Every term this field uses, defined in one passage each, with a permanent link. Where a definition rests on a measurement or a published source, that source is named in the definition rather than implied.
Definitions are the thing retrieval systems reuse most, so each entry here is written to survive being pulled out on its own. 20 terms, current as of September 2026.
Terms
Answer engine optimisation (AEO)
The work of becoming a source an AI engine quotes when it answers a question rather than returning links. Distinct from SEO, which competes for a position in a list. Kunzum measures it by asking engines the same questions repeatedly and recording what they name and cite.
Generative engine optimisation (GEO)
The academic term for the same work, introduced by Aggarwal and colleagues in a 2023 arXiv paper. In practice GEO and AEO are used interchangeably; GEO is the term in the literature, AEO the one buyers use.
Retrieval-augmented generation (RAG)
The architecture underneath every citing answer engine: fetch documents, then generate an answer from them. Described by Lewis and colleagues in 2020. It is why AEO can work at all, and it is where AEO stops — an answer generated without retrieval cannot be reached by anything published.
AI Overview
Google's generated answer above the ordinary results. Google's documentation states eligibility requires only that a page is indexed and can be shown with a snippet, with no special markup or file.
Retrieval
Whether an engine went to the web for a given answer. Not guaranteed: in Kunzum's September 2026 index, ChatGPT retrieved on 54% of answers and Claude on 65%, while Perplexity retrieved on 100%.
Citation
A source an engine links beside its answer. Kunzum logged 12,518 of them across 1,396 domains in September 2026. A citation is evidence the engine used a source, not evidence it represented it correctly.
Surface rate
How often a category appears in answers at all, separately from whether a specific product is named. In edition one, crypto surfaced in 127 of 360 unprompted answers while an actionable crypto product was named 6 times.
Prompt set
The fixed list of questions a measurement runs. Kunzum's two editions use 40 and 10 prompts respectively, published in full so the measurement can be reproduced or disputed.
Open-vocabulary scoring
Reading product names out of the responses rather than matching against a shortlist. It is slower and it is the only way to discover that a product you had not listed is winning the category.
OAI-SearchBot
OpenAI's crawler for surfacing sites in ChatGPT's search features, documented separately from GPTBot, which collects training data. A site can allow one and block the other.
GPTBot
OpenAI's crawler for content that may train its foundation models. Blocking it does not remove a site from ChatGPT search, which is governed by OAI-SearchBot.
PerplexityBot
Perplexity's crawler for surfacing and linking sites in its results. Perplexity's documentation states it is not used to crawl content for AI foundation models.
Google-Extended
Google's control over use of content for training and grounding in some of its systems, separate from Search indexing. Blocking Googlebot instead removes a page from Search entirely.
llms.txt
A proposed plain-text file describing a site for language models. Kunzum publishes one. Google's documentation states no such file is required to appear in its AI features, so it should not be sold as a mechanism.
Chunk
The passage-sized unit a retrieval system actually indexes and reuses. The practical consequence is that a claim and its evidence must sit in the same passage, because they will be separated otherwise.
Unsupported statement
An engine asserting something its own cited source does not support. Liu, Zhang and Liang measured this in generative search engines in 2023; it is why citation presence and claim accuracy have to be recorded as separate columns.
Visibility manipulation
Text added to a page in order to move how a model ranks or describes a product, rather than to inform a reader. Demonstrated by Kumar and Lakkaraju in 2024. It works, it is adversarial, and it breaks when the model changes.
Baseline
The recorded state of what engines said before any work started — prompts, answers and citations. Without one, no later change can be attributed to the work rather than to a model update.
Localisation
Whether an engine adapts its answer to the asker's location. Measured across nine markets in edition two: ChatGPT named a country-specific product in 85 of 240 answers, Google AI Overviews in 60, Perplexity in eight, and Claude in none of 180.
Where these definitions come from
The engine and crawler entries come from the engines' own documentation: Google Search Central on AI features, OpenAI’s crawler documentation, Perplexity’s crawler documentation and Anthropic’s crawler documentation, all read on 12 September 2026.
The method entries come from the literature — Aggarwal and colleagues on GEO, Liu, Zhang and Liang on verifiability, Kumar and Lakkaraju on visibility manipulation and Lewis and colleagues on retrieval-augmented generation.
The numbers come from Kunzum's own corpus: 1,500 responses and 12,518 citations, published raw. If a definition here disagrees with the data, the data is downloadable and this page is wrong.
Sources
- Google Search Central, “AI features and your website”. Google. Checked 2026-09-12.
- OpenAI’s crawler documentation. OpenAI. Checked 2026-09-12.
- Perplexity’s crawler documentation. Perplexity. Checked 2026-09-12.
- Anthropic’s crawler documentation. Anthropic. Checked 2026-09-12.
- Aggarwal and colleagues, “GEO: Generative Engine Optimization” (2023). arXiv:2311.09735. Checked 2026-09-12.
- Lewis and colleagues, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (2020). arXiv:2005.11401. Checked 2026-09-12.
Published 2026-09-12. Written by Narender Charan, who runs Kunzum.