Method

How the report is made

Four kinds of public evidence, joined on the questions buyers ask. No interviews and no access to your systems. A person checks the work at every gate.

1Four kinds of evidence

Every item of evidence is mapped to a question in the market's question map, so the four kinds can be compared directly.

Evidence	What it shows	Where it comes from
Demand	What businesses ask when they choose and use an agency, and how much each question matters	Search volumes and suggestions, forums and communities, reviews, public Q&A
Who answers	Which questions each agency answers, and which the wider web answers instead	Every agency's site and public footprint; platform guides, communities, directories and media
AI answers	What AI tools say, whom they name and which sources they cite	The questions put to several AI tools, several times each, from Australia
Tools	Which tools, AI apps and agents agency websites offer	A survey of every agency site in the market

2Which questions count

Only two kinds of evidence count as demand: what people do (searches and suggestions) and what customers say in their own words (forums, reviews, public questions). Agencies' own content is supply, and questions an AI tool makes up count for nothing.
We check who is speaking. Much of what's online is practitioners talking to each other. Each item is tagged as a customer, a practitioner or a vendor, and only customers count.
Two kinds must agree. A question is evidenced when two independent kinds of evidence agree, and partly evidenced with one. Otherwise it's marked as assumed, and the status shows on every item.
How much a question matters combines how often it's asked with how decisive it is. Decisiveness comes from what reviewers say made them choose or leave, from comparison and cost searches, and from what advertisers pay per click for the search.
Recent evidence weighs more. We use the past 12 months and weight the past six more heavily.

3How we ask AI

Questions are sampled by how much they matter, so the overall share of answer is weighted by demand. Every intent and buying stage gets a minimum.
Each question is asked three ways: short, conversational, and with a persona ("I run a café in Brisbane and need a new website").
Both ways of asking: discovery, with no agency named, and verification, naming the agency.
Several AI tools, five or more runs each, with the location set to Australia where the tool allows it.
For every run we record the agencies named and their order, each source cited with its owner and channel, and whether the tool searched or answered from memory.
A person checks a sample of what AI says about each agency against the agency's own site and listings.

4How we survey tools

Every agency site in the market is surveyed for interactive experiences: estimators, calculators, graders, audits, chat assistants, brief builders and booking agents.
Each one is classified by the four kinds of answer: a tool (no AI inside), an AI app (it reads and explains), or an agent (it acts). We also note what it needs from the buyer, and whether it uses anything only that agency has.
We report prevalence and gaps: how many sites offer each kind of tool, and which much-asked questions no tool answers.
A person checks a sample of the classifications every cycle.

5The measures

Measure	What it means
Share of answer	The share of AI answers that name an agency, given with the range across runs. Weighted by how much each question matters.
Recognised and recommended	Whether AI describes an agency accurately when asked about it by name, and whether it names the agency when the question doesn't.
Accuracy	The share of AI's statements about an agency that a person confirms are correct and current.
Coverage	How well an agency's own pages and tools answer a question, from 0 to 3, on presence, depth, evidence and distinctiveness.
Source mix	Where AI's citations come from at each buying stage: platforms, media, communities, profiles, directories, agencies' sites.
Tool prevalence	The share of agency sites offering each kind of tool, AI app or agent.
White space	Questions that matter, that no one answers well.

6Quality

Double coding. A person codes a sample of the evidence again every cycle, and we publish the agreement.
Every agency report is checked by a person before it is sent.
Every snapshot is dated, and the codebook is versioned, so quarters can be compared like for like.

7Limits

Public evidence is not everyone. Forums over-represent the articulate and reviews the finished. Nothing represents the silent, and nothing behind a login is visible. Each report says which voices it rests on.
An API is not an app. AI tools reached through their programming interfaces don't see a person's history, memory or exact location. Answers in the apps can differ.
Volumes are estimates. Search volumes are modelled, and no one outside the AI companies can count what people ask AI tools. We use volumes to rank questions, not to size them.
Small markets are thin. Where few people ask, we rank questions rather than measure them.
No study yet links share of answer to revenue, so we report it beside enquiries.

ASources

Research behind the figures and claims on this site, all from the past 12 months.

Agarwal, S. and Sen, A. (2026). Randomised field experiment on AI Overviews, 1,065 US users, January–February 2026. SSRN working paper, not yet peer-reviewed. Reported by Search Engine Journal, 27 April 2026.
Datos and SparkToro (2026). State of Search, Q1 2026. Reported by PPC Land, 27 April 2026.
Conductor (2026). On AI prompt volume estimates, updated 8 July 2026.
Agency Management Institute and Audience Audit (2026). Agency Core 2026, executive summary, May 2026.
Basis (2026). 2026 Advertising Agency Report, April 2026.
AgencyAnalytics (2026). 2026 Marketing Agency Benchmarks Report, surveyed February–April 2026.

See the method applied to your agency

Get your free report