Open an AEO report and you may find a visibility score up 18 points, 212 mentions tracked, and a citation gap closed on 14 prompts. If you learned search inside a Google Ads account, the feeling is familiar: you have seen much of this before, under different names and sometimes with a higher invoice attached.
Most AEO jargon is paid-search vocabulary with the serial numbers filed off. A prompt resembles a keyword. AI share of voice borrows from impression share. A citation gap looks a lot like a search terms report showing where you failed to appear. Those comparisons are useful until someone treats the metrics as interchangeable. This glossary follows the terms you tend to meet in a pitch, then ends with the one I would retire.
Prompt and prompt set
Plain definition: A prompt is a question someone enters into ChatGPT, Gemini, or an AI search experience; a prompt set is the fixed list of questions you track over time.
The PPC ancestor is obvious: prompt is keyword, prompt set is keyword list. The misuse starts when a vendor reports wins on questions no buyer would ask. A checker that runs buyer prompts across answer engines helps only if those prompts sound like buyer language. My practical baseline is 25 to 40 real buyer questions run on ChatGPT, Gemini, Perplexity, and AI Overviews, repeated weekly for a month. Choose the questions before you admire the chart.
Prompt gap and citation gap
Plain definition: A prompt gap is a buyer question where rivals get mentioned and you do not; a citation gap is a question where no URL from your domain appears as a source.
Think search terms report: here is the query, here is who showed up, and you were absent. The gap taxonomy vendors use adds mention, source, and narrative gaps to the list. Those labels matter only if they lead to different work. I used to tell clients a gap meant writing a page. I was wrong often enough to change the rule. A mention gap can mean other sites explain you badly. A citation gap can mean your page gives the model nothing quotable. Name the gap before you commission the fix.
Content gap vs entity gap
Plain definition: A content gap means you have no page that answers the question; an entity gap means the model does not clearly connect your business to the category.
I learned the difference the way most PPC people learn match types: by spending effort on the wrong one. A content gap calls for a direct answer on your site, in plain language, with a fact worth quoting. An entity gap sends you beyond your site. If only 1% of sources cited by LLMs come from brand-owned sites, your page claiming you are the best plumber in Austin has plenty of competition from reviews, directories, and Reddit threads.
One missing appearance tells you little. Pattern-level tools, such as the Semrush AI Visibility Toolkit drawing on 289M+ prompts, aim to separate a persistent absence from run-to-run noise. Disappear across 50 related prompts in 40 markets and I would investigate an entity problem. Disappear on two and it might just be Tuesday. Check the pattern before you rewrite the site.
Mention vs citation vs recommendation
Plain definition: A mention names your brand, a citation links to one of your URLs as a source, and a recommendation tells the reader to pick you.
You can earn one without the others. A mention without a citation can point to third-party pages that explain your brand more clearly than your own pages do. I have seen the pattern: plenty of names in answers, almost no links to the business, and cleaner facts on review pages.
Vendors express this as mention rate, the prompts where you are named divided by tracked prompts, and citation rate, runs with a link to your domain divided by total runs. Both need a fixed prompt set, geography, and platform. Without those controls, you are reporting weather. Ask for mentions and citations side by side.
Source, footnote and passage
Plain definition: The source is the page cited, the footnote is how the link appears, and the passage is the text that supports the answer.
In PPC terms, think landing page, displayed placement, and the words the reader actually sees. The analogy is not exact, which is why the distinction matters. Bing’s AI Performance report warns that citation activity does not represent ranking, authority, clicks, or traffic. A footnote shows your page was used as a source in that answer. It does not show that anyone clicked. Treat every footnote as a number-one organic ranking and you can pay for wins you never received. When a report shows a citation, ask which passage supported it.
AI Overview citation vs AI Overview ad
Plain definition: An AI Overview citation is a source link in the answer; an AI Overview ad is a paid placement labeled Sponsored that can appear above, below, or within it.
This is the confusion that costs money. Ads can appear above, below, or within AI Overviews. Existing Text, Shopping, Local, and App ads can show above or below where Overviews are available; ads inside the answer have additional relevance conditions. Google Ads Liaison has said ads appear either inside or around an Overview, not both at once. A report that blends a sponsored placement with a source citation mixes rent with ownership. Separate paid appearance from earned sourcing before you discuss performance.
The placement also has limits that pitches tend to skip:
- Ads inside Overviews are available in English on mobile and desktop in 12 countries, including the US, Canada, Australia, and India. Sensitive verticals such as finance, healthcare, gambling, and politics are excluded.
- You cannot target only AI Overviews or opt out of them. Google reports those ad impressions as Top Ads without a separate inside-the-answer view.
- The campaign levers remain familiar: broad match or keywordless targeting through AI Max, Performance Max, Shopping, or Dynamic Search Ads, alongside Smart Bidding.
If an agency adds “AI Overview ad management” as a separate line item, ask which lever it pulls that your Search campaign does not already pull.
AI share of voice
Plain definition: AI share of voice measures your brand’s mentions or citations against the total brand mentions in a tracked set of prompts.
Brand24 describes that calculation. It resembles Google Ads impression share in purpose, but the denominators are different: eligible ad impressions are not the same thing as observed brand mentions in a chosen prompt set. To compare visibility across ChatGPT, Gemini, and Google AI Overviews, run the same buyer questions repeatedly on each and show the results separately.
Then ask what moved the share. With brand-owned sites accounting for a small fraction of cited LLM sources, other people’s pages can affect your number. Share of voice may help track reputation across answers. It is a poor basis for an invoice on its own. Keep the prompt set and the denominator in view.
Answer variability
Plain definition: Answer variability is the change you see when an AI system responds differently to the same prompt across runs.
Same question, different answer, different Tuesday. That drift is not automatically a tracking bug or evidence that your latest page worked. LLMs can return different answers across runs, so a move from 17% to 19% can be ordinary variation. I keep a rule from testing ad copy: never trust one run.
Vendors have plenty of reason to display the number. One analysis of 5,699 share-of-voice mentions found AI SOV discussion running 3.4 times larger than traditional SOV discussion. Popularity does not make a single reading reliable. Repeat buyer questions before you call a movement lift.
Verified bot fetch, unverified hit and live-answer fetch
Plain definition: A verified bot fetch is a request you can identify in server logs, an unverified hit is a dashboard claim you cannot substantiate, and a live-answer fetch is a retrieval request associated with answering a user.
The mechanism lives in server logs, not in a colorful chart. Logs can show distinct jobs: live-retrieval agents such as ChatGPT-User and Claude-User, search-index crawlers such as OAI-SearchBot and Googlebot, and training crawlers such as GPTBot. I have seen reports sweep all three into “AI visibility.” That is like counting impressions, bot clicks, and view-throughs as conversions. Ask which agent requested which page, and when.
Training crawl vs retrieval fetch
Plain definition: A training crawl gathers material for model training, while a retrieval fetch requests material for an answer being assembled now.
Only the second points to an immediate opportunity to appear in that answer. Even then, log interpretation has limits. Gemini has no retrieval-specific user agent, and Googlebot hits cannot be cleanly split between Gemini and classic Search. Say you spend $20k a month and a dashboard says Gemini traffic is up 40%. Ask how it knows. If the answer is a pile of Googlebot requests, do not let index activity masquerade as demand.
Answer-first content, entity markup and llms.txt
Plain definition: Answer-first content gives the direct answer near the top of a page, entity markup labels facts with schema, and llms.txt is a Markdown map of a site placed at /llms.txt.
The PPC translation: answer-first is ad copy that meets the query immediately, markup resembles a clean product feed, and llms.txt is a sitemap built for a crawler that never asked for it. The file was proposed in September 2024 as a Markdown map, not a blocking file or an established standard. No major provider has committed to reading it for answers. One vendor found 0.1% of 60,000 bot visits touched it; another saw a similar skip pattern across 500 million events, with no link to more citations.
I have nothing against publishing one. It costs an afternoon. I object when the file becomes the strategy.
Off-site authority and citation building
Plain definition: Off-site authority is what third parties say about you; citation building is work to make your business and its facts available on pages an answer might cite.
Think backlinks, but pay attention to the sentence around the link. Directories, review pages, trade press, comparison tables: the pages your buyer may trust over your homepage can also shape how an answer describes you. The practical sequence I use comes from groas work on getting pages cited in AI Overviews: check crawling, rewrite the direct answer, clean up the HTML, then make a dated check to see whether the citation sticks. If third-party pages carry clearer facts than yours, fix your facts first. Five more posts will not correct a review page misstating your pricing.
AEO reporting mistakes
Plain definition: AEO reporting mistakes happen when a report presents one kind of appearance or activity as evidence of another.
Read enough reports and the same five substitutions recur:
- Mention as citation: The answer names you but links elsewhere.
- Citation as ranking: A footnote becomes a claimed number-one position, though citation activity is not ranking, authority, clicks, or traffic.
- Ad as citation: A sponsored appearance gets counted as earned visibility.
- Index crawl as demand: GPTBot or Googlebot requests get blended with live retrieval fetches.
- Variation as lift: A small score movement gets its own slide when normal answer variability could explain it.
Each turns a distinction PPC operators already know—impression, click, conversion—into mush. Ask what was actually observed before you pay for the next page of the report.
Visibility score
Plain definition: A visibility score is a vendor’s blended measure of appearances and prominence, calculated according to that vendor’s rules.
If I could retire one AEO term, this would be it. Not because measurement is useless, but because the label lets vendors skip the measurement. Prompt sets can be checked. Share of voice can be recalculated. Fetches can be inspected in logs. A paid ad should not be passed off as a citation, even when Google’s ad reporting does not break out every AI Overview placement.
A blended score tells you none of that on its own. It can move when the model changes its answer, rise while you are already winning, and still fail to identify the sentence worth fixing. I spent years watching Quality Score get used the same way in Google Ads: a diagnostic dressed up as a goal. Keep the prompt. Keep the gap. Keep the citation. Retire the score, and spend the money on the passage that earns the next one.