Take AEO with a grain of salt

Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs®. Each week, I dig into research on AI, knowledge management, and the future of work so you don't have to.

Last week I left a question open: who still carries the reference layer, and who gets to evaluate it. Companies chasing AEO already have an answer they like. They assume models look for the same things search engines do.

They do not. Every AI system can be trained or prompted toward different criteria. Uniformity was never the default. This week is the evidence for that, and a more practical problem underneath it. Even when you do get cited, the person on the other side of a chatbot is rarely about to visit.

Optimizing your website for generative AI features on Google Search | Google Search Central

Why It Matters: Google's official line is that you do not need a separate AEO or GEO playbook for AI Overviews and AI Mode. Those features sit on core Search ranking and the same helpful-content rules, so optimizing for generative AI is still SEO. Ignore the third-party hacks, at least if Google Search is the only engine you care about.

What bothers me is the incentive underneath the advice. Google is cannibalizing its main Search revenue while the AI side still does not have a settled way to pay the bills. Guidance that looks like best practice is self-serving in the moment. Google has to keep the old machine running while it builds the new one. Other companies can go all-in. Google cannot.

So if the vendor with the most to protect says the engine is still search, what does an independent look at a different stack actually find?

Inside ChatGPT's retrieval stack: The index, cache, and pages it actually reads | Search Engine Land

Why It Matters: This is reverse engineering from the live product, not a vendor playbook, which is why I find it more credible and more substantial. ChatGPT finds pages through its own index, reads a shared cache of pages already fetched, and only sometimes opens URLs live. Free and paid users do not even see the same corpus. Instant mode leans on OpenAI's own index and snippets. Thinking mode spends money on scraped Google results and live fetches. The citation rules are not Google SEO.

I would bet Gemini's architecture is similar in principle: an index, a cache, and a live fetch path, each with its own staleness and blind spots. That is the opposite of one ranking game everyone can optimize against.

You do not even have to leave Google to see the split. Two of its own products can agree on the answer and still pick different sources.

Are AI Mode and AI Overviews Just Different Versions of the Same Answer? (730K Responses Studied) | Ahrefs

Why It Matters: Ahrefs compared 730,000 paired responses. The two surfaces agreed semantically about 86 percent of the time and shared cited URLs only about 13.7 percent of the time. Same company, same query, different citation set.

That is less mysterious once you remember that a conversational query carries different context than a traditional search. AI Mode leans on different kinds of knowledge, including visuals, and it writes a longer grounded answer. On health-style questions, my read is that Google is choosing what it believes is absolutely trustworthy, because the user is unlikely to dig past the answer. AI Overviews still sit on a results page with a click-through path. The company is adjusting by knowledge intent, and by how it thinks people will actually consume the information. Optimizing for "being cited" is a consumption bet, not a single ranking.

Once the prize is a citation rather than a visit, the web starts to reshape itself for the bot.

Your Search Results Are Getting Sloptimized | The Atlantic

Why It Matters: Will Oremus names the scramble GEO, or sloptimization: brands publishing self-ranking listicles, seeding forums, and writing pages meant to be ingested by a model rather than read by a person. The recommendation is the point now, not the click.

I do not think sloptimization is the story so much as the symptom. We are going to need an evolution in how quality gets evaluated, the same kind of shift that produced PageRank in the early web. Links were a useful proxy for trust. They were never the whole of quality. The better frame is more holistic: who is consuming the information, what they are looking for, and whether a citation even implies anyone will show up. The evaluation layer has not caught up to conversational retrieval.

A Cambridge paper on attribution puts a number on how hard that evaluation is when the logs themselves are incomplete.

The attribution crisis in LLM search results: Estimating ecosystem exploitation | Data & Policy / Cambridge

Why It Matters: Drawing on about 14,000 LMArena conversation logs, Ilan Strauss, Jangho Yang, Tim O'Reilly, and colleagues measure the gap between pages a model used and pages it credited. The gaps are model-specific, and disclosure is selective. What looks like tight attribution may reflect what the logs show, not what the system did. It is hard to audit who was used versus who was cited.

That is why the paper supports the PageRank-evolution point. If we cannot see the retrieval, we cannot evaluate quality, credit, or harm. I am supportive of systems that improve online citations and transparency. Until search traces and citation logs are disclosed in a more standard way, AEO is a bet on a black box.

Put the five together and last week's question gets a harder answer: the evaluators do not share a ranking. Google tells you to keep doing SEO because it still has a Search business to protect. ChatGPT retrieves through an index, a cache, and live fetches that are not Google's rules, and free users do not even see the same web as paid ones. Google's own AI Mode and AI Overviews can agree on the answer and still share only 13.7 percent of cited URLs. Brands respond by sloptimizing pages for the bot. And those 14,000 logs show that "used" and "credited" are not the same event.

The walkaway is simpler than the stack. A citation is not a visit. Take AEO techniques with a grain of salt. Every model is different, and there is not much documented around any of this. Even when a chatbot cites you, the intent is different from a standard web query, so the person is less likely to click through. Evaluate the intent of the audience. Ask whether traffic via bots is actually going to matter for growth, or whether the better move is to put the energy somewhere a human still shows up.

I have been turning over a related problem: models that do not fully show their chain of thought. Astra is the one on my mind, and what that does to the next evolution of evaluation. If you cannot see how the system decided, you cannot audit the citation, let alone the visit.

Know someone treating AEO like a new SEO playbook? Forward this to them.

Have a take on any of this? Reply directly. I read every response.