Skip to content
MentionRadar

How MentionRadar measures mentions and AI visibility

MentionRadar collects public mentions from news, blogs, forums, Reddit, Hacker News, review sites, YouTube and X, removes duplicates by URL, scores sentiment with a transparent word model, runs your tracked prompts in five AI engines every week and turns both into scores you can audit. This page explains each step.

Scan my brand See plans from $24 a month
  • Mentions matched as whole words
  • One mention per normalised URL
  • Sentiment you can read and correct
  • Scores computed from stored answers

Step one, collecting public mentions

Each source has its own connector. A connector searches for your keywords, reads the results and returns a uniform record: title, snippet, link, author when public, publication time and a reach estimate.

Sources MentionRadar monitors and the reach signal used for each
SourceWhat is collectedReach signal
News sites and blogs Headline, publicly visible opening text, outlet, time, link Size of the outlet
Reddit Thread title, text excerpt, subreddit, author name, link Community size, votes and comments
Hacker News Stories and comments with the brand, author name, link Points and comments
Forums and review sites Post or review excerpt, site, time, link Site size
YouTube Video title and description, channel, link Channel and video signals
X Public post text, author handle, time, link Engagement on the post

On paid plans every workspace is collected every 15 minutes. Each request has a short time limit. When a source is slow or refuses a request, that source is skipped for the run and tried again on the next one, so one unavailable source never stops the others. Paywalled text is not copied: you get the headline, the link and whatever the publisher shows publicly.

Reach is an estimate, shown as Low (under 10,000), Medium (10,000 or more) or High (100,000 or more). It exists to rank the feed and to drive alert rules such as "negative and High reach goes to Slack". It is not an audited audience figure, and we do not present it as one.

Step two, matching and removing duplicates

Whole word match
A keyword must appear as a whole word in the title or snippet. "Notion" does not match "notional", and "Atlas" does not match "Atlassian".
Case when it matters
If you type a keyword with capital letters, such as "Notion", the match respects that capitalisation, so the brand is found and the everyday word "notion" is not. A keyword typed in lower case matches any case.
Exclusions and filters
Each keyword carries exclusion words, a language filter and source filters. A mention that contains an exclusion word is not stored.
Time window
New mentions are collected continuously. The first scan of a new keyword backfills the last 30 days so charts are useful on day one.
One mention per page
Links are normalised before storage: protocol, www, mobile and old prefixes, tracking parameters and the trailing slash are removed. The normalised link is unique per workspace, so the same article arriving through two feeds is stored once.

Step three, scoring sentiment

Sentiment is scored with a transparent word model built for brand language. Words carry weights from strongly negative ("scam", "outage", "breach") to strongly positive ("excellent", "reliable", "recommend"). The model reads each sentence in order and applies three rules that matter for real posts.

  • Negation flips meaning. "Not reliable" counts as negative, and the effect reaches the next few words.
  • Intensifiers scale weight. "Extremely slow" weighs more than "slow", "slightly confusing" less than "confusing".
  • "But" moves the weight. In "the dashboard is great but exports are broken" the second half counts more, as it does for a human reader.

The total is normalised to a score between minus one and one. Scores above a small positive threshold are labelled positive, below the matching negative threshold negative, and the rest neutral. Editors can correct any label. The social sentiment tracking page shows how teams use the weekly trend.

Step four, asking the AI engines every week

Tracked prompts run once a week in ChatGPT, Perplexity and Google AI Overviews on Starter, and in all five engines, adding Gemini and Claude, from Growth. Runs are spread across the week, and every answer is stored with its date so it can be reread later.

Mentioned

Your brand appears as a whole word in the answer. An answer that says the engine is not familiar with the brand counts as not mentioned.

Position

When the answer is a numbered list, the position of the first line that names your brand. Prose answers have no position.

Recommended

Named in the top three of a list, or described with recommendation language such as "recommend", "leading" or "widely used" in prose.

Competitors named

The other products in the list, plus any competitor you track that the answer names.

Cited sources

Pages and domains the answer cites or links, with your own domain marked. Stored per answer from Growth.

Sentiment

The same word model as for web mentions, applied only to the sentences that name your brand.

The verdict is computed by our parser from the stored answer text, not by asking a model to grade itself. When an engine cannot be reached or the run would exceed the plan or budget limit, the run is marked unavailable and no verdict is recorded. MentionRadar never fills a gap with a guessed answer.

Step five, computing the scores

AI visibility score, 0 to 100

For every pair of prompt and engine, MentionRadar takes the latest run from the last eight days and gives it points. The score is the average across all pairs, so it moves when the engines change their answers, not when you add or remove a single prompt mid week.

Points per AI answer
Answer outcome Points
Named and recommended 100
Named at position 1, not recommended 100
Named at position 2 85
Named at position 3 70
Named at position 4 55
Named at position 5 or lower, or in prose 40 minimum
Not named 0

Take 10 prompts on 3 engines, which give 30 pairs. If 6 answers recommend you, 6 name you at position 2, 6 at position 4 and 12 do not name you, the score is (6 x 100 + 6 x 85 + 6 x 55) / 30 = 48. The full view per engine is on the AI visibility tool page.

What the numbers can and cannot tell you

Reliable for

  • Trends over weeks and months
  • Comparing your brand with tracked competitors on the same prompts
  • Spotting new negative threads and high reach coverage quickly
  • Seeing which pages AI engines cite for your category

Not meant for

  • Judging one comment by its label alone, sarcasm still needs a human
  • Audited audience numbers, reach is an estimate
  • Private groups, direct messages or content behind a login
  • Predicting a single future AI answer, answers vary run to run

How the collected data is stored, retained and protected is described on the security page.

Questions about the methodology

Why does a mention appear once even when several feeds carry it?

Every link is normalised before it is stored: the protocol, www and mobile prefixes, tracking parameters and the trailing slash are removed. Two links that point to the same page therefore share one key, and only the first copy is kept in the workspace.

Can I correct a sentiment label?

Yes. Sentiment is a starting point for routing alerts and charting trends, not a verdict. Anyone with editor rights can change the label on a mention, and the sentiment index and charts use the corrected label from then on.

Why are AI prompts run weekly and not every day?

Answers vary from run to run, so a daily reading mostly records noise. A weekly run of the same prompt set produces a series you can compare month over month, and it keeps the cost of every plan fixed, which is why each plan states a number of prompts per week.

Does the AI visibility score depend on which prompts I track?

Yes, and that is intended. The score answers one question: across the prompts you chose, how often and how prominently do the engines name you. A prompt set that mirrors how buyers really ask about your category gives a score that tracks real visibility.

How is reach estimated?

Reach is an estimate of how many people could have seen a mention, based on signals each source exposes: the size of the outlet for news, community size, votes and comments for Reddit, and points and comments for Hacker News. It is shown as Low, Medium or High and used to rank and route mentions, not as an audited audience figure.

Does MentionRadar train AI models on my data?

No. Your keywords, prompts, mentions and reports are used only to run your monitoring. Prompts are sent to the AI engines you track to read their answers, the same way a buyer would ask them.

Check the method against your own brand

Run a scan and open any mention or AI answer to see the source, the match and the label behind it, before you create an account.

Scan my brand Compare plans