Key takeaways
- AI perception is how AI models describe, rank and recommend your brand. It has its own vocabulary, and most of the terms are measurable.
- Separate the terms that measure presence (visibility, share of voice, position) from the ones that measure meaning (narrative, sentiment, hallucination) and the ones that measure sources (citations, grounding, fan-outs).
- Mentions and citations are different things. A brand can be named without being cited, and the fix is different in each case.
- Numbers move month to month. Quote the period with any figure, and treat vendor benchmarks as vendor claims.
- Tools like Promptwatch track most of these metrics from the real AI interfaces, but the vocabulary matters more than the tool.
Why marketers need a shared vocabulary
Every glossary I found while researching this tells you what a term means. Few tell you what to do about it. Knowing that a citation is a linked source in an AI answer is trivia. Knowing that your competitor's product page gets 22% of those links in your category is a plan.
So each term below has a short definition and a note on the decision it should change. They are grouped into five sets: measurement, meaning, sources, mechanics and strategy. If you only have five minutes, read the measurement and sources sections.

SalesMarketing.ai groups its 25 terms into Core, Structural, Intelligence, Surfaces and Adjacent. That is a reasonable taxonomy, though I have regrouped them around how a marketing team actually works.
Quick reference table
| Term | Plain meaning | Question it answers |
|---|---|---|
| AI perception | How AI describes and positions your brand | What do models think of us? |
| Visibility Score | Composite 0-100 presence number | Are we showing up? |
| Share of Voice | Your slice of all brand mentions | How do we compare? |
| Position | Order of brands in an answer | Are we first or fifth? |
| Narrative | The claims AI repeats about you | What story is told? |
| Sentiment | Tone of those claims | Is it favorable? |
| Citation | A source linked in an answer | Where does AI read about us? |
| Grounding | Tying an answer to retrieved sources | Is this live or from memory? |
| Query fan-out | One prompt split into several searches | What does AI actually search? |
Measurement terms
1. AI perception
The overall picture AI models form and present of your brand: whether they mention you, how they describe you, who they put you next to. Refractia calls the discipline AI Perception Intelligence. The label varies by vendor, but the idea is stable. It covers presence, tone and accuracy together, so no single metric captures it.
2. AI visibility
How often and how prominently your brand appears in AI-generated answers to relevant questions. It is the umbrella term. The next few entries are the specific ways to count it.
3. Visibility Score
A single number, usually 0-100, summarizing presence. Definitions differ by vendor. In Promptwatch, each response is scored on placement, share of attention and depth, then averaged, and responses that don't mention you count as zero. See the Visibility Score definition for the exact method. Treat any proprietary score as comparable only within the tool that produced it. Do not compare a score from one vendor against another's.
4. Mention frequency
The share of tracked responses that name your brand at all. It is the crudest metric and the easiest to explain to a boss. Low mention frequency means a coverage problem: the models don't associate you with the topic.
5. Share of Voice
Your brand's slice of all brand mentions across tracked prompts. If five brands are named across 100 responses and you appear in 20 of those mentions, that is the arithmetic. It tells you about competition, not just presence. See the Share of Voice explainer.
6. Position
Where your brand lands in the order of brands inside an answer. Lower is better. Being listed third of eight reads very differently from being the first recommendation. Late position with decent frequency is a prominence problem.
7. Recommendation share
SalesMarketing.ai uses this for how often a brand is actively recommended, rather than just named. It is useful for commercial prompts like "best CRM for agencies". Their audit data claims category leaders hold a median 34% in ChatGPT, but that is a vendor claim from their own audits, so use it as color rather than a benchmark.
Meaning terms
8. Narrative
The set of claims AI makes about you, such as "affordable for small teams" or "mobile app lacks features". A narrative is made of individual claims, grouped by theme: pricing, features, support, reputation. Promptwatch's Brand Narrative extracts those claims from every tracked response and scores each theme 0-100 against competitors. The point is that you can't fix a story you haven't read.
9. Narrative alignment
How closely AI's description matches the positioning you intend. High alignment means the model repeats your messaging. Low alignment often means the models are leaning on old reviews, third-party comparisons or competitor pages instead of your own site.
10. Sentiment
The tone of what AI says about you, scored separately from visibility. In Promptwatch it runs 0-100 with about 50 as neutral. A brand can be highly visible and poorly regarded, which is a reputation problem and not a coverage one.
11. Hallucination
An AI statement about your brand that is simply false: a feature you don't have, a price you never charged, a founder who doesn't exist. Some vendors pair this with fact-checking features that let you mark a claim inaccurate and trace the page that taught the model the wrong thing.
12. Brand drift
The slow divergence between how AI describes you and who you are now. It happens after a rebrand, a pricing change or a product pivot, because models keep repeating older material. Track narrative over time to catch it.
Source terms
13. Citation
A source an AI answer links to or lists. Citations are the evidence trail. They differ from mentions: citations versus mentions matters because a brand mentioned without a citation usually has a content or crawlability gap.
Promptwatch Data shows how this plays out in ChatGPT. For the week of Sep 28 to Oct 4, 2026, product pages were 22.4% of citations, landing pages 16.8% and listicles 13.6% (citation share by content type). Reviews were only 1.7% overall but 5.6% for branded prompts.
14. Citation share
Your domain's portion of all citations within a topic or prompt set. It is to citations what share of voice is to mentions.
15. Self-citation rate
The share of citations pointing at your own domain. If AI names you constantly but cites other people's pages, you are being described by third parties. Sometimes that is fine. Often it means your site is not readable or not useful.
16. Grounding
Grounding is when a model anchors its answer in retrieved, external sources rather than relying only on training data. A grounded answer comes with citations. An ungrounded one comes from what the model absorbed during training.
This distinction drives strategy. For grounded answers you win by being retrievable: crawlable, clear, current. For ungrounded answers you win slowly, by being widely and consistently described across the web. Both exist, and tracking real interfaces next to API models shows you each.
17. Retrieval surface and training surface
SalesMarketing.ai splits where AI learns about you into the training surface (what the model memorized) and the retrieval surface (what it fetches live). It is a useful frame, with one caveat: you can only influence training indirectly and slowly.
18. Source diversity
How many different domains feed answers on your topics. Promptwatch Data reports that almost half of ChatGPT responses cite 10 or more unique domains in that Sep 28 to Oct 4 week (unique domains per response). More sources per answer means more places you can earn a slot.
Citation volume also varies by engine. Per Promptwatch's sources per response data, ChatGPT cites about 5 sources per web-search answer, while Google AI Overviews and Perplexity cite about 10. Don't mix this up with brands recommended per answer. SalesMarketing.ai says answers compress to 2-4 recommended brands, which is a different measure from sources cited.
Mechanics terms
19. Query fan-out
When an AI splits one prompt into several web searches behind the scenes. Promptwatch Data tracked an average of about 1.84 fan-outs per ChatGPT response by early March 2026, down from 2.15 in December (ChatGPT query fan-outs). The practical tip: write headings that read like search queries, and cover the sub-questions on comparisons, pricing and alternatives. More in the query fan-outs guide.
20. Prompt (and prompt volume)
The question a user types. Unlike keywords, prompts are long and conversational, and volume is harder to know. Some platforms now estimate monthly prompt volumes and difficulty. Use them to prioritize, not as exact counts.
21. Persona
A simulated buyer profile used when running prompts, since AI answers can shift with who appears to be asking. Persona-level results show where your visibility differs between, say, a CFO and a developer.
22. AI crawler
The bots (GPTBot, ClaudeBot, PerplexityBot and others) that fetch your pages for training or live answers. If they get errors, you are invisible no matter how good the content is. Check access with a free AI crawler URL inspector or server-side crawler logs. Our reference list of bots is at promptwatch.com/bots.
Strategy terms
23. GEO (Generative Engine Optimization)
Optimizing your brand and content to be found, understood and recommended by AI systems. Also called AEO, LLM SEO or AI SEO. The names compete, but they describe the same job. Start with SEO for AI search if you want the overview.
24. Entity clarity
How unambiguously AI can tell who you are: name, category, products, relationships. Schema markup, consistent profiles and a Wikidata entry help. A brand name that is also a common word needs the most work here.
25. Content gap
A prompt AI answers where you have no page that fits. Gap analysis maps your pages against tracked prompts and shows what to write. See the content gap guide.
Reading the metrics together
No single number tells you what's wrong. Read them as a set:
| Symptom | Likely problem | First move |
|---|---|---|
| Low share of voice | Coverage | Publish pages for uncovered prompts |
| Late position | Prominence | Strengthen comparison and product pages |
| Mentioned, rarely cited | Content or crawlability | Check crawler access, rewrite thin pages |
| Weak sentiment | Reputation | Review narrative themes, fix source pages |
| Wrong claims | Hallucination or drift | Update the pages models learn from |
This is the most useful table in the guide. If you remember one thing, remember that each symptom has a different fix.
Where tools fit
You can learn most of these terms without buying anything. Measuring them is another matter, since manual spot-checks in ChatGPT don't scale and answers vary by country and model.
Promptwatch covers the full set in one product: visibility score, narrative, citations, crawler logs and traffic, plus content tools that act on the gaps it finds. It collects data from the real interfaces of ChatGPT, Gemini, Perplexity and AI Overviews, rather than only APIs. Pricing starts at $95/mo (check promptwatch.com/pricing for current plans), and there is a free AI visibility report if you want a first read.

Other options worth a look, depending on your needs:
- Profound for enterprise monitoring
- Peec AI, Otterly.AI and Scrunch for lighter-weight tracking
- Waikay for a perception-focused view

For a wider shortlist, the GEO software directory lists many more.

Refractia's glossary is a good second reference for terms like narrative alignment and brand drift, though its GEO Score is proprietary and not comparable to other tools.
How to use this glossary with your team
- Pick four metrics: one for presence (share of voice), one for position, one for narrative and one for citations. Fewer is better.
- Agree on definitions in writing. Two people saying "visibility" and meaning different things is the common failure.
- Fix the prompt set and the markets before measuring. Compare like with like, and note the date range, because AI data shifts quickly.
- Review monthly, and tie each metric to an owner and an action.
A shared vocabulary doesn't win you citations. It does stop you arguing about what to measure.


