AI visibility
When a developer asks an AI assistant "what should I use for X?", the answer is a recommendation you never see. BackTalk asks those questions for you, on a schedule, across the models you bring keys for (Claude, ChatGPT, Gemini, Perplexity and more, including open models via Venice), and tracks whether the answer names you or a competitor. That slice of recommendations is your AI share of voice.
It runs on your own AI keys, like everything else in BackTalk: no markup, no metering, and the models answer exactly as they would for a real user of that provider.
How it works
- You define topics. A topic is one buying decision: the questions buyers ask (its prompts), the names your product goes by (your aliases), and the competitors to watch. Manage them under AI visibility → Manage topics.
- BackTalk asks the models. On every probe, each prompt is asked several times per model, because answers vary run to run. One-off answers are noise; repeated samples are a signal.
- Mentions are extracted. Your filtering model (the What runs each task assignment, or your default key) reads each answer and records which brands were recommended, in what order, plus any new names it saw and any citations the model returned.
- You read the standings. Share of voice is your slice of tracked-brand mentions, averaged so each model counts equally: a model you probe more often than another cannot tilt the blended number. The dashboard shows the leaderboard, the trend, and per-model splits; click any prompt to read the actual answers behind the numbers. On the trend, a day when only some of your models were probed carries a faded marker, because its overall is averaged over just those models.
Set it up
Two steps, both on the AI visibility page the first time you visit:
- Add AI provider keys in Settings. By default every saved key's model joins the comparison; curate exactly which models compete, including several from one provider, under What runs each task (the Models to compare set). More models means a fuller picture, but even one works.
- Create your first topic with a few prompts, your aliases, and the competitors you care about.
Your first results arrive on the next scheduled probe, or immediately with Probe now on Pro and Business.
Probe cadence by plan
| Plan | Scheduled probes | Manual "Probe now" |
|---|---|---|
| Solo | Weekly | Not included |
| Pro | Daily | 6 per day |
| Business | Daily | Unlimited |
Owners and admins can dial the cadence down (or off) in Settings → AI visibility to save tokens; the plan sets the fastest allowed. Solo tracks up to 3 topics; Pro and Business are unlimited.
Events in your inbox
Probe results file events into the same inbox as your listening signals, under the AI answers source:
- Rank changes, for example a competitor overtaking you on Gemini.
- Wins, when you take the top spot, and streak milestones at 7 and 30 days.
- New competitors, when models start recommending a name you don't track yet.
They obey the same triage flow as every other signal, and the AI visibility badge in the nav counts the ones you haven't seen.
What it costs
Probing spends tokens on your own provider keys, at each provider's rates, with no markup from BackTalk. A daily probe of one topic with five prompts across four providers is typically a few cents a day; the extraction step runs on your cheapest default key. Dial the cadence down any time.
Troubleshooting
- A red dot on the sync trigger means a recent probe run had a failing model. Open the dropdown, then View sync history for the per-model error, usually an expired or unfunded key.
- A model's data is stale: the dashboard says so next to the charts and keeps showing its last good numbers. Fix the key in Settings and the next probe catches it up.
- A new topic shows no numbers yet: it was added after the last probe. It joins the next scheduled pass, or press Probe now.
Export and share
Every trend chart and the Saved answers shelf carry two export buttons, on the overview and on each topic:
- Export CSV downloads the exact series the chart plots (or the saved answers) as a bare table, ready for Excel or Google Sheets.
- Export Markdown downloads the same data as a formatted document that reads on its own. A trend export carries a title, the scope, the models and window the numbers were read under, the data-through and generated dates, and a short methodology note. A saved-answers export carries a title, the scope, the answer count and the date. Paste either straight into a doc, a ticket, or a team wiki.
Both are built in your browser from what is already on screen. A trend export reflects the model filter and time window you are viewing. The Saved answers shelf is not filtered that way: it always exports every starred answer for the current scope (all topics on the overview, one topic on a detail view), whatever the model filter or window. Nothing leaves your session. For a programmatic feed, use the API and MCP tools below.
API and MCP
Everything above is available programmatically: GET /api/v1/ai/topics, GET /api/v1/ai/share-of-voice and GET /api/v1/ai/answers on the REST API, and the list_ai_topics, get_share_of_voice and list_ai_answers tools on the MCP server. Star or flag a sampled answer with PATCH /api/v1/ai/answers/:id (the curate_answer MCP tool), and trigger a fresh probe with POST /api/v1/ai/probe (the trigger_probe MCP tool) or the app's Probe now.