How Tonecast works

Tonecast is a SaaS platform that shows what AI assistants (ChatGPT, Claude, Gemini and Perplexity) say about a brand, product, person or topic, and which web pages shaped those answers. This page follows one campaign from setup to the first early signal. Setting up a campaign takes about five minutes; everything after that runs on its own.

1. You describe the subject

A campaign is about one subject: a brand, a product, a public person or a topic. You give its name; Tonecast proposes aliases, spelling variants and the exclusions needed to avoid confusing it with namesakes, checked against Wikidata. You confirm or edit them. For brands and products you can add competitors, so that share of voice can be measured.

Two more declarations are optional but make the analysis sharper: a fact sheet (verified facts such as prices, features and dates, each with a validity date) for claim check, and your owned properties (your website, product pages, FAQ, official profiles), so your own pages can be told apart from everyone else's.

That is all you declare. There are no keywords, no boolean queries and no list of websites to monitor.

2. Tonecast proposes the prompts, in each language

A tracked prompt is a question real people would ask an AI assistant about the subject. Tonecast proposes prompts by category: discovery ("best compostable coffee capsules"), comparison ("Aurora Coffee vs its competitors"), evaluation ("is Aurora Coffee reliable?"), price, support and current events. (Aurora Coffee is a fictitious brand from the demo.) You keep, edit or add prompts, and choose the languages and countries. Each prompt is tracked separately in each language, because the answers differ by language.

3. Every week, the engines answer through official APIs

Every week Tonecast sends each prompt to each engine of the campaign: ChatGPT (OpenAI API), Claude (Anthropic API), Gemini (Google API) and Perplexity (Sonar API), with web search or grounding enabled. Google AI Overviews is on the way. Each prompt is run several times a month on each engine, because answers vary between runs; your plan sets how many runs.

Every answer is stored in full, with its citations and the measured cost of the call. Transient engine errors are retried; an engine that runs out of quota is paused for the rest of the cycle rather than retried endlessly. Spending caps are checked before every call.

4. Each answer is analysed

Each answer goes through the same analysis:

• Mention detection: does the answer talk about the subject, including partial names and descriptions such as "the Italian capsule brand"? Competitors are detected the same way.
• Prominence and recommendation: where the subject appears, whether it is in a list, and whether the engine recommends it.
• Tone in AI: the tone of what is said about the subject, not the tone of the whole text.
• Claims: factual statements about the subject, extracted in a structured form and compared with the fact sheet (consistent, incorrect, outdated, unverifiable).
• Citations: every cited URL is normalised, de-duplicated and sent to the source graph as a Cited page.

5. The sources are discovered, without keywords

At every cycle, Tonecast rebuilds the campaign's source graph: the pages on which what AI says about the subject was formed. It collects candidates from:

• the pages the answers cite (the Cited tier);
• the web search results for each prompt, language and country, which are the pool that search-grounded engines read from;
• the subject's entity pages, such as Wikipedia in the campaign's languages and Wikidata;
• targeted searches on open platforms: discussion threads, videos, news feeds;
• your declared owned pages.

Pages that engines typically draw on without citing them form the Formative tier. Every page records why it was included. Collection uses official, open APIs and crawling that respects robots.txt; only the text needed for the analysis is kept, with a link to the original.

6. Each source gets an influence weight

The influence weight estimates how much a page is likely to shape a given engine's answers. It combines how often that engine cites the page, how much the engine typically draws on that kind of source, the page's best position in the search results for the prompts, and how recent it is compared with how long citations usually last on that engine. Pages whose full text could not be read weigh less. The components are stored next to each weight, weights are re-estimated every week, and old weights are kept rather than overwritten.

The pages above the weight threshold are analysed like answers: relevance to the subject, tone, claims. Their weighted tone is the Formative Sentiment Index (FSI).

7. The lead estimate tests whether sources move first

Once there are enough weeks of data, Tonecast compares, engine by engine, the source-side series (FSI, claim prevalence, visibility precursors) with the answer-side series (Tone in AI, AI visibility rate, presence of each claim). It looks for a positive lag at which the sources lead the answers, checks it with a permutation test and corrects for the number of pairs tested. The result is one of three states: none, observing or active. Only an active signal becomes an early warning. The details are on the methodology page.

What Tonecast never does

• It never scrapes the AI assistants' consumer apps. Engines are queried only through their official APIs, with the cost of every call measured.
• It never uses closed platforms. No X/Twitter, Instagram, TikTok, Facebook or private LinkedIn data, and no third-party data brokers for them.
• It never asks you to pick sources. No keywords, no boolean queries, no channel picker, no feeds to register.
• It never overwrites a score. Every score records the model and version that produced it; a better model supersedes it and the old one is kept.
• It never generates corrective content. It shows what is wrong and where; you decide what to publish.
• It never presents a forecast as a certainty. Early signals stay "observing" until the evidence is statistically reliable.

Questions about how it works