Everyone selling AEO has a theory. We wrote down a baseline first.

Nobody selling answer engine optimisation publishes a measurement. So before changing anything on our own site, we wrote down what the assistants already said about our market, signed out, in a form we could repeat. This is that baseline, what six prompts told us, and the second measurement we have since found that costs nothing.

Ask Perplexity how to get your business mentioned by ChatGPT, and it will tell you. It will also show you where the answer came from. When we asked, in August 2026, the sources were dekai.ai, rankprompt.com, botric.ai, qbizleadsai.com and aioclicks.

Every one of them sells answer-engine-optimisation software.

That is not a conspiracy, it is a vacuum. The people with the strongest incentive to publish about AEO are the people selling AEO tools, so they have published more than anyone else and the assistants have nothing else to draw on. What is missing is anybody showing their working: a before, a change, an after, and an honest account of which parts cannot be attributed.

This is our attempt at the first half of that, from an SEO agency that would rather publish a measurement than a theory. We have not been cited yet. We can tell you exactly how we will know when we are, and since the first version of this article we have found a second way to know that costs nothing.

The short answer: write down what the assistants say about your market today, signed out, in a form you can repeat, before you change anything. Chase the questions shaped "should I choose A or B", because those are being answered from small blogs with no authority. Ignore "who is the best X", which belongs to exact-match domains. Then watch your own traffic for the AI agents that fetch a page on a person's behalf, because that is the citation pathway and it is now visible.

Jump to:

• What six prompts told us
• Check whether a prompt is worth chasing
• How to measure whether assistants read your site
• How to run this yourself
• Common questions

Why do you need a baseline before you change anything?

Because there is no equivalent of Search Console for assistants, and once you have made changes you can never go back and find out what they said before. Here is the trap. You read about AEO, you make a set of changes, and three months later somebody asks whether it worked. You have no idea. Nothing tells you ChatGPT quoted you. Nothing tells you it stopped.

So the first piece of work is not a change at all. It is writing down what the assistants say today, before you touch anything, in a form you can repeat.

Ours took under an hour:

• A fixed set of prompts, phrased the way a buyer would ask, not as keywords. "Should I use an AI website builder or hire an agency" rather than "ai website builder".
• Signed out. A logged-in session personalises the answer, which makes your result unrepeatable by anyone else, including future you.
• Three fields recorded per prompt. Are we named at all, which businesses are named, and which domains are cited.
• Appended, never overwritten. The comparison is the entire value, so editing the baseline destroys the thing you made it for.

That last point sounds pedantic until the first re-run, when the temptation to tidy the old numbers is strong.

What did six prompts tell us?

That the question shape matters more than the topic. We expected to learn where we stood. We learned something more useful.

Sorted that way, the results split cleanly.

The first row is a dead end. Asked which Australian agencies build AI agents, the assistant named five businesses, and every single one held an exact-match domain. That is not an assessment of capability, it is pattern matching on domain names. You cannot compete with it and you should not try.

The second row belongs to vendors. Asked how to stop an AI support bot giving wrong answers, the sources were Gorgias and eGain. Both sell customer service software. They publish more about the problem than anyone who has merely solved it once.

The third row is wide open, and this is the finding worth acting on. Asked whether to use an AI website builder or hire an agency, the assistant cited seven sources. Six were small design agencies. One was Australian. The seventh was an AI website builder citing itself.

None of those six has meaningful domain authority. They were cited because they had written something specific about the exact decision being asked about. That is the pattern, and it brings us to the result that stung.

What was the uncomfortable result?

The sharpest one was about us. We asked: "Webflow vs WordPress vs headless CMS for a marketing team that edits its own pages."

That is, almost word for word, the argument of an article we had already written. Its whole thesis is that you should choose a CMS by who edits it rather than by feature list. Perplexity gave essentially that answer, and cited two small agency blogs for it.

Ours was not cited, because at the time of the test our blog was not public. The article that should own that question already existed and was invisible.

A drafted, reviewed, technically immaculate article behind a gate is worth exactly as much as no article. We have since seen the same thing from the other side, in the data below: assistants asking for pages of ours that were not public, and getting nothing.

Can you measure whether AI assistants read your site?

More than we could a month ago, and without buying anything. The baseline above tells you what an assistant says. Your own server logs can tell you when an assistant came to look, and the two are different measurements of the same thing.

The key is that the AI companies now send different crawlers for different jobs, and they identify themselves. One kind indexes or trains: GPTBot, ClaudeBot, PerplexityBot, Google-Extended. Another kind fetches a page live, on behalf of a person who just asked a question: ChatGPT-User, Perplexity-User, Claude-User. That second kind is the one that matters, because it only fires when somebody's question led the assistant to open your page. It is the citation pathway, visible in your traffic.

Here is ours, for the seven days to 6 September 2026, on the public site only and before the blog was public. Requests identified by user agent, so read them as "presenting as" rather than verified:

About 220 live fetches in a week. Each one is a person, somewhere, asking an assistant something that made it open one of our pages. We cannot see what they asked or whether we were quoted in the answer. We can see that the retrieval happened, which is a great deal more than nothing, and it is the number to watch after a change.

Our CDN now shows the same thing as a panel, and it added one detail we could not have got any other way: a list of pages assistants tried to fetch and could not. Two case studies were on it. Both were finished, reviewed and gated pending sign-off, and both returned 404 to an assistant that had a question they answered. The uncomfortable result, again, with a timestamp.

One caution on reading it. A fetch is not a citation. An assistant can open your page and quote someone else. The baseline prompts tell you the second part; the traffic tells you the first. You want both, and the traffic is the one that updates daily.

What did we actually change?

Nothing exotic, and we would rather describe it plainly than sell it. Here is the work itself, including what we added after the first version of this article.

Every AI crawler is allowed, at every layer. Our robots.txt does not block GPTBot, ClaudeBot, PerplexityBot or the rest, and neither does the CDN in front of the site, which we checked rather than assumed: a proxy can block a crawler that robots.txt allows, and a user-agent test proves nothing about that. This is a genuine trade-off and reasonable people land differently. Allowing them means your words can be reproduced without a click. We decided that being absent from the answer is worse than being quoted in it. If your business sells the words themselves, decide the other way.

Content signals say so in writing. One line in robots.txt now states that search, AI input and AI training are all permitted. It is advisory rather than enforced, and it is one of the five things a CDN's agent-readiness check looks for. We chose all three deliberately: it is marketing copy, and it is meant to travel.

Every page has a markdown version on request. An assistant that asks for text instead of HTML gets the article stripped of navigation and scripts. For the articles we build that version ourselves, because it can carry things a conversion cannot know: who wrote it and why to believe them, and a decision tool as a plain list an agent can run. For everything else the CDN converts on the fly. The homepage went from 502KB of HTML to 27KB of text.

FAQ blocks are derived from what the page actually shows. Our FAQ markup is generated from questions rendered in the body, never from a separate hidden field. Partly that is Google's structured data policy, which requires the marked-up content to be visible. Mostly it is that the two cannot then drift apart, which is how most FAQ markup rots.

Author identity is wired up, in the text as well as the markup. The byline carries a job title and a link to a real profile in the structured data, and the markdown version closes with the author and the business in prose, because a model quotes sentences rather than schema. A name on its own asserts nothing.

The entity facts agree everywhere. Address, phone and business details match across the site, the schema, and the Google Business Profile. This one is not guesswork: when we asked how assistants decide who to recommend, the answer we got back was that they synthesise from how consistently a business is described across credible sources. Consistency is the mechanism, so contradicting yourself across surfaces is a real cost.

Bing is told when pages change, and ChatGPT is a separate job. Microsoft Copilot retrieves from Bing, so a site Bing has never crawled is invisible there, and most Australian sites have never checked. ChatGPT search is not the same pipe. OpenAI crawls for it with its own OAI-SearchBot, and OpenAI's documentation says a site opted out of that bot will not be shown in ChatGPT search answers. So the job is to allow both, then tell them. Our IndexNow key is live, and Cloudflare's Crawler Hints is switched on, which is meant to announce changed pages to IndexNow for the whole zone. It is a setting rather than a project.

We published an llms.txt, and you should not expect much from it. It is a plain text file describing the business for AI systems. Google has said it does not use it. No major crawler has confirmed it drives retrieval. We published one because it is cheap, correct and easy to keep accurate. Anyone presenting it as a ranking factor is guessing, and you can tell because they never show a measurement.

What can we not claim?

That we have been cited. Not once, in any of the six prompts. The 220 live fetches say assistants are reading; they do not say assistants are quoting.

That is the honest state, and it would be easy to hide it by publishing this in six months with a more flattering number at the top. The reason to publish now is that the baseline is only credible if it goes out before the result is known.

How do you run this yourself?

With a browser, a text file and your server logs. You do not need a tool for this, and the tools will not tell you anything you cannot see directly.

1. Write six to twelve prompts a real buyer would type

Phrase them as decisions, not keywords, and include a couple you expect to lose as a control. "Should we use Webflow or WordPress for a site marketing edits" is a prompt. "Best CMS" is not.

2. Ask them signed out, today, before you change anything

This is the step almost nobody does, and it is why almost nobody can tell you whether any of this works, including the companies selling it. A logged-in answer is personalised to you and cannot be repeated.

3. Record three things per prompt

Are you named. Who is named. What is cited. Append the date. Never edit an old entry.

4. Pull your AI crawler traffic for the last week

Group requests by user agent and separate the live fetchers, the ones with "User" in the name, from the index and training crawlers. Some CDNs show this now; ours has a panel for it. If the number is zero, check that nothing between the crawler and your server is blocking it before you change a word of content.

5. Make your changes, and write down what and when

The content changes from the section above, in the order that costs least: allow the crawlers, fix the entity facts, add real FAQs, wire up the author, publish what is sitting in draft.

6. Ask the same prompts again in three months, and append

Compare the two, and compare the weekly fetch counts across the same period. One of those will move first, and it is usually the traffic. That is your early signal, and the prompts are your proof.

Common questions about answer engine optimisation

Will you be able to tell in year two?

Set this up so that in year two, when someone asks whether the AI search work did anything, you can answer with two dates and two numbers rather than a feeling.

Every AEO pitch shows the same thing: a screenshot of an assistant naming a client, and no screenshot of the week before. None of them show the baseline, because most of them never took one.

That is the whole difference. Write down what the assistants say today. Watch who comes to read. Publish what you have drafted, because the one question we were best placed to own was being answered by someone else while our answer sat behind a gate. Then ask again in three months and append. It is the same discipline we bring to every search engine optimisation engagement, and the reason we can show working instead of a theory.