Foundations

People increasingly ask an AI assistant questions they used to type into a search engine: which tool to buy, which agency to hire, which hotel to book. The assistant replies with a short answer that names a few brands and often a few sources. Many users never click through to a list of results.

For your client, that creates a new question: when someone asks an assistant about our category, are we in the answer, and what does it say about us? Most clients cannot answer that today. That gap is the service you can sell.

You will hear several labels for closely related work:

• GEO (generative engine optimization): improving a brand's presence in answers generated by AI systems. The term was introduced in a 2023 research paper.
• AEO (answer engine optimization): improving the chance of being the answer, whether the answer comes from an AI assistant, an AI summary in a search engine or a featured result.
• AI visibility: the outcome you measure. How often and how well a brand appears in AI answers.
• AI SEO, LLMO and similar labels: informal names for the same area. Clients may use any of them.

In practice the day-to-day work is the same. Use the client's word in conversation, and use "AI visibility" when you report results.

The unit of success changes. In SEO you track positions and clicks for keywords. In GEO you track whether the brand is mentioned, cited and described correctly across a set of questions.

Answers are composed, not ranked. An assistant combines what it learned in training with information it retrieves, then writes one answer. Several sources can influence it at once.

Third-party sources matter more. Assistants often draw on communities, video, encyclopedic pages, review sites and industry media, not only the brand's own website. Which ones matter depends on the sector.

Results vary. The same question can return different brands on different days or in different assistants. You measure patterns over time, not single answers.

Description matters, not only presence. Being mentioned with the wrong facts, or in a negative tone, is a problem you can find and work on.

Much of what you already do still applies.

• Crawlers must be able to reach and read the site.
• Content must be clear, accurate, well structured and useful.
• Authority and reputation still influence which sources get trusted.
• Technical basics, such as speed and clean markup, still help.

This is why SEO agencies are well placed to sell GEO. You are extending a discipline you already practice, with new measurement and new places to build presence.

What the model learned (its memory). During training, a model absorbs patterns from large amounts of text. This shapes how well it "knows" a brand: what the brand does, who it serves and how it compares. If a brand is rarely or inconsistently described across the web, the model's knowledge of it can be thin or wrong. CiteMind's Brand Knowledge score looks at how confident LLMs are when answering user questions about a client, segmented by category.

What it retrieves (live information). Many assistants can search the web or other sources while answering, then use the pages they find. Here, the sources the assistant picks matter, and so does whether its crawlers can access your client's pages.

How much an assistant leans on memory versus retrieval varies by assistant, by question and by settings.

The exact process differs between systems, but a typical flow looks like this:

1. The user asks a question, often longer and more conversational than a search query.
2. The assistant interprets the intent. Some systems break the question into several searches.
3. It draws on its trained knowledge and, where available, on retrieved sources.
4. It writes a single answer, which may name a few brands and cite a few sources.

Your lever is in steps 3 and 4: what the model knows about the brand, and which sources it finds and trusts.

Answers are generated, not looked up. Several things cause variation:

• Different assistants use different models, sources and rules.
• Different runs of the same question can produce different wording and different brand lists.
• Model updates change behavior. A change in a model or in how it retrieves sources can shift results overnight.
• Phrasing changes the answer. "Best CRM for small agencies" and "What CRM should a five-person agency use?" can return different brands.
• Context such as location, language or conversation history can affect results.

This is why one screenshot proves nothing. A fixed set of questions, run on a schedule across several assistants, shows what is consistent and what is noise.

No one outside the labs knows the exact rules, and they change. These factors are widely observed in practice:

• Clear, direct answers on the brand's own pages, such as comparison pages, FAQs, pricing and service descriptions.
• Consistent facts across the website, profiles and third-party mentions.
• Presence on sources assistants draw from. These often include communities, video, encyclopedic pages, review sites and trade media. The mix differs by sector, so check it for each client.
• Access. If a search-oriented crawler is blocked, the assistant may not be able to use the page.
• Freshness and accuracy of the information.

Treat these as hypotheses to test per client, not rules. CiteMind shows the sources behind answers so you can see what is being used for your client's questions.

A guarantee such as "we will get you into ChatGPT's top recommendations" ties your retainer to something you do not control. Answers vary between runs and assistants, and a model update can move results without any change on your side. One bad month under a guarantee can end the account.

The agencies that keep GEO clients set expectations on the work and the method, then show progress on a trend.

These are within your control and can be written into scope:

• A baseline. A measured starting point across the assistants that matter to the client.
• Ongoing monitoring. A fixed set of questions tracked on a schedule, with visibility, Brand Knowledge and sentiment over time.
• A prioritized plan. Specific actions ranked by expected impact and effort.
• Execution of agreed actions. Content, structure, technical fixes and presence work, delivered on a cadence.
• Clear reporting. A regular report that shows what changed, what you did and what comes next.
• Honest diagnosis. When something drops, you investigate the cause and tell the client.

Avoid these promises, in writing and in conversation:

• A guaranteed mention, citation or position in any assistant.
• A specific percentage increase in visibility by a date.
• Guaranteed traffic or revenue from AI answers.
• Control over what an assistant says.
• Claims that you know "the algorithm."

If a prospect asks for a guarantee, say what you can offer instead: a measured baseline, a plan and a monthly trend report.

Use language that describes method and outcome you can evidence. Example wording you can adapt:

Example scope sentence: "We will measure {{client}}'s AI visibility across {{assistants}} using a fixed set of {{number}} questions, deliver a prioritized action plan, implement the agreed actions each month and report trends in visibility, Brand Knowledge and sentiment."

Example expectations sentence: "AI assistants generate answers that vary by run and change as models update. We report trends over time and do not guarantee specific mentions or positions."

Example success measure: "Success is a measurable improvement in the trend of visibility and sentiment across the tracked questions, alongside completion of the agreed actions."