AEO: what actually makes AI assistants cite you

We sell this service, so start with the part that argues against us.

Almost every number you will read about answer engine optimisation was published by a company selling answer engine optimisation. Those numbers contradict each other, they are rarely dated, and the ones that are real move faster than the blog posts quoting them. We went looking for what can actually be demonstrated, found less than the category claims and more than the cynics say, and this is the honest version.

If you only take one thing: the work that moves the needle is mostly the work you already know about, done properly, plus one thing nobody sells you.

The category has a measurement problem

Here is the shape of it. Cloudflare publishes how many pages each AI platform crawls for every visitor it sends back — real traffic data, from a company with no product in this category.

Now compare that with what is in circulation. While writing this we found published claims that Anthropic sits near 4,580 pages per referral and OpenAI near 848. On the day we read the live dashboard it was close to the reverse, and OpenAI’s ratio had moved more than 122.1% inside a single week.

Neither the blogs nor the dashboard is lying. The blogs are quoting a window from months ago as though it were a constant. That is the category in one example: numbers that were true once, repeated until they become folklore.

The same volatility applies to share of crawl. On the week we checked, 23.9% of AI bot traffic was Googlebot and 13.1% was ClaudeBot — but ask again next month and expect something else.

What Google says, which nobody quotes

Google runs the largest AI answer surface in the world. It publishes guidance for site owners about exactly this. It is the most-ignored document in the category, presumably because of what it says:

optimizing for generative AI search is optimizing for the search experience, and thus still SEO

And on the specific tactics that AEO packages are built from:

You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search

There’s no requirement to break your content into tiny pieces for AI to better understand it.

You don’t need to write in a specific way just for generative AI search.

Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.

seeking inauthentic “mentions” across the web isn’t as helpful as it might seem

That is the company with the most to gain from a thriving optimisation industry telling you there is no separate discipline here. You should weigh it accordingly — Google is describing its own products, not Claude or Perplexity, and it has an interest in you not chasing other platforms. But it is a primary source, and the people selling against it are not.

The llms.txt question, settled

Every AEO proposal we have seen includes an llms.txt file. It is the signature deliverable of the category: cheap to produce, impressive to a non-technical buyer, and completely measurable — which turned out to be its problem.

Ahrefs went through the server logs of 137,000 domains. Of the valid files they found, 97% received no request at all during the month. Of the handful that were fetched, AI retrieval bots accounted for about 1.1% of requests; the biggest single requester was SEO audit tooling checking whether the file existed. A separate study of around 300,000 domains found no statistically significant relationship between having the file and being cited in AI answers.

We maintain one anyway, and we will set one up for you. Not because it drives citations — the evidence says it does not — but because it costs an hour, it breaks nothing, and coding agents genuinely do read them. That is the whole case for it. If a proposal leads with llms.txt, you are looking at a deliverable chosen because it is easy to show you, not because it works.

What AI answers do to your clicks

This is the part that makes the subject urgent rather than interesting.

The Pew Research Center watched real browsing behaviour rather than asking people what they do. Users clicked a traditional result in 15% of searches with no AI summary, and 8% when a summary appeared — a relative fall of about 47%. Links inside the summary were clicked in 1% of visits. Sessions ended with no click at all 16% of the time normally, and 26% of the time with a summary.

Google disputes the study, and the caveats on that figure are real: the sample is small against Google’s query volume, and the data is now more than a year old. Treat it as direction, not as a measurement of your site. The direction is still unambiguous, and it is the reason being in the answer now matters more than ranking beneath it.

What the one peer-reviewed study found

There is exactly one piece of academic work most people in this field cite, and almost nobody reads: a paper from a Princeton-led team at KDD 2024 that coined “generative engine optimization” and built a benchmark to test tactics against.

What raised visibility: citing sources, adding quotations, and including statistics. What did not: keyword stuffing. The headline is a lift of up to 40%, and that number deserves its context — it is a best case, on the authors’ own benchmark, against the models of the time. It is not a result anyone can promise you.

But the finding underneath it is robust and it matches what you would expect from how these systems work. A model assembling an answer needs something quotable: a specific claim, attributed, with a number attached. Pages written as atmosphere give it nothing to lift.

That is the actual content strategy, and it is unglamorous. Answer real questions directly. Put the number in the sentence. Say where it came from. Date it.

The audit we ran on ourselves

We ran this method on our own site in September, before selling it to anyone. It found real problems.

The one that mattered most was not technical. Our site said we were founded in 2020, our Clutch profile said 2022, and our homepage claimed “5+ years” — three sources, three answers. Our structured data placed us in “Ontario” with no city, while our LinkedIn page said Kitchener. Several unrelated companies trade as “Ingenious” something.

Put that together and an assistant asked about us cannot work out which company it is looking at, so it describes one it can pin down instead. No amount of content fixes that. Agreeing with yourself does.

The last row is the one we would point a sceptic at: crawler access — the thing most AEO packages lead with — was the one thing we already had right. It took an afternoon years ago and it has needed nothing since.

What to actually do

In order of how much it moves, for most businesses:

1. Make your facts agree everywhere. Your site, your schema, your LinkedIn page, your directory profiles, your Google Business Profile. One founding year. One address. One description of what you sell. This is the highest-value work in the category and nobody sells it, because it is tedious and it is finished when it is finished.
2. Get into the sources answers are assembled from. Assistants quote directories, review sites, roundups and community threads. If your category’s “best of” lists do not have you, on-site work will not put you in the answer. Earn the placements legitimately — Google explicitly warns that inauthentic mentions are less useful than they look.
3. Check that crawlers can reach you. One afternoon. Read your robots.txt, check your logs for who actually visits, and decide deliberately which agents you allow.
4. Write pages a machine can lift an answer from. Questions as headings, answers immediately underneath, numbers with sources and dates.
5. Keep doing technical SEO. Indexable architecture, structured data that matches your visible content, internal links, speed. Google’s position is that this is the work, and for its surfaces it is right.
6. Set up the reporting that exists. Search Console has a generative-AI performance report. It covers Google only, and it is the one real measurement available.

How to buy this without being sold to

There is a clean test for any proposal, including ours.

Ask which column each line item sits in. Work on the left can be shown to you as evidence: here is the robots.txt before and after, here are the four sources that disagreed and here they are agreeing, here is the Search Console report. Work on the right can only be narrated.

Two specific things to refuse. A ranking report for ChatGPT — nobody outside OpenAI can produce one, because answers vary by user, session and day, and a screenshot is a sample of one. And credit for a rise in AI referrals without a control: assistant usage is growing fast enough to produce that curve on pages nobody touched. The best-known success story in this field was re-examined against its own control pages, and most of the growth turned out to be the platform, not the work.

If you want this done by people who will show you the evidence and tell you which parts cannot be proven, that is what our AI search visibility service is. You get a written baseline first, so that whatever happens next can be attributed to something.

All figures in this post were read on October 15, 2026 and every source is linked below. The crawl ratios in particular will have moved by the time you read this — which is rather the point.