Glossary
Forty-two terms. Twenty-eight earn a page of their own. The rest are defined here and nowhere else, because they could not carry a worked example without padding.
Every term here has to survive one question: what would settle it? Name the check, the reading, the log line or the run that would show the claim false, and the term has substance. Most of this vocabulary cannot answer that, which tells you something about the field it came from, and each page below carries a mark saying which kind it is.
That is the same question criterion 5 puts to every agency on the leaderboard, turned on our own vocabulary. It would be a poor look to score companies on whether their claims could be shown false while never asking it of our own.
| Term | Status | What would settle it |
|---|---|---|
| Answer engine | Checkable | Run one question through three surfaces and look at the shape of what comes back. A list of links is a search engine. A... |
| Answer engine optimisation (AEO) | Contested | By its method rather than by its name. A fixed prompt set, named engines, a stated cadence and a written failure... |
| Generative engine optimisation (GEO) | Contested | The same way as anything else here: prompt set, named engines, cadence, verbatim capture. The label on the invoice does... |
| Retrieval augmented generation (RAG) | Checkable | Citations are the visible evidence. An answer that links sources you can open has retrieved. An answer with no sources... |
| Grounding | Checkable | Open the cited source and check whether it says what the answer claims. That is the entire test, it costs two minutes,... |
| Citation | Checkable | Count them. For a fixed set of questions, record every domain cited across every run. The list is short and it repeats. |
| Chunk retrieval | Correlation | Count the words before the answer to the page's own question. Then read the answering paragraph alone, out of context,... |
| Query fan-out | Checkable | Read the answer and count the things it told you that you did not ask. Those are the sub-questions. Do it across ten... |
| Prompt variation | Checkable | Write five phrasings of one need, run each ten times, and compare. If your appearance rate swings widely across... |
| Share of voice in AI answers | Checkable | Count every company name returned across the whole prompt set. Yours divided by the total. Report the competitor... |
| Featured snippet | Checkable | Search the query and look. The box is either there or it is not, and if it is, the text is either yours or somebody... |
| People Also Ask | Mis-sold | Search your target query and read the questions. Then check whether your pages answer them in plain language. That is... |
| Entity | Checkable | Search your exact company name plus your town. Does a knowledge panel appear? Do the listings that come back all refer... |
| Entity salience | Checkable | Run the page text through an entity analysis API and read the salience scores. No interpretation needed, and it is one... |
| Knowledge graph | Checkable | Search your business name. A knowledge panel appearing is the visible sign. For a definitive answer, query the Knowledge... |
| sameAs | Checkable | View your page source, find the structured data, and check whether sameAs exists and what it contains. Then open every... |
| Author entity | Checkable | Does the page name a person? Does Person markup exist and validate? Does the bio state a real credential with an issuing... |
| E-E-A-T | Contested | You cannot measure it. You can inventory the assets: is there a named author with a bio, a verifiable licence, a... |
| robots.txt for AI crawlers | Checkable | Visit yoursite.com/robots.txt and read it. Look for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended.... |
| ClaudeBot | Checkable | Two checks. Grep your server access logs for the user agent string to see whether it has visited and how often. Read... |
| Training data cutoff | Checkable | Ask the model its cutoff, then test it with something you know postdates that. If it knows a fact from last month,... |
| Hallucination | Checkable | Run your own business through the engines, read every factual claim, and check each against what is true. Log what is... |
| Zero-click search | Checkable | Search Console gives impressions and clicks for the same queries. Falling clicks against flat impressions and flat... |
| Extractive vs generative answer | Checkable | Copy a distinctive sentence from the answer and search it in quotes. If it appears on a page, the answer is extractive.... |
| Citation rate | Checkable | Fix a prompt set. Run each prompt several times per period, logged out, from the market you serve, against each engine... |
| Brand mention rate | Checkable | Same runs as your citation rate, counted differently: search each recorded answer for your company name whether or not a... |
| Non-determinism | Checkable | Test it yourself in five minutes. Open a logged-out session, ask who the best heating contractor in your town is, and... |
| Prompt set | Checkable | Ask for the list in writing at the start, then check every report covers the same list. If questions appear, disappear... |
Defined here, with no page of their own
These could not carry the five-part test. Two of them are not standardised vocabulary at all, and saying so is more useful than a page pretending otherwise.
- Source attribution
- Too close to citation to survive a separate worked example. Where it differs, it means the naming of a source rather than the link to it, which is brand mention rate.
- Embedding
- The numerical representation of text a model works with. Real, and it cannot carry an HVAC worked example without becoming a machine-learning tutorial and drifting off what this site is about.
- Vector similarity
- How close two of those representations sit. Same reason.
- Synthetic query
- Overlaps query fan-out with no separate test of its own.
- Knowledge panel
- The visible box. It is the artefact of the knowledge graph, measured the same way, with the same example.
- AI Overview
- Has a pillar page. A term page would compete with it.
- AI Mode
- Covered on the AI Overviews page, which is also where the two get told apart.
- Schema markup
- Has a pillar page.
- Organization schema
- Covered on the same pillar, where the specific types are listed.
- llms.txt
- Covered at whether llms.txt does anything, which is the honest length for it.
- GPTBot
- Covered on the ChatGPT page, alongside OAI-SearchBot, because the pair only make sense together.
- PerplexityBot
- Covered on the Perplexity page.
- Model grounding window
- Not standardised. It appears in vendor material meaning roughly how recent an engine's material is, but no engine publishes anything by that name. Treat a proposal using it as using its own private vocabulary, and ask what it is measured against.
- Consensus answer
- Not standardised. Used loosely for the impression that several sources agree. No engine publishes a mechanism by which one is computed, so there is nothing here to measure.