What should you measure in AI answers?
Five measurements matter; everything else on a dashboard is decoration.
- Mentions. Does the answer name your brand at all? Mentions divided by total answers tracked is the core of share of AI voice.
- Citations. Does the engine link a source, and is it yours, a third party's or a competitor's? Citations show which documents the engine trusts, which is exactly where you can intervene.
- Sentiment. Recommended, neutral, hedged ("some users report...") or negative. LLM sentiment is cruder than social-listening sentiment but directionally usable.
- Position. Where you appear in a list-style answer. "One of the top options" and "also worth considering" are different commercial outcomes.
- Factual errors. The share of answers that contain wrong facts about you: founding year, discontinued products, invented pricing, a confused merger. For regulated industries this matters most, and generic dashboards surface it worst.
If a tool cannot tell you which of these it measures and how, treat that as information about the tool.
How do you build a free 20-prompt baseline?
Write 20 fixed prompts, run them on four platforms in the same week each month, and log six columns per answer. One person, one spreadsheet, about two hours.
The 20 prompts:
- 5 category prompts: what a buyer asks before knowing names ("best [category] for [use case]", "top [category] providers in [market]").
- 5 comparison prompts: "[you] vs [competitor]", "alternatives to [market leader]".
- 5 brand prompts: "what is [brand]", "is [brand] legitimate", "who founded [brand]".
- 5 adverse prompts: "[brand] problems", "[brand] complaints". You want to see what the engine reaches for when the question turns hostile.
Platforms: ChatGPT, Perplexity, Gemini and Grok, in a logged-out or clean session where possible. Keep the prompts frozen; the value is in the time series, not in prompt-writing.
Columns: mentioned (y/n), position (1st / 2nd-3rd / later / absent), sentiment, domains cited, factual errors (verbatim), date. Screenshot anything surprising, because answers are not reproducible.
After two or three months you know your baseline share of voice, which engines already cite you, where competitors out-mention you and whether anything said about you is false. That is the information a purchase decision needs.
Which AI monitoring tools exist, and what do they cost?
Five names cover most buying scenarios. We list them alphabetically, without ranking, and we do not state prices: plans and limits change often, so use the figure published on each vendor's site.
| Tool | Type | Typical fit | Pricing |
|---|---|---|---|
| Ahrefs Brand Radar | Module inside an SEO suite | SEO teams already on that suite | Published on the vendor's site |
| Otterly.AI | Standalone monitoring | Small teams starting structured tracking | Published on the vendor's site |
| Peec AI | Standalone monitoring | Mid-market teams tracking several engines and markets | Published on the vendor's site |
| Profound | Enterprise platform | Large brands that need governance, integrations and API access | On request |
| Semrush AI Toolkit | Module inside an SEO suite | Marketing teams in that ecosystem | Published on the vendor's site |
How to choose between them:
- Suite add-ons make sense if you already pay for the parent suite; check the prompt allowance at the entry tier, which is usually thin.
- Standalone tools give finer prompt-level tracking and are the usual first paid step.
- Enterprise platforms mean a procurement cycle, not an experiment. Buy one only when manual logging has stopped being viable.
Before any purchase, ask the vendor four questions: Which engines and countries are covered? How many prompts can I run per day? How is "share of voice" calculated? Can I export raw answers? A vendor that cannot answer the last one gives you a number you cannot audit.
What none of them sells is the fix. A dashboard can show you losing every comparison prompt to a competitor; it cannot write the sources that change the answer.
How should you read AI monitoring data?
Treat a snapshot as noise and a three-month movement on two engines as a trend. The most common failure in AI monitoring is not picking the wrong tool but over-reading week-one data.
Citation patterns are real but unstable. 5W's Citation Source Index (May 2026) put Wikipedia at 13.15% and Reddit at 11.97% of US ChatGPT citations, the two largest sources, while stressing that behavior is volatile and engine-specific. Each engine leans on a different source mix, and a model update can reshuffle all of them overnight.
Practical rules:
- Never report week-over-week share of voice to anyone who allocates budget.
- A trend is three or more months in one direction on at least two engines. That is the bar for celebrating or escalating.
- Watch the stable layer. Engines change weights but keep returning to canonical sources: encyclopedic entries, structured data, high-trust communities. Source presence decays slowly; answer phrasing flickers daily.
- Act on one sighting of a false claim about your company. Everything else needs a trend line first.
Which fix matches which gap?
Each finding maps to a source-side lever, because engines are downstream of their sources.
| What the data shows | The lever that moves it |
|---|---|
| Absent from category and comparison answers; competitors cite Wikipedia and you have no page | Notability assessment first, then Wikipedia page creation if the sourcing supports it. Not every brand qualifies yet; start with a notability audit or the free notability index |
| Engines state wrong facts: founding year, ownership, products | Correct the records engines treat as ground truth: Wikidata and the knowledge graph, plus the source articles the wrong claim traces back to |
| No community proof; category threads on Reddit and Quora never mention you | Disclosed, compliant participation; see Reddit, Quora and AI visibility |
| You are cited, but from thin pages | Machine-readable depth: structured documentation, llms.txt, FAQ schema, an LLM Hub |
| Mentions exist but decay or get vandalized at the source | Ongoing source monitoring: Wikimonitoring |
Two caveats. Levers are slow: a Wikipedia page or a corrected knowledge-graph record typically shows up in answer behavior over months, and no one can promise what an assistant will say. And sequencing beats volume: fixing one false claim at its source usually outperforms publishing ten new assets nobody cites.
Do AI monitoring tools cover languages other than English?
Mostly as an afterthought. Ask the same engines the same question in German, Polish or Ukrainian and you get different answers built from different sources: local Wikipedia editions, local media, local forums. A brand dominant in English answers can be invisible in Polish ones.
Some tools let you run prompts in other languages, but their benchmarks and citation indices remain US-centric. The fix is procedural: run the full 20-prompt baseline separately in every language where you earn revenue, each with its own gap-to-lever map. See our multilingual Wikipedia strategy for why language editions differ.
When do you not need a tool yet?
A paid dashboard is the wrong purchase in four situations:
- Your category barely exists in AI answers. If engines refuse to name any vendor for your prompts, there is no share of voice to win. Re-run quarterly and spend the budget on citable sources.
- You are pre-traction. Monitoring measures the footprint of evidence. With no customers, coverage or community, a dashboard reports zero.
- Volume is tiny. Twenty buying-relevant queries a month do not justify always-on tracking.
- You have not run the baseline. Two months of DIY data turns a leap of faith into a sized decision.
Frequently asked questions
Continuously. Answers vary between sessions on the same day, and model updates can reshuffle sources overnight. That is why monthly fixed-prompt sampling and three-month trend reading beat daily dashboard-watching for most teams.
There is no deletion request that edits a model's answer. What works is source-level correction: fix the claim where the engine learned it (a news article, a Wikipedia entry, a Wikidata record, your own documentation) and answers follow as systems re-retrieve. Expect weeks to months, and verify with your prompt log.
Not yet. Every vendor computes it from its own prompt panel, so figures are not comparable across tools. Treat it as an internal time series: same prompts, same engines, measured against your own baseline and named competitors.
There is no single best. Start with the free 20-prompt protocol for two months. If you outgrow it, compare the standalone tools on the engines and languages you actually care about, and check each vendor's published pricing. The commitment to fixed prompts and a monthly cadence matters more than the brand name.
Mentions, citations, sentiment, position and factual errors. For regulated industries, factual errors come first, because a wrong claim about your company can be repeated to many buyers before you see it.
Twenty fixed prompts are enough for a baseline: five each for category, comparison, brand and adverse questions. Paid tools scale to hundreds, but volume only helps once you know which questions drive revenue.
It can help: Wikipedia is among the most-cited sources across assistants. But no one can promise an outcome, and a page exists only if independent sources support notability. See why Wikipedia matters.
How WikiBusines helps
We do not sell monitoring dashboards; we work on the sources they measure. Paid editing on Wikipedia is always disclosed in WikiBusines' name, and we never guarantee what an AI will say.
- AI visibility audit, €490. How engines currently see you, which sources they use and which gap to close first.
- AI visibility packages, from €800. The source-side work that follows the audit.
- Wikipedia notability audit, from €490, credited toward the page if you order it within one month. Page creation from €2,400 for English, 4-5 weeks, with 90 days of monitoring and guarantee after publication.
- Wikidata, from €680. Correct entity records for the facts assistants get wrong.
You bring the dashboard; we move what it measures. Start with a conversation.
Authors and editorial policy
Bohdan DubylovskyiCo-founder and CBO, WikiBusines→
Roman MelnykCo-founder & Chief Editor, WikiBusines→The article was prepared by the WikiBusines co-founders. We sell paid Wikipedia services, so we disclose this interest: paid editing is always declared on Wikipedia in WikiBusines' name, and claims about the rules are backed by links to primary sources.
Editorial policy: payment disclosure, conflicts of interest, sources