What generative engine optimization is, and what it cannot promise

What is generative engine optimization? Generative engine optimization, or GEO, is the work of getting your brand named inside an AI answer, rather than ranked on a page of blue links.

I am Kevin Pantanella. Syndral is my AI visibility agency, part of Suzaku Productions, which I founded in Bangkok in 2014. AEO, answer engine optimization, is the older neighbouring label, and the market now uses the two almost interchangeably.

Most pages that define GEO need the answer to be complicated, because they are selling it. The base layer is the SEO you already know. What changes is the question asked of your pages, where the result shows up, how you find out, and what anyone can honestly promise about it.

Generative engine optimization is the practice of making a brand nameable and citable inside the answer an assistant writes. The generative engines that matter today are ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude. Each one composes an answer instead of listing pages.

A search engine returns a page of links and you compete for a position on it. An assistant returns one paragraph or one table, and three to seven companies get named. If you are not in it, you are not on page two. You are nowhere.

That second difference gets missed. A buyer rarely opens with your category. They ask whether their own company shows up in AI answers. Then how that gets fixed. Then whether anyone does this for a living, what it costs, and whether anyone does it near them. Five turns, one thread, and your brand can be named at any of them or at none.

That is what you can measure. There is rarely a position to look up. There is whether you were named, and how often.

GEO was named in an academic paper, not by an agency. Pranjal Aggarwal and five co-authors published GEO: Generative Engine Optimization in 2023 and presented it at KDD 2024. They reported that their methods "can boost visibility by up to 40%", measured on GEO-bench, a benchmark the same authors built for the experiment.

That last clause is the one the category drops. The experiment optimized source documents against a benchmark. It did not take a live brand into a live market against live competitors.

The paper is also more careful than its headline. The authors write that the effectiveness of these strategies "varies across domains", which argues against any portable number. What worked in their tests was unglamorous: citing sources, adding quotations, adding statistics.

So the mechanism is real. Nobody can sell you the 40 percent.

Google publishes its position and almost nobody in this category quotes it. Its AI features guidance, updated 10 July 2026, states that "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO". It adds that you do not need new machine readable files, AI text files, markup or Markdown to appear in Search.

Read that with the interest attached. Google has nothing to gain from your visibility inside an assistant it does not own, and "you need nothing new" is the advice that suits the engine already holding the traffic. It is still worth reading, because Google is the only one publishing a position at all.

Google also backs its claim with a tool. The same page points you to the Generative AI performance report in Search Console, for content appearing in its AI features.

That report exists for Google and for nothing else. A surface is just a place an answer appears: AI Overviews, ChatGPT, Perplexity, Gemini. Google hands you its own numbers on the surfaces it owns, and ChatGPT, Perplexity and Claude hand you nothing.

Take llms.txt as the test case. No engine has confirmed it does anything, Google says it does not support it, and Syndral publishes one anyway because it costs an hour. That is the right posture under uncertainty. It is not the same as calling it a strategy.

The model does not decide who gets named. The retrieval chain in front of it does, the machinery that goes out and fetches pages before a word is written, and each provider builds its own. OpenAI documents a search crawler, OAI-SearchBot, separate from the one that gathers training data. Anthropic and Perplexity document the same split. Different pipelines, different sources at answer time.

What one question returned on three Google surfaces

Here is what one question looked like across three surfaces. Syndral ran it on 13 August 2026 in Bangkok, with the query written down before it was asked, and Syndral appears in none of the three answers.

Query: AI visibility Agency Bangkok. One pass per surface, signed in to a Google account, so personalization is not excluded. Counting Inspira Digital Agency and Inspira as one company, and the two IBEX entries as one, eleven distinct companies were named. One appeared on all three, ANGA.

Now read that as an illustration and not as a measurement, for a reason this article supplies further down. One pass cannot separate a difference between surfaces from the ordinary variation inside one of them. We ran AI Mode again seventy-five minutes later, signed out, and it shared three of its ten names with the first reading.

Which is the same lesson arriving from the other direction. The work has to be checked per surface, and more than once. An agency that promises you "AI visibility" without naming the surface, or without saying how many times it asked, is promising nothing you can check.

Being named is not one skill. Syndral splits it into three layers, Reach, Answers and Corroboration, and they fail independently. Pass two and fail the third and you are still absent, which is why a single fix so rarely moves anything. These are the layers our own ScoreCard scores, and you can check all three yourself this week without buying anything.

Reach is the one people get wrong, because it is invisible from a browser. Syndral's own site returned 403 to AI crawlers for weeks and nobody noticed. HubSpot's free AI Search Grader scored it identically before and after the fix. The full test is in the AI visibility score field study.

Every major AI company gives you separate robots.txt controls for answers and for training, and blocking the wrong one removes you from answers while doing nothing about training. The documentation is public and dated. Read it before you touch a toggle.

These are controls, not always separate visits. OpenAI says that if you allow both bots, it may use one crawl for both. Apple is blunter: Applebot data may also train its foundation models, and Applebot-Extended opts out of that alone.

Disallow Applebot-Extended and your content stays discoverable in Siri, Spotlight and Safari. Two different decisions, one of which people make by accident.

What actually visited syndral.co in 24 hours

Now the part only a log can tell you. Syndral read its own traffic in Cloudflare AI Crawl Control on 5 August 2026. The tool logged 45 crawler requests in 24 hours, all allowed. Googlebot made 27 of those, and Googlebot serves classic Search, so 18 came from AI-specific agents. ClaudeBot made 13. GPTBot made 3, BingBot 2, and PerplexityBot none at all.

Zero, on a site we interrogate on Perplexity every month. What visits you is not what you assume visits you.

Nobody can promise you a position, because positions mostly do not exist here. In January 2026 SparkToro and Gumshoe published a study of how consistent AI recommendations are. Six hundred volunteers ran 12 prompts through each of three tools, ChatGPT, Claude and Google's AI Overviews, for 2,961 runs in total.

They tested ChatGPT, Claude and Google's AI Overviews. For ChatGPT and Google, under a 1 in 100 chance of returning the same list of brands twice. Claude was only slightly better. Rand Fishkin flags his own conflict of interest, his co-author works for an AI tracking vendor, and publishes the method anyway.

A screenshot of a good answer is therefore not a measurement. It is one draw.

Some surfaces cannot be measured today, which is different from being unreachable. Your pages stay discoverable inside Siri, and Apple publishes no report of what Siri says about you. So an agency selling Siri optimization is selling a result that, as things stand in August 2026, neither of you can verify. That may change, and if Apple ships a reporting surface we will say so.

Worth watching rather than acting on. Siri's next version runs on Apple's own retrieval chain, and it arrives by default on a very large number of phones. A surface that is almost silent today becomes a large one quickly, without ever becoming measurable. Naming the unmeasured as unmeasured is the difference between a method and a pitch.

The honest unit is how often you get named across repeated questions, compared between two dates with the same instrument. Not a screenshot, not one good answer, a rate. It is the only number that survives the instability above. Four steps, and you can run them yourself before you hire anyone.

1. Freeze a set of buyer questions, written down before you ask them.
2. Ask each question several times on each surface, in a temporary or incognito chat with personalization cleared, or your own history answers for you.
3. Record how often your brand is named, as a rate and not as a screenshot.
4. Re-run the identical set later, same questions, same surfaces, same method.

On Google, read the Generative AI performance report in Search Console alongside that. Everywhere else, the manual run is the only instrument you control.

One action, if you take one. Run the free AI visibility ScoreCard. Eleven questions, about two minutes, and it tells you which of the three conditions above you fail. It refuses to hand you a visibility number, because a form cannot measure what an assistant says about you.

For the difference between the three labels, read SEO vs GEO vs AEO. The rest of the writing lives in resources.