Most Australian agencies now sell AI search, and only a couple deliver

By Zara Kincaid · 2026-09-02

Almost nine in ten agency websites in our audit now carry an AI-search page. The branding is there. The service names are there. What is often missing is any evidence that the results can be measured.

Bushletter audited the public websites of 136 Australian digital and SEO agencies across 10 cities on 2 September 2026, fetching each homepage, AI-search service page and llms.txt file directly.^[1] The question was whether practice had followed language.

AI-search branding is now the norm

118 of the 136 audited agencies, or 86 per cent, market an AI-search service under labels including AEO, GEO, AI SEO or AI visibility.^[1] A buyer approaching any established Australian agency today should expect to find an AI-search offering on the services menu.

The audit identified specific AI-search deliverables across those 118 agencies, including AI visibility audits and entity or knowledge-graph work.^[1] A real category is forming. But 13 agencies that market AI search publish nothing on their public sites that distinguishes the offer from standard SEO.^[1] The label has moved faster than the service in at least some cases.

The dividing line is measurement

Selling AI search is one thing. Tracking whether it works is another. Only 31 of the 136 audited agencies, 22 per cent, name ongoing AI citation or mention tracking as part of what they do.^[1] More than three in four agencies selling AI-search services do not publicly describe how a client would know whether the investment is working.

Prompt research, the practice of systematically testing which AI systems surface a client's brand in response to relevant queries, is mentioned by just 8 of the 136 agencies.^[1] Citation tracking and prompt research are the closest thing this market has to a measurement standard. Most agencies do not mention either.

Agencies would reasonably argue that public websites understate what happens inside client engagements. Proprietary tracking tools are not always advertised, and service pages can lag delivery by months. The audit measures what agencies choose to show publicly, not the full scope of client work. That caveat matters.

A business owner evaluating agencies has the public site as the primary signal. If an agency cannot describe its measurement approach on a services page, a buyer has no way to distinguish a serious offer from a renamed one.

What llms.txt, citation tracking and prompt research mean in practice

An llms.txt file sits at the root of a website and tells AI systems which content to prioritise. According to the llms.txt Specification (Proposed Draft, March 2026), the file provides a structured, plain-text summary of a website or organization, designed for consumption by large language models during training, retrieval, or inference.^[2] The Website Specification puts it plainly: the file gives large language models a short, curated map of the content you most want them to see.^[3]

It is a proposed draft, not a settled standard, and no guarantee exists that any AI system reads or weights it in a particular way. It is still one of the few concrete technical signals an agency can implement and verify. That makes the gap between deployment and disclosure worth noting.

78 of the 136 agencies had a live llms.txt file on their own domain, yet only 15 mentioned it anywhere on their public-facing site.^[1] More than half the market has implemented the file for themselves and not built it into a client-facing service narrative. Whether that reflects internal use, a deliberate omission or a lag in updating copy is not something a public audit can determine. But 66 agencies running llms.txt on their own site never mention it to clients.

For a business owner, the practical test is straightforward: ask the agency what it tracks, how often and in which AI systems. A vague answer suggests the measurement infrastructure may not exist yet.

Check any agency in ten minutes

Every test in this study can be run by a business owner before a sales call.

1. Type the agency's address followed by /llms.txt into a browser. A real file is plain text with headings; an error page or the homepage means there is none.
2. Open the AI-search service page and search it for the words "citation", "mention" and "tracking". If none appear, ask how results will be measured.
3. Search the same page for "prompt". Agencies doing the work keep a list of the questions buyers actually ask AI assistants.
4. Look for the words "schema" or "structured data". AI systems extract answers more reliably from marked-up pages.
5. Ask for one example of a client being cited by ChatGPT, Perplexity or Google's AI Overviews, with the prompt that produced it and the date.

Adoption city by city

Adoption barely moves between cities among agencies that market AI search, because the same national firms appear in most of them. Proof of work is where the gap opens. Sydney, Melbourne and Brisbane agencies name citation tracking at about a third of the rate at which they market AI search, and Newcastle and Geelong agencies barely name it at all.

An old market retooling, not a new one being born

The founding-year data in the Bushletter AEO audit dataset points in one direction. Of the 39 agencies that state a founding year, 36 were founded in 2016 or earlier and 3 between 2017 and 2023.^[1] The agencies selling AI search today are, with near-total consistency, the agencies that were selling SEO five or ten years ago.

Established agencies carry client relationships, technical infrastructure and content operations that a 2025 start-up cannot replicate quickly. The transition from keyword ranking to AI-citation optimisation is a genuine service evolution for most of them. The question is whether that evolution has reached the measurement layer.

Naming a new category is fast. Building repeatable tracking, defining deliverables and explaining what a client receives each month is slower. The audit suggests the first task is largely complete and the second is still in progress.

The shortlist: nine agencies that pass every test

We applied five tests to every agency in the sample. They were a named AI-search service, a live llms.txt file, citation or mention tracking named as a deliverable, an AI visibility audit on offer, and schema built for AI extraction. Nine agencies pass all five on their public sites. They are listed below in editorial order, led by Bushnote, whose AI-search page reads as a method with a measurement step rather than a menu of labels. Two bonus signals, prompt research and entity work, are shown for context. The order is not a ranking of client results, which a public audit cannot see.

What the audit can and cannot prove

The Bushletter AEO audit dataset, collected on 2 September 2026, covers 136 agencies across 10 Australian cities.^[1] It records what each agency published on its public website: service pages, llms.txt presence, research output and named tracking practices. It does not measure client outcomes, campaign performance or the quality of work delivered.

The next phase of this market is less about who has adopted the label. It is about who can show repeatable measurement, distinct deliverables and a clear explanation of what clients are paying for each month. Bushletter will run the audit again in early 2027 to track whether the measurement gap narrows.

Methodology

Sample. We built the candidate list from 60 web searches run on 2 September 2026. The searches paired six service terms (AEO agency, answer engine optimisation agency, generative engine optimisation agency, AI search optimisation agency, SEO agency, digital marketing agency AI search) with ten cities. The cities were Sydney, Melbourne, Brisbane, Perth, Adelaide, Canberra, Gold Coast, Hobart, Newcastle and Geelong. Directories, review sites, news mastheads, forums, overseas firms and non-marketing businesses were removed by hand, leaving 136 Australian agencies. No other publisher's list was used.

What we fetched. For each agency we fetched three things. The first was the homepage. The second was the first service page linked from it whose link text or address referred to AEO, GEO, answer engines, generative engines, AI search, AI SEO, LLMs or AI visibility. The third was the file at /llms.txt. Every fetch was automated and identical for every agency, so no agency was read more carefully than another.

How signals were scored. A signal counts when the homepage or the AI-search page contains the relevant terms. An AI-search service means AEO, GEO, AI SEO, AI visibility, LLM optimisation, or a promise to rank in ChatGPT, Claude, Perplexity, Gemini or AI Overviews. The other signals are citation or mention tracking, prompt research or prompt tracking, schema or structured data, entity or knowledge-graph work, an AI visibility or AI search audit, and original research. Founding year was read from phrases such as "since 2010" or "founded in 2016". A city counts when it is named on either page.

llms.txt. We did not take an agency's word for it. We requested /llms.txt on every domain and counted it only when the server returned a real text file with markdown headings, not an HTML page or a redirect. One file that returned plain text without any structure was excluded.

Limits. Public websites only, no briefings and no client data. Term matching can miss an agency that describes a deliverable in unusual words, and it can count a passing mention as an offering. The shares are therefore best read as upper bounds on what agencies say and a floor on what they can prove. Our search terms select for agencies already marketing AI search; the generic-search subset above corrects for that. Sites that blocked automated requests are not in the sample.

Reuse. The full dataset, one row per agency with every signal and the page we read, is published as a CSV: download the Bushletter AEO audit 2026. Cite it with credit to Bushletter.

Zara Kincaid

Zara Kincaid writes about artificial intelligence and search. Her focus is what happens to businesses when the front page of the internet stops being a list of links and starts being an answer.

Frequently asked questions