The language of AI search

Asking a generative AI tool a question and getting a written answer, rather than a list of links. The answer is often built from web pages the engine looks up, and may link to them. ChatGPT, Gemini, Perplexity, Claude, Microsoft Copilot and Google’s AI Overviews and AI Mode are all forms of AI search.

The practice of improving how a brand shows up in AI answers: whether it is named, how it is described, and which sources the engines rely on when they talk about it. GEO builds on SEO (search engine optimisation), but the goal is to be cited and recommended inside an answer rather than ranked in a list of links.

Also called AEO (answer engine optimisation) or AI search optimisation.

How often, and how favourably, a brand appears in AI answers to the questions its customers ask. It covers whether the brand is named, how it is framed and whether its own website is cited. On this site, the closest measures are Focus brand named, Focus brand’s own site cited and Brand framing.

The AI search audit from We AI & Data Advisors at We Communications. Each audit tracks a set of questions about a brand and its category across AI engines over time. The GEO Compass dataset on this site pools those audits, anonymised, into one reference set.

A product that answers questions with generative AI, such as ChatGPT, Gemini or Perplexity. The site uses “engine” for the product people use, and compares engines side by side.

Engines tracked →

The AI model that writes the answer, such as OpenAI’s GPT models, Google’s Gemini or Anthropic’s Claude. An engine wraps a model with extra features such as web search and shopping cards, so the same model can behave differently in different products.

A specific release of the model behind an engine, such as Claude Sonnet 4.6 or Claude Sonnet 5. A new version can change which sources an engine prefers.

Model updates on the Trends page →

The AI-written summary that appears above Google’s regular search results for many searches.

Google’s conversational AI search: a separate mode where you can ask follow-up questions and get a longer answer with sources.

Direct access to OpenAI’s models for developers, without the ChatGPT app around them. It is tracked separately from ChatGPT because it searches and cites differently.

A question put to an AI engine. Each audit tracks a fixed set of prompts that real people ask about an industry, and runs them repeatedly. Prompt wording is never published on this site.

Also called a question.

One response from one engine to one prompt on one run. “AI answers analysed” counts these.

Also called a chat.

A question that names the focus brand or one of its products, such as “Is [brand] reliable?”.

Branded vs. Unbranded →

A question about a need or a category that names no brand, such as “best travel insurance for families”. Unbranded questions show whether a brand is recommended when nobody asks for it by name.

Branded vs. Unbranded →

The brand an audit was built around. Its own websites, including product, campaign and market sites, are grouped as “Focus brands’ own websites”, so no individual brand can be identified.

What a question is about. Every question is classified into one of 12 themes, such as Recommendations & “best of”, Brand trust & reputation, and Price & value.

All 12 themes →

The country a question was asked from. Engines tailor their answers and sources to location, so the same question can get different answers in different markets.

The language the question was asked in. Answers in local languages often draw on a different set of sources from English answers.

The sector an audit covers, from Automotive to Travel & Tourism. Each industry pools audits for at least two organisations.

Industries →

A website an engine retrieved or cited while answering. Sources are counted by domain, and sub-domains roll up, so pubmed.ncbi.nlm.nih.gov counts as nih.gov.

A website address such as reddit.com or nih.gov. The site counts sources at this level.

An engine retrieves a page when it reads it while building an answer, and cites it when it points to it in the answer, usually as a link or footnote. Unless a chart says otherwise, the site counts both.

A source behind an answer. Each site counts once per answer, however many of its pages were used. “Source citations classified” is the total across the dataset.

Answers that retrieved or cited at least one web source. Most percentages about sources use these answers as the base, because some answers use no sources at all.

Also called source-using answers.

The average number of distinct websites behind an answer that uses sources.

Of all the sources in a selection, the percentage that belong to one source type or one site. It describes the overall source mix.

The percentage of answers with sources that cite a source type or a site at least once. It describes reach: a site can have a small share of citations but still appear in many answers.

The share of answers with sources that cite the focus brand’s own website. It shows whether the brand’s own content is shaping the answer.

The breakdown of citations by source type. It shows where an engine gets its information for a selection of answers.

Explore the source mix →

Every domain is assigned to one type: brand & corporate, earned media, social & UGC, reference, government & public sector, academic & research, institutions & associations, marketplaces & aggregators, or other. The main types are defined below.

Websites run by brands and companies, including product and investor pages.

A brand’s own site is its owned media.

Coverage a brand does not pay for or control: news, magazines, trade press, blogs and review publications.

Reddit, YouTube, forums, LinkedIn and other platforms where people publish their own posts, videos and reviews.

UGC stands for user-generated content.

Wikipedia, dictionaries, encyclopaedias and data reference sites.

Government, regulator and public-body websites.

Universities, journals, research institutes and medical literature such as PubMed.

NGOs, trade bodies, standards bodies and professional associations.

Retailers and comparison, booking and listing sites, such as online travel agencies and car marketplaces.

Reddit, YouTube, Wikipedia and LinkedIn are tracked one by one, because engines rely on them heavily and their use changes over time.

Platforms over time →

Companies, products, public bodies and other organisations that an answer names. Spelling variants are merged, so “Apple” and “Apple Inc.” count once.

The average number of tracked brands an answer names. It shows how crowded the answers in a category are.

The share of answers that name the focus brand. It is the simplest measure of AI visibility.

Whether the sentences that name brands in an answer are positive, neutral or mixed, or negative. It is scored for English-language answers only, with brand names masked, so a name cannot sway the score. Read it as a comparison between groups rather than a verdict on any one brand.

Also called tone or sentiment.

How tone is scored →

When an engine searches the web before answering, instead of relying only on what its model learned in training. Answers that use web search can cite current sources.

Web search on the Trends page →

The text a model learned from before it was released. Answers that do not use web search rely on it, so they can be out of date and usually cite nothing.

Product listings, usually with images, prices and retailer links, shown inside an AI answer.

Also called shopping or product cards.

Shopping on the Trends page →

Paid placements shown inside or alongside an AI answer.

Ads on the Trends page →

The same set of long-running audits, followed month by month. The Trends page uses a fixed panel, so changes reflect the engines rather than a change in who was measured.

One GEO Compass audit: a set of prompts tracked across AI engines for one brand, in one or more markets, over a period of time.

Also called an audit.

The client company behind one or more studies. A company with several studies counts once when minimum sizes are checked.

No client, study or question text is published. Results are pooled across studies, and each focus brand’s own websites are grouped, so no individual client can be identified.

Results are hidden when a selection covers fewer than 200 answers. Lists of sites appear only when a selection draws on studies for at least two organisations.

Anonymity rules →

The date stamp of the dataset behind the site, shown on the Methodology page. It changes when new studies are added.