How Saidly works

Which AI models does Saidly ask?

Saidly queries the four flagship consumer assistants people use to look things up: Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), and Grok (xAI). By default every query runs with live web search on, so the answers mirror what a customer, reporter, or investor would see today. Each report names exactly which models answered, so the score is never a black box. You always know what it is based on.

What is the difference between grounded and model-only?

A model can answer two ways, and the difference changes what you measure. Saidly supports both.

Grounded (web search)

The model searches the live web before answering and we keep the pages it cited. This reflects what is published about you right now. Grounded is on every plan and is the default.

Model-only

We turn web search off, so the model answers from its training alone, with no live pages and no citations. This is the AI's built-in impression of you, which can lag reality by months because training data has a cutoff date. Model-only is on Pro and Business.

Compare

We run the same question twice, once grounded and once model-only, and show both scores plus the gap. The gap is the useful signal: a warmer grounded answer means the web has caught up but the model's memory has not. Compare is on Pro and Business.

The model-only and compare lenses are the advanced sentiment tools. For a plain-English walkthrough of when to use each, see the FAQ.

How does Saidly compute the 0-100 sentiment score?

For each model, Saidly reads the full answer about your name and rates the overall sentiment, from clearly negative to clearly positive, on a 0 to 100 scale. The rating weighs the tone of the answer, the framing of recommendations or warnings, and whether the model voices doubts or caveats. A score near 50 is neutral or mixed. A score near 80 is consistently favorable. A low score flags answers that repeat complaints or steer people away.

Each of the four models gets its own score, because they often disagree. One assistant can call you a strong choice while another repeats a stale complaint. The report shows the per-model scores side by side and an overall figure that rolls them up, plus the change since your last report so you can see what moved.

The same scoring runs whether a report is scheduled or on demand, so two reports a month apart are directly comparable. You can see a worked example in the sample report.

How does Saidly measure share of voice?

Share of voice measures how often the models mention you versus the competitors you list. When you add a name, you can name the competitors that matter. Saidly then asks category questions ("best alternatives to X", "top tools for Y") and counts which names each model brings up, and in what order. Your share of voice is your mentions as a portion of the total across all tracked names.

This answers a different question than sentiment. Sentiment is "when AI talks about me, is it positive?" Share of voice is "does AI bring me up at all when buyers ask about my category?" A brand can score well on sentiment yet rarely surface in category answers, which is its own problem to fix. For a deeper treatment, see the guide to competitor share of voice.

Where do the cited sources come from?

When a report runs grounded, each model returns the web pages it read before answering. Saidly captures those citations and lists them in the report, grouped by model. This is the most actionable part of a report: if an assistant cites a 2023 forum thread for a complaint you fixed, you know the exact page to address. Model-only reports have no citations, because the model never looked at the web.

Cited sources are also how you connect a low score to a cause. A score is a symptom. The sources behind it are the lead. Tracking them over time shows whether a fix you published has been picked up by the models yet.

What are themes and representative quotes?

Beyond the number, each report pulls a representative quote from every model, a short passage that captures how that assistant actually talks about you, and a set of recurring themes, the points that show up again and again across models and across reports. Themes cover both the flattering claims and the complaints. The quote tells you the tone in the model's own words. The themes tell you what to reinforce or correct at the source.

How often do reports run?

Reports run on a schedule you choose, monthly, weekly, or daily depending on your plan, and you can run an on-demand report at any time, for example right after a launch or a news story. You can set custom schedules and pause or reactivate tracking whenever you want. Scheduled reports arrive in your inbox and stay in your dashboard, so you keep a full history to chart.

What are the honest limitations?

We would rather be plain about this than oversell it. AI answers are generated fresh and vary slightly between runs, so a score can wobble a few points without anything real changing. The trend across reports matters more than any single snapshot. Models also get retrained and swap underlying versions, which can shift scores on their own, which is exactly why tracking over time is the point. Saidly measures what the models say, not whether it is true. Use the cited sources to judge that yourself.

For the bigger picture of why this measurement exists and how to act on it, read the AI visibility monitoring guide. For a company built on this method, see about Saidly, operated by Woodfire Digital, LLC. New to the terms? The glossary defines sentiment, share of voice, grounding, and the rest.