Should an Australian company train its own team on AI search or buy it in, and the four traps in building it yourself

Train the judgement and buy the measurement. That split works for most Australian companies because the two halves have opposite economics: the judgement is cheap to teach and expensive to outsource, and the measurement is cheap to buy and expensive to build. A company that gets it backwards either pays a consultant to form opinions it should own, or spends three months building a measurement harness it could have rented for a few hundred dollars a month.

What sits on each side of the line

Train in house	Buy in
Deciding which buying questions matter to your business	Running those questions across several engines on a schedule
Judging whether an answer is good or bad for you	Recording which sources each answer leaned on
Deciding what to publish next	Keeping the readings consistent and dated between runs
Writing the pages, or briefing whoever does	Coverage of surfaces you would not otherwise check
Reading the result and arguing about it internally	The engineering of fetching, parsing and storing answers

The left column is domain knowledge about your own buyers and it does not transfer to a supplier. The right column is plumbing and it is the same plumbing for everybody.

Why building the measurement yourself is harder than it looks

Three traps, each one measured, and we can be specific because we build this.

The country parameter does not work. The search interface we use accepts a country parameter, a language parameter and a location parameter, returns HTTP 200 for every one, and ignores all three. A homemade harness that sets a country field and trusts it produces readings for nowhere in particular while reporting a country in the output. Our own market readings write the market into the question text for exactly this reason.

A citation is not always a publisher. In our Australian set, one returned host, ccn.web.tr in 8 readings, is our own data provider's fetch infrastructure rather than a source, and has to be resolved back to the page it serves before counting. Another, staging.clutch.co in 2 readings, is a directory's staging environment. A homemade counter lists both as publishers, and a reader who checked would find a proxy and an unfinished page in a table of sources.

Counting is a definition rather than an operation. A citation means one domain in one reading, deduped. Counting raw url occurrences for the same dataset in an earlier audit of ours gave 36,576 instead of 20,143. That is not a correction, it is a different measurement wearing the same word, and two teams using the two definitions cannot compare notes.

A fourth trap specific to reading the sources

Some of the hosts the engines cite cannot be read from an ordinary request. On 22 October 2026 we fetched several of the Australian sources our own audit named: asic.gov.au returned HTTP 200, accc.gov.au returned HTTP 403 on two guidance paths, and legislation.gov.au returned HTTP 200 at 1,358,447 bytes carrying only the arrangement of provisions. One industry association page returned HTTP 404 while serving 473,318 bytes of fully styled page, which is the most instructive of the four: anything that treats a non empty body as a successful read has already accepted a not found page as a source. A homemade harness has to check the status code on every fetch and nothing else will save it.

What each side costs

The training is a procedure rather than a course. A fixed question list, the market written into each question, the same engines every time, the date recorded, named and cited counted separately. Any numerate person can run it in a fortnight of part time attention, and our page on running a repeatable check on whether a business appears in an AI answer works as the training document.

The measurement, priced with the market named and read from aiknowsus.com on 22 October 2026. Outside the United States and the United Kingdom, Growth is $199 a month for 100 buyer questions and 1,600 credits across all five engines, and Scale is $399 for 300 questions and 4,667 credits. Growth is the cheapest plan on that ladder that reads all five. SEO Start at $99 and SEO Scale at $469 read zero engines and are not AI visibility plans. Australia takes that column rather than the United States and United Kingdom one. A free scan runs without a sign up if you want to see the output shape first.

The case for buying the judgement too, and when it holds

It holds in one situation: when nobody internally has time to form an opinion and the alternative is nothing happening. A consultant forming your judgement is worse than you forming it and better than it going unformed. If you take that route, buy it as a fixed scope piece with an output you keep rather than as a retainer, and transfer the judgement to somebody internal within a quarter.

What an Australian team should do in its first month

1. Write 20 buying questions in the words your buyers use, each with Australia in it. This is the part only you can do.
2. Take a baseline reading and date the file. Record for each answer whether a business was named, which one, and which sources appeared.
3. Separate named from cited. They fail for different reasons and a merged number hides which problem you have.
4. Pick the questions where nobody was named and write those pages. In our 64 Australian readings that was 39 of them, so the list is usually long.
5. Retake the reading at the end of the month. Same questions, same engines, new date.

Questions about training versus buying for AI search in Australia

Can we just add this to our agency retainer

You can, and the agency then holds the measurement of its own work. That is arithmetic rather than an accusation, and the cleaner arrangement is that the reading sits with you and the agency gets access.

How technical does the in house person need to be

Numerate rather than technical. The skill is reading a result honestly and deciding what to publish, not engineering.

Is there a certification worth having

Not yet, and treat anybody selling one with the scepticism you would apply to any new category credential. Ask for a dated reading instead.

What is the biggest risk in doing it all in house

Inconsistency between runs, which quietly destroys the comparison. Same questions, same engines, same definitions, every time, or the second reading means nothing.

How many questions does an Australian business need

Twenty is enough to act on and a hundred lets you break the result down by buyer type. Start at twenty, because a short list you actually run beats a long one you do not.

The trap that will not announce itself

Of the four, the status code trap is the one that produces confident wrong answers rather than visible failures. On 22 October 2026 an Australian industry association's codes page returned HTTP 404 while serving 473,318 bytes of fully styled page with working navigation, the consumer regulator returned HTTP 403 twice, and the statute register returned HTTP 200 at 1,358,447 bytes carrying only an arrangement of provisions. A homemade harness that checks for a non empty body collects all three as sources and reports a successful read. Nothing throws, nothing is logged, and the wrong answer looks exactly like the right one, which is the shape of every measurement bug this programme has recorded. So whatever you build or buy, the question to ask it is what it does with a status code. Then keep the judgement in house, because deciding which questions matter and what to publish next is knowledge about your own buyers that no supplier holds. AI Knows Us supplies the plumbing across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, checks a status code on every fetch, and hands back a file your own people argue about. In Australia the first argument is usually about the 28 readings out of 64 where nobody was named at all.