AEO and GEO misconceptions: what changes and what needs evidence

Filed under Search & AI Discovery

In this article

I’m interested in AEO. I’m also getting slightly tired of how confidently people explain it.

A particular file is essential. Backlinks are dead. Your content needs rebuilding around an entirely new set of rules.

Some recommendations are sensible. Some deserve testing. Others turn a useful suggestion into a technical requirement without explaining how we got there.

I work with lead generation. I like building tools, testing ideas and finding out whether they help. I also have a backlog. It has survived several rounds of “this changes everything” remarkably well.

Every new essential task competes with something else we could spend that time on. For marketers and web editors, that might mean improving a product page, updating an outdated guide or fixing a form that loses leads.

My starting position is that good SEO, including the technical work that makes content accessible to search engines, remains a foundation for visibility in AI search. A foundation is not a guarantee of being recommended. What I want to understand is which extra work produces a useful result.

Which brings me, somewhat unexpectedly, to office cats.

The cats passed the test

SEO consultant Mark Williams-Cook created cats.txt, a satirical standard for declaring a website’s cats, their jobs and their enthusiasm for being stroked.

In his account, AI bots fetched the files, Google indexed them, AI answers repeated their contents, and ChatGPT recommended cats.txt as a possible way to improve visibility.

Those were observations he had seen used to promote llms.txt, a proposed file for directing AI systems towards useful website content.

By those criteria, we should apparently have added the office cats to the SEO roadmap.

But the evidence only takes us so far:

What happened What it does not prove
A bot downloaded the file The system supports it as a special standard
Google added it to its index It improves visibility for the site’s other pages
An AI answer cited its contents The format works better than an ordinary page
ChatGPT recommended it Anyone measured an improvement

If the file helps an AI assistant find the right instructions in a large documentation site, that is useful. It still needs to be evaluated separately from the claim that adding the file will make a brand appear more often in AI search answers.

AEO, GEO and a little acronym housekeeping

AEO stands for Answer Engine Optimisation. GEO stands for Generative Engine Optimisation. Both are used for work intended to improve visibility in AI answers. For this article, I’ll discuss them together.

The labels are less useful than understanding how the system gets its information:

  • Training: Information is used to build or update the model. That does not guarantee the model will later repeat or cite it.
  • Search: The assistant looks up information to help answer a question.
  • A requested page visit: Someone asks the assistant to read a particular URL.

ChatGPT can use web search to find current information. Not every answer involves search.

Where search is involved, sources have to be found and selected before the answer is written. SEO helps with visibility in that process. Being found still matters, even when the final result is a paragraph rather than a list of links.

That’s why advice should name the platform and feature it applies to. “AI needs this” is usually too broad to be useful.

1. “A special file is required for AI visibility”

Google’s optimisation guidance says Search does not use llms.txt or require special AI markup.

That answers the question for Google Search. It does not describe every assistant or software integration.

The important distinction is whether a system deliberately uses llms.txt as a guide, or simply finds a public text file through search.

There is a similar issue with schema, or structured data: code that labels information such as a product’s price or an article’s author. Google says no special schema is required for its generative search features.

Use structured data for its documented purposes.

For an editor, the practical starting point is accurate, accessible content. A proposed file format needs a clearer justification before it becomes an urgent task.

Seer Interactive’s tests offer a useful example. In their two formal llms.txt tests, neither client logged a single ChatGPT request for the file. Their tests of ordinary web pages summarising company facts were more promising: AI answers picked up information from the pages for three of four clients. None of those information-page tests had a control group. I’d consider testing one where AI gets company facts wrong, with those errors as the starting point for measurement.

2. “AI search requires a completely new content setup”

Some pages load their main content through JavaScript after the initial page arrives. If a bot cannot run that code, it may miss the information.

That is a real issue. It also existed before AEO became a familiar acronym.

Google’s JavaScript SEO documentation already explains this problem. It recommends approaches that make content available earlier and notes that not all bots can run JavaScript.

The practical question for your web team is whether the important content is accessible to the systems you want to reach. Several technical approaches can achieve that.

For editors, the equivalent question is whether the content is understandable.

Clear headings, specific explanations and sufficient context help. But “every section must stand alone” is too absolute, including when I say it.

A comparison table should identify what it compares. A section about pricing should explain which product it covers. That does not mean every paragraph needs to reintroduce the entire company.

Readers will eventually notice.

Google also says its AI search features do not require content to be divided into tiny chunks. Structure the page around what the reader needs to understand. Test additional formatting ideas without treating them as universal rules.

Backlinks are links from other websites to yours.

The claim that they no longer matter skips over how sources become available to AI answers.

Google still documents link analysis and PageRank as part of its core ranking systems. Its AI search guidance describes using a range of models and techniques to identify supporting pages during query fan-out.

My inference is that links can still influence which sources reach the stage where an answer is written. Their weight across platforms, and any relationship between backlink counts and citations, are separate questions.

An assistant using a different search index may select sources differently. An index is the collection of information a search system can search. Building a separate one does not establish which signals it uses.

A relevant industry article might provide a link, mention a company and explain what it does. Those benefits can coexist. I would assess the value of that coverage before replacing it with an obituary for backlinks.

Good rankings do not guarantee a place in the answer

An AI system may turn one question into several related searches. Google calls this query fan-out.

That means the question someone asks is not necessarily the exact search your page needs to answer.

In his discussion of AI visibility KPIs, Wil Reynolds shows examples where those searches already include particular brand or expert names. These examples raise a useful question: is the system investigating your company as an option at all?

For illustration, a buyer might ask for software suitable for a small team. The assistant could investigate several named products, compare their pricing and check their limitations. A page ranking for the buyer’s original wording would not automatically be selected for those searches.

This does not establish that brands are permanently “hardwired” into a model, or that we know why each name was selected. But it does challenge the assumption that a good ranking must become an AI recommendation.

For marketers, accurate information about who a product suits matters alongside whether the page can be found. Independent coverage may help explain that fit too.

OpenAI documents separate controls for its training bot, GPTBot, and its search bot, OAI-SearchBot.

A site can allow the search bot while blocking GPTBot. For OpenAI, training access is therefore not a prerequisite for search eligibility.

If your goal is visibility in current answers that use search, search access is more directly relevant. A training-bot visit does not establish that your brand will appear in a future model.

Ask your web team to check access for the search systems you want to reach. “Let all the AI bots in” needs a better explanation of what each permission is supposed to achieve.

5. “More bot activity means the strategy is working”

A bot request, a citation and a customer are different events.

An AI answer might mention your company as an example of what to avoid. That is technically a mention. I would hesitate to put it on the wins slide.

An article can also be cited without the company being recommended. A recommendation can influence a buyer who never clicks its source.

Reynolds makes this distinction concrete with an answer that mentions brands while directing the user to look elsewhere. Counting the names alone would give a misleading result.

For the questions buyers ask, distinguish a neutral mention, a citation to your content, a recommendation and a warning against choosing you. Track the relevant buying context too: a recommendation for a small team might do little for an enterprise campaign.

A visibility score needs a definition too. Does it cover a selected set of questions, assess the writing on a page or count observed citations? A score of 85 means little until we know what was scored.

An increase after an edit is worth investigating. It does not, by itself, establish that the edit caused the increase. The AI product may have changed, competitors may have updated their content, or the questions tested may be different.

Repeat checks using the same relevant questions, with comparable settings, and record the platforms and dates. Small changes can be variation rather than momentum, so avoid drawing conclusions from a handful of answers.

Server logs show page requests; website analytics show visits and actions. Qualified leads and sales answer another question. We can measure those outcomes while acknowledging that some influence will remain difficult to attribute.

Research can help. The original GEO paper reports improvements in its evaluation, with differences across domains. That gives us ideas to investigate. It does not establish a universal recipe for every current product.

Fast growth still needs a starting number

AI referrals are visits from people who click links in an AI assistant’s answer. They can grow rapidly while remaining a small share of traffic.

For scale, the Ahrefs traffic dashboard, checked on 11 November 2026, reports that Google accounted for 24.74% of tracked traffic in October 2026, compared with 0.46% for AI assistants. On those rounded shares, Google accounted for roughly 54 times as much traffic.

That is a comparison from websites using Ahrefs’ analytics. It is not a representative measurement of the entire internet or a forecast for your site.

Google traffic can include its own AI search features, so “Google versus AI assistants” is more accurate than “traditional search versus AI.”

Direct referrals also miss some influence. Someone can see an AI recommendation and later search for the brand. Missing information about where a visit came from can affect its classification too.

A small channel can produce valuable leads. Report volume, growth and conversion quality together.

Going from two visits to six is 200% growth. It is also four extra visits. Both facts belong in the meeting.

What I would put on the backlog

First, check whether relevant systems can access the important pages. Fix confirmed problems with the web team.

For a practical review of an existing page, my AEO content audit guide covers structure, context and evidence.

Then improve content that helps people make decisions: clear product information, current facts, useful comparisons and evidence behind claims. Explain who the product suits and where its limitations are. Check that a visitor can find the answers without needing a guided tour of your navigation.

Measure how the brand is represented, including recommendations and warnings, alongside visits, enquiries and lead quality. Give new experiments a defined outcome and a sensible limit on the time spent.

Before adding another AEO task, I would ask:

  • Which platform and feature should this affect?
  • What evidence supports the recommendation?
  • Does the evidence show access, visibility or a business result?
  • What else could explain the change?
  • What result would make the work worthwhile?

“AEO is just SEO” can be too convenient as well. Being found, being considered and being recommended are different outcomes. Good SEO supports the first; it does not guarantee the others. The searches an AI system chooses and the way it represents a brand deserve attention.

But “wrong,” “not yet demonstrated” and “worked in this particular test” are different conclusions. We should keep them separate.

My backlog can accommodate uncertainty. It has less room for another mandatory task supported by a confident sentence.

The cats have already shown how easily those get through.

More writing connected by category or shared topics.