AI SEO, or GEO, or AEO, apparently has to be rebuilt from scratch: that is what you read almost everywhere. If you run an SME, someone has probably already offered you an audit or a tool that promises to get you cited by ChatGPT. Before signing anything, I look at what is actually documented.
I have no figures to show you on being cited by an AI, and I am wary of anyone whose numbers look too tidy. What I did was read the primary sources, Google, OpenAI and two studies, then hold them up against the rules of my own content machines.
01What Google says when you ask it about GEO
The answer sits in its guide to generative search: from its point of view, working for these features means working for the search experience, "and thus still SEO" (Google Search Central, guide to generative AI features). The same guide explains why: AI Overviews and AI Mode fetch their pages from the index, using the usual ranking systems.
Danny Sullivan says much the same in an episode of Search Off the Record, whose transcript I read. He treats AEO and GEO as a subset of SEO, and points out that he is only talking about Google (Google Search Central, episode on SEO, AIO and GEO). That caveat matters. Google is speaking for its own shop, and it says nothing about what other engines do.
There is still one new thing to understand, "query fan-out". The model runs several related searches to build its answer. A question about a lawn full of weeds also becomes a search about herbicides and another about prevention. So your page can be picked up for a question nobody typed in those words.
02What the research measured, and what it did not
The paper that launched the term was presented at KDD in 2024. Its authors tested page rewrites on an engine they simulated, using the top five Google results and the gpt-3.5 model. Citing sources, adding quotations and adding statistics improve one of their visibility measures by 30 to 40%. Keyword stuffing adds almost nothing (Aggarwal et al., GEO: Generative Engine Optimization, KDD, 2024).
I keep three reservations when I read that figure.
- The engine is simulated. The Perplexity test gives 22% and 37% depending on the measure, on 200 examples, with text supplied as file uploads rather than pages found on the web.
- The gain moves around. In their table, one method gains 115.1% for the site ranked fifth and loses 30.3% for the first.
- The authors themselves write that they did not evaluate how these methods affect search rankings.
A more recent study is less enthusiastic. Across several tasks and domains, most of the methods tested turn out to be "largely ineffective", and often negative for the document's rank. Classic strategies, the ones that improve the source's rank in the context given to the model, are clearly more effective, and the gains shrink as competitors adopt the same method (Puerto et al., C-SEO Bench, NeurIPS Datasets and Benchmarks, 2025).
These two studies do not contradict each other as much as it seems: they test neither the same methods nor the same engines. What I take from them is modest. A page that shows its evidence seems to have an edge, and I would not go further on this basis alone.
03What I would not do
I would not create an llms.txt file hoping for an effect on Google. Its documentation says it ignores these files, that they neither help nor harm, and that you can keep one for other services (Google Search Central, guide to generative AI features). This website does not have one.
I would not chop my text into small pieces to please the models. Danny Sullivan is very clear, "we don't want you to do that", and he adds the argument I find most useful: suppose it works today, the systems will change, and you will have reorganised all your content for an edge that can vanish (Google Search Central, episode on SEO, AIO and GEO).
I would not publish one page per sub-question to cover the whole fan-out. Google considers that doing so, when the main aim is to manipulate the answers, breaches its scaled content abuse policy. This concerns me directly, since I write with a machine. It is a limit I keep in mind.
Finally, I would not pay for a score. Google points out that no third-party tool has access to its internal ranking data, and lists "AEO" or "GEO" tools among those promising more than they can guarantee (Google Search Central, guidance on third-party tools). A tool can be handy for tracking answers. It does not tell you why you appear in them.
04What I check on a website, in order
None of what follows is spectacular, and that is rather the point.
Access
- The robots.txt file and the CDN let Googlebot through. Google specifically says "and by any CDN or hosting infrastructure", because a block at that level does not show up in robots.txt (Google Search Central, AI features and your website).
- Important pages are indexed and can show a snippet, with no `nosnippet` set by mistake. That is the condition for appearing as a link in these answers.
- In Search Console, the Search generative AI control is set to include. That is the default, and it does not touch model training, which goes through Google-Extended (Search Console Help, Search generative AI control).
- OAI-SearchBot is allowed if you want to appear in ChatGPT search. A site that blocks it does not appear in those answers, and GPTBot, which concerns training, is set separately (OpenAI, overview of its crawlers). So you can refuse one and accept the other.
The page
- The information that matters is in text, not only in an image or a video. Structured data says the same thing as the visible page.
- The page offers something you cannot find elsewhere: an example, a sourced figure, a mistake described honestly. Google talks about "non-commodity" content, and the 2024 study points the same way for evidence.
Measurement
- I look at the dedicated Search Console report, open to all websites since the end of August 2026. The help page describes impressions, along four dimensions: pages, countries, dates and devices (Search Console Help, Generative AI performance report). It lists no queries, so I would not make it say what it does not say.
- I cross-check with conversions in analytics. Google claims that clicks from pages with AI Overviews are higher quality, without publishing a figure. That is its claim, and on your own website you can check it. I would add this to the monthly SEO report that sends itself, rather than open yet another dashboard.
05What my content machine already does, and what I do not know
Reading these sources, I recognised things I built for a different reason. The content machine behind this blog refuses a brief without sources that have been read, each with an identified title and the claim it supports. Every figure must have its link in the same paragraph, and the FAQ starts from questions people really ask, noted with their origin. A machine of the same kind publishes the articles for Moonworkers, my company.
I wrote these rules to stop text that makes things up, not to get cited by an AI. I find that a sounder motive. If the advantage for generated answers disappears tomorrow, a human reader still gets sources they can check.
Now for what I do not know. I do not know whether these rules get me cited more, as I have no clean test to show. Nor do I know how ChatGPT chooses its sources, and the OpenAI page I read covers access, not selection. And this subject moves quickly: Google's guide was updated in July 2026, and everything I write here will need rereading.
06Where to start on Monday
Take the five pages that matter most to your business and go through the list above, starting with access. That is where a simple settings mistake makes you invisible, and it is also the cheapest thing to fix.
If someone offers you a GEO service, ask the questions Danny Sullivan suggests for any SEO advice: where does Google say that, and could it cause a spam action? Then ask what will be measured, and how you will know it is not working. A good answer exists, it will just be less spectacular than the promise.
If you would like to look at this on your own website, content and SEO work is part of what I describe in my services, and we can talk about it.