Docs

When you track a question such as “what's the best tool for X”, the sampled answer that comes back either names you or it doesn't - and there is no page two. Builder radar asks up to 10 provenance-separated, web-grounded answer paths across 9 families on plan-aware schedules, retains substantial sanitized answer text for the declared history window, and turns the results into four things you can inspect:

• Where you stand - recommended, merely named, or absent, per question and engine, versus your competitors.
• Factual conflicts - substantial sampled claims that conflict with an approved brand fact, with the exact sampled quote and confidence evidence.
• Why - pages cited in the same sampled answer, so a loss carries inspectable citation context instead of a vague “do more content.” A citation is not presented as proof of causation.
• What changed - when comparable substantial samples move, you get the before/after receipt: both sanitized answer texts, with the same engine and dates.

One honesty rule shapes the whole product: we show you changes, we never claim your fix caused them. Answer-derived verdicts, losses, and movements require substantial stored evidence; coverage and heuristic figures identify their separate inputs.

1. 1 · Sign up and create a project. Enter your product's URL. We read your homepage to suggest a category, likely competitors, and tracked questions for you to review.
2. 2 · Review the generated buyer prompts. These are the questions we'll keep asking the engines. Edit them freely - the closer they match what your real buyers ask, the more the report is worth. Add your own; remove any that don't fit.
3. 3 · Confirm your competitors. We suggest them from your homepage and from who actually appears in AI answers; you can add or remove tracked competitors at any time.
4. 4 · Start the first snapshot. New projects attempt the missing Gemini snapshot cells immediately. Valid answers appear as the provider returns them; partial or failed attempts remain visible and retryable.
5. 5 · Read the Overview. Your observed visibility, coverage, qualifying competitor wins, and proof-backed factual conflicts. A zero is shown only from substantial observed cells; missing paths are reported as coverage instead.

Prefer to look before signing up? The sample report is a full illustrative walkthrough of every section below.

A project is one brand or product being tracked. A buyer promptis one question eligible paths sample on their disclosed cadence - e.g. “best uptime monitoring for indie SaaS” or “is Acme worth it”. Prompts are the unit everything else hangs off: verdicts, receipts, alerts, and source maps are all per-prompt.

A good prompt set covers four intents:

• Category shortlists - “best tools for [job you do]”, where recommendation lists form.
• Problem questions - “how do I [problem you solve]”, where category winners get named in passing.
• Comparisons - “[you] vs [competitor]” and “[competitor] alternatives”.
• Verification - “is [brand] worth it”, “what does [brand] cost” - where sampled factual conflicts and identity inconsistencies can surface.

Changing a prompt starts its history fresh - a receipt only makes sense against the exact same question - so settle your prompt set early and resist rewording prompts to flatter your brand. The plan sets your prompt budget: 25 on Solo, 75 across 3 brands on Agency.

We support up to 10 answer paths across 9 vendor/model families. OpenAI Search API and the ChatGPT consumer UI stay separate because API and browser observations can differ; Gemini, Perplexity, Claude, Grok, Google AI Overviews, and search-grounded open models retain their own provenance too. Each path discloses whether it came from an official API, consumer/search UI, or a raw model supplied with live retrieval context, rather than presenting those transports as interchangeable.

Paths are not sampled at identical rates. Their assigned mode depends on the plan and transport: daily, weekly, on-demand, snapshot, or historical. The dashboard reports:

• live / on schedule - an automatic path produced a usable sample inside its cadence window.
• overdue - an automatic path missed that window; existing history stays visible while current coverage remains missing.
• on demand / snapshot / historical - no daily freshness promise applies. A requested attempt with no valid answer is reported separately as incomplete.

Engines degrade occasionally - that's the reality of measuring live AI surfaces, and the badge is how we keep it honest. Answers also pass a quality filter before they count, so page junk and error screens never become verdicts.

Each project has six main surfaces, in the order you'll usually work them:

• Overview - the brief. Observed visibility with coverage-sensitive history, cadence-aware path status, qualifying competitor wins, and the highest deterministic review priorities from the other surfaces.
• Answers - the ledger. Stored valid samples keep sanitized answer text and are searchable, plus receipts for comparable changes. This is the evidence behind answer-derived verdicts.
• Competitors - who appears in your category's answers: an independent presence score per competitor, which winnable prompts each one wins, and exact-run detail. Presence scores do not need to sum to 100.
• Sources - observed citation context. Pages cited in substantial sampled answers, ranked by prompt and engine breadth, including citation gaps where your owned domain is absent. The ranking is a work-priority heuristic, not causal attribution.
• Fix Queue - qualifying factual conflicts and winnable losses turned into evidence-pinned suggestions, with QA-gated draft starting points where supported (more below).
• Activity - the changelog: answer transitions, new alerts, and shipped fixes in time order, so you can catch up after a week away.

A “More” group holds supporting views - discovered prompts worth tracking, the full AI accuracy list, demand signals, and social mentions. Their highlights also surface on the main pages, so you don't need to patrol them daily.

Each substantial sampled answer can receive one of three verdicts for your brand and tracked competitors. Degraded outputs are excluded:

• Recommended - the answer actively suggests you for the buyer's question.
• Named - you appear, but as one option among others or in passing.
• Absent - you don't appear at all. We show this bluntly, including a score of zero.

The visibility score applies recommended = 1, named = 0.5, absent = 0 to the latest substantial winnable prompt-by-engine cells, scaled to 0–100. Dashboard, emails, and client reports use the same points model while disclosing their window, expected-engine set, observed coverage, and whether a prior period is comparable.

Two reading rules save a lot of false alarms: a single run is a data point, never a trend - AI answers vary run to run, so judge weeks, not screenshots. And a low first score is normal; most brands discover they're absent from questions they assumed they owned. That reading is the starting line, not a verdict on your product.

A receipt is the before/after record of a tracked answer that changed: the same prompt, the same engine, both sanitized answer texts, with dates. Improvements and declines both get receipts - the point is a record you can act on (and hand to a client), not a highlight reel.

Change alertsfire on observed categorical transitions between substantial answers on the same question and path: you or a competitor entering or leaving an answer, or a new citation appearing. Both endpoints must be substantial answers on the same question and engine. One-off wording wobbles don't alert - answers are non-deterministic, and alerting on noise would train you to ignore the alerts that matter. Each transition alerts once. Where evidence is too thin to call, the product says inconclusive instead of guessing.

And the rule that runs through everything: when an answer changes after you shipped a fix, the receipt says the answer changed. It never says your fix caused it - nobody can prove that, so we don't sell it.

An accuracy alertfires when an engine states something about you that contradicts your own published facts: wrong pricing, a feature you don't have, the wrong category - or the quiet killer, entity confusion, where the engine describes a similarly-named company under your name.

Each proof-backed alert shows the detected claim, an excerpt from the exact substantial evidence run, the approved fact when available, severity, occurrences, and last-seen time. Independent model-family checks add a separate confidence label only after they complete; they never replace or retroactively invent the evidence run.

From any proof-backed alert you can jump to the Fix Queue for the exact answer receipt, an owned-page clarification, and any citation observed alongside it. Third-party outreach is withheld unless the disputed value is found on that cited page.

The Fix Queue turns qualifying evidence into a suggested move. For a winnable loss, it shows the exact evidence run, any source cited in that same competitor-winning answer, how actionable that source type is for you, and a QA-gated draft starting point when the evidence supports one - a pitch to the listicle's editor, a transparent contribution to the thread, a review profile to complete, or a comparison page to publish yourself. For a factual conflict: a polite, sourced correction draft plus a clarification for a page you own.

• Drafts with placeholders, unsupported facts, or missing evidence are withheld. A missing draft is safer than invented copy in your name.
• Nothing sends automatically. You copy the draft out, review it, and ship it yourself - it's your name on the outreach.
• Mark a move shipped and the queue keeps re-sampling that prompt, reporting whether the answer changed - up or down.

Set the horizon accordingly: there is no reliable schedule for an answer to change, and some never do. The queue is designed for that uncertainty - a short list of supported actions, followed by later comparable sampling rather than an outcome promise.

• Weekly analyst email - reports matched-cell movement only when comparable cells exist, same-answer citation context, coverage caveats, and evidence-backed next actions. A provider gap is reported as coverage, not a visibility decline.
• Movement alerts - answer transitions as they're confirmed. Agency plans can route these to Slack as well as email.
• Monthly client report (Agency) - an unbranded summary suitable for forwarding straight to a client. It removes BuilderRadar branding but does not add a custom customer logo.

Every email has a working unsubscribe link, and notification preferences live in Settings.

Free is a one-time snapshot: reviewed tracked questions sampled once on Gemini - observed visibility, any qualifying competitor wins, and a read-only preview of factual conflicts and winnable losses. No card, no ongoing tracking.

Solo ($49/mo) is recurring tracking for one brand: up to 10 provenance-separated paths across 9 families subject to their disclosed transport/cadence/availability, 25 tracked prompts, the receipts ledger for the declared history window, change and accuracy alerts, the Fix Queue, the source map, and the weekly email.

Agency ($129/mo) is everything in Solo for 3 brands and 75 prompts, plus unbranded reports, CSV export, API keys, webhooks, an MCP server, 3 team seats, Slack alerts, per-competitor battlecards, and 60-day history.

Lifetime plans (when offered) are scan-based: you trigger checks on demand from a monthly pool, plus an automatic weekly Gemini heartbeat attempt. Provider availability and the remaining pool still determine whether a usable sample is produced. Full details on the pricing section; all paid plans carry a 7-day money-back guarantee per the terms.

• Unbranded client reports - BuilderRadar branding is removed from a shareable, printable report. Answer-derived verdicts trace to substantial stored samples; coverage and heuristic figures state their separate inputs. Custom customer logos and themes are not implemented.
• CSV export - the answer ledger as data, for your own reporting stack.
• API keys & webhooks - pull findings into your systems, or get pushed when answers change. Keys are managed in Settings; treat them like passwords and rotate any key you suspect exposed.
• MCP server - connect Builder radar to AI assistants that support the Model Context Protocol, so you can query your own visibility data in natural language.
• Team seats - invite up to 3 teammates; access is per-account, so never share logins.

If you're building a monthly reporting practice on top of this, the playbook in Reporting AI Visibility to Clients Without Overclaiming is the recommended structure.

• What we store: the AI answers we sample (they're what you're paying to audit), your project settings, prompts and competitors, and the facts you publish that we check answers against.
• What we don't do: no tracking scripts on your website, no access to your analytics, no crawling behind your logins. Everything we measure is what a public AI assistant says.
• Deleting a project removes its tracked data; account deletion and data-export rights are covered in the privacy policy.

• “My score is zero - is the tool broken?”
  First check coverage and the Answers ledger. A zero is an observed finding only when substantial eligible samples exist and your brand is absent from them; missing or degraded paths must not be read as absence.
• “I asked ChatGPT myself and got a different answer.”
  Expected. Answers vary run to run, by account history, and by whether browsing was on. That's why we sample repeatedly and judge transitions over weeks - one chat (yours or ours) is a data point, not the truth.
• “An automatic path is overdue or incomplete.”
  That label means the path has not produced a usable sample inside its disclosed cadence, or a requested attempt returned no valid answer. It does not identify the external cause; the badge keeps missing coverage from masquerading as brand absence or a fresh result.
• “I shipped a fix - when will the answer change?”
  There is no reliable schedule, and an answer may never change. Keep the move marked shipped and let later comparable samples report what actually happened; do not accept a timing guarantee.
• “Can I change my prompts?”
  Yes, anytime - but a changed prompt starts its history fresh, since a before/after receipt is only meaningful against the identical question.
• Something else?
  Email support@builderradar.pro - a human reads it.