How do we get seen by AI?

This is the question we get asked most at the moment. It usually arrives the same way: someone typed a question into ChatGPT that their business should have been the answer to, watched four competitors get named, and did not see themselves.

So here’s an honest answer, including the parts that are inconvenient.

First, the thing nobody wants to hear

Most of getting seen by AI is just being findable and being good.

AI answers are assembled from the web. The models are reading the same pages search engines have always read, using largely the same signals to decide which ones are worth trusting. A site that is slow, thin, badly structured or unclear about what the business even does was invisible before, and it is invisible now, for the same reasons.

If someone offers to make you visible to AI without touching any of that, they are selling you a hat.

Now the thing that genuinely changed

For a while, the shortcut answer was “rank well and you’ll get cited”. That was broadly true. It is not true any more.

Ahrefs looked at around 4 million AI Overview citations across 863,000 keywords. The share of cited pages that also ranked in Google’s top 10 for the same query fell from 76% to 38% in roughly six months. More than a third of citations now come from pages that do not rank in the top 100 at all.

That is a big shift and it has a specific cause.

Query fan-out

When you ask a complicated question, Google no longer just searches for it. It decomposes the question into a set of smaller related questions, runs searches for all of them, and builds the answer from sources across the whole set.

Ask “is it worth getting solar with a battery in Adelaide” and behind the scenes it is separately asking what batteries cost, how feed-in tariffs work in South Australia, what payback periods look like, whether batteries are worth it generally, and half a dozen other things. The pages that get cited are the ones that answered those sub-questions well. Several of them will not rank for your original question at all.

The practical consequence is the useful bit: covering a topic thoroughly now beats optimising a single page perfectly. A page that answers one question brilliantly can lose to a site that answers the twelve questions surrounding it competently.

What to actually do

1. Check what you are blocking

This is where we find real damage most often, and it is the fastest thing to fix.

AI companies run two different kinds of crawler and the names are not obvious:

• Training crawlers collect content to train models. GPTBot for OpenAI, ClaudeBot for Anthropic.
• Retrieval crawlers fetch pages live to answer a question right now. OAI-SearchBot and ChatGPT-User for OpenAI, Claude-SearchBot for Anthropic, PerplexityBot for Perplexity.

Blocking the first group keeps your content out of model training. Blocking the second group removes you from the answers. Plenty of sites have blocked the second by accident, having read an article about “blocking AI scrapers” and pasted in a list.

Decide which you actually want. They are different decisions with different consequences.

2. Stop trying to opt out of AI Overviews

You cannot, not without leaving Google Search.

AI Overviews are built from Google’s normal search index, crawled by ordinary Googlebot. There is no separate crawler to block.

Google-Extended, which a lot of people block believing it covers this, governs whether your content is used for Gemini model training. It does not affect AI Overview citations at all. If there is a specific passage you genuinely do not want appearing in an AI answer, the right tool is a data-nosnippet attribute around that section, which leaves the rest of the page working normally.

3. Write things that survive being quoted

An AI lifts a passage and presents it without the paragraph before it. So write in a way that survives that.

Put the question in a heading and answer it directly underneath. Keep the answer self-contained, so it makes sense with no context. Front-load the actual answer instead of building to it. Include the specifics, the numbers and the caveats, because those are what make a passage worth quoting.

This is exactly why every service page on this site has a proper questions and answers section, and why the answers are written to stand alone.

4. Map the sub-questions, not just the keyword

Given fan-out, keyword research on its own is no longer enough. For any topic that matters to your business, work out the surrounding questions a buyer would ask before deciding, and answer each properly.

The fastest way to find them is to ask an AI your own commercial question and look at what it cites, then go and see which questions those pages actually answer.

5. Be legible as a business, not just as a website

Models need to understand what you are. That means consistent business details everywhere they appear, structured data that ties the pieces together, and a presence in the places that get quoted about your industry: the directories, the review platforms, the association listings that already have authority.

Being mentioned matters even where it does not come with a link. Co-occurrence is a signal in a way it never quite was for classic SEO.

What to ignore

llms.txt. This is a proposed file that gives models a clean map of your site, and it is currently the most over-sold thing in the category. No major provider has committed to reading it. Google has explicitly said it does not use it. Analysis across roughly 300,000 domains found no measurable effect on citations, and one model actually predicted citations better with the variable removed.

It costs nothing and it might matter later, so publishing one is fine. Paying someone for it is not.

Anyone guaranteeing placement. There is no submission form, no paid slot and no visible algorithm. Answers vary between users and change week to week. A guarantee is either meaningless or a lie.

Doing it instead of SEO. Every study points the same direction: the pages that get cited are mostly the pages that deserved to be. AI search raises the value of being genuinely useful and lowers the value of being merely optimised. That is roughly the opposite of a reason to stop doing the fundamentals.

How to tell whether it is working

Referrals from chatgpt.com, perplexity.ai and similar show up in your analytics as ordinary referral traffic. Watch that as a channel. It is usually small and unusually good, because someone arriving from an AI answer has already been told you are worth a look.

Beyond that, check directly. Ask the questions your customers would ask, in the tools they would use, on a schedule rather than once, and record whether you get named. It is manual and a bit rough, and it is still the most honest measurement available.

That roughness is worth accepting. This is a young channel and the measurement will improve. The businesses that get named in a year are the ones being useful now.

---

If you would rather not work through all of that yourself, that’s what we’re for.