. This publication does not produce a ranking, an index score or a "GEO score" for any business or platform. It publishes frequencies from defined prompt sets on defined dates, with the runs behind them, and leaves the reader able to check.

Versioning

Any change to prompt sets, platforms, session profiles, recorded fields or classification rules increments the version. Observations carry the version they were collected under and are never re-labelled.