The words behind search and AI Search

How often AI assistants name, recommend and cite your brand when buyers ask questions about your category that do not mention you. The word is often used for a single score, while three different questions sit underneath it.

At Recommendable: We always report visibility as three separate numbers per topic and per assistant: mentioned, recommended and cited. Questions that already contain your brand name are measured, but kept out of the totals.

The share of answers in which the assistant names your brand, anywhere in the answer. This measures presence and says nothing yet about preference.

The share of answers in which the assistant puts your brand first: the option its advice settles on, or number one in an order it gives as a preference. You can appear in a list and still never be the recommendation.

At Recommendable: This is the number we give the most weight, because customers come from a recommendation.

The share of buying conversations that end with your brand as the choice. The buyer keeps asking until one option is left, and that option is chosen.

At Recommendable: We measure chosen in Conversations: a second model plays a buyer from your audience and keeps asking until it decides. Together with mentioned and recommended it forms the ladder from the first answer to the choice.

A citation is a source an assistant links to in its answer. Cited is the share of answers that link to a page on your own site as a source.

At Recommendable: Citations only come from live answers, since a model gives no sources when it answers from memory. The list of cited domains per topic works as an outreach list.

A tool that sends a fixed set of prompts to AI assistants every day and counts whether your brand appears in the answer. That measures the outcome and says little about the cause.

At Recommendable: Recommendable also measures mentioned, recommended and cited, repeatedly and with a band, and also measures what drives that outcome: your body of evidence, what the model looks up and how the conversation runs.

The range the true rate most likely sits in, given the number of answers. A language model answers the same question differently each time, so one measurement says almost nothing; a narrow band means a solid figure.

At Recommendable: Next to every rate we show the band of the Wilson score interval at 95 percent. Below five usable answers we show a count instead of a rate.

The label next to a figure says how much you can lean on it and follows from the width of the band. Robust is a narrow band you can build strategy on, Adequate gives a direction, Minimal is too little signal to base advice on.

At Recommendable: A statement that is an interpretation and not a measurement gets a fourth label: Interpretation. That way you do not mix the two up.

Measuring the same topics again later on the same instrument, to see what really changed. A change only counts once the bands of the two measurements no longer touch.

At Recommendable: The AI Search Strategy includes one remeasurement after three months. When a provider changes model, the new measurements start their own series and are never averaged with the old ones.

A multi-turn conversation with an AI assistant, from the first question to a choice, through exploring, comparing and choosing. The decision is rarely made at the first answer.

At Recommendable: In Conversations a second model plays a buyer from your audience and keeps asking until it chooses. You see where you enter, where you drop out, to whom and for what reason.

An AI answer uses your content as a source but recommends another brand. Being cited is not the same as being recommended.

At Recommendable: Because we measure mentioned, recommended and cited separately, you see when you are the source but not the recommendation.

Searching and choosing through AI assistants such as ChatGPT, Claude, Gemini and Perplexity, and the discipline of getting a brand recommended there. SEO optimises a page for a query; AI Search optimises a brand for a recommendation.

At Recommendable: We measure AI Search at four assistants: ChatGPT, Claude, Gemini and Perplexity, through the providers' APIs.

A common name for optimising content and brand for search engines and assistants that write their own answer. The term comes from academic research; on this site we mostly say AI Search.

An assistant answers from memory, meaning what it learned in training, or live, where it may search the web and base its answer on the sources it finds (grounding). Live offers search without forcing it, so a live answer can still come from memory.

At Recommendable: We measure both modes separately and show per topic in what percentage of live answers a search actually happened. Perplexity always searches, so there we measure live only.

Before an AI search engine answers, it splits one question into several sub-questions and searches each of them. Those sub-questions decide which pages get read.

At Recommendable: We capture the sub-questions from live answers and cluster them into the subjects the models actually research. Perplexity does not disclose its sub-questions.

What an assistant looks up before it answers: the sub-questions it searches and the pages it reads. Part of those searches targets one specific site, to check a candidate.

At Recommendable: We test whether your pages contain the words assistants use to check your site, and put the missing words per page in your action plan.

The AI-written summary Google shows at the top of a results page. It draws on a small number of pages.

At Recommendable: Per topic we read whether Google shows an AI overview, which sites it cites and whether your site is among them. AI Mode is kept separate, because it cites other sites. Every reading is kept, so you see the line over time.

What someone wants to achieve with a search or a question to an assistant. The classic model has four kinds: informational, navigational, commercial and transactional. The Search Intent Framework distinguishes nine.

At Recommendable: Every topic in an audit carries an intent from the Search Intent Framework.

Our model of search behaviour: nine intents, 31 subcategories and fifteen platforms, scored on how well they fit each intent. The intent decides the platform, from Google and AI assistants to YouTube, TikTok and Reddit.

Searching inside platforms such as YouTube, TikTok, Instagram and Reddit instead of a search engine. What is said about a brand there, AI assistants also read as a source.

At Recommendable: We read YouTube, TikTok and Reddit as sources for what is said about your brand. What people search for inside those platforms we do not measure yet.

Questions and recommendations in closed channels such as WhatsApp, Slack and Discord. They steer many choices, but no tool can measure them.

Everything a model finds about your brand: what you claim, what your site proves and what others say about you in reviews, trade press, forums and video. A model recommends what it keeps running into, and most of the evidence sits outside your own site.

At Recommendable: An audit measures that body of evidence, what the model has made of it, and only then the answer and the conversation.

The situation that makes someone start thinking about your category, such as a move, a contract running out or a problem that keeps coming back. People ask a question because something happened.

At Recommendable: Topics maps the buying moments in your market and groups the topics we measure by moment. Each moment gets a weight and a link to an ICP.

A subject buyers in your category ask search engines and AI assistants about, with the questions that fall under it. Some topics are looked up with a fixed search term, others are a challenge without a fixed term.

At Recommendable: Per topic you see how much demand there is, whether you have a page for it and how your brand does in live AI answers. The questions we measure come from these topics.

The conditions on which a model tips its advice, such as budget, maintenance or service network. They decide which brand becomes the recommendation for which buyer.

At Recommendable: We extract the decision axes from the answers and show whether you are named on each axis and whether your site answers it.

A thing that can be identified without ambiguity, such as a brand, person or product, with fixed properties and relations. Language models and search engines organise knowledge around entities; if your brand is not clearly defined online, a model cannot connect the dots.

At Recommendable: In Reputation you see what each assistant associates your brand with, set against your brand profile. Rounds in which an assistant confuses you with another organisation are flagged and left out of the scores.

The ideal customer profile (ICP) describes who you are the best choice for. A persona makes that concrete as a typical buyer with a situation and priorities.

At Recommendable: You set down ICP personas in the brand profile. They steer which buying moments we research and which buyers talk in the simulated buyer conversations.

Experience, Expertise, Authoritativeness and Trustworthiness: the framework from Google's quality rater guidelines for judging whether content can be trusted. It covers signals such as who the author is, how current a page is and who links to it.

At Recommendable: The AI Search Website Audit checks E-E-A-T signals on your editorial and commercial pages, such as author bylines with credentials and dates.

Code on a page, usually JSON-LD following Schema.org, that tells machines what is there: an organisation, a product, a price, a review. A search engine or model then does not have to infer it from the running text.

At Recommendable: The SEO layer of the AI Search Website Audit checks your structured data. The tool has ready-made files for missing markup.

A text file at the root of your site that gives language models a short description and your most important pages. It is a proposed standard; not every AI provider says it uses it.

At Recommendable: The AI Search Website Audit checks whether your site has an llms.txt and whether a page is listed in it.

Whether AI assistants and AI agents can reach, read, use and cite your site. An agent acting for a customer also needs to get things done: search, fill in a form, check out.

At Recommendable: The Agents layer of the AI Search Website Audit looks at AI crawler access in robots.txt, readability without a browser, how much text an agent has to work through, protocols such as MCP, security signals and citability.