AI Search Optimization: The Complete Guide to GEO and AEO

What is documented, what is inferred and what is hype

Most of what is sold as AI search optimization is ordinary SEO, plus crawler access, plus measurement. A small part is documented by the companies that run the engines. A larger part is inferred from studies of what the engines output. The rest is unproven, and some of it is false.

We sort every claim into one of those groups. This table is the quickest way to judge a proposal, a tool or an article, including ours.

Treat the status column as a snapshot. We checked these sources in November 2026, and several of them were updated in the previous six months.

How the main assistants find and use your pages

Each engine tells site owners something different about access, and the differences matter more than any tactic.

Two practical consequences follow. First, a single blanket rule that blocks "AI bots" can remove you from search features you wanted, because training bots and search bots are separate user agents. Second, Google-Extended, the token people mention most, does not affect inclusion in Google Search, according to Google's crawler documentation. It governs Gemini training and grounding, which is a different decision.

The work, in the order we would do it

This is the sequence we would follow for a small or mid-size business. It is deliberately dull, because the documented parts are dull.

1. Fix crawling and indexing first. If a page is not indexed, no AI feature built on the search index can link to it. Our technical SEO guide covers the checks.
2. Check bot access. Read your robots.txt and your CDN or firewall settings. Look for rules that block OAI-SearchBot, Claude-SearchBot or PerplexityBot by accident, and decide on training bots separately.
3. Answer the question in the first sentence of each section. Assistants quote passages, so each passage has to make sense alone. This helps human readers equally.
4. Add facts that only you can supply. Prices, process steps, local detail, your own data. Google's guide asks for non-commodity content, and the Princeton paper found that added statistics and quotations helped on its benchmark.
5. Keep your business facts consistent. Name, address, services, opening hours and prices should match across your site, your Google Business Profile and the directories you appear in.
6. Earn real mentions. Google's guide says manufacturing mentions is less useful than it looks. Customers, partners and local organizations that genuinely refer to you are the durable version.
7. Measure with a fixed prompt set. Pair it with first-party reports. How to track your brand in AI answers sets out the method.
8. Repeat monthly and expect noise. AI answers vary between runs, so a single check proves nothing.

What the research says about getting cited

Three studies carry most of the weight, and each has limits you should know before repeating its headline.

The Princeton-led GEO paper (2023). Aggarwal and colleagues built a benchmark of 10,000 queries across 25 domains and tested nine ways of rewriting content. Adding quotations, statistics and cited sources performed best, with gains of up to about 40% in their visibility measures. Keyword stuffing scored below the unoptimized baseline. The main experiments used a model from 2023, drew on the top five Google results and covered two engines in total. The authors say methods may need to change as engines change, and that they did not measure effects on search rankings.

Ahrefs on AI Overview citations (March 2026). Across about 863,000 keyword results and 4 million cited URLs, 37.9% of cited pages also ranked in the top 10, 31.2% ranked 11 to 100 and 31.0% were outside the top 100. Ahrefs' earlier study (July 2025) found about 76% in the top 10. Ahrefs lists a newer AI Overviews model, improved parsing and a dataset more than twice the size as differences, so the two figures are not directly comparable. The data comes from Ahrefs' own tool. The reading is that ranking well is no longer enough on its own, and that pages with a precise answer can be cited without ranking first.

Pew Research on clicks (2025). In browsing data from 900 US adults, 18% of 68,879 Google searches showed an AI summary. Visits with a summary led to a click on a regular result 8% of the time, against 15% without one. Only 1% of visits led to a click on a link inside the summary. This is one country, one engine and March 2025, so it shows direction and not a current rate.

None of these tells you what any one assistant will do for your pages next month. They support a modest claim: clear, sourced, specific content is a better bet than clever formatting.

What we would not pay for

Some offers are not worth money, whoever is selling them.

• An llms.txt package. Google's guide says its Search does not use the file, and none of the OpenAI, Anthropic or Perplexity crawler documentation we read says their search crawlers use llms.txt. If you want one, a plugin makes it in minutes. See what llms.txt is and whether you need one.
• Special "AI schema". Structured data is useful for rich results and for helping machines read your page, but Google says AI features need no special markup.
• Pages written for every fan-out query. Google warns that mass-producing pages for query variations to influence AI answers falls under its scaled content abuse policy.
• Anything described as direct access to Google's or OpenAI's systems. Google's guide states that no third-party tool has access to its internal ranking or AI systems.
• Citation guarantees. Nobody can offer one honestly. Google itself says indexing and serving are not guaranteed, even for pages that meet every requirement.

Does this change what a Cyprus business should do?

It changes the checklist more than the strategy. In our October 2026 sample of 33 commercial and informational search results, Google AI Overviews appeared on all 33, including "seo limassol". A re-test of 10 of them showed a real AI answer first each time. The sample was small and built around our own topics, so it shows that local and service searches are not exempt, and it is not a market-wide rate. We have not measured Greek-language results separately.

For a local business, the practical order is: Google Business Profile accurate and complete, a site that states services, areas and prices in plain text, and consistent details everywhere you are listed. Our local SEO guide for small businesses covers that base, and it is the same base AI features draw on.

Demand is also shifting in kind. In our October 2026 pull of DataForSEO's AI prompt estimates for the US, "is seo dead" came to about 8,650 modelled prompts a month against about 590 Google searches, and "seo cost" about 7,900. These are modelled estimates available only for top terms. They suggest people put questions to assistants that they do not type into Google.

Four questions to put to any AI search claim

Use these on an article, a tool's marketing page or a proposal. A claim that survives all four is worth acting on.

1. Who says so, and how do they know? The best answer is the company that runs the engine, in its own documentation. The next best is a named study with a stated method. A blog post that repeats a number with no method is the weakest.
2. How big and how old is the sample? The Princeton paper used a 2023 model. Pew's data is from March 2025. A figure can be real and still out of date.
3. Does it measure what the headline says? A correlation between having a file and being cited is not proof that the file causes citations. A gain in "visibility" on a research score is not extra customers.
4. What would change my mind? If nothing could disprove the claim, it is a belief and not a finding. Real claims name conditions under which they fail.

Applied to the table at the top of this page, these questions are why llms.txt sits at "no confirmed effect" and why crawler access sits at "documented".

Your first 30 days

Here is a plain schedule for a small business. It assumes no tools beyond the free ones.

Expect the second log to look much like the first. AI answers vary, and the changes you made in week two may take longer to be picked up. The first month builds a baseline and removes blockers. Judging results takes a second and third month of the same measurement.

Guides in this topic

Start with the definitions if the vocabulary is new, then move to the task you need.

Understand the terms and the argument

• What is generative engine optimization (GEO)? covers the origin of the term and the evidence.
• What is answer engine optimization (AEO)? covers answer-first page design.
• GEO vs SEO vs AEO compares what each label changes in practice.
• What is LLM SEO? explains training, retrieval and live fetching.
• Is SEO dead? takes a position with the data.

Google

• What are Google AI Overviews and how do they work?
• How to appear in Google AI Overviews
• How to optimize content for AI Overviews
• How to track AI Overviews

ChatGPT and other assistants

• How to rank in ChatGPT
• How to rank higher in ChatGPT search
• How to track your brand in AI answers
• How to use ChatGPT for SEO

llms.txt

• What is llms.txt and do you need one?
• How to add llms.txt on WordPress

Tools and buying help

• Best AI visibility tools compared
• Best answer engine optimization tools
• Best AI SEO tools
• AI search engine optimization tools: what you actually need
• AI SEO agency: what they do and how to choose one
• GEO services: what is included and what it costs

Where we fit

Sempor Labs is a studio in Limassol. What we offer is the process described above, written down, with the evidence labelled and the limits stated. Our AI search optimization service starts with a visibility audit, and we quote in writing after seeing your site.

The next useful step costs nothing. Open your robots.txt, search it for the bot names in the table above, and write down what you find.