Track where your brand is mentioned, cited and recommended inside AI answers — on a fixed prompt set, per engine, over time. The measurement layer for an answer engine optimisation programme.
AI search monitoring is the practice of running a fixed set of prompts through answer engines — ChatGPT, Perplexity, Google AI Overviews and AI Mode, Gemini, Microsoft Copilot — on a schedule, and recording whether your brand is mentioned, which sources are cited, how you are described, and how all of that changes over time. It replaces the rank position with three numbers: mention rate, citation share and sentiment.
It is a different job from search monitoring because the output is different. An answer engine does not return ten links; it synthesises one answer, cites a handful of sources, and moves on. There is no page two. The answer also varies with prompt wording, engine, region and the day it was asked, so a single manual check tells you almost nothing. Monitoring means the same prompts, repeated, controlled and logged, so that a change in the numbers can be trusted and traced.
Asva's AI search monitoring feature runs your prompt set on each engine separately, stores every raw response, extracts brand and competitor mentions, classifies the sentiment of each one, and captures the cited URLs. The Brand Visibility Tracker and Citation Intelligence are views built on this data; this page describes the collection layer underneath them.
Seven things recorded on every run, for your brand and every competitor in the set
Whether your brand appears in the answer body at all, and where: named first, listed among alternatives, or mentioned in passing. Aliases and product names are matched, not just the company name.
The URLs each engine attaches to the answer — yours, your competitors', and the third-party pages the model trusted. This is the raw material for Citation Intelligence.
How the model describes you: recommended outright, recommended with a caveat, a neutral list entry, or a warning. The descriptive phrases are captured, not just a score.
Your mentions as a proportion of all brand mentions across the prompt set, against a competitor set you define. Competitor mentions are extracted in the same pass.
The same prompt set run separately on ChatGPT, Perplexity, Google AI Overviews and AI Mode, Gemini and Copilot. Results differ materially by engine, so nothing is averaged across them by default.
A fixed, versioned set of prompts per topic — unbranded category questions, comparisons, problem-led queries and branded checks. Fixed so that the trend line is comparable week to week.
Every run is stored. Mention rate, citation share and sentiment are charted over time with annotations for content and PR changes, so a movement can be tied to a cause.
Five steps from an empty workspace to a trend line you can report on
Enter your brand name, aliases and product names, then the competitors you want measured against and the topics you want to be recommended for. Topics are the unit everything else reports on.
Asva proposes prompts per topic, phrased the way people actually ask answer engines. Edit them, add your own, and keep a mix of unbranded, comparison and branded prompts. Once agreed, the set stays fixed so runs are comparable.
Each prompt is sent to each engine separately and the raw response is stored — text, cited URLs, and the engine and region it came from. Daily is the default cadence; the point is repetition, not a single reading.
Entity extraction matches your brand and competitor aliases in each response, resolves and de-duplicates cited URLs, classifies the sentiment of every mention, and records the position of your brand in the answer.
Dashboards break the numbers out by engine and topic. Alerts fire on a drop in mention rate, new negative framing, or a competitor picking up a citation you used to hold. Findings route into content, technical and PR work.
Monitoring can only report what an engine is able to read. Retrieval-side user agents — OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot for Perplexity, ClaudeBot for Claude — are the ones that decide whether a page can be cited. Training-side agents and tokens — GPTBot, CCBot, Google-Extended — are a separate policy decision that does not change whether you appear in an answer.
A large share of "we are never cited" cases turn out to be a blocked retrieval agent or a managed robots.txt rule injected by a CDN. Check that before rewriting content; the llms.txt generator and the answer engine optimization guide cover the mechanical fixes.
The same prompt set, five different retrieval systems. This is why results are reported per engine.
Answers can come from the model's training data or from live web retrieval, and the two behave differently. Retrieval is performed by OAI-SearchBot (which builds the search index) and ChatGPT-User (which fetches pages on behalf of a user mid-conversation). GPTBot is the training crawler — blocking it does not stop ChatGPT citing you, but blocking OAI-SearchBot does. Responses vary heavily with prompt wording, which is why a fixed prompt set with several phrasings per topic matters most here.
Retrieval-first: almost every answer carries numbered citations, so citation share is the most useful metric on this engine. Its crawler is PerplexityBot. It leans towards recent, well-structured pages and community sources, and the cited domain list for a category is usually different from what a vendor would guess. Track which domains are cited, not only whether you are mentioned.
Both sit on Google's ordinary search index, so Googlebot access is what governs eligibility. Google-Extended is a separate control for Gemini training and grounding and does not change whether you appear in an AI Overview. Overviews only trigger for a subset of queries, so whether one appears at all is itself a data point; AI Mode is the conversational surface where follow-up questions happen. Asva records both trigger rate and inclusion.
Combines model knowledge with Google Search grounding. Google-Extended governs whether your content is used for Gemini training and grounding. Descriptions tend to be more hedged than on other engines, so sentiment framing — recommended versus recommended-with-caveats — carries more of the signal.
Grounded in the Bing index, with citations shown as footnotes. Bing Webmaster Tools and IndexNow are the practical levers for getting pages indexed. Because Copilot lives inside Windows, Edge and Microsoft 365, work-context and B2B queries are over-represented relative to consumer engines — worth knowing if you sell software.
The same data, read three different ways
Reputation now includes what a model says when asked. Monitoring gives you the exact descriptive language per engine, alerts on new negative framing, and the sources feeding it — so a correction can target the page a model is quoting rather than a press release nobody cites.
This is the tracking layer for an answer engine optimisation programme. Mention rate and citation share per engine, on a fixed prompt set, reported next to rankings and clicks — the number that shows whether the crawler-access, rendering and structured-data work changed anything.
Referral data tells you what happened after a click; monitoring tells you what the answer said before it, and most AI answers produce no click at all. Pair it with the AI Traffic Decoder to connect upstream presence to downstream sessions.
There is no position one in an AI answer, so the reporting set changes. These are the seven numbers the feature produces, in the order most teams baseline them. Each is available per engine, per topic and over time.
It borrows vocabulary from all three and replaces none of them
GA4 sees a visit after a click. AI search monitoring sees the answer before there is a click — and most AI answers produce none, because the user got what they needed in the response. Analytics also under-attributes the AI traffic that does arrive, since several assistants strip or generalise the referrer, which is what the AI Traffic Decoder corrects for. The two are complementary: monitoring is the upstream measure of presence, analytics is the downstream measure of sessions and conversions.
A rank tracker records a position for a keyword on a results page that is roughly the same for everyone. An AI answer has no position, is generated fresh for each request, and can differ between two runs of the same prompt a minute apart. So the unit of measurement changes: instead of one rank for one keyword, monitoring records mention rate, citation share and sentiment for a prompt set, over repeated runs, per engine. A rank tracker pointed at an answer engine would report noise.
Social listening tracks what people say about you on public channels. AI search monitoring tracks what a model says about you when asked — a synthesised, machine-authored statement that draws on reviews, community threads, press and your own site. The sentiment in an AI answer is downstream of those sources, which is why the citation data matters: it shows which of them the model is actually reading.
A single check is a snapshot on one engine with one phrasing, in a session that may already be personalised to whoever ran it. It answers "are we mentioned today?" and nothing else. Monitoring is the series: the same prompts, every engine, on a schedule, with the responses logged — the only way to say whether anything you did changed the outcome.
Terms used on this page — mention rate, citation share, share of voice — are defined in the AEO glossary.
A drop in mention rate on any engine, a new negative phrase, or a competitor taking a citation you held — delivered when it happens, with the response attached.
Every response is stored verbatim, so a number on a chart can always be traced back to the exact answer that produced it.
Versioned prompt sets managed in Prompt Intelligence, so a change in the trend reflects the engines, not a change in what you asked.
AI search monitoring is the practice of running a fixed set of prompts through answer engines such as ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot on a schedule, and recording whether your brand is mentioned, which sources are cited, how you are described and how that changes over time. It replaces the rank position with mention rate, citation share and sentiment.
ChatGPT, Perplexity, Google AI Overviews and AI Mode, Gemini and Microsoft Copilot are the core set, each run separately on the same prompt set. Additional assistants are added as they earn a meaningful share of answers. Results are reported per engine because they differ materially and averaging them hides the differences that matter.
On a schedule, and daily is the default. Every run is stored, so the trend line is built from repeated readings rather than a single snapshot. Higher-priority prompt sets can be run more often; the important property is consistency, not raw frequency.
Answer engines are not deterministic. The same prompt can return a different set of brands and citations minutes apart, and retrieval-backed answers change as the index does. This is exactly why monitoring reports rates over repeated runs rather than a single result, and why a manual check in a browser tells you very little.
Enough to cover each topic you want to be recommended for with several phrasings — typically tens of prompts per topic, mixing unbranded category questions, comparisons and problem-led queries. More is not automatically better: near-duplicate prompts inflate the count without adding signal, and a bloated set is harder to keep fixed.
Yes. Competitor mentions, citations and sentiment are extracted in the same pass as yours, which is what makes share of voice possible. You define the competitor set per topic, and the reports show who is named where you are absent.
Partly. The cited-source data shows which pages the engine trusts for your category, and zero citations combined with occasional mentions usually points to a mechanical problem — blocked retrieval crawlers, JavaScript-only rendering, missing structured data — rather than an editorial one. The fixes are covered in the answer engine optimization guide; monitoring is how you confirm they worked.
A manual check is one engine, one phrasing, one moment, in a logged-in session that personalises results to whoever ran it. Monitoring runs the full prompt set on every engine on a schedule, from a clean session, logs every response, extracts sentiment and citations automatically, and alerts you when something moves.
No. GPTBot is the training crawler, alongside CCBot and the Google-Extended token for Gemini. Live retrieval uses different agents — OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot for Perplexity, ClaudeBot for Claude — and mentions drawn from training data persist regardless. Allow the retrieval agents if you want to be cited; training access is a separate policy decision.
Monitoring is the data-collection layer. The Brand Visibility Tracker is the mention-rate view built on it, Citation Intelligence is the source view, Prompt Intelligence manages the prompt sets, and the AI Traffic Decoder connects what the answers said to the sessions that followed.
Baseline mention rate, citation share and sentiment on every engine this week. Set up in minutes.
Get Started Free