Help & FAQ
Common questions about results, failed collections, credits and the AI Recognition Check.
Results and answers
Why do some AI answers contain URLs or bracketed links?
Models that search the web usually append their references at the end of an answer. This is the model's own citation text, not a system error. Answers are displayed as Markdown, so links are clickable; summary views such as the overview show a plain-text preview. The sources a model actually retrieved are listed separately under "Citations" at the bottom of each answer.
Why does the same question get different answers on different runs?
AI answers are inherently non-deterministic, and the web pages retrieved during search change over time, so different answers and rankings for the same question are normal. Look at the trend of the visibility score over a period instead of a single run.
Why do results differ with web search on versus off?
With search on, the model looks up pages before answering, so results depend on what is on the web right now and come with cited sources. With search off, it relies on training knowledge only, so results reflect what the model already knows about your brand. The two measure different things: do not compare their scores directly, and keep the mode constant when comparing trends. Whether search is actually used also depends on whether the selected model supports it, and each answer records whether search was used.
Read the blog: Web Search vs No Search: Why the Same Question Gets Different AI AnswersWhy do models return different numbers of citations?
Models search and cite differently: some search actively and return several sources, others answer from built-in knowledge and cite few or none. The number of citations does not indicate answer quality.
The answer clearly mentions my brand, so why is it marked "Not mentioned"?
Mentions are determined from the structured brand list returned by the model, using name matching. If the answer uses an abbreviation, alias or a different-language name, it may not match. Add the brand's common names in brand settings and run the collection again.
What the metrics mean
How is the GEO score calculated?
The GEO score is out of 100 and has three parts: mention rate up to 60 points, average rank up to 20 points (rank 1 earns the full 20, rank 10 earns 0, linear in between), and sentiment up to 20 points. If the brand is never mentioned, the rank and sentiment parts score 0. The score is an observation at the time of collection, so look at the trend over a period.
Read the blog: AI Visibility Score ExplainedWhat does the sentiment score mean?
It measures whether the AI speaks about your brand in a positive or negative tone. Each answer that mentions the brand is rated positive (100), neutral (50) or negative (0), and the sentiment score is the average of those ratings; answers that do not mention the brand are excluded. For example, two mentions, one positive and one neutral, give 75%. Closer to 100% means the AI describes you more favorably, around 50% means mostly neutral descriptions, and clearly below 50% means many negative remarks, which you can read under Negative mentions. With very few mentions the score swings a lot, so read it together with the mention count. If the brand is never mentioned it shows 0%, which means no data, not negative sentiment.
What are mention rate, average rank and share of voice?
Mention rate is the share of questions in which the AI mentions your brand; the overview shows the active mention rate, which counts only questions that do not name the brand directly. Average rank is the brand's average position in the answer's recommendation list, where a smaller number is better. Share of voice is your mention rate as a share of the combined mention rate of you and all competitors, showing who the AI mentions more often for the same set of questions.
Why is there no "vs last period" comparison in my report?
The comparison appears only when data from a previous period exists. The first collection has no baseline, so no change is shown. After each later collection, the score is compared with the previous period. Collection has natural variation, so a small rise or fall does not prove that your optimization worked or failed.
Failed collections and retries
Why does a question fail to collect?
Common causes are a model API timeout or temporary outage, and a model not returning a parseable structured result. The system retries once automatically. If it still fails, the item is marked as failed and nothing incomplete is stored, so visibility scores are not affected.
How many times can I re-run a failed item?
On the History page, each task can be re-run manually up to 3 times. Only your manual clicks are counted; automatic system retries are not. A re-run does not charge credits again. Once the limit is reached, please contact support.
Am I charged credits for failed collections?
A manually triggered collection reserves credits when the task is created, based on the planned number of questions and models. Failed items can be filled in with a re-run without being charged again. If AI question generation fails, the credits spent are refunded automatically.
AI Recognition Check
How is the AI Recognition Check different from regular monitoring?
Regular monitoring simulates customer questions with web search and checks whether your brand is recommended. The Recognition Check works offline and asks each model whether it knows the brand, measuring the native awareness in its training data. They are two independent mechanisms, and Recognition Check results are not included in the visibility score.
How is the Recognition Check priced?
Each participating AI model costs 2 credits. When you click "Re-run check", a confirmation dialog shows the estimated credits, and nothing runs until you confirm. The page shows the time of the most recent check.
The check says a model does not recognize my brand. What can I do?
It usually means your brand appears rarely in the model's training data, which is common for new or niche brands. Publish informative content and earn coverage on authoritative media and industry sites; awareness improves gradually as models are updated.
Reports and sharing
What are "Negative mentions" in a report?
They are the records from the report period where an AI model spoke negatively about your brand, listing the model, the question and the answer. The list shows the first 200 characters by default; click "Expand" to read the full answer.