Understanding results

Documentation Index

Fetch the complete documentation index at: https://docs.onsomble.ai/llms.txt

Use this file to discover all available pages before exploring further.

Every completed scan produces four views of the same underlying data. Each view is designed to answer a different question, and they’re most useful when you read them together. This page explains what each view shows, what to look for, and where attention usually pays off.

​

The four views

​

Citation benchmarks

What it shows: how often your business appears across the prompts in the scan, compared to each competitor. The view is usually a table or bar chart with one row per business (yours plus competitors) and counts per model. A healthy result looks like consistent citation across prompts and across models. A weak result looks like a business that only surfaces for a few specific prompts, or that’s mentioned by some models and ignored by others. What to look for:

• Flat-out absence. If you’re not cited for prompts where competitors are, that’s the single highest-priority gap. The business isn’t even in the conversation.
• Model-specific blind spots. Businesses often show up in two models but are invisible in the other two. This usually points to representation in a specific data source.
• Trend over time. If you’ve been running weekly scans for a month, compare this scan’s citation count to last week’s. Rising is good; steady is acceptable; falling needs investigation.

​

Share of voice

What it shows: of the total words or sentences each AI answer produced, how much was about your business vs. each competitor. Share of voice is distinct from citation count. A business can be cited in 80% of prompts but with only a one-line mention each time — meanwhile a competitor might be cited in just 50% of prompts but get a whole paragraph of description each time. The competitor has lower citation count but higher share of voice. What to look for:

• Low share of voice despite high citation count — the business is being mentioned as an afterthought rather than as a primary recommendation. That usually means content and positioning need work, not visibility.
• High share of voice for competitors on prompts you care about — a flag that those competitors are currently winning the moment of decision in that specific category.

​

Sentiment

What it shows: whether descriptions of the business across the scan are positive, neutral, or negative. Most sentiment will sit in neutral territory — AI assistants tend to describe businesses factually. What you’re looking for is the distribution and the outliers. What to look for:

• Any negative sentiment. Track it down to the specific response. Negative mentions often come from outdated or inaccurate source material and are worth addressing directly.
• Significantly less positive sentiment than competitors. If competitors are consistently described more favourably, look at what they’re doing differently — case studies, reviews, distinctive positioning — that the business isn’t.
• Drift over time. Sentiment that’s been gradually slipping is an early warning, even if the absolute level still looks fine.

​

Individual responses

What it shows: the actual text each AI model returned for each prompt in the scan. This is the raw material behind the other three views. The numbers and charts summarise what’s happening; this view lets you read exactly what was said. Use it when:

• A number on another view surprised you and you want to see what’s behind it
• You want to understand why a particular prompt is producing weak results
• You’re preparing a client-facing report and want specific quotes to show them

See Reviewing AI responses for more on navigating this view.

​

Reading the views together

The four views are most useful when you triangulate across them. A typical first-scan read might go:

1. Start with citation benchmarks. Identify prompts or models where the business is absent. That’s the headline.
2. Drop into individual responses for one or two of those missing-citation prompts. See what the models are saying instead — which competitors are being recommended, and in what framing.
3. Check share of voice for prompts where the business does show up. Even when cited, is the business getting meaningful airtime or a throwaway mention?
4. Skim sentiment last. Flag any negatives for deeper review. If the distribution is roughly in line with competitors and mostly neutral, move on.

From there, the next step is turning what you’ve seen into action — see Acting on recommendations.

​

Tracking over time

Every scan is preserved. Once you have two or more completed scans, comparison views show how each metric has moved since the previous scan and over a longer trend window. When you make a change — publishing new content, updating messaging, adjusting structure — the next one or two scans are where you’ll see whether it moved the metrics. Model updates from providers can also shift numbers independent of anything you did, which is why week-on-week trends are more informative than any single reading.