Skip to main content
What others say about you is harder to influence and has less impact than what you say about yourself. Recent research shows that 94.7% of AI citations come from corporate content (the pages brands publish themselves), with UGC (Reddit, forums, Medium) accounting for less than 6% across most verticals (Kevin Indig, The Science of What AI Actually Rewards, March 2026, analysing 98,217 ChatGPT citations across seven verticals). The leverage sits on your own pages, because they are the part you directly change and where change lands immediately. That is where zued focuses its analysis and recommendations.
In most verticals, no. Kevin Indig’s March 2026 analysis of 98,217 ChatGPT citations found corporate content accounts for 94.7% of all citations, with Finance at just 0.5% UGC and Healthcare at 1.8%. The “Reddit effect” that reshaped organic search in 2024 and 2025 has not translated proportionally to AI citations. Three verticals did show measurable UGC value: Crypto (9.2%), Product Analytics (6.9%) and HR Tech (5.8%), where community content answers a meaningful share of queries. Outside those, invest in your own authoritative content first. zued focuses exclusively on what you can change yourself, which is also where the data says the leverage sits.
No. The most consistent finding from recent citation research is that AI citation is a content architecture problem, not a content quality problem. Thin single-intent pages are structurally locked out of AI citations regardless of how well they are written. Word count, list density and named entity counts are flat or even negative predictors at the aggregate across verticals (Kevin Indig, The Science of What AI Actually Rewards, March 2026). What zued analyses (content architecture, chunk-level alignment, technical signals, schema coverage per page type, internal linking between related URLs) is exactly where the actual leverage is.
Audit prompts are designed to be repeatable and neutral, not to replicate any individual user’s phrasing. Real users have chat history, memory, and personalisation that produce unique, non-reproducible responses. Even if you had the exact text of a real user’s prompt, running it in a neutral session wouldn’t produce the same answer. zued’s prompts run in consistent sessions so the same prompt produces a comparable result every week. That consistency is what makes trend tracking reliable.
ICPs are extracted from your page content, not from your CRM or external data. zued clusters persona signals from all your URLs into 5 to 6 profiles that represent how your existing pages position themselves. If your content addresses a broader range of audiences, you can edit the generated ICPs manually to better reflect them.
AI responses are non-deterministic: the same prompt can produce different answers across runs. This is a fundamental property of how AI engines work, not a measurement issue. Small week-to-week swings are usually sampling noise, not real change. Treat any single snapshot as a point estimate with a margin of error around it, and look for movement that persists across multiple snapshots before reading it as a trend. The larger the jump and the more consistent it is week over week, the more likely it reflects a real shift in how AI engines describe your topic.
No. Each engine uses its own crawler with its own user agent. Blocking GPTBot in robots.txt has no effect on PerplexityBot or Bingbot. But blocking any single one means zero visibility for that engine’s users. zued’s bot accessibility check shows exactly which crawlers are blocked, per URL.
Nothing is deleted. Your URLs, analysis results, and trend history stay in your account and remain visible. The project keeps its plan until the end of the period you paid for. Then weekly snapshots stop and the project reverts to the free demo tier, so resubscribing later picks up right where you left off.
Each AI engine retrieves and interprets content differently. Your page might align well with Gemini’s consensus but drift significantly from what ChatGPT tells users about the same topic. zued scores each engine separately so you can see exactly where and why the gaps differ.
Prompt tracking tools check if your brand appears in AI responses: they answer are you mentioned? zued goes deeper: it compares what AI engines say against what your pages actually contain, identifies specific content gaps, and tells you what to change so your content aligns with the AI consensus and adds information gain. It’s the difference between monitoring whether you show up and knowing what to fix so you show up accurately.
The AI consensus on a topic doesn’t shift day-to-day. What changes meaningfully between weeks is whether your content supports that consensus, and whether your fixes moved the needle. Weekly cadence gives you time to implement recommendations and measure their impact without noise from daily fluctuations.
No. The architectural signals that drive AI visibility compound across engines: better structure, stronger claims, and clearer entity definitions help every engine. The differences between engines are matters of degree, not kind. Which specific signals weigh most does vary by vertical (see Kevin Indig, The Science of What AI Actually Rewards, March 2026 for a cross-vertical analysis showing signal relevance is more vertical-specific than the AEO industry assumes), so the right order is to cover the architectural fundamentals first, then tune for your vertical. Optimizing for a single engine in isolation is the wrong move in every case we have seen. See AI Engines for more.
Matching is exact on purpose. A name like “Sparkasse Bremen” or “KfW 300” might be your brand spelled differently, a product of yours, or a different organisation that shares the name, and only you can say which. Counting it as yours automatically would move your numbers on a guess. The Brands screen lists every such name under your brand and under each tracked competitor, with its own Share of Voice, appearance and share when present and a short note on what it looks like. If one is yours, count it as yours there and it joins your brand from then on; everything else stays separate.The same holds for a shorter name your own website uses for you, such as “Weinmann” for “Weinmann Emergency”. The same word can also be another company, a network or a common surname, so zued asks on the Brands screen whether it counts as yours and counts it only after you say yes. Answers you already have are then counted again with it.Reading which names belong together needs no decision from you, so zued does that part on its own. Every name the engines used is grouped into a family: a brand, its spellings and compounds, and its products or services. Names the engines use for a kind of thing or a place are kept out of the brands list and can be seen under their own tab. By default the whole platform compares brands, so a family is one row with one figure and a product is never counted twice. Switch to specific names when you want every name on its own, for example your product against the products competitors offer for the same prompts. Any name can be opened by itself, tracked or not, to see its engines, its prompts and how often it is named without its parent. If a grouping is wrong for your project, split the name off there and it stays split for your project only. A name read as the wrong kind of thing, a city listed as a brand, can be set right for your project alone, so it leaves or enters the count without any answer being fetched again. A name that landed in the wrong place, or on its own, can be moved under the brand it belongs to in one motion, with the effect on that brand’s figure shown before you confirm, and one Undo takes it back. Every decision keeps who made it and when, shown beside the name, so a team can see why a number reads as it does.Two kinds of names get a decision of their own. Members of a network, such as the regional institutions of one banking group or the local organisations of a franchise, count as your brand only when you say so, in one reversible decision for the whole project, made beside the level picker; a single brand can differ from it in its own panel. Until then the network is one collapsed row beneath its brand with its own figure, and each institution inside it stays a row of its own at the product level either way. A name that appears only inside a cited link is not a mention and is left out. A network here means institutions that divide one brand between them, where a customer deals with the one their region or their branch decides; a company that owns another brand a customer picks between is two brands, and each keeps its own numbers. And when the engines write one institution in several ways, those spellings are read as one name at every level, so a regional bank does not appear twice because it was named in two word orders. A product with a name of its own, one that shares no word with its brand, is recognised from the answers that name it beside its brand, so it is grouped with that brand rather than listed as a brand of its own.
The Brands screen lists every name the engines used in any snapshot of your project. Its figures come from the snapshot you pick, or the newest one when you pick none. That is the snapshot the Overview and the competitive landscape read, so with no filter set a brand reads the same on each screen. A name the engines did not use in that snapshot reads 0% and is marked “Not named in this snapshot.” A name they used only in answers to prompts that name your brand also reads 0%, because these figures leave those prompts out, as every visibility figure does by default, and the screen says so. Either way the name stays listed because answers change from week to week and the name may return. Any decision you made about it, such as tracking it or counting it as yours, still applies when it does.
By default, yes. A portal that recommends accounts, a magazine that publishes a test, or a public body the engines name takes the same airtime a competitor would, so it is measured like one and listed on its own tab. If that is wrong for your project, each group can be left out in one reversible decision. The numbers are then recomputed from what was already measured, so nothing is dispatched again and the change is visible in minutes.