The 5 things an AEO agency in Sydney does first

An AEO agency in Sydney will start you on the same 5 jobs, and you can finish all of them this morning. An llms.txt file. A filter on your UTM reports for AI traffic. A scheduled prompt run against your own brand. A Q&A block on every blog post. A robots.txt check. None take longer than an hour, and all of them are one-time work that keeps paying.

Most teams put this off because answer engine optimisation feels more like art than science. That is founded on something real: a model will not give you the same output twice, and the tracking is nowhere near what SEO has. But hard to measure does not mean hard to influence. Brands are already winning queries inside every flagship model that they have no business winning, because they put these pieces in while everyone else waited for the tracking to catch up.

1. Write an llms.txt file and make it discoverable

An llms.txt file is your best shot at controlling the narrative models repeat back. What your company is, what it is not, your positioning, your credibility signals, and instructions for assistants reading the site. The spec is markdown at the root of your domain: an H1 with the company name, a blockquote summary, then H2 sections of curated links. Reference it in your sitemap so crawlers find it rather than guess the path.

Now the part most people selling AEO will not tell you. Google's documentation says there are no additional requirements to appear in AI Overviews or AI Mode, and that you do not need to create new machine readable files or AI text files (Google Search Central). So llms.txt does nothing for Google. Its value is with agents that fetch your site directly, and with the discipline of writing your positioning down in the words you want repeated. Worth an hour. Not worth a retainer.

2. Filter your UTM reports for AI search

Build a channel group for AI traffic and have it run weekly. In GA4, match referral sources containing chatgpt.com, perplexity.ai, claude.ai, gemini.google.com and copilot.microsoft.com. You want 3 things: how many sessions arrived from a model, which pages they landed on, and where it is exposed, the query.

Read it as a trend, not a count. A share of AI sessions arrive with the referrer stripped and get filed as direct, so what you see is a floor rather than a total. Week on week is useful. The absolute figure in a board pack is not. The landing page data is the part worth having: it tells you which of your pages models actually pull from.

3. Run a cron job on your own brand

Same prompt, fixed interval, multiple flagship models, piped into Slack. Run it against ChatGPT, Claude and Perplexity separately, because they retrieve differently and averaging them hides the thing you are looking for.

Run 2 kinds of prompt, not one. A topical prompt describes the problem without naming you ("who builds attribution systems for Australian B2B firms") and tells you whether you get surfaced at all. An evaluation prompt names you and tells you what gets said. The first is a visibility check, the second a reputation check.

Never change the prompt, and never react to a single run. A reworded prompt resets the baseline, and the same prompt twice will not give you the same answer. You are watching for narrative drift over a month, not a bad result on a Tuesday.

4. Add a Q&A section to every blog post

Your old posts were optimised for how people type into Google, not how they ask a model. Someone typing searches "cold email reply rate benchmark". Someone asking says "what's a decent reply rate if I'm only sending 200 a week". Different shape, different answer, and your existing headings match the first one.

Pose the question as a heading and answer it in 2 to 4 self-contained sentences. Self-contained is the constraint that matters. A model lifting your answer will not carry the 3 paragraphs above it.

Do it retroactively, then keep publishing. Connect Claude Code to your CMS and it will work through the back catalogue in an afternoon. Cadence counts for more here than in search: across 16.975 million cited URLs, Ahrefs found AI assistants cite pages averaging 1,064 days old against 1,432 days for Google's organic results, with Perplexity and ChatGPT ordering references newest first (study). We wrote up the self-learning brand voice system that keeps that cadence running.

5. Check your robots.txt is not blocking AI crawlers

Blocking AI crawlers looks reasonable until you realise it is the same as blocking yourself from search results. OpenAI alone runs 4 user agents and each setting is independent (documented here). OAI-SearchBot is what surfaces you in ChatGPT search, and sites that block it do not appear in those answers. GPTBot crawls for model training. ChatGPT-User fires when a person asks ChatGPT to go and look at a page, and because it is user-initiated, robots.txt rules may not apply.

Blocking training and blocking retrieval are two different decisions. Most blanket disallows conflate them, which is how a brand goes invisible in ChatGPT search while trying to keep its content out of training runs. Give any fix a day: OpenAI notes it can take around 24 hours for a robots.txt change to be picked up.

Building it yourself

The monitoring job in point 3 is the only one of the 5 that needs building rather than configuring. The starting spec:

Build a scheduled workflow that runs every Monday at 8am. For each of 3 models (ChatGPT, Claude, Perplexity), send 2 fixed prompts: one topical prompt describing the problem we solve without naming us, one evaluation prompt naming us directly. Store each raw response with a timestamp, the model name and the prompt ID. Post one Slack message containing, per model, whether we were mentioned in the topical response, which competitors appeared alongside us, and any change against last week's stored run. Link to the raw text rather than summarising it.

Store the raw responses. The comparison against last week is the entire product, and a summary destroys the thing you needed to compare.

What to ask an AEO agency in Sydney before you hire one

Ask which of these 5 they are charging you for. All 5 are configuration. If a proposal is built around setting up llms.txt and checking robots.txt, you are paying a retainer for an afternoon. The work worth paying for sits underneath: a publishing cadence you could not hit by hand, monitoring that still runs after the setup call, and attribution that connects an AI referral to a booking.

Ask what they will show you in week 6. There is no rank to screenshot, so the honest answer is a mention trend and a referral line, both directional. Anyone promising a position is describing something that does not exist.

And the case where you should not buy any of it. If you publish twice a year and have no case studies, AEO is not your constraint. Models cite what exists. There is nothing to optimise into a recommendation until you have made something worth citing, and that is a go-to-market foundations problem rather than a file at the root of your domain.

Do the 5 things this morning. Then go and write something.

Common questions

What is answer engine optimisation?

Getting your brand recommended inside assistants like ChatGPT, Claude and Perplexity, rather than ranked on a results page. It covers what a model can crawl, what it finds when it does, and how consistently your positioning is described across the web. The setup is one-time. The publishing that feeds it is not.

Is AEO different from SEO?

They overlap more than most AEO pitches admit. Crawlability, clear headings and useful content serve both, and Google says no special optimisation is needed to appear in its AI features. The differences that matter: freshness counts for more, question-shaped headings get lifted more readily, and there is no rank position to measure against.

Is generative engine optimisation the same as AEO?

Effectively yes. Generative engine optimisation, GEO and AEO all describe getting cited by models rather than ranked by search engines. The vocabulary has not settled. If someone draws a hard distinction between them in a pitch, ask what they do differently under each name.

How do I know if AI search is sending me traffic?

Build a custom channel group in GA4 matching referral sources from the main assistants and run the report weekly. Expect it to undercount, because some AI sessions arrive with the referrer stripped and land in direct. Use it to see whether the line is moving and which pages get pulled from.

Automation Corner is where I break down the growth systems worth building, and how to actually build them. If you would rather have it built for you, book a 15 minute call.